Somebody on your team types the question a customer would type. Best clean protein powder for someone lifting four days a week. Best carry-on under $300. The assistant gives back four brands and a short reason for each. You are not one of them.
The first instinct is to rewrite the product page, because the product page is the part you own. That instinct is not wrong, and it is nowhere near enough. Most of what decided that answer happened somewhere you do not control, and it happened before the question was asked.
Here is what the data says about where citations actually come from, and what a direct to consumer brand can do about it.
Almost all of it is earned, and almost none of it is bought
Muck Rack has been running the same study since July 2025. The May 2026 edition looked at more than 25 million cited links in responses from ChatGPT, Claude, and Gemini, across 17 industries.
Earned media accounted for 84% of citations. Journalism on its own was 27%. Paid and advertorial content came to 0.3%.
The number that matters is not 84%. It is that the figure has barely moved. Across three editions over ten months, earned media has stayed between 82% and 89%, and journalism between 25% and 27%. That is not a quirk of one model or one crawl. It is how these systems are built.
The other useful detail: the models do not behave alike. ChatGPT attached sources to 96% of its responses, Gemini to 82%, and Claude to only 55%. Claude answers plenty of questions about your category without showing its work at all.
Being cited and being named are two different wins
This is the part most teams miss, and it changes what you should be measuring.
Semrush, working with Kevin Indig, tracked 3,981 domain appearances across 115 prompts, 14 countries, and four AI platforms. They separated two outcomes: the domain showing up as a source link, and the brand name showing up in the text of the answer.
61.7% were citations with no brand mention. The link was there in the sources. The name was not in the sentence.
Nobody buys from a footnote. A shopper reads the four brands the assistant named and closes the tab. So the goal is not to be a source. The goal is to be the answer, with the source underneath you.
The same study found the gap moves with the question. Informational prompts named a brand 18% of the time. Comparative prompts named one 43.3% of the time. Short conversational questions produced far more brand mentions than long structured ones.
That is good news for D2C. "Best running shoe for flat feet" is a comparative question asked in eight words. It is the most name-friendly shape of prompt there is, and it is most of your category's search demand.
Your competitor is not outranking you. They are being named while you are being footnoted.
The four things that decide it
We work on citation as four separate gates: clarity, trust, machine readability, and domain authority. The first three are the foundation under all of this, and the case for them is laid out in what AI search actually took. What follows is the citation specific cut of each one: not whether a fact survives being put in someone else’s words, but whether your page is the one a model reaches for in the first place. The fourth gate is what citation work adds, and it is the one most brands have never looked at.
They are separate gates because a brand can pass three and fail one, and the one failure is enough. Most brands we look at are failing two or three at the same time without knowing which.
Clarity: give the model something it can lift
SE Ranking studied 129,000 domains and 216,524 pages across 20 niches to see what predicted a ChatGPT citation. The content findings are consistent and unglamorous.
- Pages carrying an expert quote averaged 4.1 citations. Pages without averaged 2.4.
- Pages with 19 or more data points averaged 5.4. Pages with almost no data averaged 2.8.
- Sections of 120 to 180 words between headings averaged 4.6. Sections under 50 words averaged 2.7.
- Pages updated within three months averaged 6. Pages left alone averaged 3.6.
Read those together and they describe one thing: a fact that can be pulled out of your page and dropped into an answer without losing its meaning. "Fast, free shipping" cannot survive that trip. "Free shipping over $50, delivered in three to five business days" can, because it is still true and still specific after it leaves your site.
Every vague line on your page is a line no model can use to defend you.
Trust: someone else has to say it too
The same study measured what happens when third parties talk about a brand. Domains listed on several review platforms averaged 4.6 to 6.3 citations. Domains absent from all of them averaged 1.8. Domains with heavy Reddit presence averaged 7 against 1.8 for domains with almost none. Quora showed the same shape.
Ahrefs put a number on the same effect from a different angle. Across 75,000 brands, branded web mentions correlated with AI visibility at 0.664. Backlinks correlated at 0.218. The three strongest signals in that study were all off your site.
A decade of SEO taught marketing teams to think in links. These systems think in agreement. The question they are answering is not whether reputable sites point at you. It is whether the rest of the internet says the same thing about you that you say about yourself.
For a D2C brand, the practical version of that is short. Claimed and current profiles on the review platforms your category actually uses. Real presence in the forums where your customers compare products. Press coverage that repeats your specific facts, not your tagline. A founder or expert who is quotable by name.
Machine readability: the parser reads first
This is where most advice goes wrong, so it is worth being precise.
There is a finding here that gets quoted badly, so it is worth handling properly. In the SE Ranking data, pages with FAQ schema averaged 3.6 citations against 4.2 for pages without, and llms.txt files showed almost no effect at all. That gets passed around as proof that answer content does not work for AI search. It is not that, and the reason sits in what the number counts. It counts citations, meaning how often a page gets pulled in as a source. It says nothing about what happened next.
A page that settles the question is a different kind of source from a page that half settles it. Perplexity stacks around 19 sources into a single answer. When your page ends the question on its own, it gets used once and cleanly. When it leaves gaps, the model goes and fills them, and every one of those fetches counts as a citation for somebody. Being sampled more often is not the same as being the answer more often.
The tag is still not the lever. A line of markup announcing that your page contains an answer does not create the answer, and llms.txt is a note left for a reader that mostly does not read notes. Write the answer. Mark it up too, because it is cheap and it is correct. Just do not read a lower citation count as a failure, because it is not counting the thing you are trying to win.
What did track was whether the content arrived at all. Pages with a First Contentful Paint under 0.4 seconds averaged 6.7 citations. Pages over 1.13 seconds averaged 2.1. A crawler working through thousands of candidates does not wait for a slow page to finish assembling itself. It takes whatever loaded. The specific ways a retail page fails that test are in product pages AI cannot read.
Machine readability is not decoration on top of the page. It is whether your price, your policy, your ingredients, and your sizing survive the trip from your server into a model's context window. If the size chart is an image, it does not exist. If the return policy loads after a click, it does not exist. If the spec table is built by a script that a crawler never runs, it does not exist. Mark it up as well, by all means. Just do not mistake the label for the thing.
Domain authority: a gate, not a leaderboard
Authority is real and the numbers are blunt about it. In the SE Ranking data, domains with a trust score below 43 averaged 1.6 citations. Domains scoring 97 to 100 averaged 8.4. Referring domains showed a threshold effect, with average citations stepping from 2.9 to 5.6 once a domain passed roughly 32,000.
If you are doing $30 million a year, that reads like a wall. It is not, and two other studies show why.
Evertune looked at 200 million prompts over five months and found that even the most-cited domain on any given platform rarely passes 5% of total citations. The remaining 95% is spread across thousands of domains. There is no top ten to be locked out of.
Orbit Media tracked 13,184 citations across 1,765 answers from ChatGPT, Claude, Gemini, and Perplexity. All four models cited the same domain for the same question 30 times, which is 1.7% of the set.
So authority works as a filter on whether you are eligible, not as a ranking that hands the answer to whoever is biggest. You do not have to out-authority Amazon. You have to be one of a small number of credible sources that clearly answers a narrow question, in enough places that agreement is visible.
That is a winnable game for a mid-sized brand and an unwinnable one for a brand that only talks about itself on its own website.
Count the arrivals, not the citations
If citation volume is the wrong scoreboard, here is a better one. Who showed up, and what did they do when they got there.
A member of Stripe’s go to market team gave us the headline version of it in an interview: when someone uses an AI agent to shop, they convert four times more than a regular ecommerce shopper. That is a read from the company that settles the payment, not a survey asking shoppers what they intend to do.
The settled revenue supports the shape of it. Attrifast benchmarked 168,000 Stripe payment events across 200 Stripe-connected sites and 41.2 million sessions, in the thirty days to 15 May 2026. Three numbers from the ecommerce cut are worth your time.
- First orders ran 43% larger. First-time buyers arriving from AI engines averaged $112.40 from Perplexity and $87.40 from ChatGPT, against $61.20 from Google organic.
- They sent less of it back. Thirty day refund rates were 3.8% for AI-referred orders, against 6.1% for organic search and 8.4% for paid search.
- They skipped the front door. AI traffic landed on a deep page 64% of the time and on a homepage 12% of the time. Google organic landed on a homepage 24% of the time.
That last one is the tell for anyone wondering whether their answer content is doing any work. The assistant did not drop the shopper on your homepage to start looking around. It sent them to the page that answers the question, because it had already read the answer and built the shortlist.
The refund rate is the one to sit with. People who arrive already briefed buy the right thing the first time. They also argue less about price, which is what we found in one store’s own order data, where AI referrals took the smallest discount of any channel.
Stripe’s four times and Attrifast’s numbers are measuring different populations, and the gap between them is instructive rather than awkward. Stripe is describing someone shopping through an agent. Attrifast is measuring someone who clicked a link out of an answer. The further the assistant carries the decision before handing it over, the warmer the buyer lands.
None of this proves that one FAQ page caused one sale. These are separate datasets on separate sites, and anyone selling you that line is selling you something. What it does establish is which way the error runs. Count citations, and a page that ends the question scores worse than a page that prolongs it. Count arrivals, and it scores like what it is.
A citation is a fetch. An arrival is a buyer. Only one of the two turns up in Stripe.
Why this is infrastructure and not a campaign
One more finding from Orbit Media, and it is the one that should shape your budget.
Week to week, Perplexity kept 65% of the URLs it had cited before. Gemini kept 38%. Claude kept 30%. Most of what gets cited about your category this week is not what got cited last week.
Put that next to the 1.7% agreement rate and the shape of the work becomes clear. There is no list to climb and no single answer to win. A campaign that wins you a good week is a campaign that has to be run again.
What holds is the substrate underneath. A set of facts about your brand that are specific, consistent everywhere they appear, corroborated by people who do not work for you, and readable by a machine on the first pass. That gets re-derived every time a model refreshes, on every platform, for questions nobody has thought to track yet. It compounds. A campaign decays.
This is the same reason the work does not finish. Your catalog changes, your policies change, and the models re-crawl on their own schedule. Citation is a system you run, closer to inventory management than to a launch.
Where we come in
eLLMo builds this as infrastructure, mostly for brands between $10 million and $150 million a year. That band is the interesting one. You are big enough to have real product data, real customers with opinions, and press worth earning. You are small enough that no model has memorized you, so nothing is decided yet.
We run the four gates as four workstreams, in that order, because they build on each other. Clear facts give third parties something specific to repeat. Repetition creates the agreement that reads as trust. Machine readability makes sure the facts survive the crawl. Authority decides how far all of it travels. Brands we run this for have typically seen their visibility to AI agents rise two to three times within 60 days, and the reason it moves that fast is unflattering: almost nobody is failing at one gate. They are failing at three, so the first fixes are cheap.
The starting point is finding out which gate is yours. See an example AI search audit to see how a brand's current standing gets pulled apart, or read what makes an AI agent pick your brand for how the choice gets made on the other side.
The brands that win the next few years of discovery will not be the ones with the loudest campaign. They will be the ones a machine can read, check against three other sources, and repeat with confidence.
*Related Links: What Is AI Reading? May 2026 (Muck Rack), The Ghost Citations Study (Semrush and Kevin Indig), Top Factors Influencing ChatGPT Citations (Search Engine Journal, on SE Ranking data), An Analysis of AI Overview Brand Visibility Factors (Ahrefs), LLM Citation Study (Orbit Media), The 2026 AI Search Revenue Benchmark (Attrifast, on 200 Stripe-connected sites), A third of retail product pages cannot be read by AI.*