Research
AI shelf study: 500 shopping questions, 3 AI assistants
We asked ChatGPT, Perplexity and Google AI Overviews 500 shopping questions across 10 categories and counted every product they recommended. Here is what 1,479 answers show.
01
Key findings
We asked ChatGPT, Perplexity and Google AI Overviews 500 shopping questions across 10 e-commerce categories and counted every product they recommended. Five things stood out.
- 01
Change the wording, change the winner. We asked about each shopping need five different ways. For 84 of 96 needs (88%), ChatGPT's #1 brand changed at least once.
- 02
The assistants disagree. ChatGPT and Perplexity named the same first pick in 42% of questions. All three agreed in 22%.
- 03
The shelf is wide open. 1,315 different brands were recommended, the 10 most-named took only 12% of mentions, and half were named exactly once.
- 04
Each assistant reads different sites. ChatGPT cited retailers most, Google AI Overviews cited YouTube, Facebook and Reddit, and Perplexity cited retailers and publishers.
- 05
Every category has a few brands AI keeps coming back to, like CeraVe in skincare (105 recommendations) and Lodge in kitchen (62).
02
About the data
Real answers from three AI assistants, collected on October 26, 2026, to questions written the way US shoppers ask before they buy.
500
Shopping questions
10 categories, 10 shopping needs each, every need asked 5 different ways. Most include a price.
3
AI assistants
ChatGPT and Perplexity answered all 500. Google AI Overviews answered 479; Google shows no overview for some searches.
1,479
AI answers
5,958 product and brand recommendations in total, about four per answer.
1,315
Brands named
Distinct brands recommended across all the answers.
The categories: coffee, pet, skincare, activewear, footwear, home, supplements, kitchen, outdoor and jewelry. For each answer we recorded every product and brand recommended, in order, and the sites it cited. Generic advice that named neither a brand nor a product ("a fragrance-free cleanser") was left out.
Limits: each question was asked once on each assistant, on one day. AI answers change from one run to the next, so a single pick is a snapshot, and part of the change between wordings below is the AI's own randomness. The conclusion holds either way: one question asked once tells you very little.
03
Finding 1: Change the wording, change the winner
Shoppers ask for the same thing in many ways: with a price, with a use case, or "which brand makes the best". We asked about each shopping need five ways and compared the first pick.
of shopping needs got a different #1 brand from ChatGPT across the five wordings.
84 of 96 needsProdukto data, 500 ChatGPT answersdifferent brands took ChatGPT's top spot per shopping need, on average.
Out of five wordingsProdukto data, 500 ChatGPT answersof shopping needs got a different #1 brand from Perplexity too.
83 of 94 needsProdukto data, 500 Perplexity answers
Real ChatGPT answers, recorded October 26, 2026. The price is blurred because the session ran from Europe.
What it means
Checking one prompt tells you almost nothing about whether AI recommends your product. Measure across the many ways your customers actually ask, and look at how often you're named, not whether you were named once.
04
Finding 2: The assistants disagree
Same question, same day, three assistants. We compared which product or brand each named first, and how many brands two answers shared.
| Compared | Questions | Agreed |
|---|---|---|
| ChatGPT and Perplexity named the same first pick | 498 | 42% (210) |
| All three named the same first pick | 427 | 22% (92) |
| Brands shared by a ChatGPT and a Perplexity answer, on average | 498 | 34% |
Being recommended by one assistant says little about the others. A product can lead in ChatGPT and be missing from Perplexity's answer to the same question. Each assistant needs to be checked on its own.
05
Finding 3: The shelf is wide open
A few big names show up often, but they don't own the answers. Most recommendations went to a long tail of brands.
different brands were recommended across the 500 questions.
3,743 distinct productsProdukto data, 1,479 answersof all brand mentions went to the 10 most-named brands.
687 of 5,895 mentionsProdukto data, 1,479 answersof the brands were named exactly once in the whole study.
642 brandsProdukto data, 1,479 answersWhat it means
AI recommendations aren't locked up by a handful of giants. A smaller brand can get named on the questions where its product fits best, which makes it worth knowing exactly which questions those are.
06
Finding 4: Each assistant reads different sites
The sites each assistant cited most often in its answers, out of 8,833 citations.
| # |  ChatGPT |  Google AI Overviews |  Perplexity |
|---|---|---|---|
| 1 | target.com 162 | youtube.com 170 | target.com 223 |
| 2 | walmart.com 132 | facebook.com 95 | amazon.com 179 |
| 3 | amazon.com 95 | reddit.com 81 | nytimes.com 161 |
| 4 | chewy.com 82 | amazon.com 42 | rei.com 127 |
| 5 | nordstrom.com 77 | nytimes.com 25 | everydayhealth.com 122 |
| Citations | 2,321 | 1,512 | 5,000 |
ChatGPT leaned on retailer listings, Google AI Overviews on video and social, and Perplexity on a mix of retailers and publishers such as The New York Times and Good Housekeeping. In our sample, ChatGPT's answers showed no Reddit citations; that describes the links it displayed, not everything it may read.
Where your product is listed, reviewed and described outside your own site shapes what each assistant can find about it.
07
Finding 5: Who AI recommends, by category
The three most-recommended brands and the most-recommended product in each category, across all three assistants. 50 questions per category.
| Category | #1 brand | #2 brand | #3 brand | Most-named product |
|---|---|---|---|---|
| Coffee | Peet's 31 | Counter Culture 27 | Lavazza 23 | Stumptown Hair Bender 7 |
| Pet | Nutramax 30 | Ruffwear 28 | Frisco 25 | KONG Extreme 8 |
| Skincare | CeraVe 105 | La Roche-Posay 51 | Vanicream 27 | CeraVe Moisturizing Cream 12 |
| Activewear | Nike 49 | Under Armour 49 | lululemon 46 | REI Co-op Trailmade Pants 8 |
| Footwear | HOKA 61 | Brooks 51 | New Balance 51 | HOKA Bondi SR 14 |
| Home | Quince 34 | Brooklinen 22 | Silk & Snow 16 | Pottery Barn Hydrocotton Quick-Dry Towel 6 |
| Supplements | Thorne 28 | Nature Made 27 | Optimum Nutrition 25 | Optimum Nutrition Gold Standard 100% Whey 13 |
| Kitchen | Lodge 62 | Ninja 26 | Tramontina 22 | Victorinox Fibrox Pro 8-inch Chef’s Knife 9 |
| Outdoor | REI Co-op 57 | Osprey 33 | Patagonia 32 | REI Co-op Flash 22 9 |
| Jewelry | Kendra Scott 30 | Mejuri 20 | Quince 18 | Tissot PRX 25mm 3 |
Numbers are how many times each brand or product was recommended in that category's answers.
08
What to do with this
A check you can run on your own best seller today, in about 15 minutes.
Write down five ways a customer would ask
Use the patterns from this study: with a price ("under $40"), with a use case ("for a dog that pulls"), and "which brand makes the best". Leave your brand name out.
Ask each one in ChatGPT, Perplexity and Google
Use a fresh chat each time. Note whether your product is named, in what position, and which product comes first.
Count, don't screenshot
Fifteen answers tell you far more than one. Record how many of them name you, then repeat on another day to see how stable it is.
Study the product that wins most often
Open its page and the sites the assistant cited. Note the facts it states that your page doesn't: materials, sizes, who it's for, how to use it.
09
Frequently asked questions
Does the way a shopper words a question change which product AI recommends?
Often. When we asked about the same shopping need five different ways, ChatGPT's first-pick brand changed for 84 of 96 needs. Part of that is randomness between runs, since each wording was asked once, but either way one question asked once is a poor guide.
Do ChatGPT, Perplexity and Google AI Overviews recommend the same products?
Not usually. ChatGPT and Perplexity named the same first pick in 42% of questions, and all three agreed in 22%. On average only 34% of the brands in a ChatGPT answer and a Perplexity answer to the same question overlapped.
Do only big brands get recommended by AI?
No. 1,315 brands were recommended across the study and the 10 most-named took only 12% of mentions. Half of all brands were named exactly once, usually on a specific question their product fit well.
Which websites does ChatGPT cite for shopping questions?
In this study, mostly retailers: Target, Walmart, Amazon, Chewy and Nordstrom were the five sites ChatGPT cited most. Google AI Overviews cited YouTube, Facebook and Reddit most, and Perplexity cited retailers and publishers.
How was this study run?
500 shopping questions across 10 categories, each asked once on ChatGPT, Perplexity and Google AI Overviews on October 26, 2026. Every product and brand recommended in each answer was recorded in order, along with the sites cited.
Method
- 500 questions: 10 categories (coffee, pet, skincare, activewear, footwear, home, supplements, kitchen, outdoor, jewelry) × 10 shopping needs × 5 wordings, written the way US shoppers ask, most with a price in dollars.
- Answers collected by Produkto on October 26, 2026 from ChatGPT (web search on), Perplexity and Google AI Overviews, one run per question per assistant. ChatGPT 500 answers, Perplexity 500, Google AI Overviews 479.
- Recommendations were extracted from each answer in order, with brand, product and link. Mentions with neither a brand nor a product name were treated as general advice and excluded. Brands were compared case-insensitively.
- "First pick" is the first product or brand an answer recommended. The wording comparison groups the five wordings of each shopping need and counts needs where the first-pick brand was not the same in every answer (needs with at least five answers that named a brand).
- Citations are the links each assistant displayed with its answers, counted by website.
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.