The short version
- In a test of three commercial AI systems, product facts such as rating, price and reviews explained 82.4% of how products were ranked, and brand name only 1.2% (Chu and Hou (opens in a new tab)).
- Rewriting listings in an advertising style dropped products 1.82 places on a ten-product list on average; the best automated rewrites instead kept facts and organized attributes for comparison (Bagga and colleagues (opens in a new tab)).
- Longer is not better: one of the strongest rewriters shortened listings by about 14 words on average, while one that added about 170 words ranked near the bottom (Bagga and colleagues).
- In a hotel test across twelve AI systems, a top guest rating raised the chance of being recommended by 31.6 percentage points, while replies to reviews made no detectable difference (Baig and colleagues (opens in a new tab)).
- Only 61.9% of software plan prices quoted by AI assistants were fully faithful to the vendor’s pricing page (our pricing study).
What does product content that ranks well look like?
It preserves facts, names concrete attributes, organizes them for comparison and matches buyer language. That is the conclusion of Bagga and colleagues (opens in a new tab), who built the largest shopping test so far. They used 13,747 real shopping requests drawn from Reddit, each paired with 10 Amazon listings.
The requests were long, averaging about 59 words. They stated budgets, past experiences and must-have features. A simulated AI shopping assistant then ranked the ten products, and the researchers rewrote one listing to see whether it moved up.
They let an automated system search for the best way to rewrite listings, starting from 15 different styles. Almost all ended up in the same place. The winning instructions told the writer to open with a summary and use labeled sections and short bullets. They also asked for real use cases, answers to likely buyer questions and strictly accurate facts.
This playbook worked broadly. After the automated search, 63 of 75 combinations of writing style and AI ranker improved. Results also improved on GPT-5 and Claude, which were never used during the search, though most gains there were modest.
Do the usual copywriting tricks help?
Mostly not, and several hurt badly. Bagga and colleagues tested 15 common rules of thumb for AI-friendly writing, from an authoritative tone to adding an FAQ. Only four matched or slightly beat a plain one-line rewrite instruction.
| Writing style | Average places moved on a ten-product list |
|---|---|
| Advertisement-like copy | down 1.82 |
| Foreign-language flourishes | down 1.66 |
| Cut to a single sentence | down 1.49 |
| Creative short story (a deliberate bad example) | down 4.36 |
Length did not explain success either. GPT-5, one of the strongest rewriters, shortened listings by about 14 words on average.
GPT-4.1 added about 170 words and ranked near the bottom. Bulleted lists gave only a small lift, about 0.23 places.
The authors draw a blunt conclusion: “manually instilling ‘GEO knowledge’ through prompt design can backfire.” Testing beat intuition.
How much do hard facts like price and ratings matter?
More than anything else tested. Chu and Hou (opens in a new tab) gave three commercial AI systems lists of ten skincare products: one famous brand and nine invented ones. They varied rating, price, review count and ingredients.
Product facts explained 82.4% of how the AI ranked products. Position in the list explained 6.5%, and the brand name only 1.2%. When an invented brand had clearly better specifications, the AI recommended it about 96% of the time.
Tiny real differences were enough. An invented brand won half the time with just a 0.075-star rating edge or 1.6 times as many reviews. Brand name mattered most when product information was ambiguous, which is when AI falls back on names it knows.
Which signals do AI assistants weigh differently from shoppers?
Some signals matter far more or less to AI than marketers expect. Baig and colleagues (opens in a new tab) asked twelve AI systems to choose among five hotels whose details were randomly varied.
| Signal | Change in chance of being recommended |
|---|---|
| Top guest rating (4.7 versus 3.9 stars) | up 31.6 points |
| High price | down 30.0 points |
| Eco-certification | up 11.6 points |
| High review volume | up 8.3 points |
| Visible replies to reviews | up 0.1 points (no detectable effect) |
The authors note that replying to reviews is a tactic “the optimization industry actively promotes.” In their test, AI ignored it. Eco-certification counted for far more than a human-focused reputation playbook would expect.
List position also mattered, though it says nothing about the hotel. Being placed higher in the candidate list was worth about $12 per night. And the reasons the AI gave for its choices did not fully match what actually drove them. Our guide to why AI picks one hotel over another covers the hotel results in full.
Where else do AI assistants get product information?
Mostly from third-party sites, which vary by assistant. Uberti-Bona Marin and colleagues (opens in a new tab) audited ChatGPT, Gemini and Google’s AI Overviews on 117 real product questions, each asked three times.
Editorial and product-review sites made up 56.7% of the domains ChatGPT displayed. AI Overviews, the AI summary at the top of Google’s results, most often showed Reddit (35.0%) and YouTube (30.0%). For the same question, ChatGPT and Gemini shared only 5.4% of the domains they displayed.
The assistants also framed advice differently. ChatGPT gave a first-person pick, such as “my pick would be,” in 79% of answers that recommended products. Gemini did so in 7% and AI Overviews in 2%.
So your product page is one input among many. Reviews, comparison articles, forums and videos often shape what the assistant says.
Does consistency across your pages matter?
Yes: AI answers often quote whichever version of a fact they find. In our pricing study, 61.9% of quoted plan prices were fully faithful to the vendor’s pricing page. A further 3.8% had the right amount but dropped a condition, usually presenting an annual-billing price as monthly.
Our business facts study found the same pattern with phone numbers. Where a business’s Google profile number did not appear on its own website, 30.6% of the numbers AI gave differed from the profile. Where it did appear, only 1.6% did.
The lesson is simple. If your site, marketplace listings and profiles disagree, AI may pick the wrong one.
What should you do about it?
Write product content for comparison, not persuasion, and keep every fact consistent everywhere it appears.
- Lead with specifications buyers compare. Price, ratings, dimensions, materials, certifications and compatibility should be stated plainly and labeled.
- Answer the real questions. Shopping requests in the largest test averaged about 59 words, full of use cases and constraints. Address those situations directly.
- Cut advertising language. Advertisement-style rewrites lost 1.82 places on average in testing.
- Earn genuine proof. Ratings and review volume moved AI choices; replies to reviews did not, in the one test that measured them.
- Make facts match everywhere. Align your site, marketplace listings, pricing pages and profiles, including billing terms.
- Check third-party coverage. Review sites, forums and videos feed many answers. Make sure they have accurate information to work from.
- Test, don’t guess. Rules of thumb often backfired. Measure how assistants describe your products before and after changes.
If you want help improving how AI assistants present your products, see our generative engine optimization service.
What does the research not tell us yet?
How these preferences play out inside live shopping assistants, end to end. The gaps:
- The largest test simulated only the ranking step, over ten Amazon listings, in English, with mostly North American buyers.
- The skincare and hotel tests used invented products and fixed short lists; real listings have photos, full reviews and many more competitors.
- None of these studies measures sales or clicks, only rankings and recommendations.
- AI systems change often. The hotel authors warn their results describe specific model versions and may drift.
- No study yet compares the same product content across Amazon’s, Google’s and OpenAI’s live shopping features.
Frequently asked questions
Does product description length affect AI recommendations?
Not on its own. In the largest test, one of the strongest rewriters shortened listings by about 14 words on average, while one that added about 170 words ranked near the bottom.
Should product pages include an FAQ for AI search?
It may help a little but is no shortcut. An FAQ-style rewrite was one of only four of 15 common styles that matched a plain rewrite, and the best automated rewrites included buyer questions among other changes.
Do AI shopping assistants favor big brands?
Only when products look identical. In one test, the famous brand won every trial with identical specifications, but brand name explained just 1.2% of rankings once real differences existed.
Does structured, bulleted product copy rank better with AI?
Slightly. Bulleted lists added about 0.23 places on average in one large test, and the best-performing rewrites organized attributes into labeled sections.
Sources
- Bagga, Farias, Korkotashvili, Peng and Wu (2025), E-GEO: A Testbed for Generative Engine Optimization in E-Commerce (opens in a new tab), arXiv:2511.20867.
- Chu and Hou (2026), Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems (opens in a new tab), arXiv:2606.17443.
- Baig, Gillani and Ali (2026), Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection (opens in a new tab), arXiv:2606.16344.
- Uberti-Bona Marin, Bertaglia, Astante, Rijsbosch, van Dijck, Hannák, Spanakis and Kollnig (2026), "If I Had to Buy Just ONE: Galaxy S26 Ultra": Auditing AI-Generated Product Recommendations (opens in a new tab), arXiv:2609.18729.
- Underneath (2026), How faithfully do AI assistants quote software prices?
- Underneath (2026), Do AI answers match a business’s Google profile?