How AI agents actually choose products: 12 ChatGPT, Claude, Perplexity and Gemini conversation teardowns

The setup: how we ran the study

We picked twelve product categories with an active mid-market: running shoes, wireless earbuds, a linen blazer, a coffee maker, a standing desk, a moisturiser, a children's car seat, a robot vacuum, a kitchen knife, a yoga mat, a suitcase, and a protein powder. For each category we wrote a single natural-language shopping prompt ("I need X, my budget is Y, which brand should I buy?") and ran it, without modification, across ChatGPT Plus, Claude.ai, Perplexity, and Gemini Advanced.

Every conversation was screen-recorded. Every cited URL was logged. Where the agent surfaced a product name without a URL, we traced it back to the retrieval source using the agent's "search results" panel or by re-running the query with the browsing source visible. The annotations ran to about 80 pages across the twelve teardowns. The wider pattern across engines is in our 1,200-prompt citation study.

A quick disclaimer before the findings: twelve conversations is not a panel. It is a structured anecdote set. The numbers that follow are observations from this corpus, flagged where they match or diverge from the broader 1,200-prompt dataset we run continuously. Do not treat the twelve conversations as a statistical benchmark; treat them as an annotated specimen set.

Which PDP components actually get quoted?

Across all twelve conversations and all four engines, a clear hierarchy emerged for which PDP components made it into the answer. Verified review excerpts led by a wide margin, appearing in full or near-full quotation in eight of the twelve final recommendations. Specific factual descriptors in the product title or opening paragraph came second. Material and spec lists came third, but only when they were server-rendered rather than behind a JavaScript accordion.

Marketing copy performed poorly at every engine. Phrases like "elevate your run" or "experience unmatched comfort" appeared in zero final recommendations across all twelve conversations. When an agent retrieved a page that leaned heavily on aspirational language, it either paraphrased a spec buried deeper in the page or moved to a different brand entirely. The pattern was consistent enough to feel like a rule rather than a tendency.

The Perplexity teardowns were the most instructive. Perplexity surfaces its citations explicitly in the interface, so you can see which exact URL produced which sentence in the answer. In seven of twelve runs, the brand that was cited was not the brand with the highest SEO ranking for the category keyword. It was the brand whose page had a Review JSON-LD block with verified-buyer attribution and at least one photo or video attached to the review record.

Engine-by-engine breakdown

ChatGPT Plus

ChatGPT cited a specific brand in 4 of 12 conversations. In the other 8, it gave a category-level recommendation ("look for a shoe with a wide toe box and carbon-fibre plate") without naming a brand. When it did cite, the cited URL was almost always a review aggregator (a Wirecutter or similar), not the brand's own PDP. The exception: the linen blazer query, where the brand's own PDP was cited directly because the first 300 words of the page read like a spec sheet.

Claude.ai

Claude cited a brand page in 5 of 12 conversations. Claude was most likely to quote a specific verified review excerpt and attribute it to the brand by name. In three of the five brand citations, the review that was quoted was the first server-rendered review on the PDP, not the highest-rated one, suggesting Claude reads the first available review rather than sorting by score. That is an implementation detail worth knowing: put your most quotable review first, not your most positive one.

Perplexity

Perplexity cited a brand in 7 of 12 conversations, the highest of any engine. It is also the only engine where the citation source was fully transparent in the UI. The seven brands that were cited all had one thing in common: their Review JSON-LD block included at least one review with a datePublished field within the previous 18 months. Brands with outdated or missing datePublished fields were never cited by Perplexity in this corpus.

Gemini Advanced

Gemini cited a specific brand page in 3 of 12 conversations, the lowest rate. Gemini showed the strongest preference for large, authoritative domains: Wikipedia, major review publications, and brand sites with high Merchant Center product feed quality. In the three cases where Gemini did cite a brand's own page, the brand had a complete Google Merchant Center feed with all required fields populated, including gtin, condition, and age_group where applicable.

The three patterns that cut across all engines

Verified-buyer attribution is the strongest signal. A review labelled "Verified Purchase" or "Verified Buyer" with a date and at least one photo gives an agent far more to verify than a text-only review without verification metadata. This is the single highest-leverage intervention available to most PDPs right now.

Server-rendering decides retrieval. Agents do not wait for JavaScript to execute. If your reviews, specs, or comparison content loads via a client-side fetch, the agent does not see it. The three brands in this corpus with the lowest citation rates all shared the same problem: their review blocks were rendered by JavaScript after page load. Switching to server-side rendering for the first five reviews is a one-day engineering task with an outsized impact on agent visibility.

Comparison paragraphs are the dark horse. Only two of the twelve PDPs in the corpus had a comparison paragraph ("how this compares to X"). Both were cited by multiple engines. The comparison paragraph gives the agent something to quote when a shopper asks "X or Y, which should I buy?" Without it, the agent goes to a third-party comparison article and cites that instead.

The conversation that surprised us most

Conversation 9 was a standing desk query. The brand that was ultimately recommended by three of the four engines had a PDP that, by traditional SEO standards, was underperforming. Its domain authority was below the category median. Its backlink count was low. Its meta title was not keyword-optimised.

What it did have: a 2,400-word PDP with a factual opening sentence, a server-rendered block of 12 verified reviews averaging 340 characters each, a comparison paragraph addressing the specific question of "desktop vs. L-shaped, which for a small home office?", and a structured Q&A block with 8 questions answered in plain prose with measurements. The agents quoted all four of those components in their recommendations.

The lesson is not subtle. The page that wins an agent recommendation is not the page with the highest authority. It is the page that answers the shopper's actual question with enough specificity and attributed evidence that the agent can re-state the answer without paraphrasing. That is a content and schema problem, not a link-building problem.

What should I do with this?

The teardowns point to a short, prioritised list. Treat it as a two-week sprint for one person, not a quarter-long roadmap. If you want the full audit framework first, run the 42-signal AEO scorecard against your top PDPs.

1. 1Server-render the first five reviews on every top-revenue PDP. Pick the reviews by quotability (specificity + verified-buyer attribution), not by recency or rating.
2. 2Add a Review JSON-LD block with datePublished, author, reviewRating, and at minimum one image URL per review. Validate with Google's Rich Results Test before publishing.
3. 3Rewrite the opening sentence of each top PDP to be a fact, not a headline. One sentence, all specs, no adjectives.
4. 4Add a comparison paragraph to any PDP that competes on category keywords with multiple alternatives. Two to four sentences. Position the product against the obvious alternatives honestly.
5. 5Add a five-question Q&A block (FAQPage schema) using real customer service questions, answered in plain prose with numbers and units.

None of these are glamorous. All of them show up in the teardowns as the difference between being cited and being skipped. The sprint takes less than a quarter. The citation lift shows up in two to four weeks.