AI recommends consumer brands on other people's pages: Babel42's 11,147 picks, read for CPG

When an AI assistant tells a shopper to buy your moisturizer, what page is it pointing to as proof? A study published on November 4, 2026 gives the clearest answer we have seen so far, and for consumer brands it is mostly not your page.

Babel42, a company that sells AI visibility tracking, asked five assistants the same 257 US English questions once a day on November 1, 2 and 3, 2026. The assistants were ChatGPT, Perplexity and Gemini, queried through each company's API with web search on from a US location, plus Google AI Overviews and Google AI Mode collected from US Google results. That produced 3,776 answers. The team recorded every site each answer cited and, for the 188 buying questions that named products or companies, every brand each assistant recommended.

The headline number: for 11,147 brand recommendations where Babel42 knew the brand's official site, the same answer cited that site 15% of the time. Even the brand named first had its own site cited only 21% of the time. The study's own summary of what fills the gap is blunt: the rest of the time, the evidence behind the recommendation was someone else's page.

The consumer categories are at the bottom

The 15% average hides a split that matters to anyone selling a physical product. Babel42 covered 10 industries. Business software, HR and legal services brands had their own site cited in roughly one recommendation in five. The four categories closest to CPG sat at the bottom of the table.

Show the data

Industry	Own site cited
HR and recruiting	23.5%
Business software	23.3%
Legal services	21.7%
Personal finance	19.4%
Travel and hotels	16.4%
Marketing services	16.2%
Health and supplements	8.6%
Sportswear and fitness	4.9%
Gifts	4.3%
Beauty and fragrance	1.6%

Source: Babel42, AI visibility study 2026, Nov 4, 2026. Brands named in buying answers with a known official site. Consumer categories highlighted.

Beauty and fragrance is the extreme. Of 799 beauty recommendations, the brand's own site was cited 1.6% of the time. Put plainly, when ChatGPT, Gemini or Google tells someone to try a cleanser, it almost never links to the cleanser maker as its evidence. Health and supplements did a little better at 8.6% of 952 recommendations, and sportswear and fitness sat at 4.9% of 1,216.

Why the gap? Babel42 does not test causes, and we should not pretend to know them. A plausible reading is that a payroll company's own site is often the most detailed public description of its features and pricing, while a moisturizer's own site mostly repeats what the box says. Reviewers, dermatologists and retailers add things the brand cannot credibly say about itself: comparisons, test results, prices across stores, and other people's experience.

Which pages do get cited

If the brand's page is not the evidence, what is? Babel42 published the top five sources per assistant per industry. For beauty, the lists look nothing like each other.

Assistant	No. 1	No. 2	No. 3
Google AI Overviews	YouTube, 76%	Reddit, 31%	Byrdie, 26%
Google AI Mode	YouTube, 27%	Reddit, 23%	New York Times, 15%
Perplexity	Allure, 52%	Byrdie, 51%	Everyday Health, 47%
Gemini	YouTube, 16%	Ulta, 11%	Dermondemand, 7%
ChatGPT	American Academy of Dermatology, 63%	FDA, 17%	DermNet NZ, 5%

Source: Babel42, November 4, 2026. Beauty and fragrance answers, pooled over three days; 62 to 75 answers per assistant.

Three different strategies fall out of that table. Google's two AI answers lean on video and forums. Perplexity leans on beauty editors. ChatGPT, in this API setup, leans on medical and regulatory bodies. A beauty brand that only does influencer video is well placed for AI Overviews and absent from Perplexity's evidence. A brand that only does magazine placements has the reverse problem. And none of those placements does anything for ChatGPT if the claims on the product contradict what the dermatology bodies say about the ingredient.

Health and supplements shows the same pattern with different names. Babel42 found ChatGPT cited the US National Institutes of Health in 68% of its supplement answers and NSF in 17%, while Perplexity cited Healthline in 65% and Google AI Overviews cited YouTube in 84% and Amazon in 43%. For a supplement brand, third-party testing and an accurate Amazon listing look like evidence the assistants actually read.

The assistants agree on brands, not on sources

The most useful finding for planning may be this one. On 344 occasions all five assistants named brands for the same buying question on the same day. They had at least one brand in common 266 times, which is 77%. In contrast, when all five cited sources for the same question, only 8% of the time was there a single site all five cited.

Show the data

Measure	Share of occasions
At least one brand in common	77%
Same first-named brand	27%
At least one cited site in common	8%

Source: Babel42, AI visibility study 2026, Nov 4, 2026. Brand overlap: 344 occasions when all five named brands. Site overlap: 286 occasions when all five cited sources.

So the assistants reach similar shortlists from very different reading lists. Babel42 frames being on the shortlist as the shared prize, and being named first as unstable: the same first brand across two assistants happened only 40% to 56% of the time depending on the pair, and Gemini and Google AI Mode changed their first pick between days for about half of all questions.

For a CPG team this argues against chasing one assistant's sources in isolation. The brands that show up across all five are, we think, the ones with a broad base of independent coverage that each assistant can find in its own preferred places. The SISTRIX citation study we covered in October pointed the same way from a different angle: page features matter less than whether the page says something specific enough to cite.

Who made the shortlist in two CPG categories

Babel42 also published the three most-named brands per industry across all five assistants. Two of them are consumer packaged goods.

Industry	Buying answers	No. 1	No. 2	No. 3
Beauty and fragrance	260	CeraVe, 24%	La Roche-Posay, 20%	Neutrogena, 14%
Health and supplements	285	Thorne, 30%	Transparent Labs, 30%	Optimum Nutrition, 30%

Source: Babel42, November 4, 2026. Shares as published (rounded).

We added both lists to our AI Shelf Award today, with the method and source attached. Notice what the beauty winners have in common: dermatologist positioning, wide drugstore distribution and a lot of editorial coverage. Those are exactly the kinds of pages the table above says the assistants read. That is a correlation, not proof, but it is consistent with the citation data.

How this compares with other new studies

Babel42's 15% is not the only number in circulation this week, and the others measure different things.

Tilio, a UK agency, analysed 44,587 AI answers for four anonymised UK client brands between October 5 and November 3, 2026. Brand websites made up 3.5% of all 270,861 citations. But when an answer named one of the four brands, it cited that brand's own site 80% of the time. Those brands are not identified by industry, and they are clients tracked on questions chosen for them, so the samples are not comparable with Babel42's high-volume consumer questions.

QuQi checked 3,906 AI answers for 21 business websites from late August to October 21, 2026. When a Google AI Overview appeared, it named the business in 10.2% of cases and listed one of its pages as a source in 6.5%.

We wrote a separate think piece today on why these numbers disagree. The short version: the share depends heavily on who the brands are, what the questions are and what counts as a citation. For consumer products on generic buying questions, Babel42 is the closest match we have, and its answer is that the brand's own site is rarely the evidence.

What we would do with this, as a CPG brand

Find the cited pages in your own category. Take 20 buying questions your shoppers actually ask ("best fragrance-free moisturizer for sensitive skin", "best creatine for women"). Run them in each assistant you care about and write down every cited source, not only whether you are named. You are building a reading list: the five to ten domains each assistant trusts for your category.

Check your product's facts on those pages. For each cited domain that covers your category, search it for your product. Is it there? Is the price, size, formula and claim current? A review written before your reformulation is still being read. The tip on fixing wrong answers applies directly: the fix usually lives on someone else's page.

Treat retailer listings as evidence, not only as shelves. Amazon showed up in 43% of Google AI Overviews for supplement questions, and Ulta in 11% of Gemini's beauty answers. An incomplete retailer listing is an incomplete citation. Our earlier note on retailer listings beyond Amazon covers the basics.

Do not abandon your own site. Babel42 found the brand's own site was cited more often by Google AI Overviews (26.2% of its 2,053 recommendations) than by Gemini (6%). The site is also where reviewers, retailers and assistants pick up facts they later restate. If it is thin, everyone downstream repeats the thin version.

Measure across assistants, repeatedly. Babel42 found Perplexity cited 91% of the same sites again the next day, but Gemini only 40%. One check tells you very little. This is the kind of monitoring XENA Intelligence runs for consumer brands in SearchPanel, asking assistants shopper-style questions and scoring how each product shows up; a weekly spreadsheet of 20 questions will get a small team most of the way.

The limits of the study

Babel42 is a vendor, and the study promotes its product, though the method is published in detail with a CSV download. The ChatGPT, Perplexity and Gemini answers came from APIs, which the authors note can cite differently from the consumer apps. The questions came from high-volume US Google searches, rewritten for chat, so they are not a random sample of what people ask. Three days is short. And within each industry, the sample per assistant is small: 62 to 81 answers.

None of that changes the direction. When five different assistants, asked about skincare, almost never cite the skincare brand's own site, the work of being recommended is mostly the work of being described well by other people. Our older how assistants pick sources post makes the same case with earlier data. This is the most direct evidence yet for consumer categories.