Nobody has a complete answer yet. There is enough real data now to rule out the comfortable ones.

The comfortable answer is that the biggest brand wins

It does not. An agent has no brand loyalty, no habits, and none of the twenty years of advertising that make a person reach for a familiar box. It has a request from a human, a set of candidates, and whatever it can read about each one.

That is a harsher market than most large brands are used to. Scale stops being a moat. What replaces it is a narrower question: of the candidates, which one can the agent describe most confidently to the person who asked?

What agents are actually doing right now

Start with what is measurable. Salesforce found that consumer reliance on AI assistants as the first stop in shopping grew 200% between May 2025 and May 2026. Half of shoppers now report using an AI assistant somewhere in the buying journey, up 67% in a year. Traffic from AI chats grew between 150% and 428% year over year across the quarters they measured, while overall traffic grew in the single to low double digits.

Then look at where those people went. Adobe Analytics, working from more than a trillion visits to US retail sites, found that in March 2026 AI-referred traffic converted 42% better than traffic from every other source, and produced 37% more revenue per visit.

Put those two facts together and the shape appears. The agent does the finding. The human still does the buying. The agent narrows the world to two or three options, hands them over with reasons attached, and the person clicks through to check.

So the answer is not really about agents at all

It is about what happens in the ten seconds after the handoff.

An agent that recommends you has made a claim on your behalf. It told someone your return window is generous, or your product suits sensitive skin, or you ship in two days. That person is now on your page looking for the sentence that confirms it.

The agent creates the intent. Your page decides whether the intent survives contact.

If the claim is easy to find and clearly true, the visit closes. If it is buried, hedged, or contradicted by the page, you have done something worse than lose a sale. You have taught a buyer that the agent was wrong about you, and that is the one lesson you cannot afford them to learn.

Three things that decide whether an agent can pick you

• Can it read you? Adobe found roughly 34% of retail product pages inaccessible to AI systems, along with about a quarter of homepage and category content. A page an agent cannot parse is a candidate it cannot shortlist. This is the cheapest problem on the list and the most commonly ignored.
• Can it verify you? Agents work from claims they can point at. A specific, checkable statement travels through a recommendation. An adjective does not. Premium quality survives no journey at all. Ninety day returns, free both ways survives every journey.
• Can it defend you? The person receiving the recommendation looks for the reason behind it. If that reason is not visible on your page within a screen or two, the recommendation quietly stops working.

What this means for a brand with real budget

Salesforce expects 20% of all 2026 holiday ecommerce traffic to arrive from AI chat agents, and one in three ecommerce sites to have its own shopper agent live by Cyber Week. Of shoppers surveyed, 74% said they trust product recommendations from AI chat.

That is not a future state to schedule a workshop about next year. It is this quarter's traffic mix.

The strategic move is not to build an agent. It is to make sure the thing an agent reads about you is worth reading. Most brands spent a decade optimizing pages for a human eye scanning a layout. The new reader has no eye. It has a parser, a shortlist, and a person waiting for an answer.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying, including the moment a claim fails to land. For the research on how AI buyers weigh what they read, see AI buyer behavior and structural bias.

The trillion dollars is not a prize for the biggest brand. It is a prize for the most legible one.

*Related Links: US agentic commerce revenue forecast to reach $1 trillion by 2030 (Retail Dive, on the ICSC and McKinsey report), Shopping's New First Step: Agentic Search Grows 200% (Salesforce), AI traffic to US retailers rose 393% in Q1 (TechCrunch, on Adobe Analytics data), 2026 Holiday Predictions (Salesforce).*

The traffic mix changed underneath you

Salesforce predicts that 20% of all 2026 holiday ecommerce traffic will originate from AI chat agents. Not 20% of discovery. Twenty percent of visits.

That is a channel that barely registered two seasons ago. Adobe measured a 693% jump in AI-driven traffic to retail sites across November and December 2025, on the way to a record $257.8 billion in online spending for the period, up 6.8% year over year. On Cyber Monday alone, AI traffic was up 670%.

The reason to care is not novelty. It is that this traffic behaves differently. Adobe found AI-referred visitors spend 48% longer on site, view 13% more pages, and convert 42% better than everyone else. They are your best visitors, and they arrive with expectations somebody else set.

The economy underneath is not generous

Salesforce's numbers on shopper sentiment are worth reading twice. Consumer pessimism is up 16% year over year. Around half of low-income shoppers and more than a third of middle-income shoppers report actively buying less. Thirty-five percent of holiday shoppers say they are trading down to cheaper alternatives.

At the same time, global digital traffic grew 18% in the second quarter while order volume grew 1%. More people are looking. Roughly the same number are buying.

More traffic and flat orders is not a demand problem. It is a conversion problem wearing a demand problem's clothes.

Cart abandonment sits at 82%. The cost of the usual fix went up too: brands are expected to spend an additional $3 billion globally subsidizing free shipping this season, about 7% more than last year.

What to fix, in order

1. Make the page answer the claim the AI made

A shopper arriving from an AI chat has been told something specific about you, and they are on your page to confirm it. If your return policy lives in the footer, your shipping cutoff lives in a FAQ, and your material detail lives in a collapsed accordion, you have turned a verification visit into a search task.

This is the highest-value fix available before the season, because it costs copy and placement rather than engineering.

2. Fix what an agent cannot read

Adobe found roughly 34% of retail product pages inaccessible to AI systems, and about a quarter of homepage and category content not readable by language models. If a fifth of your holiday traffic comes through AI chat, an unreadable product page is not a technical debt item. It is a delisted product.

3. Write for the gift buyer, who is not your usual customer

Most product pages are written for someone buying for themselves. During the holidays a large share of your traffic is buying for someone else and evaluating completely different things: is this appropriate as a gift, what happens if the size is wrong, is there a gift receipt, will it arrive in time, can the recipient return it without learning what it cost.

A page that never answers those questions loses a high-value segment silently, because gift buyers rarely contact support. They leave. We covered the pattern in the gift buyer conversion gap.

4. Put the price justification next to the price

With a third of shoppers trading down, the discount alone is not doing the work you think it is. Value-conscious buyers need a reason to believe the thing is worth what it costs even at a discount: cost per use, size comparison, the original price stated plainly, what is included in the box.

5. Decide what happens on mobile

Salesforce puts 75% of online traffic on mobile. Adobe measured 56.4% of holiday transactions on smartphones in 2025, rising to 66.5% on Christmas Day. Every fix above is a mobile fix first. A trust signal that sits above the fold on a laptop can be four scrolls down on a phone.

The part that is genuinely new this year

Salesforce expects one in three ecommerce sites to have its own shopper agent live by Cyber Week, and reports that retailers running branded AI shopper agents in 2025 saw 59% higher holiday sales growth than those who did not, at 6.2% versus 3.9%. Of shoppers surveyed, 41% said a brand-owned AI assistant made them much more confident in a purchase, with another 36% somewhat more confident.

If building one is not on your roadmap for this season, the fallback is the same work in a different order: make the information an agent would surface easy to find on the page itself.

The window closes in about a month

Every item on this list is a copy, placement, or structure change. None of them require a replatform. All of them get harder once traffic is live and every change is a risk to a season you cannot rerun.

eLLMo runs test buyers matched to your real customers against your product pages and returns a ranked list of what stops people from buying, before the traffic arrives to find out. If the holiday season matters to your year, the useful time to look at the page is now, not in the post-mortem.

*Related Links: 2026 Holiday Predictions: A Unified Yet Fractured Journey (Salesforce), Adobe: Holiday Shopping Season Drove a Record $257.8 Billion Online, AI traffic to US retailers rose 393% in Q1 (TechCrunch, on Adobe Analytics data).*

Read those together and the conclusion is uncomfortable but simple. The place where people find out your product exists is moving away from anything you own.

This is a bigger shift than the last one

When search rankings decided discovery, at least the destination was yours. Someone typed a query, saw a list, clicked, and landed on your site. You could measure it, optimize it, and buy your way into it when the ranking would not come.

What replaced it works differently. Salesforce measured 200% growth in consumer reliance on AI assistants as the first stop in shopping over a single year, with traffic from AI chats growing between 150% and 428% year over year across the quarters they tracked. Overall traffic grew in the single to low double digits over the same period.

The comparison is the point. AI chat traffic is not growing with the market. It is taking share from it.

Caila Schwartz, who leads shopper insights for agentic commerce at Salesforce, put the change plainly: customers are not starting with ten blue links any more. They are asking an AI.

What happens to your site when discovery moves upstream

Your website does not become worthless. It changes job.

It used to be the place where a stranger learned what you sell. Now it is increasingly the place where someone who has already been told what you sell arrives to check whether it is true.

The behavioral data supports that reading. Adobe Analytics, across more than a trillion visits to US retail sites, found AI-referred visitors spend 48% longer on site and view 13% more pages than other traffic, with an engagement rate 12% higher. Those are not the numbers of a casual browser. They are the numbers of somebody comparing what they were told against what they can see.

A visitor who arrives already convinced is not an easier sale. They are a stricter one.

And the strictness pays when the page holds up. The same Adobe data has AI-referred traffic converting 42% better than everything else in March 2026, with 37% higher revenue per visit.

The uncomfortable part for large brands

Falling discovery on brand-owned properties hits big brands hardest, because big brands built the most on owned property. Twenty years of investment went into making the site the destination: the content hub, the buying guide, the quiz, the loyalty program, the app.

None of that reaches a shopper who never arrives at the site to see it. If the AI assistant answering their question has not read it, or cannot read it, that investment is invisible at the exact moment the decision gets made.

That is not hypothetical. Adobe found roughly 34% of retail product pages inaccessible to AI systems, and about a quarter of homepage and category content not readable by language models. A meaningful slice of the internet's product information is currently invisible to the thing shoppers are asking.

What actually transfers

Three things travel through an AI recommendation. Everything else stays home.

• Specific, checkable facts. Dimensions, materials, return windows, shipping times, certifications, what is included. These survive being summarized because they can be restated without losing meaning.
• Answers to real questions. Content structured as a question and its answer is content an assistant can lift and attribute. A buying guide written as brand narrative is not.
• Anything a third party said about you. Reviews, press, testing results. An assistant weighs outside evidence differently from your own description of yourself, and it is generally more willing to repeat it.

Brand voice does not travel. Layout does not travel. The careful sequence you built to walk someone from problem to solution does not travel, because the assistant does the walking now and hands over a conclusion.

What to do about it this quarter

The instinct is to chase visibility inside the assistants. That is worth doing, and it is also a moving target across a fragmenting set of platforms.

The more durable work is the other half: assume the shopper arrives pre-briefed, and build the page for verification rather than persuasion. That means the claims an assistant is most likely to repeat about you should be the easiest things to find on the page, stated in the same terms, without hedging.

It is a smaller change than a replatform and it compounds, because the same clarity that satisfies a skeptical human satisfies a parser.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying. The buyer who arrives from an assistant is a distinct kind of visitor, and the page either finishes the job the assistant started or interrupts it. For more on why different buyer types read the same page differently, see preference heterogeneity and persona validity.

*Related Links: Shopping's New First Step: Agentic Search Grows 200% (Salesforce), AI traffic to US retailers rose 393% in Q1 (TechCrunch, on Adobe Analytics data).*

The funnel shows you the exit, not the reason

Analytics are very good at one thing: telling you where people stopped. They are structurally incapable of telling you why, because the reason formed in the shopper's head a moment before they left and was never recorded anywhere.

So teams end up reasoning backwards from a drop-off number. The page must be too slow. The price must be too high. The photography must be weak. Those are hypotheses generated by a room full of people looking at a chart, and they compete with each other on confidence rather than evidence.

Baymard Institute puts the documented average cart abandonment rate at 70.22%, across 50 studies. The reasons shoppers themselves give are specific and mostly fixable: unexpected costs, slow delivery, no trust in the site with a card, a forced account, a policy they could not find. None of those show up in a funnel report as anything other than a number getting smaller.

The gap between present and findable

When we run test buyers against a page, the most common finding is not that something is missing. It is that something exists in a place the buyer never reaches while they still care about it.

Recent runs turned up the same shape repeatedly. Warranty terms written into the middle of a long description. A return policy with no presence anywhere near the buy button. A review count set in type quiet enough that a scanning buyer read the page as having no reviews at all. Certifications and guarantees sitting below the fold, three screens past the decision.

Present is a fact about your page. Findable is a fact about your buyer. Only one of them converts.

Two pieces of published research explain why this is so easy to get wrong.

Nielsen Norman Group's eyetracking work, drawn from over 130,000 eye fixations across 120 participants, found that people spend about 57% of their page-viewing time above the fold, and 74% within the first two screenfuls. Attention does not stop at the fold, but it falls away sharply and keeps falling. Anything you place on screen four is competing for the last quarter of a buyer's attention.

Baymard's product page benchmark, built from more than 30,000 manually reviewed product pages across 155 sites, found that 44% of sites keep return policies out of the main content, and 67% give no total cost estimate. Only 48% of desktop product pages and 38% of mobile ones rate as decent or better overall.

That is not a content problem. Those brands wrote the policy. It is a placement problem.

Different shoppers stall on different things

The reason this is hard to fix by intuition is that there is no single missing sentence. What one buyer needs to proceed is not what the next one needs.

In a run against a furniture retailer's page, one buyer stopped over dimensions, because the room was small and the listing gave a footprint but no sense of scale. Another moved past dimensions without pausing and stalled on assembly, because they lived alone and needed to know if one person could manage it. A third read both without concern and stopped at the return terms, because the item was expensive and the decision felt irreversible.

Three buyers, one page, three different stopping points. The aggregate conversion rate for that page is a single number that contains all three and describes none of them.

This is why a ranked list beats a score. When the same hesitation appears across several buyer types, you have a page problem. When it appears in one, you have a segment problem. Those need different fixes, and a funnel cannot distinguish them.

What to do with the finding

The work that follows is mostly rearrangement, which is the cheapest kind of work there is.

• Move the reassurance to the decision, not the page. Return terms, shipping timing, and guarantees belong adjacent to the add to cart control, not in a policy page and not in a footer. If the buyer has to leave the decision to find the reassurance, many will not come back to it.
• Say it twice when it matters. Repeating a return window near the price and again near the button is not redundancy. It is meeting two different scan paths.
• Give quiet facts visual weight. A review count in small grey type reads as no reviews. The information is technically present and functionally absent.
• Pull the buried specific out of the paragraph. A warranty stated in the fourth sentence of a description is not a claim the buyer will find. Lift it into its own line.
• Check it on a phone first. Mobile compresses everything below the fold. A trust signal that sits comfortably above the fold on a laptop can be four scrolls down on a phone, which is where most of your traffic is.

Ahead of the season

Every fix above is copy and placement. None require engineering, and all of them get harder once peak traffic is live and every change is a risk to the quarter that pays for your year.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying, in the buyer's own words. When the answer is that a fact was present but unreachable, that is exactly what comes back. See a live run while rearranging a page is still cheap.

*Related Links: 50 Cart Abandonment Rate Statistics and The Current State of Ecommerce Product Page UX (Baymard Institute), Scrolling and Attention (Nielsen Norman Group).*

Swings that size usually mean somebody changed how they measure. This one does not appear to be that. The volume moved with it: AI traffic to US retailers grew 393% year over year in the first quarter of 2026.

So something real changed. Working out what is the difference between treating this as a headline and treating it as a plan.

What did not change

Your page did not get 80 points better. Neither did anyone else's.

The channel did not become more valuable through some property of the channel. AI referrals are still a small slice of total retail visits. What changed is who is inside that slice, and what they know when they arrive.

What did change

A year ago, AI shopping assistants were mostly answering questions. People used them the way they used a search box, then went off to do their own research. The click that followed was early-stage and exploratory, which is exactly the kind of traffic that converts badly.

Now the assistant does more of the work before the click. It compares options, weighs reviews, checks specifications, and hands over a short list with reasons attached. By the time a person reaches your page, the comparison has already happened somewhere else.

The engagement numbers show it. Adobe found AI-referred visitors spend 48% longer on site, view 13% more pages, and post a 12% higher engagement rate than other traffic. Revenue per visit runs 37% higher.

A year earlier, that same comparison went the other way: regular human traffic was worth 128% more per visit than AI traffic.

The traffic did not improve. The brief the buyer arrives with improved.

Why more time on site is not the good news it sounds like

It is tempting to read 48% longer as deeper engagement, which sounds like affection. It is more likely to be verification.

Somebody arrives having been told three specific things about your product. They are looking for those three things. Longer sessions and more pages viewed are consistent with a person working through a checklist, not a person enjoying your photography.

That reframes what the page is for. A visitor who already decided to consider you does not need to be persuaded that the category matters or that your brand has a story. They need the specific claim confirmed, fast, without hunting.

The failure mode this creates

There is a new way to lose a sale that did not exist three years ago: the assistant says something true about you that your page does not obviously support.

It says the return window is 90 days. Your page mentions returns once, in the footer, without the number. It says the product is fragrance free. Your ingredient list is behind a tab labeled Details. It says two day shipping. Your shipping page says two to five business days depending on location.

In each case nothing is a lie. But the person came to confirm, found ambiguity, and now has a reason to doubt not just the claim but the recommendation that carried it. That doubt is expensive, because it arrives at the moment they were closest to buying.

What to do with this

Three things follow, and none of them require a new platform.

• Find out what assistants are saying about you. Ask them, in the way a customer would. The claims that come back are the claims your page now has to support on demand.
• Put those claims where a verifier looks. Near the price, near the add to cart, above the fold on a phone. Not in a policy page, not behind an accordion, not implied.
• Match the wording. If the assistant says free returns and your page says hassle-free returns process, a scanning buyer has to do a translation step. Small friction, applied at the exact moment of highest intent.

The wider point

Every channel eventually gets efficient enough that the page becomes the bottleneck. Paid search went through it. Social went through it. AI referral traffic just went through it in a single year, which is faster than any channel before it.

The brands treating this as an acquisition story are optimizing to be recommended. The brands treating it as a conversion story are making sure the recommendation survives the landing. The second group gets the 42%.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying, in the buyer's own words. If your assumption is that traffic arriving with high intent will convert itself, that assumption is worth testing before the next quarter's spend goes out.

*Related Links: AI traffic to US retailers rose 393% in Q1, and it's boosting their revenue too (TechCrunch, on Adobe Analytics data), Shopping's New First Step: Agentic Search Grows 200% (Salesforce).*

One in three product pages. Not slow, not badly designed, not poorly worded. Unreadable by the thing that increasingly decides which products a shopper hears about.

Why this is worse than it sounds

An unreadable page is not a page that ranks lower. It is a page that does not enter the consideration set at all.

A human who lands on a badly structured page can still squint at it, scroll, and work out what the product is. An assistant building a shortlist does not squint. If it cannot extract the product, the price, the specification, and the terms, it moves to a competitor it can extract them from. There is no partial credit and no second chance in that turn.

Now put that against the traffic trend. Adobe measured AI traffic to US retail sites up 393% year over year in the first quarter of 2026, and that same traffic converting 42% better than every other source in March, with 37% higher revenue per visit.

The fastest growing and best converting source of traffic in retail cannot read a third of the pages it is being asked about.

What makes a page unreadable

The specifics vary, but the patterns repeat.

• Content that only exists after JavaScript runs. If the meaningful text is assembled in the browser, a crawler or model working from the raw document sees an empty shell where your product description should be.
• Facts locked inside images. Sizing charts, ingredient lists, comparison tables, and specification panels shipped as pictures. A human reads them fine. A parser gets nothing, unless the alternative text carries the content, which it almost never does.
• Information hidden behind interaction. Tabs, accordions, and modals that only load their content on click. The detail is on the page in a sense that matters to a designer and not in a sense that matters to a machine.
• Missing or thin structured data. Product markup that omits price, availability, review count, or return policy leaves the assistant guessing at exactly the fields a shopper asks about.
• Claims with no anchor. Dermatologist tested with no study, sustainably made with no standard named. These are weak with skeptical humans and useless to a machine that is looking for something citable.

The reason this persists at large brands

It is rarely negligence. It is usually the accumulated result of good decisions made for a different reader.

Tabs and accordions exist because a product page with everything visible looks overwhelming. Image-based sizing charts exist because they were easier to produce and looked better than a table. Client-side rendering exists because it made the site feel fast. Every one of those choices optimized for a human eye on a screen, which was the correct target right up until it was not the only one.

The result is that the brands with the most sophisticated front ends are sometimes the least legible to the new reader, while a plain competitor with boring HTML gets read cleanly and cited.

How to find out where you stand

You can get most of the way with an afternoon and no budget.

• Read your own page as a machine would. Fetch the raw HTML of a top product page and look at what is actually in the document before any script runs. If the description, price, and specifications are not there, an assistant may not see them either.
• Ask the assistants about your product directly. Not your brand, your product, the way a customer would ask. Note what comes back wrong, vague, or missing. Those gaps are usually extraction failures rather than opinions.
• Check your five best sellers first. This is not a whole-catalog project to begin with. The revenue concentration in most catalogs means a handful of pages carry the outcome.

What to fix first

Fix the facts a shopper asks about before you fix anything else: price, availability, sizing, materials or ingredients, shipping time, return terms. Get those into the document as plain text and into your structured data accurately. Everything else on the page can wait.

This is unglamorous work and it competes badly for attention against a redesign. It is also one of the few remaining places where a modest engineering effort changes whether you are considered at all, rather than changing how you look once you are.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying. A claim a buyer cannot find and a claim a machine cannot extract are usually the same claim. For the research on how AI buyers weigh page content, see AI buyer behavior and structural bias.

*Related Links: AI traffic to US retailers rose 393% in Q1, and it's boosting their revenue too (TechCrunch, on Adobe Analytics data), Shopping's New First Step: Agentic Search Grows 200% (Salesforce).*

That number is not ours. Baymard Institute puts the documented average cart abandonment rate at 70.22%, drawn from 50 separate studies. The same body of work estimates the average large ecommerce site could gain a 35.26% lift in conversion from better checkout design alone.

The gap between those two facts is the whole problem. You can see the loss. You cannot see the cause.

Five findings, five different categories

What follows is drawn from recent runs across several brands. None of these pages were broken in a way a dashboard would flag. Traffic was healthy, sites were fast, nothing was throwing errors.

1. A subscription the buyer could not price

A specialty coffee brand offered a bag one-time and on subscription at the same visible price. The savings were real but arrived later, spread across the year, and the page did not do the arithmetic. Test buyers took the one-time option while saying some version of: I cannot tell what I get for committing.

Baymard finds 12% of abandoning shoppers cite being unable to calculate the total cost upfront. A subscription that hides its own value is that problem with better graphic design.

2. A commitment with no visible exit

A supplements brand explained delivery cadence and savings in its subscription module and never mentioned that a shopper could skip, pause, or cancel. Buyers with any sensitivity to ongoing charges read the silence as a trap, because that is what silence means when money recurs.

3. A mobile default nobody chose

On a personal care brand's mobile page, the fast path to add to cart interacted with the subscription module such that a buyer moving quickly could commit to a recurring order without registering it. Nothing was deceptive by intent. It was a layout consequence, and it is the kind of thing that produces cancellations a month later that nobody traces back to a product page.

4. Reviews about a different product

A kitchen appliance brand pooled reviews for several blender models into one feed. A buyer looking at the compact model read reviews written by someone who bought the professional model for a different job. The rating was high and the content was useless, which is the worst available combination: it looks like social proof and works like noise.

5. Two bugs worth real money

At an outdoor gear brand, a link labelled Reviews sent shoppers to the FAQ page. And the abandoned cart email arrived with its product image not clickable, so the highest-intent shopper in the funnel hit a dead end.

Klaviyo's benchmark across more than 143,000 abandoned cart flows puts the average placed-order rate at 3.33% and average revenue per recipient at $3.65, with the top decile at 7.69% and $28.89. An email that cannot be clicked competes in that market with one hand tied.

Every one of these is invisible in aggregate data and obvious to a buyer trying to complete a purchase.

Why analytics structurally cannot see this

Three reasons, and none are fixable with a better dashboard.

• Analytics record behavior, not reasoning. A drop-off event tells you where someone stopped. The reason happened in their head one moment earlier and was never transmitted.
• Aggregate numbers average away the segment that is failing. If a page works for four buyer types and fails for the fifth, conversion moves slightly and the cause hides inside a stable-looking average.
• Nobody reports a confusing subscription module. Baymard separates out that 42% of abandonment is people who were only browsing. The rest rarely complain. They leave, and they are counted as demand that did not convert.

What this looks like as a working method

A run puts a set of buyers, each with a stated personality, budget, and reason for being on the page, through the real journey: product page, subscription choice, cart, checkout, and the follow-up email. What comes back is not a score. It is a ranked list of the moments where a specific buyer stopped, in their own words, with the reason attached.

The output is a brief your design and copy team can act on this week, not a metric you have to interpret.

Why the timing matters

Every fix above was copy, placement, or a broken link. None needed engineering beyond an afternoon. All of them get more expensive once peak traffic arrives, because then every change is a live experiment on the quarter that pays for your year.

The brands that go into the holidays knowing what is wrong with their pages are not the ones who test more. They are the ones who looked before the traffic did.

See a live run on a real product page, or bring us yours before the season starts.

*Related Links: 50 Cart Abandonment Rate Statistics (Baymard Institute), Abandoned Cart Benchmark Report (Klaviyo).*

Which makes it strange how often the product page throws the subscription away.

The pattern

We ran test buyers against a specialty coffee brand's product pages ahead of peak season. The subscription offer was genuinely good. The attach rate was not.

The cause turned out to be one design decision. The page showed the same upfront price for a one-time bag and for the subscription. The subscription saved money over time, across repeat deliveries, over a year. At the decision point, both cost the same today.

Test buyers hit that screen and reasoned the way people do: if the price is identical right now, why agree to something ongoing? Several chose one-time while saying they would probably have subscribed if they understood the benefit.

The offer was not weak. The arithmetic was missing.

The buyer is doing math you did not show them

A subscription asks a shopper to trade flexibility for value. That trade only looks good if both sides are visible in the same moment.

Most pages show the commitment clearly and the value vaguely. Save 15% is a percentage of a number the buyer has not calculated. Every 3 weeks is a cadence, not a benefit. Meanwhile the cost of committing is perfectly clear, because it is a recurring charge on their card.

Baymard Institute finds 12% of abandoning shoppers cite not being able to calculate the total cost upfront. Subscriptions concentrate that problem, because the total cost is a stream rather than a number.

Four things that fix the arithmetic

• Show the annual number. Not the percentage. The dollars saved over a year, next to the dollars spent. A buyer who can see $58 saved does not need to compute 15% of anything.
• Price the consumption, not just the unit. For anything bought repeatedly, the honest comparison is what a year costs. If a bag lasts three weeks, the page should say what twelve months looks like, in the place where the buyer is deciding.
• Make the first-order difference visible. If subscribing changes today's price at all, that belongs next to the button. If it does not, the page has to work harder on the other three.
• Name the frequency in the buyer's terms. Every 21 days is a schedule. Arrives before you run out is a reason.

The part most teams get backwards

The instinct when attach rates disappoint is to increase the discount. That is expensive and usually does not work, because the problem was never that 15% was too small. It was that the buyer never converted 15% into a reason.

The cheaper fix is to state value in the same units as cost. Dollars against dollars, year against year. A buyer who can see both sides of the trade will make it more often, at the discount you already offer.

Why the season makes this urgent

Peak season traffic is the most expensive traffic you buy all year, and it skews heavily to first-time buyers. A first purchase is the only moment a one-time buyer becomes a subscriber without a second acquisition cost.

A subscription module that loses the argument in November loses it on every one of those buyers, and you pay for the second chance in January.

eLLMo runs test buyers matched to your real customers against your product page and returns a ranked list of what stops people from buying, in the buyer's own words. Subscription hesitation is one of the clearest signals it surfaces, because buyers explain the trade out loud when they refuse it. See a live run before your peak traffic arrives.

*Related Links: 50 Cart Abandonment Rate Statistics (Baymard Institute).*

We ran test buyers against a supplements brand's product pages before peak season. The subscription block was clear about two things: when the shipment arrives, and how much it saves. It said nothing about whether a buyer could skip a delivery, pause the plan, or cancel.

The buyers who cared most about that omission were the ones most anxious about ongoing commitments. They did not go looking for the cancel option. They picked the one-time purchase and moved on.

Silence is not neutral on a recurring charge

When a page does not mention something, buyers fill the gap from experience. For subscriptions, the experience of a large number of consumers is a plan that was easy to start and unpleasant to stop.

So the absence of an exit does not read as no comment. It reads as evidence, and the evidence points the wrong way.

On a recurring charge, an unstated cancellation policy is not missing information. It is a claim, and the claim is bad.

Regulators have been circling exactly this. The Federal Trade Commission finalized its Click-to-Cancel rule in October 2024, and the rule's status has since been contested in court. In March 2026 the FTC issued an Advance Notice of Proposed Rulemaking seeking comment on amendments to its Negative Option Rule. Whatever the final shape, the direction is not ambiguous: clear disclosure of terms, real consent, and a cancellation path that works. The FTC also continues to act against deceptive subscription practices under its general authority regardless of the rule's status.

The compliance reading of that is defensive. The conversion reading is more interesting.

The exit is what makes the entrance possible

Anxiety about commitment is not irrational and does not respond to enthusiasm. It responds to a visible way out.

This is the same mechanism that makes free returns work on apparel. Nobody buys a jacket because returns are free. They buy it because the return policy made the purchase reversible, which lowered the cost of being wrong to almost nothing.

Baymard's abandonment data has an unsatisfactory returns policy at 13% and lack of trust at 19% among the stated reasons shoppers leave a checkout. Subscriptions raise both stakes at once, because the buyer is not deciding about one purchase. They are deciding about an open-ended series.

What to write, and where

• Put it inside the module, not in a policy page. The sentence has to be visible at the moment the choice is made. A link to terms is not a statement.
• Name all three actions. Skip, pause, and cancel are different reassurances for different worries. Skip handles I have too much already. Pause handles I am travelling. Cancel handles I do not want to be trapped.
• Say where it happens. Manage everything from your account, no phone call needed is stronger than cancel anytime, because it answers the question the buyer was about to ask next.
• Do not bury it in small grey type. A reassurance in 10px grey reads as a disclaimer, and disclaimers make people more suspicious, not less.

The counterintuitive part

Teams resist this copy because it feels like advertising the exit. The worry is that reminding people they can cancel will make them cancel.

What we see in simulation is the opposite sequence. The buyer who never sees the exit does not subscribe at all, so there is nothing to cancel. The buyer who sees it subscribes, because the decision stopped being permanent. You cannot lose a subscriber you never acquired.

Before your peak season

This is the cheapest fix on any subscription page. One sentence, one place, no engineering. It is also the fix most likely to still be missing when your most expensive traffic of the year arrives.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying. Commitment anxiety shows up loudly, because buyers say the quiet part out loud when they decline. See a live run while there is still time to act on it.

*Related Links: Negative Option Rule (Federal Trade Commission), 50 Cart Abandonment Rate Statistics (Baymard Institute).*

The short-term data usually says yes. Defaults win. That is one of the most reliable findings in behavioral science, and it holds on product pages the way it holds everywhere else.

The problem is that the short-term data is measuring the wrong thing.

What we saw on mobile

We ran test buyers against a personal care brand's product pages, including the mobile experience, ahead of peak season. The team was actively working out whether the subscription module should be open, defaulted, or both.

On mobile, the fast path to add to cart interacted with the subscription module such that a buyer moving quickly could commit to a recurring order without registering it. Nothing was designed to deceive. It was a consequence of a small screen, a sticky button, and a module that scrolled.

That is the expensive version of this problem. Not a dark pattern anyone chose. A layout that produces one.

Why mobile changes the calculation

Mobile buyers move faster, see less at once, and complete purchases with fewer confirmation steps. Mobile conversion rates run materially below desktop, and mobile cart abandonment runs higher, roughly 77% against 65% on desktop in recent benchmarks.

A default a desktop buyer notices and evaluates is a default a mobile buyer may simply inherit. Same setting, different consent.

A default the buyer did not see is not a decision. It is an outcome you have to defend later.

The bill arrives in a different department

Here is what makes this hard to manage. The gain shows up in one team's numbers and the cost in another's, a month or two later.

The product page team sees attach rate rise. Finance sees the second charge trigger cancellations. Customer service sees the tickets. Payments sees the disputes. Nobody connects the three, because the events are weeks apart and sit in different systems.

Involuntary and early churn behave very differently from ordinary churn. Across direct-to-consumer subscription programs, monthly churn commonly runs in the 5% to 8% range for replenishment categories, and a meaningful share of total churn is involuntary. A subscriber who did not mean to subscribe is not a retention problem you can email your way out of.

The regulatory floor is rising

There is also an external constraint. The Federal Trade Commission finalized its Click-to-Cancel rule in October 2024, the rule's status has been litigated since, and in March 2026 the FTC issued an Advance Notice of Proposed Rulemaking on its Negative Option Rule. The consistent themes across every version are clear disclosure of material terms and affirmative express consent to the recurring charge.

Affirmative express consent is a hard phrase for a defaulted checkbox to satisfy on a phone screen where the checkbox was never visible.

A workable position

The choice is not binary between defaulted and buried. Three things move the trade in your favor.

• Open the module, do not pre-select it. Visibility does most of the work a default does, without borrowing the consent. Buyers who want the subscription can see it exists.
• If you default, confirm on the same screen. The subscription terms have to be visible in the same viewport as the button that accepts them. On mobile that is a real constraint and it should drive the layout.
• Test the cancellation rate, not just the attach rate. An attach rate that rises while 60 day retention falls is not a win. It is a loan against next quarter at a bad rate.

Before the season

Peak season concentrates every risk here. Traffic skews to first-time buyers, mobile share hits its annual peak, and shoppers move faster than at any other point in the year. A default that mostly works in June can produce a January of disputes.

eLLMo runs test buyers matched to your real customers against your page, including the mobile path, and returns a ranked list of what stops people from buying and where they commit to things they did not intend. See a live run before the traffic makes the decision for you.

*Related Links: Negative Option Rule (Federal Trade Commission), 50 Cart Abandonment Rate Statistics (Baymard Institute).*

We ran test buyers through the full journey for an outdoor gear brand before peak season: product page, cart, checkout, and the follow-up email. Two of the findings were not design opinions. They were bugs.

Bug one: the link that lied

A link labelled Reviews sent shoppers to the FAQ page.

It returned a 200. It rendered fine. Analytics recorded a click and a pageview, and both looked like engagement. To a buyer at the exact moment they went looking for social proof, it was a dead end that also felt like being ignored.

The cost is invisible in aggregate because the shopper does not bounce in a way that stands out. They read a FAQ they did not want, then leave. That looks like normal browsing.

Bug two: the email nobody could click

The abandoned cart email arrived with its product image not clickable.

Consider what that costs. Klaviyo's benchmark across more than 143,000 abandoned cart flows puts average revenue per recipient at $3.65 and the top decile at $28.89, on an average placed-order rate of 3.33% against 7.69% for the best performers. Abandoned cart email reaches the highest-intent audience you have: people who chose a product and stopped.

The main visual element in that email was a wall. The shopper who wanted to return to their cart had to go find it themselves.

Neither of these throws an error. Both fail silently, on your highest-intent traffic, every day until somebody walks the journey.

Why this class of bug survives

Three structural reasons, and they apply to every brand past a certain size.

• Nothing in the stack is looking for it. Uptime monitoring checks whether pages respond. Error tracking checks whether code throws. Neither has an opinion about whether a link goes where its label promises.
• The failure looks like normal behavior. A shopper who hits a dead end and leaves produces the same data as a shopper who was not that interested. There is no distinct signature to alert on.
• The journey spans systems nobody owns end to end. The product page belongs to one team, the email to another, checkout to a third. The bug lives in the seam, and the seam has no owner.

Baymard puts website errors and crashes at 17% of stated abandonment reasons, level with checkout complexity. That is the visible tier of this problem. The invisible tier is larger, because a broken link never registers as an error at all.

The cheap discipline that catches them

Walk the journey as a buyer, end to end, on the devices your buyers use, before every peak. Not a QA pass against a spec. A purchase attempt with a goal.

That is what a simulation run does at scale: a set of buyers with stated intentions moving through the real journey, reporting where they stopped and why. When the reason is a link that went to the wrong page, it comes back as a finding rather than as a slightly lower conversion rate.

The economics are hard to argue with. Both bugs above were minutes of work to fix. Both had been live long enough to be part of the baseline everyone was optimizing around.

Do this before your traffic peaks

Every bug of this kind is cheapest to find now and most expensive to find in December, when the traffic is bought, the team is frozen, and a broken link in a cart email is running against your best audience of the year.

See a live run, or bring us your journey and we will walk it before your customers do.

*Related Links: Abandoned Cart Benchmark Report (Klaviyo), 50 Cart Abandonment Rate Statistics (Baymard Institute).*

The confusion comes from treating design quality and purchase clarity as the same achievement. They are not, and on some pages they actively compete.

What polished pages get right, and what they skip

Strong visual design does real work. Baymard's research finds that image quality and product variation UI are among the better-performing parts of the average product page.

The weakness sits elsewhere. Their benchmark, drawn from more than 30,000 manually reviewed product pages across 155 sites, rates only 48% of desktop product pages and 38% of mobile ones as decent or better overall. The failures cluster in the unglamorous parts: specifications, shipping and returns, reviews, and the buy section itself.

Some specifics from that benchmark are worth sitting with. Thirty-seven percent of sites provide no image that shows the product in scale. Sixty-seven percent give no total cost estimate. Forty-four percent keep return policies out of the main content. Eighty-one percent omit price per unit.

None of those are design failures in the aesthetic sense. Every one of them is a buyer left holding a question.

A page that looks expensive and answers nothing is not a premium experience. It is a beautiful dead end.

The text-in-image problem

The most common version of this we see in runs is information baked into a picture. Dimensions on a lifestyle shot. A comparison chart exported as a JPEG. Care instructions set as type inside an image. Assembly steps as a graphic.

It looks great, and it fails three audiences at once.

• Screen reader users get nothing. The 2026 WebAIM Million study of the top one million home pages found 16.2% of images missing alternative text, an average of 10.8 per page, affecting 53.1% of pages. More than one in four images carry missing, questionable, or repetitive alt text. Image counts rose 13.6% in a year while detected accessibility errors rose 10.1%.
• AI systems get nothing. Adobe Analytics found roughly 34% of retail product pages inaccessible to AI systems. If a shopper asks an assistant whether your chair fits a small room and the dimensions live inside a photograph, the assistant cannot answer and recommends someone it can read.
• Scanning buyers usually get nothing. People skim text and look at images. They rarely read images. A specification hidden in a picture is discovered by the buyer who was going to buy anyway.

The questions a beautiful page tends to leave open

Across recent runs on visually strong pages, the same categories of question kept stopping buyers.

• Does it fit my space? Not the dimensions, which are usually present. Whether those dimensions mean it works where the buyer intends to put it. Baymard's finding that 37% of sites show no in-scale image is this exact gap.
• Can I set it up? Assembly, tools, whether it needs two people. An apparel brand does not face this. Anything shipped flat does.
• Will it be comfortable, and for how long? Comfort cannot be photographed. It has to be described, or evidenced through reviews from people with similar needs.
• Does it adjust to me? Height, firmness, tension, fit. Adjustability is a feature that gets shown in motion and rarely explained in text.
• What happens if I am wrong? Warranty and returns, close enough to the decision that a buyer can see them without leaving it.

Design is how the proof arrives

The fix is not less design. A page stripped back to a specification sheet converts worse, not better, and it undermines the brand the photography was built to establish.

The fix is to stop treating information as the thing design has to work around. The best pages use hierarchy, spacing, and type to make the proof easy to absorb: dimensions rendered as a diagram with real text beside it, a scale reference in the gallery, return terms sitting beside the price in a weight a buyer notices, reviews filtered to the use case the buyer arrived with.

That is design doing the job. It is more work than a mood board and it is what separates a page that photographs well from a page that sells.

Before the season

Peak traffic is the worst time to discover that a beautiful page leaves questions open, because the buyers arriving are disproportionately first-time and gift buyers with no prior experience of your brand to fill the gaps with.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying. On a well-designed page, the findings are rarely about the design. They are about the question the design never answered. See a live run before the season answers it for you.

*Related Links: The Current State of Ecommerce Product Page UX (Baymard Institute), The WebAIM Million (WebAIM), AI traffic to US retailers rose 393% in Q1 (TechCrunch, on Adobe Analytics data).*

What we found

We ran test buyers against a kitchen appliance brand's product pages ahead of peak season. Reviews for several blender models were pooled into one shared feed.

So a buyer evaluating the compact model read reviews written by people who bought the professional model, for a different job, with different expectations. The rating was strong. The reviews were about something else.

Test buyers noticed and said so. Several described scrolling for something relevant, not finding it, and losing confidence rather than gaining it. One effect of pooled reviews is that a shopper cannot tell whether the product in front of them is any good, and now also cannot tell whether the brand is being straight with them.

A five star average built from the wrong product is not social proof. It is a rating you cannot use.

What the research says the reviews are doing

Reviews carry more conversion weight than almost anything else on the page. Northwestern University's Spiegel Research Center, analyzing around 57,000 products, found that adding the first review lifts conversion by roughly 65% on average, rising to about 190% for products over $100 and around 45% for products under $25.

The same research found the highest converting star ratings are not perfect ones. The sweet spot sits between 4.2 and 4.5 stars, and ratings above 4.7 convert slightly worse, because near-perfect scores read as filtered.

Both findings point at the same mechanism. Reviews work when they look like evidence from someone in the buyer's situation. They stop working when they look like marketing.

Pooling breaks the mechanism in three ways

• Fit questions go unanswered. For anything sized, specified, or capacity-bound, the buyer's question is not is this good. It is is this right for my kitchen, my batch size, my counter. A pooled feed cannot answer that, because the answer depends on which model the reviewer bought.
• The rating stops meaning anything. An average across models measures the portfolio, not the item. A weak product hides behind a strong one, which is good for exactly one quarter and bad afterwards, when returns arrive.
• Careful buyers detect it. The buyers most likely to read reviews closely are the ones most likely to notice a review describing a different product. What they learn is not about the blender.

What to do instead

• Scope reviews to the model by default. Give the buyer the option to widen to the family, clearly labelled, rather than defaulting to a merged feed.
• Show which product the reviewer bought. If any pooling remains, label every review with its model. A small piece of metadata restores most of the lost trust.
• Let buyers filter by their situation. Use case, household size, frequency. Relevance beats volume once a product has enough reviews to be credible at all.
• Check the anchors. While you are in there, verify that every link labelled Reviews actually goes to the reviews. We have found one that went to the FAQ.

The season argument

Peak season traffic contains an unusually high share of first-time buyers and gift buyers, and both lean harder on reviews than a returning customer does. They have no experience with your brand to fall back on, so reviews do the entire job of establishing whether this specific product is the right one.

A pooled feed underperforms all year. In November it underperforms on the traffic you paid the most for.

eLLMo runs test buyers matched to your real customers against your page and reports what they went looking for and did not find. Review relevance surfaces constantly, because buyers narrate the search. See a live run before your peak traffic starts reading.

*Related Links: 50 Cart Abandonment Rate Statistics (Baymard Institute). Review conversion findings are from Northwestern University's Spiegel Research Center analysis of approximately 57,000 products.*

They call it an R2S2R engine: real-to-sim-to-real. The idea is exactly what it sounds like. Instead of deploying a robot into a warehouse and watching it fail until it learns, you build a generative simulator of the physical world. The robot acts inside that simulation, fails safely, and learns what works. Only then does it act in the real world.

The result: robots that learn complex manipulation tasks with zero real-world training data, and teams that can predict which robotic policies will succeed or fail before deploying a single physical unit.

The simulator is the linchpin.

That was Fei-Fei's framing, and we think she's right. We think the same logic applies to commerce — a point our CTO, Armando Murga, raised on LinkedIn, the day the announcement went out.

The same principle, one layer up

World Labs is simulating the physical world so robots can predict real-world outcomes before acting. At eLLMo, we simulate the human buyer — so DTC brands can predict conversion outcomes before spending real ad dollars.

The parallel is a thesis, not a claim of technical equivalence. World Labs builds embodied AI for warehouses, labs, and homes. We build synthetic buyer agents for product pages and ad content. But the underlying logic is identical: simulate the world, predict what will happen, then act.

For a robot, the world is a shelf, a conveyor belt, a box that needs to be picked and packed. For a $30M DTC brand, the world is a buyer — a real person with a specific intent, navigating your product page, reading your copy, deciding whether to add to cart or leave. The question both teams are trying to answer is the same: will this work, before we commit to it?

Why DTC brands haven't had this — until now

For the last decade, DTC brands have operated in what we'd call the react-and-learn loop: launch the campaign, drive traffic, read the bounce rates, A/B test the headline, repeat.

Every signal comes from the real world, after the money is already spent. You find out a product page doesn't convert after you've driven $50,000 in traffic to it. You find out buyers are dropping at the third scroll after your ad agency has already wrapped the campaign. The feedback loop is slow, expensive, and backwards.

This wasn't a choice. Until recently, there was no way to simulate a buyer — to model how a specific human with specific intent would navigate your site and your content before any real dollars moved. AI changes that.

What buyer simulation actually means

When we say synthetic buyer simulation, we mean a digital twin of your customer: an agent trained on real buyer behavior, tuned to the commerce context, that navigates your site the way a real person would.

It reads your product page the way a buyer reads it — not the way your CMO reads it. It processes your ad copy the way someone on a phone, in 30 seconds, distracted, would process it. It surfaces where buyers get stuck, what signals change their answer, what would tip them from maybe to add to cart.

This is not analytics. Analytics tells you what happened. Simulation tells you what will happen — before it happens. The simulation runs before you touch the product page, before the campaign launches, before the spend goes out.

What we're seeing in practice

A Canadian healthcare company ran our buyer and AEO simulation to understand their content. AI brand visibility climbed 14%, surpassing competitors they'd been chasing for a year.

A DTC premium meat company ran it on a product page. Product clicks up 15.4%. Conversion up 10%. Bounce down 8.2%.

These are not A/B test results. They came from running a simulation first, identifying exactly what was broken, and fixing it before the traffic arrived.

The world Fei-Fei is describing

In her announcement, Fei-Fei noted that World Labs' simulator is policy- and embodiment-agnostic — meaning it doesn't care what kind of robot you have, or what task it's trying to perform. The simulator is the generalized layer beneath them all.

That's the right architectural instinct. The simulation layer is the most valuable part because it removes the dependency on real-world feedback cycles that are expensive and slow.

For commerce, the equivalent of policy-agnostic is brand-agnostic. The simulation runs against any product page, any ad creative, any buyer persona. The underlying layer — a commerce-tuned persona corpus trained on real buyer behavior — is what makes the simulation accurate at the individual intent level.

Gabriel Millien, commenting on Fei-Fei's post, raised the test that matters: does this hold up for hours, under real operational conditions, when things drift slightly from training? It's the right pressure point for robots. It's the right pressure point for buyer simulation too. A snapshot simulation run once before a campaign is useful. A continuously updated simulation that re-runs as your catalog changes, your copy updates, and your buyer context shifts — that's the linchpin.

What this means if you're a DTC brand spending on ads right now

You are already making decisions under uncertainty. The question is whether you want to reduce that uncertainty before you spend.

AI search has changed the buyer journey. Buyers are now forming brand preferences inside ChatGPT, Perplexity, and Google's AI Overviews before they ever click to your site. By the time they land on your product page, they have expectations set by an AI they interacted with two minutes ago.

Your page either confirms those expectations, or it doesn't. Simulation tells you which one it is before a real buyer finds out.

Frontier AI labs are converging on a single idea: you don't need to act in the real world to learn what works in the real world. Simulate the world. Predict the outcome. Then act. World Labs is proving it for robots. eLLMo is built on the same thesis — for the buyer standing between your brand and your next conversion.

*Related Links: World Labs' R2S2R engine announcement, via Armando Murga on LinkedIn, and the original announcement thread.*

For D2C brands, that layer is unavoidable: 35% of US consumers say they use AI tools at the product discovery stage, versus 13.6% who use search. AI has officially overtaken search as where people start looking. The bigger problem? The current ways of measuring success no longer describe what's actually converting.

Here are some of the key highlights we think are worth mentioning from their 2026 report:

The market fragmented, and the winners aren't who you'd guess

ChatGPT still holds 79% of generative AI traffic as of January 2026, but its share is declining as Gemini and Claude gain ground. Gemini grew 157% in five months to 1.1 billion monthly visits; Claude's unique visitors grew roughly 180% H2 vs. H1 2025. There is no longer one "AI search" to optimize for, and a strategy built around any single platform is already out of date.

Inside that fragmentation, Similarweb's Brand Visibility Index that benchmarks AI citation share across Finance, Travel, Consumer Electronics, Beauty, Fashion, and News, found that AI visibility rank and brand size rank consistently diverge. Apple leads consumer electronics, but specialist retailers like Adorama, Crutchfield, iFixit, and Backmarket are cited by AI far more than their market share predicts. CeraVe outranks bigger beauty brands; NerdWallet and Travelmath outrank category giants in finance and travel. The pattern holds across Similarweb's findings: being the best answer to a specific question beats being the biggest brand in the room. The early-adopter advantage isn't ad spend, but structured, question-answering content published early enough to be indexed and trust before competitors knew there was a layer to win.

Citation and traffic are landing in different places

SEO strategist Aleyda Solis's contribution to the report found that 65% of URLs ChatGPT cites sit two or three folders deep in a site's structure, while 58.8% of AI referral traffic that does click through lands on homepages. AI reads your deep, specific content to build its answer, then sends the few users who click somewhere else entirely. This is a clear sign that GEO strategy can't stop at the homepage.

Clicks are also getting scarcer, not from declining relevance but from AI succeeding at what search increasingly does too. Similarweb's synthesis shows AI platform visits climbing toward 1.5 billion monthly while AI referral traffic has sat flat between 240–280 million a month since mid-2025. Chatbots have become closed-loop research environments where users compare and decide without clicking out, the same zero-click pattern behind 68% of Google searches. Google's referral rate holds at 17–19%; AI Mode's sits at 1.6–2.5%.

The funnel didn't shrink. It went dark.

Similarweb's Downstream Impact study tracked users who got a ChatGPT brand recommendation, left without clicking, then were followed for seven days. Brands ChatGPT mentioned were 2.5x more likely to get a site visit in that window than brands it didn't mention, but 55.9% of that traffic arrived through branded search, not an AI referral link. The conversation that created the intent left no trace; Google got the credit. Those visitors also convert better: roughly 7% versus 5% for other traffic on transactional sites, with nearly double the pages viewed and time on site. The intent AI creates is real and high quality, and it's invisible to a dashboard built to credit the last channel touched.

What this means for D2C marketing

D2C brands are being asked to define "performance" on platforms that fragment faster than any team can instrument, using attribution built for a funnel with visible steps. Brand size is no longer a hedge: the specialists Similarweb highlights prove AI rewards structured authority over scale.

That's the layer eLLMo AI operates in. eLLMo is the agentic layer that helps D2C brands understand and improve how they show up inside agentic commerce, not by guessing at prompts, but by testing pages against real-world psychographic buyer profiles the way AI agents actually evaluate them: what claims get trusted, what's missing, where a segment's hesitation starts. Instead of waiting on a referral line Similarweb's own data shows is structurally flattening, brands get a direct read on their visibility and credibility where discovery now happens first.

The brands that win this next era won't have the biggest budgets. They'll be the ones who treat citation the way they used to treat search rank: a metric to own and compound, starting now, while the rest of the category is still measuring a funnel that's already moved.

*Sources: The 2026 Generative AI Landscape Report, The 2026 Generative AI Brand Visibility Index, AI Search Trends: What Changed In The Past Year, AI-Recommended Brands Saw 2.5x More Site Visits: Similarweb (Search Engine Journal).*

Those are different questions, and a five star average cannot tell them apart.

The change worth copying

A sleep and bedding brand we ran test buyers for is adding a concern selector to its reviews. When a reviewer writes, they say what brought them to the product: back pain, sleeping hot, side sleeping, a partner who moves, a bad mattress they are replacing.

That single field converts an undifferentiated feed into something a buyer can navigate. Someone whose problem is overheating can read reviews written by people with the same problem and skip the ones about lumbar support.

The buyer does not want more reviews. They want the three reviews written by someone like them.

Why this works better than more volume

Reviews carry real conversion weight. Northwestern University's Spiegel Research Center, studying around 57,000 products, found the first review lifts conversion roughly 65% on average, and about 190% for products above $100. The same research found perfect ratings convert worse than very good ones, with the sweet spot between 4.2 and 4.5 stars.

Those two findings sit together because reviews persuade by looking like evidence rather than promotion. Anything that makes a review feel more specific and less curated increases its power. Anything that makes it feel generic reduces it.

A concern tag is close to the cheapest specificity you can add. It does not require the reviewer to write better prose. It records the context that makes their prose usable.

Where the mechanism pays off

• It answers fit rather than quality. Fit is what stops considered purchases. A buyer who believes the product is good but is unsure it addresses their situation does not buy, and never tells you that was the reason.
• It makes small review counts work harder. A newer product with 40 tagged, filterable reviews can out-persuade an older one with 400 in a single pile, because the buyer can find the relevant ones.
• It reduces returns. A buyer who bought after reading someone with the same problem had their expectations set by a realistic account rather than a marketing claim.
• It generates first-party data you can use. Concern tags tell you which problems your buyers think you solve, in their words, at scale. That is a segmentation input, a merchandising input, and an ad copy input.

How to implement it without a rebuild

• Ask at review time, not at signup. One required selector in the review form, drawn from a short fixed list. Free text here defeats the purpose.
• Keep the list under about eight options. Long lists get skipped and fragment the data into groups too small to be useful.
• Surface the filter above the reviews, not inside a menu. If the buyer cannot see that filtering is possible, it does not exist.
• Show counts on each tag. Nine reviews from people who sleep hot is a specific promise. A tag with no count is a gamble the buyer may not take.

Ahead of the season

Peak season brings a surge of gift buyers and first-time buyers, and both are buying for a situation they may only partly understand. Someone buying for a relative with a bad back is not going to read forty general reviews. They are looking for one review from someone with a bad back.

If your reviews cannot be filtered to that, the gift buyer either guesses or leaves. Both cost you, and one comes back as a return in January.

eLLMo runs test buyers matched to your real customers against your page and reports what each one went looking for and whether they found it. Review relevance is one of the most consistent findings, because buyers describe the search as they do it. See a live run while there is still time to ship the change.

*Related Links: 50 Cart Abandonment Rate Statistics (Baymard Institute). Review conversion findings are from Northwestern University's Spiegel Research Center analysis of approximately 57,000 products.*

The implementation

A pet nutrition brand we ran test buyers for handles its recurring orders through a workaround familiar to anyone who has built subscriptions on a standard commerce platform. The recurring item is represented as a product variant priced at zero, which is substituted for the real item during fulfillment.

It works. Orders flow, food ships, subscribers stay subscribed. As an engineering answer to a platform limitation it is entirely reasonable.

The cost is not operational. It is strategic, and it shows up as a list of things the team cannot easily test.

What a zero-price variant forecloses

• Showing the real economics on the page. If the recurring item is modelled at zero, the page has no clean number to display for what the subscription actually costs over a year. That is precisely the number a buyer needs to evaluate the offer, and its absence is one of the most common reasons attach rates disappoint.
• Testing the subscription price independently. When the price is derived from a swap rather than set as a price, a straightforward pricing test becomes a fulfillment change. The experiment now needs operations sign-off, which usually means it does not run.
• Discounting the first shipment. A common high-performing subscription pattern is a larger discount on the first order, to lower the cost of trying. A zero-price variant makes that awkward, because there is no first-order price to discount.
• Clean reporting. Revenue attribution across a swapped variant tends to need manual reconciliation, so the numbers arrive late and get trusted less.

The architecture did not just make some experiments harder. It quietly removed them from the list anyone proposes.

Why this matters more than it looks

Baymard finds that 12% of abandoning shoppers cite being unable to calculate the total cost upfront. For a subscription, calculating the total cost is the entire decision. If the platform representation cannot express the price, the page cannot explain the value, and the buyer defaults to the one-time purchase.

The chain runs from a data model, through a page, to a buyer's decision. Most teams only look at the last link.

How to work the problem

• Write down the experiments you are not running. Not the backlog. The ideas abandoned in the first five minutes because someone says the platform cannot do that. That list is the real cost of the architecture, and it is invisible because nothing was ever ticketed.
• Separate presentation from fulfillment. The page does not have to display what the order object contains. The annual cost, the per-shipment cost, and the savings can often all be presented accurately from configuration, without touching how the order is built.
• Price the change against the season. A platform change before peak traffic is risky. A copy and presentation change that solves most of the problem is not. Know which one you are proposing.
• Test the buyer's understanding first. Before rebuilding anything, establish whether buyers actually fail to understand the offer. If they do, you have a business case. If they do not, you have saved a migration.

Ahead of the season

This is work that never feels urgent until the quarter that matters. Peak season is when subscription attach rate does the most for lifetime value, because it is when the largest volume of first-time buyers passes through the page.

You are unlikely to re-architect anything before then. You can find out, in days, whether your subscription offer is being understood, and fix the presentation layer where it is not.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying. When the blocker is that the offer cannot be understood, the buyers say so directly. See a live run before the season closes the window.

*Related Links: 50 Cart Abandonment Rate Statistics (Baymard Institute).*

The constraint

A home cleaning brand we ran test buyers for sells its refill system across all of those surfaces. Any visible price change on the direct site propagates into problems elsewhere.

Two distinct mechanics are at work, and they get confused constantly.

Minimum advertised price governs what a reseller may advertise. It is a contract between a brand and its resellers. A marketplace does not enforce a brand's policy on its behalf, because doing so would resemble resale price maintenance, which carries antitrust exposure.

Price parity is different, and it is what actually bites. A marketplace's automated pricing systems monitor prices elsewhere, including a brand's own domain. A lower price on the direct site can cause the marketplace listing to be suppressed or to lose its default buy option. A brand can have a well-written, well-enforced advertised-price policy and still get caught by parity, because parity is not asking whether anyone broke a rule.

The advertised discount is the one lever that touches every channel at once, which is why it is the one lever you often cannot pull.

What you can still do

The useful reframe is that these mechanics govern the advertised list price. They say much less about value delivered through means that are not a public price cut.

• Non-public offers. Coupons, codes, and account-level pricing operate differently from a public list price change. This is well-trodden ground and worth taking to your counsel rather than your growth team.
• Value in kind rather than value off price. Free shipping on recurring orders, an extended warranty for subscribers, priority replacement, an extra refill on the first shipment. These change the buyer's arithmetic without changing the number a parity system reads.
• Convenience framed as the product. For replenishment, the strongest argument is often not price at all. It is never running out mid-job. Nobody else in your channel mix can sell that, because nobody else knows when the last refill shipped.
• Bundle composition. A subscription that includes something the marketplace listing does not include is not the same product, which changes what parity is comparing.

The messaging problem underneath

Constraint aside, there is a second failure entirely within your control. When brands cannot lead with a discount, they often lead with nothing, and the subscription module becomes a delivery schedule with a small percentage attached.

That is the state we most often find, and it fails for a reason Baymard's data points at directly: 12% of abandoning shoppers cite not being able to calculate the total cost upfront. A subscription whose value is expressed only as a percentage has not been calculated for the buyer at all.

If the percentage cannot move, the presentation has to do more work. Dollars over a year rather than a percentage. Cost per use rather than cost per bottle. The week they will run out rather than a delivery cadence.

Ahead of the season

Peak season is when channel conflict is at its worst, because every surface is discounting at once and parity systems are at their most active. It is also when the largest share of your first-time buyers decides whether to subscribe.

That combination is why presentation work matters more than usual this quarter. The lever you would normally pull is least available exactly when the decision volume is highest.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying, including the moment a buyer cannot work out what an offer is worth. See a live run before your peak traffic decides for itself.

*Related Links: 50 Cart Abandonment Rate Statistics (Baymard Institute).*

The published research is less equivocal on this than the marketing around it. Work out of Stanford in 2024 found that a language model given a rich individual profile could reproduce that specific person's survey responses with roughly 83% to 86% reliability. Other research is blunt about the failure mode: when synthetic respondents stand in for a representative survey population, variance collapses and a large share of statistical relationships shift.

Both findings point at the same conclusion. The quality of a simulated person is dominated by the quality of the description they were built from.

A simulated buyer is only ever as specific as the data used to describe them. Thin input produces confident, useless output.

What thin looks like

Most persona work in commerce is thin in a particular way. It describes a demographic and calls it a person: women 25 to 44, urban, health conscious.

That is a segment, not a buyer. It says nothing about what this person is worried about at the moment they reach a product page, what would make them leave, or what evidence would settle the question. Run a simulation on that and you get output that sounds reasonable and predicts nothing, because there was no signal in the input.

What actually adds signal

A home fitness brand we ran test buyers for is enriching its customer profiles with data that changes the picture: household income band, discretionary spend, and whether there are children in the home.

Consider what each does to a decision about a piece of equipment with a membership attached.

• Presence of children changes both the space calculation and the time calculation. It moves the objection from is this good to where does this live and when would I actually use it, and the page has to answer a question it was not written to answer.
• Discretionary spend determines whether the recurring membership is background noise or a monthly decision that gets re-examined. It predicts subscription retention better than income alone.
• Income band against product price sets how much justification the page has to supply. The same $2,000 machine is an easy purchase for one household and a considered one for another, and the considered buyer needs a cost-per-year framing the easy buyer skips past.

None of that is guesswork. It is first-party data the brand already holds, plus enrichment, turned into the specification that makes a simulated buyer behave like a real one.

The practical sequence

• Start from your actual customers, not your target market. Who buys is frequently not who the brand imagines, and the gap between the two is often the most valuable thing in the exercise.
• Add the fields that change decisions, not the ones that are easy to get. Household composition, replacement cycle, the problem that brought them in. Skip anything that does not plausibly alter what a buyer notices on a page.
• Include the reason for the visit. The highest-signal field is why this person is on this page today. Replacing something broken, acting on a New Year intention, buying a gift, and researching for later are four different buyers who look identical in a demographic profile.
• Keep it inside your privacy commitments. This is aggregate specification work, not individual targeting. Describe segments, not people.

Ahead of the season

Peak season traffic is not your usual traffic. It is heavier on first-time buyers, heavier on gift buyers, and heavier on mobile than any other period of the year. A persona set built only from your existing customer base will under-represent exactly the buyers who arrive in November.

Which is an argument for doing this now rather than in the middle of it. The data you need already exists in your customer records. Turning it into a set of buyers you can test a page against is a matter of weeks, not quarters.

eLLMo runs test buyers matched to your real customers against your product page and returns a ranked list of what stops people from buying. The matching is the part that decides whether the list is worth acting on. For the research on why buyer specification dominates outcomes, see human heterogeneity in agentic markets and preference heterogeneity and persona validity.

See a live run, or bring us your customer profile and we will build the panel before your season starts.

*Related Links: Synthetic respondent reliability findings are from Stanford University research published in 2024 on generative agent simulations of individual survey responses.*

The list is always the same: rewrite the descriptions, add reviews, clarify shipping, fix the navigation, reshoot the imagery, surface the guarantee, rework the calls to action. Every item is plausible. Every item has an internal advocate. Capacity is four of them this quarter.

The base rate nobody plans around

Here is the uncomfortable part. Ronny Kohavi, who built the experimentation platform at Microsoft and later led work at Amazon and Airbnb, puts it plainly: over two thirds of ideas fail to move the metrics they were designed to improve.

It gets starker at the top end. At Airbnb, 250 ideas were tested for search improvement and 20 succeeded. Over 90% failed. Those 20 were worth a 6% improvement in booking conversion and hundreds of millions of dollars, which is the point: the failures are the cost of finding the winners.

But a small team does not get 250 attempts. If you can run four experiments a quarter and two thirds fail by default, the quality of your selection is not a detail. It is most of your result.

More testing is a strategy for a team with unlimited traffic and time. Better selection is the strategy available to everyone else.

Repeated friction is signal. A single comment is noise.

This is where a simulation run earns its place, and where it is easy to misuse.

A run produces a lot of observations. If you treat all of them as findings, you have replaced a backlog of thirty guesses with a backlog of thirty observations, which is not progress. The useful discipline is to look for repetition across different buyers.

One buyer confused by the shipping module is a data point that might be about that buyer. Five buyers with different priorities, budgets, and reasons for visiting, all stopping at the same shipping module, is a page problem. The repetition is the signal, not the eloquence of any individual complaint.

That distinction also tells you what kind of fix you need. A hesitation that shows up across every buyer type is usually structural: something absent, buried, or contradictory. A hesitation concentrated in one type is usually a segment gap: the page serves most people and loses a specific group, often the most risk-sensitive or the least familiar with the category.

Separate the quick fixes from the strategic gaps

Two piles, and they should never compete for the same slot in a roadmap.

Quick fixes are clarity problems. The return window is not stated near the price. Shipping timing is a range with no cutoff date. The review count is styled too quietly to register. These are copy and placement changes, they ship in days, and they are usually the highest ratio of impact to effort available.

Baymard's checkout research estimates the average large ecommerce site could gain a 35.26% lift in conversion from better checkout design alone, and a large share of that is this kind of work rather than anything architectural.

Strategic gaps are positioning problems. Buyers cannot tell why this product differs from the cheaper one. The page speaks to a use case that is not the one most buyers arrive with. The product genuinely does not fit a segment you are paying to acquire. These need research, merchandising decisions, sometimes product changes. They matter more and they do not belong in the same sprint as a copy fix.

The failure mode is putting both on one list sorted by enthusiasm.

Write each finding as a hypothesis

A finding is an observation. A hypothesis is something you can be wrong about, which is what makes it testable.

The translation is mechanical. Take the friction, name the segment, state the expected direction:

• Finding: Several buyers could not tell what the materials were without opening a tab.
• Hypothesis: Surfacing material details above the fold will reduce hesitation for quality-focused buyers and lift add to cart on this page.

That version tells you what to build, who it is for, and what would count as it having worked. It also tells you what would count as it having failed, which is the part teams skip and then argue about afterwards.

Simulation prioritizes. Live traffic decides.

The last step matters and it is the one most easily overstated. A run is a way to choose what to test. It is not a substitute for the test.

The sequence that works: run buyers against the page, collect the friction that repeats, sort it into quick fixes and strategic gaps, write the top few as hypotheses, ship the quick fixes, and put real traffic behind the ones that carry a real bet. You are using simulation to raise the quality of what enters the funnel of experiments, in a world where two thirds of what enters it will not work.

That is a smaller claim than replacing experimentation, and a more useful one. Better inputs, same rigor.

Before the season

This is the moment in the year when selection matters most. Between now and peak traffic there is time for a handful of changes, and after that every change is a risk taken against the quarter that pays for the year.

eLLMo runs test buyers matched to your real customers against your page and returns a ranked list of what stops people from buying, ordered by how widely each blocker appears across the panel. That ranking is the input to the roadmap, not another dashboard. See a live run and bring back a shortlist rather than a backlog.

*Related Links: 1,000 Experiments Club: A Conversation With Ronny Kohavi (AB Tasty), 50 Cart Abandonment Rate Statistics (Baymard Institute).*

Our CEO and Co-Founder, Adrienne Murga, is on the ground at the festival this week — learning from brands, agencies, and talent alike about the shift in how the industry talks about AI and automation. After attending sessions, meeting with clients, and talking with industry leaders, we've summarized the key themes that will define what agentic commerce means for brands thus far, and why understanding the buyer before the agent does is now the work.

Want to see how eLLMo AI helps brands connect with customers — and the agents now shopping on their behalf? Explore the platform.

From AI Hype Cycle to Operating Model

AI has only recently dominated the Cannes conversation, but the trajectory has been visible for three years running:

• 2023: AI hype begins.
• 2024: AI is cautious experimentation — pilots, governance committees, legal review.
• 2025: Generative AI becomes the connective tissue across content, media, and measurement, stitched together by foundation models.
• 2026: The recurring language isn't generation. It's agents, interoperability, and control.

The phrase agentic commerce is being used loosely on the Croisette as ad-tech announcements cluster around adjacent topics such as agent-to-agent conversations, and the governance layer that decides what an agent is allowed to do on a brand's behalf. The hype-cycle framing is being retired in public, on stage, this week as new conversations emerge around what agentic commerce actually means.

What Is Agentic Commerce? The New Frontier in Consumer Decision-Making

Agentic commerce is the model in which an autonomous system — typically built on a foundation model and connected to commerce infrastructure — does three things in sequence, without continuous human prompting:

• Interprets what a customer is trying to accomplish.
• Decides which products and offers best satisfy that intent.
• Purchases on the consumer's behalf across whichever surface is most efficient.

What this means for brands

In an agentic model, the consumer may never visit the page. An agent reads structured information about the product, weighs it against alternatives, and completes the transaction inside a different interface.

With discovery and checkout collapsing into two steps of a single agent turn, a brand's job shifts from winning the click to being legible to the agent deciding on the consumer's behalf.

As agents take on more of the transactional surface, the question of what they are allowed to recommend, refuse, escalate, or negotiate on a brand's behalf becomes existential.

The eLLMo AI advantage: As DTC brands grow, it becomes more important not just to differentiate, but to be discoverable and purchasable by both customers and agents. eLLMo AI Search helps ensure brands land in those top answers — accurately positioned, and driving measurable revenue.

The Customer Intelligence Opportunity in an Agentic World

A recurring note across the festival — yes, even on the entertainment stages — is to build a real model of your audience before the audience shows up. NBCUniversal's Cannes stage discussed global fandom and franchises with Rachel Zoe, Seth Meyers, and Colin Jost. The conversations landed in the same place: a property that scales worldwide has, in effect, simulated its audience well enough to predict how a joke, drama, or special will be received.

Predictive intelligence is the muscle every commerce brand now needs. If an autonomous agent is about to decide on behalf of your customer — on a surface you don't own, against alternatives you can't see — the only durable advantage is knowing how your customer decides before the agent gets the question.

The eLLMo AI advantage: Brands that can express their commercial logic in a form an agent can reliably follow will outperform brands that cannot — but understanding your buyer before an agent does should be the priority. eLLMo AI Simulation is built on the Big Five (OCEAN) personality traits:

• Openness governs how a buyer processes novelty and mechanism.
• Conscientiousness governs demand for completeness.
• Extraversion and Agreeableness shape the weight a buyer places on social proof and community signals.
• Neuroticism governs risk-signal detection.

Conditioning a simulated buyer on OCEAN dimensions isn't a matter of appending a personality description to a prompt. Each dimension is operationalized as a set of behavioral parameters that the empirical literature attributes to that trait level — so the simulated buyer behaves and reacts the way your real customer would, in real time, as they experience your brand.

Understand the Consumer Journey Before You Spend

For two decades, brands have learned what blocks conversion the same way: ship the page, buy the traffic, watch the funnel, run an A/B test, repeat. Every lesson is paid for in live spend, and it arrives weeks — sometimes months — late. That sequence does not survive an agentic surface. By the time you've learned from live traffic that your product page is losing a specific archetype on a specific objection, an agent has already routed that archetype somewhere else hundreds of times.

The eLLMo AI advantage: eLLMo AI Simulation is built for exactly that gap. It runs your product page and creative against hundreds of AI-simulated buyers — synthetic personas scored on the OCEAN Big Five model and shaped to your real buyer profile. Simulation returns actionable insights marketers can deploy the same day.

Simulation defines what will convert and what will stall a customer before purchase — all before any live traffic is deployed — so marketing teams are empowered to make the right creative decisions before the agent shows up. Learn more about Simulation.

Redefining Brand Discovery and Performance in Agentic Commerce

Festivals reward narrative. 2024 brought generative AI to the villa; 2026 brought agentic orchestration. The narrative this week is that the marketing industry has publicly accepted the next operating layer as autonomous and transactional.

Under this narrative, marketers can build a real model of the buyer — the archetypes, the objections, the decision shape — and show up early to the agent layer of their category. Or they can wait to be discovered through someone else's interpretation of them.

Ready to see what your category looks like under agentic conditions before you spend next quarter's budget? Run your landing page through our Simulation.

Here’s the question Cannes doesn’t ask directly: does award-winning creative actually convert the buyer it was made for? Cannes celebrates effectiveness — the Grand Prix Effectiveness Lion exists for exactly this reason — but the measurement happens after the campaign runs. The missing conversation is pre-spend: what does the buyer experience when they land on the page that campaign sends them to? Creative that wins a Lion but sends buyers to an unconverted landing page isn’t doing its job.

AI is being discussed at Cannes primarily as a creative production tool — faster copy, generative imagery, AI-assisted ideation. That’s real, and it’s valuable. But the more transformative AI use case for brand marketers is on the demand side: what does an AI buyer panel tell you about whether this creative actually converts before you commit the production budget? For brands with CTV campaigns on the slate, that question is especially urgent — a CTV storyboard can be pressure-tested against a simulated ICP panel before a frame is shot. The same applies to paid social concepts, native display, and landing pages. You can generate 100 ad variations with AI today. You can also simulate how a panel of your actual ICPs responds to each one, before you spend a dollar on production or media. Most brands are doing the first. Almost none are doing the second.

The gap matters because creative effectiveness and conversion effectiveness are not the same thing. A campaign can be emotionally resonant, culturally relevant, and beautifully crafted — and still send buyers to a page that fails the moment they arrive. The hero copy speaks to the wrong job-to-be-done. The trust signals don’t address the buyer’s specific anxiety. The CTA is there but the product story hasn’t earned it yet. These aren’t creative failures — they’re buyer simulation failures. You can’t fix them with better creative. You fix them by simulating the purchase decision before the campaign runs.

eLLMo Simulation deploys a curated panel of AI buyer agents — calibrated by personality (OCEAN model), purchase motivation, and ICP profile — against your landing page, CTV video ad storyboard, paid social creative, or product story. Each agent returns first-person findings: what stopped them, what built trust, what they wanted to see that wasn’t there. It’s not a focus group and it’s not a survey. It’s behavioral simulation of the purchase decision at the individual level, before you spend on production or traffic to find out the hard way.

The brands leaving Cannes with a competitive edge won’t just be the ones who saw the best creative. They’ll be the ones who ask the next question: how do I know this creative actually works for the buyer I’m targeting, before I run it? That’s the question buyer simulation answers. And it’s available now — not in a future AI roadmap, but in a platform DTC brands and agencies are using today to de-risk their spend before every campaign launch.

If you’re at Cannes and want to talk buyer simulation, Adrienne is there. The conversation about AI and creative effectiveness is just beginning — and the most important part of it hasn’t made it to the main stage yet.

The results overturn the comfortable assumption. The intuition was that once everyone delegates to the same handful of models, outcomes would converge: less haggling, less human messiness, a cleaner market. They did the opposite. Agentic markets inherited the human behind every prompt and the model behind every agent — and in measurable ways, they amplified both.

For anyone whose revenue depends on a buyer saying yes on a page, this is not abstract. The buyer on the other side of your funnel is starting to be an agent, and the rules that govern what it does are not the rules you optimized for.

The frictionless agent is a myth

The cleanest evidence comes from a study of AI-mediated negotiations, where hundreds of people wrote the instructions for buyer and seller agents that then negotiated on their behalf. Every agent ran on the same model with the same objective — maximize surplus. Standard theory predicts near-identical outcomes. That is not what happened.

Roughly 73% of the variation in who got the better deal traced back to the individual who wrote the prompt — not to the model's randomness. The agentic negotiations were 16.5% more dispersed than the same negotiation run between humans. And the fairness norms that usually anchor human deals eroded: humans split the surplus evenly about 35% of the time; their agents did it only 14% of the time.

The mechanism is the prompt. Instructions are not neutral pipes for an objective — they carry the author's assumptions, risk posture, and habits straight into the agent's behavior. Delegation did not wash out human difference. It transmitted it, then stripped away the social norms that used to keep it in check. We go deeper on this in Human Heterogeneity in Agentic Markets.

When the model and the objective are held constant, the instruction is what moves the outcome. Who writes the prompt matters more than which model runs it.

Which agent is reading your page?

If the prompt is one half of the story, the model is the other. A separate benchmark put nine different AI agents on both sides of consumer transactions — buyer and merchant — and let them negotiate price and close deals with no human in the loop.

Outcomes were strongly model-dependent. Stronger reasoning models adapted to budget constraints and adjusted strategy to the negotiation; weaker ones did not. The weakest agents broke their own stated budgets in more than 10% of deals, with one breaching its limit in roughly 18.5% of tight-budget scenarios. Some agents closed more deals but at margins so thin they left money on the table; others held firm and walked away more often.

Read that as a marketer and the implication lands hard: the same page, shown to a strong agent and a weak one, produces different selections and different willingness to pay. Which agent is doing your customer's shopping is now a variable in your conversion rate — one you don't control and mostly can't see. More on this in Agent-to-Agent Commerce and Model-Dependent Outcomes.

Machine fluency is the new conversion skill

Here is the part that should reframe how you think about your page. In the negotiation study, observable traits — demographics, personality, risk tolerance — explained only about 17% of the variation in outcomes. The majority was something the researchers named machine fluency: the skill of getting an agent to actually pursue your objective through natural language.

Machine fluency is a real, unevenly distributed form of human capital. Some people instruct an agent and it does what they meant; others get a plausible-looking result that quietly misses the point. On the selling side, your landing page is your half of that exchange — the input an agent buyer reads to decide what you are offering and whether it is worth it.

What that looks like on a page

Agent buyers reward the same things a skeptical human does, only faster and less forgivingly:

• Lead with verifiable value, not adjectives. Agents discount self-promotion and weight specific, checkable claims.
• Make your constraints and terms explicit. Price, return policy, shipping, guarantees — ambiguity reads as risk, and risk-sensitive agents back away from it.
• Front-load third-party credibility. Editorial endorsements, named sources, and verified reviews carry more weight with an agent than a superlative ever will.
• Put the value before the ask. An agent that hits a price anchor before it understands the value treats it as a manipulation and discounts accordingly.

None of this is new copywriting. It is the discipline of writing for a reader who will not give you the benefit of the doubt — and increasingly, that reader is an agent.

The human buyer is already AI-shaped

Even when the buyer is still a person, the person has changed. In the largest study of its kind, an AI interviewer spoke with roughly 81,000 people across 159 countries in a single week about how they use AI and how they feel about it.

The picture is one of paired tensions: productivity against pressure, support against dependency, opportunity against job loss. People rated themselves substantially more productive — an average of 5.1 on the study's scale — but about one in five described a treadmill, where the time AI saved was immediately eaten by higher expectations. Optimism and anxiety live in the same person, often in the same sentence.

That is the buyer arriving on your page: fluent with AI, hopeful about it, and primed to scrutinize. They research faster, compare harder, and extend less trust to claims that are not backed. A page written for an unhurried, credulous reader misjudges who is actually reading it. See The AI-Shaped Buyer for the full picture.

Why this is an argument for testing before you spend

There is a reason all of this is happening now, and it is the same reason it is newly practical to do something about it. AI has made complex, judgment-heavy work dramatically faster. Analyzing 100,000 real AI conversations, Anthropic estimated that tasks which used to take about 90 minutes were completed roughly 80% faster with AI — and the hardest tasks compressed the most, up to 12x.

Reasoning through how a skeptical buyer reads your page is exactly that kind of task: complex, judgment-heavy, and until recently too slow to run before every campaign. Now it is not. The expensive way to learn that your refund policy is buried or your value lands after your price is to run the campaign and read the post-mortem. The cheap way is to simulate the buyer first. See The Productivity Compression.

What to do this quarter

You do not need to predict the agentic economy to prepare for it. A few moves compound:

• Simulate your page against AI buyers before you spend. Treat it as a pre-flight check, not a post-mortem input.
• Write for verifiability. Replace asserted superlatives with claims an agent — or a skeptical human — can check.
• Order the page for a fast, distrustful reader. Value before price. Credibility before claims. Terms in plain sight.
• Treat the buyer's model as a variable. Pin the model version behind any simulation so your results reflect buyer psychology, not which model happened to run.

The shift, stated plainly

The funnel used to end at a human. Increasingly it ends at an agent acting for a human — and both the agent and the human have been shaped by AI. The research is consistent on what that means: outcomes will not homogenize into a clean equilibrium. They will track the quality of the instruction on one side and the model on the other, with fewer norms to smooth the extremes.

The brands that learn to read that surface — before agent buyers are the majority of their traffic — will be making decisions while everyone else is still guessing. That is the entire case for pre-spend simulation, and the research just made it concrete.

Self-purchasers already know why they want the product. They're evaluating trust signals, price justification, and delivery confidence. The page is usually built for them. Gift-buyers are evaluating something else entirely: will this feel thoughtful? Will the recipient love it? Is there a gift message option? Can I see how it ships?

A product page optimized for self-purchasers is often actively confusing for gift-buyers. The copy talks about your personal experience with the product. The CTAs say 'treat yourself.' The imagery shows the customer using the product, not giving it. None of that speaks to the gift-buyer's job-to-be-done.

In our simulations, gift-buyer conversion intent scores are typically significantly lower than self-buyer scores on pages that aren't gift-aware — even when the product itself is highly gift-appropriate. The moment we recommend one change — adding a gift message field, changing the hero copy to acknowledge gifting, adding a 'perfect for' section — gift-buyer scores jump meaningfully.

The reason this matters: gift-buyers often have higher AOV, lower return rates, and are more likely to introduce a brand to a new household. They're a high-value segment that most CRO programs miss entirely, because standard A/B testing requires you to segment traffic — and most brands don't have enough gift-flagged traffic to reach statistical significance.

eLLMo Simulation doesn't need live traffic. It runs the gift-buyer cohort as a defined ICP segment and returns first-person findings from inside that buyer's perspective. What stops them. What reassures them. What they're looking for that isn't there.

If you sell anything that could conceivably be given as a gift, your product page probably has a hidden conversion gap. The fix is usually cheaper than you think — a few copy changes, an imagery swap, a gift option in the checkout flow. But you can't fix what you haven't measured.

But A/B testing has a structural blind spot, and it is the most expensive one in marketing: it can only tell you things after you have already spent. Every test needs a live page and real traffic, which means the budget is committed before the insight arrives. For decisions you make before a campaign launches, the gold standard is not available — so teams either guess, or they wait.

Here are five things A/B testing cannot tell you before you spend, and what pre-spend simulation gives you instead.

1. It can't tell you why

An A/B test tells you that variant B converted at 3.2% against variant A's 2.8%. It does not tell you which element drove the difference, which buyer it moved, or what psychological mechanism was at work.

That gap matters because it makes the next test slow. You won — but you do not know why, so your follow-up hypotheses are guesses dressed as strategy. You end up testing your way toward an answer one expensive iteration at a time. The what is settled; the why is still a mystery, and the why is what compounds across every page you build after this one.

2. It requires live traffic

You cannot A/B test a campaign that has not launched. Every test is, by definition, a post-spend analysis: the page is live, the media is running, the money is moving. If the test reveals that the landing page has a fundamental problem, you have already paid to send buyers to it.

For a high-stakes launch — a new product, a seasonal push, a paid campaign with real budget behind it — that timing is backwards. The most valuable moment to learn that your page does not land is before the traffic hits it, not after the spend is gone.

3. It takes weeks

Statistical significance at typical traffic volumes takes two to six weeks. Founders, campaign managers, and growth leads rarely have weeks between brief and launch. The test is often still gathering data when the campaign it was meant to inform has already ended.

Speed is not a luxury here. A diagnostic that arrives after the decision is made is a historical record, not an input. The cost of slowness is every decision you had to make on instinct while you waited for the data.

4. It can't segment by buyer psychology

Standard A/B tests segment by what they can readily capture: traffic source, device, geography. They cannot segment by the variables that actually predict conversion — personality, decision style, risk tolerance, purchase motivation.

That is a real limitation, because the same page rarely fails the same way for everyone. A risk-averse buyer abandons over a missing return policy. A detail-oriented buyer stalls on an unsubstantiated claim. A gift-buyer cannot find the one reassurance they need. Aggregate conversion data averages all of that into a single number that hides the actual problem — and the segment that is quietly being lost.

5. It requires an existing asset

You can only test something that already exists. A new landing page, a new product concept, a campaign creative that is still a storyboard — none of it can be validated with a testing framework, because there is nothing live to split traffic against.

So the highest-impact decisions — the ones made at the concept and pre-launch stage, when changes are cheap — are exactly the ones A/B testing cannot touch. By the time the asset exists and has traffic, the cheap window to change it has closed.

What pre-spend simulation does instead

Buyer simulation is built for the gap A/B testing leaves open. Instead of waiting for live traffic, it runs a calibrated panel of AI buyer personas against your page and returns first-person findings: what each buyer noticed, what built confidence, what created doubt, and where they hesitated.

Because the personas are calibrated by personality and purchase motivation, the output is segmented by psychology, not just traffic source. Because it does not need live traffic, it runs before launch — in a day, not a quarter. And because every finding carries the reasoning behind it, you learn the why, not only the what.

A/B testing tells you which version won. Simulation tells you why a version is losing — before you have paid to find out.

None of this replaces A/B testing. Once a page is live and converting traffic, testing is still how you settle the close calls. Simulation answers the questions that come first: what to fix, for whom, and in what order, before the budget is committed. For the research behind why this works, see Simulation Validity and Calibration.

The teams that win the pre-spend window are not the ones who test more. They are the ones who walk into the test already knowing what is wrong.

eLLMo calibrates each synthetic buyer persona across each OCEAN dimension. These calibrations aren't labels — they're behavioral parameters that shape how a persona processes information, what they trust, and what makes them anxious or confident. That creates measurable, predictable differences in how different buyer types respond to the same landing page.

Openness predicts how a buyer responds to novelty. High-Openness buyers are drawn to brands that name novel mechanisms, unique aesthetics, and innovation language. Low-Openness buyers want familiarity, proof, and category leadership claims. If your hero image is minimalist and abstract, high-Openness buyers may love it; low-Openness buyers may not understand what they're looking at.

Conscientiousness drives attention to detail, skepticism of claims, and demand for evidence. High-Conscientiousness buyers read every line of your product description. They notice when a claim is unsubstantiated. They want certifications, specifics, and third-party validation. Low-Conscientiousness buyers may make faster, more emotional decisions — but they're also less patient with dense copy.

Extraversion is less about sociability and more about stimulation preference. High-Extraversion buyers respond to energy, social proof, and community signals. Low-Extraversion buyers are often distracted by busy layouts, multiple competing CTAs, and aggressive social proof widgets.

Agreeableness shapes how buyers respond to persuasion tactics. High-Agreeableness buyers are receptive to testimonials, social proof, and community belonging. Low-Agreeableness buyers are more skeptical of social pressure and respond better to objective evidence, comparisons, and direct value arguments.

Neuroticism is the dimension that predicts anxiety and risk-aversion. High-Neuroticism buyers are highly sensitive to return policy ambiguity, unclear shipping timelines, and missing trust signals. A landing page that doesn't prominently feature a clear return policy, shipping guarantee, and recognizable payment methods will dramatically underperform with this segment — even if everything else is right.

In eLLMo Simulation, each persona's OCEAN profile shapes how it reads your page, what it notices, what it questions, and whether it converts. The findings are traceable back to specific dimensions — which means the fix recommendations are psychologically grounded, not generic CRO advice.

The pattern is consistent: a brand runs a significant paid traffic campaign for BFCM, sees a conversion rate lower than expected, and spends the post-mortem trying to reconstruct what went wrong from aggregate analytics data that can't explain the why. By then, the budget is gone.

The brands breaking this pattern are running pre-spend simulation 60–90 days before BFCM. The process is straightforward: submit your hero PDP, your primary landing page, and your key email-to-page flows. The simulation deploys a curated ICP panel — including the gift-buyer segment, which spikes dramatically during BFCM — and surfaces ranked conversion blockers before you commit a dollar to traffic.

What simulations consistently find in BFCM prep runs: The gift-buyer segment is almost always underserved. Most DTC product pages are written for the self-purchaser. During BFCM, a significant portion of buyers are purchasing for someone else — and they're evaluating different signals: is this gift-appropriate, what's the return policy, is there a gift message option? Pages that don't address this lose a high-AOV segment at scale.

Return policy anxiety goes up at BFCM. High-Neuroticism buyers — those most sensitive to risk — are especially active during gifting season. They scrutinize return windows, shipping guarantees, and refund policies in ways that calmer buyers don't. A return policy buried in the footer is a conversion blocker for this segment in December. During BFCM, that becomes an expensive problem.

Price anchoring matters more at BFCM, not less. The conventional wisdom is that discounts sell themselves. But value-conscious buyers still need context. Cost-per-use, size comparison, or a clear 'originally $X, now $Y' framing all function as cognitive anchors that help buyers justify the purchase. Pages that rely on the discount alone leave conversion on the table.

The simulation doesn't predict your exact BFCM conversion rate — it identifies which page elements are suppressing it and ranks them by projected impact. That gives your design and copy team a ranked, evidence-backed brief in days, not after a campaign has already run. If BFCM matters to your brand, the 90-day prep window is now.

In eLLMo Simulation, the two OCEAN dimensions that most predict trust signal sensitivity are Conscientiousness and Neuroticism. High-Conscientiousness buyers are skeptical of claims — they want evidence, specificity, and substantiation. High-Neuroticism buyers are anxious about risk — they want protection, guarantees, and certainty that things won't go wrong. These are different needs, and they require different signals.

What works for high-Conscientiousness buyers: Third-party certifications, specific ingredient origins, and clinical study citations all work well because they're substantiated. Generic claims like 'dermatologist-tested' with no named dermatologist, or 'premium quality' with no definition of premium, actively reduce trust with this segment. The signal reads as marketing noise, not proof.

What works for high-Neuroticism buyers: Prominent return policy (above the fold or adjacent to the CTA, not in the footer), free shipping threshold clearly stated, recognizable payment options, and a clear customer service contact point. This segment is not looking for quality claims — they're looking for exit ramps. They want to know that if they buy and it doesn't work, the process of returning it is painless and fair.

Where most brands go wrong: The trust signal stack is usually built around the brand's own confidence in their product, not around buyer anxiety. High-quality product photography signals confidence. A 'made with love' story signals authenticity. These build brand connection — but they don't address the pre-purchase anxiety that prevents skeptical buyers from clicking 'add to cart.'

The 'trust badge' trap: Generic trust badge icons — padlock icons, 'secure checkout' text, anonymous review stars — are often ignored by high-Conscientiousness buyers because they provide no substantiation. A four-star average with no review count, no verified purchase flag, and no response from the brand is weaker than 47 reviews with the brand's response to the two negative ones. Volume and specificity matter more than polished presentation.

In simulation, the trust signal fixes are almost always the fastest, cheapest improvements available. Copy changes, CTA placement adjustments, making the return policy more prominent — these are days of work, not weeks. And they consistently produce the largest intent jumps for the segments most likely to abandon. If your page hasn't been audited through the lens of buyer skepticism, it almost certainly has trust signal problems your analytics can't see.