Your website is the receipt, not the reason: why AI citation studies disagree about brand sites

If you read AI search research this week, you could come away with two opposite conclusions. One says AI cites your website 80% of the time. Another says 15%. A third says 6.5%. A brand manager trying to decide whether to spend next quarter on the website or on getting reviewed could reasonably throw up their hands.

We don't think the studies disagree. We think they are measuring two different jobs your website does for AI, and the gap between the numbers is the most useful thing in them.

Three studies, three numbers

On November 3, 2026, the UK agency Tilio published 30 days of tracking for four anonymised UK client brands, each in a different industry: 44,587 answers from ChatGPT, Google AI Overviews and Google AI Mode for all four, Gemini for three and Claude for one, to questions that do not mention the brand by name. Two findings sit side by side. Brand websites were 3.5% of all 270,861 citations. But when an answer named one of the brands, it cited that brand's own site 80% of the time, from 71% to 90% by brand.

On November 4, Babel42 published three days of answers from ChatGPT, Perplexity, Gemini, Google AI Overviews and Google AI Mode to 257 US questions in 10 industries. For 11,147 brand recommendations, the same answer cited the brand's own site 15% of the time. In beauty and fragrance it was 1.6%. We break that study down in today's data post.

Also on November 4, QuQi published 3,906 answers for 21 business websites from late August to October 21. In 922 Google AI Overviews, the business was named in the text 10.2% of the time and one of its pages was a source 6.5% of the time.

Show the data

Study and measure	Own site cited
Tilio: answers naming the brand	80%
Shero Commerce: brand recommendations (via Tilio)	31%
Babel42: brand recommendations, 10 industries	15%
QuQi: AI Overviews, any answer	6.5%
Tilio: share of all citations	3.5%
Babel42: beauty recommendations	1.6%

Source: Tilio (Nov 3, 2026), Babel42 (Nov 4, 2026), QuQi (Nov 4, 2026). Not like-for-like. Tilio: 4 UK client brands. Babel42: 5 US assistants, buying questions. QuQi: 21 business sites, Google AI Overviews only. Shero Commerce figure as reported by Tilio (159 recommendations).

Tilio, to its credit, lists other studies next to its own. It cites Shero Commerce, which found the brand's own page cited in 31% of 159 brand recommendations across 60 product categories, and Scrunch, which put brands' own sites at 5.95% of more than 442 million citations. Yext, counting listings, reviews and social profiles as brand-managed, got a very different 44% for brand websites alone. Everyone is right about their own sample.

What separates the high numbers from the low ones

Look at what each high number is conditioned on. Tilio's 80% only counts answers that already named the brand. Once an assistant has decided to talk about you, it often goes and gets your page to confirm the details. That makes sense: if the answer says, to take a made-up example, "Brand C offers same-day delivery in London," the most reliable source for that sentence is Brand C.

Babel42's 15% is conditioned on something different: a brand being recommended in answer to a generic buying question. "What's the best reef-safe sunscreen?" The assistant is choosing, not confirming. And for that job, it leans on people who compare products: editors, dermatologists, video reviewers, retailers. Babel42 found beauty answers in Perplexity cited Allure 52% of the time and Byrdie 51%, while ChatGPT cited the American Academy of Dermatology in 63%.

QuQi's 6.5% and Tilio's 3.5% are conditioned on nothing at all: every answer in the category, including the large majority that never mention the brand. Tilio found 86% of its answers did not name the brand, and in those, the brand's site was cited fewer than 4 times in 100.

So there is a pattern. The closer a study gets to "the AI has already picked you," the more often your site is cited. The closer it gets to "the AI is picking," the less often.

The receipt and the reason

Here is the frame we would use. Your website is the receipt: the page an assistant pulls to confirm what it is about to say about you. The reason you are in the answer at all is mostly somewhere else.

For a software company or a law firm, the receipt and the reason often live on the same site, because the company's own pages are the most detailed description of what it does. Babel42 found HR and business software brands had their own site cited in about 23% of recommendations. For a consumer brand, the reason almost always lives off-site. Nobody believes a moisturizer's own claim that it is the best moisturizer. They believe the dermatologist, the editor and the reviews.

Babel42's per-assistant numbers fit the same idea. Google AI Overviews cited the brand's own site in 26.2% of its 2,053 recommendations; Gemini in 6%.

Show the data

Assistant	Own site cited
Google AI Overviews	26.2%
ChatGPT	23.8%
Google AI Mode	16.1%
Perplexity	11.2%
Gemini	6%

Source: Babel42, AI visibility study 2026, Nov 4, 2026. Recommendations with a known official site: AI Overviews 2,053, ChatGPT 1,187, AI Mode 2,096, Perplexity 2,243, Gemini 3,568.

Even the most generous assistant cites the brand's own site in roughly one recommendation in four. Most of the time the receipt is someone else's page too.

Why this matters for how CPG teams spend money

We see two common mistakes, and the studies suggest both are expensive.

Mistake one: treating the website as the whole AI strategy. A team rewrites product pages for "AI readiness," adds schema, adds FAQs, and waits. Those are good things to do, and we have recommended many of them. But if the brand is not already being named, the site work mostly improves the receipt for an answer that is not being written. The tip on what content gets cited is still right; it just is not the whole job for consumer products.

Mistake two: ignoring the website because "AI cites other sites." The opposite reaction is just as risky. Tilio's data says that once you are named, your site is very likely to be read. If that page says something vague, outdated or different from your retailer listings, the assistant either repeats the error or looks elsewhere for the fact. Our older post on the website as a reference book for robots is about exactly this job, and tip 5 on consistent facts is the short version.

The work that follows from the receipt and reason frame is simple to state:

• For the reason: get accurately described on the pages that make the choice in your category. For beauty that means editors, dermatology sources and video. For supplements, Babel42's data points to health publishers, the NIH and third-party testing like NSF, and Amazon. For food, it is likely recipe sites, retailers and nutrition sources, though Babel42 did not study food. The tip on getting into cited sources has a method.
• For the receipt: make the facts on your own pages complete, specific and identical to what the retailer and the pack say. Ingredients, sizes, certifications, claims you can back up.

A test you can run this week

You can see both steps for your own brand in under an hour. Write ten open questions a shopper would ask without knowing your name ("best unscented laundry detergent for babies") and ten that include it ("is Brand X detergent safe for baby clothes"). Ask each in two assistants. For every answer, note two things: whether your brand is named, and whether your own site is among the sources.

If the studies are a guide, the open questions will rarely cite your site even when they name you, and the branded questions will cite it often. If your branded answers do not cite your site, your receipt has a problem: the page may be blocked, thin or contradicted elsewhere. If your open answers never name you at all, the problem is the reason, and no amount of on-site work will fix it alone. Reading these studies together, we would expect most consumer brands to find the second problem, even if they have been working on the first.

What we would not conclude

It would be easy to turn this into a slogan, "your website doesn't matter," and we think that would be wrong. The studies are small in places. Tilio's four brands are anonymous clients in unnamed industries. Babel42 used API versions of three assistants and only three days. QuQi's 21 businesses chose their own keywords. None of them tells us what happens to sales.

We also do not know which way the cause runs. A brand that is already widely reviewed probably also has a well-built site. The studies show what goes together, and every one of them says so.

What they do show, together, is that "how often does AI cite brand websites" is the wrong question. The better question is: at which step? At the choosing step, consumer brands mostly depend on other people's pages. At the confirming step, they depend on their own. A brand that only works on one step will see one of these numbers and think it has the full picture. XENA Intelligence looks at the choosing step for consumer products, asking assistants the open questions shoppers ask and scoring the answers in SearchPanel; any team can do a smaller version by asking ten generic questions and ten branded ones and noting which pages each answer cites.

The receipt matters. It just does not get you into the store. For that, as we argued in famous is not the same as recommended, you need other people to say it first.