• AI visibility

Guide · AI search

What actually drives which product an AI assistant recommends?

Product data drives the pick when the assistant can see it: in a controlled test of three AI assistants, rating, price and reviews explained 82.4% of how products were ranked, and brand name only 1.2%. A famous brand still wins when the products look identical. The bigger risk for most brands is missing, thin or wrong product data, not a famous rival.

By Underneath Team Guide Updated

The short version

  1. In a skincare test of three AI assistants, product data explained 82.4% of the rankings and brand name 1.2% (Chu and Hou (opens in a new tab)).
  2. When every product had the same specs, the well-known brand won all 670 valid trials in the same study, so brand works as a tiebreaker.
  3. A small edge flipped the result: an invented brand with a slightly better rating, price or review count won 64–80% of the time instead of 3.6–6.0%.
  4. Real assistants are less tidy: for the same shopping questions, ChatGPT and Gemini showed only 5.4% of the same source websites on average (Uberti-Bona Marin and colleagues (opens in a new tab)).
  5. Assistants also lose detail: only 61.9% of software plan prices quoted by four assistants were fully faithful to the vendor’s pricing page (our pricing study).

How much does product data matter compared with brand name?

Far more: in one controlled test, product data explained 82.4% of rankings and brand name only 1.2%.

Chu and Hou (opens in a new tab), of Trine University and Texas A&M, asked three AI assistants (GPT-4o-mini, Claude Sonnet and Gemini 3 Flash) to rank lists of ten skincare products. Each list held one real brand, such as CeraVe, and nine invented ones. Across thousands of trials they varied rating, price, review count and the order of the list.

Product details (rating, price and reviews) explained 82.4% of the variation in rankings. Position in the list explained 6.5%. Brand identity explained 1.2%. The rest came from the factors acting together.

Two caveats matter. The product details were typed into the question through the companies’ developer access, so this is not the consumer app searching the web. And skincare is a category where buyers lean on brand. A smaller repeat with USB-C cables and AA batteries (3,840 calls) found the same pattern.

When does a famous brand still win?

A famous brand wins when nothing else separates the products, or when the product data is unclear.

In the same study, when all ten products had identical rating, price, reviews and description, the real brand was recommended in all 670 valid trials. Not one invented brand was picked. The authors call this a conditional monopoly: total, but only while nothing else differs.

Brand also mattered most in the murky middle. With clearly good or clearly poor specs, real and invented brands ranked the same. With middling specs, the real brand’s average rank was 1.70 against 5.49 for an invented brand with the same numbers. When the data gives the assistant no reason to choose, it falls back on the name it knows.

Live assistants show the same default. Chen and colleagues (opens in a new tab) at the University of Toronto asked ChatGPT and Perplexity 50 unbranded questions about cola. Major brands took 62.2% of brand mentions and niche brands 9.0%. Those questions gave the assistant nothing to compare, so it named the market leaders.

How small an edge is enough to win?

Very small: a rating less than a tenth of a star higher was enough to win half the time.

Chu and Hou then gave the invented brand a growing advantage. With identical specs, it won only 3.6–6.0% of the time. With the smallest advantage tested, its win rate jumped to 64–80%, depending on which detail was better. Larger advantages added little after that.

The halfway point was strikingly low: a 0.075-star rating advantage, 1.6 times as many reviews, or a 7.3% price discount. That is less than the gap between a 4.3 and a 4.4 star rating.

The three assistants did not behave alike. Claude was the hardest to move: with the smallest rating edge, the invented brand won 11% of the time on Claude, against 94% on GPT-4o-mini and 88% on Gemini. Which assistant your buyers use changes how much a small edge is worth.

Does the way a product is described matter too?

Yes, but evidence-style wording moves assistants while sales pressure does not, and invented claims are a legal risk.

In a second experiment, Chu and Hou kept the specs identical and changed only the wording. Copy that looked like evidence, such as clinical-trial claims or customer testimonials, broke the famous brand’s hold 50–73% of the time. Pressure tactics like “limited stock” moved it only 10–13% of the time. The clinical claims in the test were invented on purpose to find the upper limit. The authors class invented claims as potential false advertising and limit their advice to real certifications and published evidence. Our guide on gaming AI shopping rankings covers those risks in more depth.

Other controlled tests agree that substance beats style. A team from MIT and Columbia built E-GEO (opens in a new tab), a test bed of 13,747 shopping questions paired with real Amazon listings, ranked by five AI models acting as shopping assistants. Making a listing longer did not help. The rewrites that worked kept the facts, led with a summary, listed concrete features and use cases, and answered likely buyer questions.

Researchers at Sprinklr, a software vendor, ran 252,000 head-to-head trials in which two pages differed in one detail. A page without a price was far less likely to be cited first in every model they tested. Formatting changes did little: seven of their 18 factors (39%) had weak or no effects, including how the text was laid out. As a vendor study in a two-page setup, treat it as directional.

Do real AI assistants behave the same way?

Partly: real assistants also lean on concrete facts, but their answers vary and draw on other people’s pages.

Uberti-Bona Marin and colleagues (opens in a new tab) put 117 real shopping questions to ChatGPT, Gemini and Google’s AI Overviews from the Netherlands in September 2026. ChatGPT stated a personal pick (“my pick would be…”) in 79% of its product answers, against 7% for Gemini and 2% for AI Overviews. For the same question, ChatGPT and Gemini shared only 5.4% of their source websites on average.

Those sources are mostly third parties. Editorial and product-review sites made up 56.7% of the domains ChatGPT displayed. So the product data an assistant compares often comes from reviewers and retailers, not from your own page.

Our own studies point the same way:

What we measuredResult
Software plan prices quoted fully correctly by four assistants61.9% (pricing study)
Odds of being recommended for each tenfold rise in independent sites naming a brand4.7 times (brand entity study)
Share of ChatGPT’s brands across five runs shown in a single answer57.8% (consistency study)

Independent coverage was the strongest predictor of being recommended in our brand study, ahead of a Wikipedia article. And because one answer shows only part of the picture, a single check of “are we recommended?” is a weak measure.

What should you do about it?

Make your product data complete, accurate and easy to compare everywhere an assistant might read it.

  1. Publish the numbers assistants compare. Put price, rating, review count and key specs on the product page in plain text. A missing price was one of the strongest reasons a page lost in controlled tests.
  2. Fix the data on other people’s pages. Review sites and retailers supply most of what assistants cite. Give reviewers current specs and prices, and correct listings that are out of date.
  3. Earn the small edges honestly. A slightly higher rating or more reviews was enough to beat a famous brand in testing. Ask every satisfied customer for a review.
  4. Back claims with real evidence. Name real certifications, tests and awards. Never invent them. In one test, assistants told to watch for manipulation flagged every listing rewritten with extreme superlatives.
  5. Check what assistants say, repeatedly. Ask the same buying questions several times across assistants and record the prices and reasons they give.

If you want help building that measurement, see our generative engine optimization service.

What does the research not tell us yet?

The controlled evidence is strong on how assistants weigh data they are given, and weak on everything before that.

  • In the main experiments the product details were typed into the question. Real assistants search the web first, and whether your data is found at all is a separate step these tests skip.
  • The core study used skincare and three assistant versions available in 2026; other categories and newer versions may weigh data differently. Our guide to sportswear brands in AI shopping answers looks at one such category, where questions turn on the activity and specific product attributes.
  • The wording experiments were not repeated for cables and batteries, so the effect of evidence-style copy outside skincare is unknown.
  • No study here links better product data in AI answers to sales.
  • Real-world audits are snapshots from one place and month, and assistants change often.

Frequently asked questions

Do AI assistants favor big brands?

Only when they have nothing else to go on. With identical specs, the known brand won all 670 valid trials in one study, but a slightly better rating, price or review count let an unknown brand win most of the time.

Does a better star rating help a product get recommended by ChatGPT?

In controlled tests, yes. A rating advantage of 0.075 stars was enough for an unknown brand to win half the time, although Claude was much harder to move than GPT-4o-mini or Gemini.

Should I rewrite product descriptions for AI search?

Rewrite for clarity and completeness, not length. In tests on Amazon listings, longer copy did not help, while factual summaries, concrete features and answers to buyer questions did.

Does listing a price matter for AI answers?

It appears to. In a vendor study of 252,000 trials, pages without a price were much less likely to be cited first by every model tested.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work