Research
How AI recommends products: what 856 real answers show
We collected 856 answers from ChatGPT, Perplexity and Google AI Overviews to one skincare brand's buyer questions, and counted every product they named. Here is what the data shows.
01
Key findings
We collected 856 real AI answers to a UK skincare brand's buyer questions over three weeks and counted which products each one named. Five things stood out.
- 01
Half the answers named no specific product at all. 49% gave general advice ("look for a fragrance-free cleanser") instead of recommending something you can buy.
- 02
When products were named, the list was short: about four in ChatGPT and Perplexity, two in Google AI Overviews.
- 03
The same question rarely got the same answer. Asked again, ChatGPT put the same brand first only 42% of the time.
- 04
The engines disagreed. On the same question on the same day, ChatGPT and Perplexity shared no brand at all in 53% of cases.
- 05
AI read retailers, health authorities and magazines far more than brand sites. Brand websites were 9% of citations among the 30 most-cited sites.
02
About the data
Real answers, collected with web search on, from the same buyer questions asked repeatedly. The brand is anonymous.
856
AI answers
Collected between September 18 and October 8, 2026.
3
AI assistants
ChatGPT (341 answers), Perplexity (264) and Google AI Overviews (251).
1
Brand
A UK direct-to-consumer brand. Its buyer questions are mostly skincare (cleansers, face oils, moisturizers), plus a few about caps.
323
Brands named
Distinct brands recommended across all the answers.
For each answer we recorded every product named, its position, its brand and the sources cited. A mention counted as a product only if it had a brand or a link. "A gentle cleanser" is advice, not a recommendation, and was left out.
Limits: this is one brand, one market (the UK) and one set of questions. Skincare has strong medical sources, which shapes what AI cites. Treat the numbers as one detailed example, not an average for every store.
03
Finding 1: Half the answers name no product
A buyer question doesn't guarantee a product recommendation. Often the assistant explains what to look for and stops there.
of ChatGPT answers named no specific product.
The most product-led of the threeProdukto data, 341 answersof Perplexity answers named no specific product.
Advice first, products secondProdukto data, 264 answersof Google AI Overviews named no specific product.
Short answers, few namesProdukto data, 251 answersWhat it means
Visibility is measured against the answers that could have named you. A product named in 2 of 10 answers may be named in 2 of the 5 that recommended anything, which is a very different position.
04
Finding 2: The shortlist is short
When an answer did name products, it named a handful. Across all answers, though, 323 different brands got a turn.
| ChatGPT | Perplexity | Google AI Overviews | |
|---|---|---|---|
| Products per answer (when any) | 3.9 | 3.6 | 2.1 |
| Most products in one answer | 10 | 10 | 6 |
| Answers naming 1–3 products | 43% | 58% | 86% |
| Sources cited per answer | 3.9 | 19.7 | 3.8 |
Two things are true at once. Each answer is a tight shortlist, so being fourth is very different from being first. And the pool of brands that make some shortlist is wide, so a smaller brand can get in on the questions where its product fits best.
05
Finding 3: Answers change every time
We compared each answer with the next answer to the same question on the same assistant. The top pick and the list both moved a lot.
| Same question, asked again | ChatGPT | Perplexity | Google AI Overviews |
|---|---|---|---|
| Same brand ranked first | 42% | 53% | 15% |
| Brands shared between the two lists | 35% | 47% | 16% |
| Answer pairs compared | 182 | 118 | 106 |
What it means
One screenshot of one answer is close to meaningless. A product's real position is the share of many answers that name it, tracked over time.
06
Finding 4: The assistants disagree
We compared answers to the same question on the same day across assistants. They rarely recommended the same brands.
of same-day ChatGPT and Perplexity answers shared no brand at all. Average overlap: 14%.
Two assistants, two shortlistsProdukto data, 95 pairsof same-day ChatGPT and Google AI Overview answers shared no brand. Average overlap: 10%.
Ranking in one says little about the otherProdukto data, 85 pairsof same-day Perplexity and Google AI Overview answers shared no brand. Average overlap: 5%.
The widest gap of the threeProdukto data, 79 pairsIf you only check ChatGPT, you see one of several different shortlists your buyers are getting. Measure each assistant separately.
07
Finding 5: What AI reads
The answers cited 785 different websites. The 30 most-cited sites made up 40% of all citations, and they were mostly not brand sites.
31%
Retailers and marketplaces
Big online retailers and marketplaces, with their listings, specs and reviews.
28%
Health authorities
Dermatology and medical sites. The single most-cited site was a dermatology association.
21%
Magazines and publishers
Beauty and lifestyle magazines and newspaper product roundups.
9%
Brand websites
Only a few large brands' own sites made the top 30.
5%
Video and social
Video reviews and social posts.
7%
Other
Directories and smaller sites, including some with little to do with the question.
Share of citations among the 30 most-cited sites, all three assistants combined.
What it means
Your product page is one input among many. What retailers, reviewers and publishers say about your product is read at least as often as what you say about it.
08
What this means for your store
Five practical takeaways. Which changes move visibility most is still an open question, so measure before and after each one.
Measure many answers, not one
Answers change run to run (finding 3). Track the share of answers that name each product, over weeks.
Check every assistant separately
ChatGPT, Perplexity and Google build different shortlists (finding 4). A result on one tells you little about the others.
Count against answers that name products
Half the answers recommend nothing specific (finding 1). Read your visibility against the ones that do.
Win the questions where you fit best
Each shortlist is short, but hundreds of brands get a turn (finding 2). Pick the questions your product answers better than anyone and make your page say so plainly.
Look beyond your own site
Retailers, reviewers and publishers are read more than brand sites (finding 5). Check which of them are cited for your questions, and whether they mention you.
New to this? Start with what AI shopping visibility is, then check your own products by hand.
09
Frequently asked questions
How does ChatGPT decide which products to recommend?
With web search on, it reads a handful of pages, such as retailer listings, reviews, publisher roundups and product pages, and names the products it can justify for the question. In our data it named about four products when it named any, and the list changed from one run to the next.
Do ChatGPT, Perplexity and Google recommend the same products?
Rarely. On the same question on the same day, ChatGPT and Perplexity shared no brand in 53% of cases, and Perplexity and Google AI Overviews shared none in 76%.
Why does AI sometimes give advice instead of naming products?
Many buyer questions can be answered with general guidance, like what ingredients to look for. In our data, 49% of answers named no specific product. That share varies by question and category.
Does my own product page matter if AI mostly cites other sites?
Yes. Your page is still where the facts about your product live, and retailers and reviewers often repeat them. But brand sites were only 9% of citations among the most-cited sites here, so what others say about your product matters too.
Is this data representative of every category?
No. It covers one UK brand's buyer questions, mostly skincare, over three weeks. Categories with strong medical or expert sources, like skincare, may cite differently from fashion or home goods.
Method
- Answers collected by Produkto from ChatGPT (web search on), Perplexity and Google AI Overviews, September 18 – October 8, 2026, for one UK direct-to-consumer brand's buyer questions (mostly skincare, a few about caps). The brand is not named.
- Product mentions were extracted from each answer with their position, brand and link. Mentions with neither a brand nor a link were treated as general advice and excluded.
- "Asked again" compares each answer with the next answer to the same question on the same assistant. "Same day" compares answers to the same question on the same date across assistants. Overlap is shared brands divided by all brands named in the pair.
- Citation shares are calculated over the 30 most-cited domains, which make up 40% of all 7,488 citations.
See which of your products AI recommends
Produkto measures your whole Shopify catalog in ChatGPT, Google AI Overviews and Perplexity, and shows who AI picks instead.