We Analyzed 10,000 AI Agent Conversations: Here's What They Actually Recommend (And Why)

The Experiment: 10,000 Queries Across 5 AI Platforms

Over early 2026 we informally worked through roughly 10,000 product recommendation queries across ChatGPT, Perplexity, Claude, Gemini, and Copilot. The queries spanned a dozen product categories - from electronics and beauty to supplements and home goods - and were structured to mimic how real consumers ask AI agents for purchase advice. "Best budget laptop for students," "top moisturizer for dry skin," "safest pre-workout supplement" - the kinds of queries that increasingly bypass Google entirely and go straight to conversational AI.

For each query, we recorded the top 3 recommended products, the reasoning the agent provided, the sources it cited (if any), and whether the recommendation included caveats or disclaimers. We then matched each recommended product against its actual product page to see what signals the AI agent was likely pulling from. The result is an informal but wide-ranging read on how AI agents make product recommendations in practice - and the patterns are not what most marketers expect.

The methodology matters because AI agent recommendations are not search results. There's no bidding, no ad placement, no SEO keyword stuffing. The agent decides based on its training data, real-time web access (where available), and whatever structured signals it can extract from product pages. Understanding those signals is the new competitive advantage - and most brands are completely blind to them.

Finding #1: The "Invisible Ranking Factors"

When we correlated product attributes with recommendation frequency, six factors stood out as the strongest predictors. But the ranking of these factors contradicts conventional marketing wisdom. In our analysis, structured data markup - the technical metadata most brands ignore - looked like the single most influential factor, directionally around a quarter of the signal. Review quality came next, then brand authority (measured by mention frequency across authoritative sources). Price competitiveness, the factor most brands optimize relentlessly, ranked only fourth.

This hierarchy makes sense when you understand how AI agents process information. They can't "see" your product the way a human shopper does. They parse structured data, extract entity relationships, and cross-reference claims against their training corpus. A product with clean Schema.org markup, legitimate reviews, and consistent brand mentions across the web gives the agent high-confidence data to work with. A product with a great price but sparse metadata is essentially invisible.

Finding #2: Reviews Are the New SEO

In traditional search, reviews matter for conversion but not for ranking. In AI agent recommendations, reviews looked like the second most influential factor - and the quality of reviews mattered more than the quantity. Products with fewer than 50 reviews were recommended far less often than products with 50+ reviews, regardless of star rating. But here's the nuance: products with hundreds of generic "great product" reviews were recommended less often than products with a smaller body of detailed, specific reviews that mentioned features, use cases, and comparisons.

AI agents appear to extract semantic information from review text. When a user asks "best headphones for running," the agent isn't just checking if the product has high ratings - it's scanning review text for mentions of running, sweat resistance, fit during exercise, and similar contextual markers. Products whose reviews naturally contain use-case-specific language get recommended for those specific queries. This is fundamentally different from how Google ranks products, and it means review generation strategy needs to shift from "get more stars" to "generate reviews that describe specific use scenarios."

The implication for brands is significant. The review profile that helps you rank on Amazon may not help you get recommended by AI agents. Agents appear to weight review depth, specificity, and recency. A product with a few hundred recent, detailed reviews can outperform a product with thousands of reviews that are mostly one-liners from years ago.

Finding #3: Structured Data Is the New Bidding Strategy

The largest factor - and the one with the most actionable gap - is structured data. Products with complete Schema.org Product markup (including brand, price, availability, rating, review count, and description) were recommended far more often than products without structured markup; in our analysis, on the order of several times as often. This is the AI equivalent of bidding in Google Ads: if you're not in the system, you don't get shown.

The reason is mechanical. When AI agents access product pages (via web browsing tools or cached training data), structured data is the fastest and most reliable way for them to extract product attributes. Without it, the agent has to parse unstructured HTML, which introduces errors and reduces confidence. Most agents err on the side of recommending products they can confidently describe - and structured data provides that confidence.

Perhaps most surprising: incorrect or outdated structured data appeared to perform worse than having no structured data at all. Products with schema markup showing a different price than the page displayed, or listing features that contradicted the description, were recommended notably less often than products with no markup. AI agents appear to penalize inconsistency, likely because conflicting signals reduce the agent's confidence in the accuracy of any product information.

Finding #4: Brand Mentions Compound - First-Mover Advantage Is Real

Brand authority - which we treated as the frequency and quality of brand mentions across authoritative web sources, review sites, comparison articles, and expert recommendations - was the third-heaviest factor in our analysis. But the distribution was heavily skewed: directionally, a small group of top brands in each category captured the lion's share of AI agent recommendations. This creates a compounding loop that's extremely difficult for newcomers to break into.

The compounding works like this: brands that were frequently recommended by AI agents early on generated more online mentions (from users sharing AI-recommended products), which increased their presence in training data and web sources, which made them more likely to be recommended later. Brands that were recommended in top-3 results early on tended overwhelmingly to stay there over the following year - even when objectively better products had entered the market.

This has profound implications for competitive dynamics. In traditional advertising, any brand can outbid competitors for visibility. In AI agent recommendations, the first-mover advantage compounds over time, creating a moat that money alone can't bridge. The brands investing in GEO (Generative Engine Optimization) today aren't just optimizing for current recommendations - they're building training-data moats that will persist for years.

Finding #5: Price Is the 4th Factor, Not the 1st

This finding will be uncomfortable for brands competing primarily on price. Across our queries, price competitiveness ranked only fourth - below structured data, reviews, and brand authority. When users asked AI agents for "the best" product in a category, the cheapest option rarely landed in the top 3. The agent's definition of "best" consistently weighted quality signals over price.

Even when users explicitly mentioned budget constraints ("best laptop under $500"), AI agents frequently recommended products at the higher end of the stated range, provided those products had stronger review and authority signals. The agent's reasoning typically included phrases like "while slightly more expensive, this offers better long-term value" or "the marginal price increase is justified by significantly better reviews." This suggests AI agents are tuned to optimize for user satisfaction, not minimum price - a fundamentally different incentive than comparison shopping engines.

The exception is Gemini, which showed notably higher price sensitivity than other platforms. In our sample, Gemini recommended the lowest-priced option far more often than ChatGPT or Claude. This platform-level variation means brands may need different optimization strategies depending on which AI agents their target audience primarily uses.

The Hidden Data Sources AI Agents Pull From

Understanding where each AI agent gets its recommendation data reveals why platforms diverge in their recommendations. ChatGPT with browsing pulls heavily from major review aggregators (Wirecutter, RTINGS, Tom's Guide), Amazon's product ecosystem, and brand websites with strong structured data. Perplexity is the most source-transparent, typically citing several sources per recommendation, with a strong preference for recent editorial reviews and comparison articles.

Claude showed the most conservative recommendation pattern, particularly in categories where health or safety claims are involved. In our supplements category, Claude recommended specific products in only a minority of queries - often responding with educational content about ingredient research rather than product recommendations. Gemini pulled heavily from Google Shopping data and merchant feeds, which helps explain its price sensitivity and tendency to recommend products with active Google Shopping listings.

The practical implication: optimizing for AI agent recommendations isn't a single strategy. Brands that want to appear in ChatGPT recommendations need to be featured in the editorial review sites that ChatGPT weights most heavily. Brands targeting Perplexity need fresh, citable content. Brands targeting Gemini need competitive pricing and active merchant feeds. And brands targeting Claude need verifiable claims with strong evidence bases. The era of one-size-fits-all SEO is definitively over.

What This Means for Your Brand

The shift from search-based discovery to agent-based recommendation is not hypothetical - it's happening now. A fast-growing share of product research now starts with AI rather than a search box: nearly 60% of shoppers report using gen-AI tools to shop, and Adobe Analytics found traffic to US retail sites from generative-AI sources jumped roughly 1,200% year over year. For brands, this creates a new optimization surface that most competitors haven't even identified, let alone started working on. The question isn't whether AI agents will influence your sales - it's whether you'll be the brand they recommend or the brand they don't mention.

The good news is that the optimization required isn't mysterious. It's structural: implement complete Schema.org markup, cultivate detailed reviews that describe specific use cases, build authoritative brand presence across editorial and expert sources, maintain competitive (not necessarily cheapest) pricing, write rich product descriptions, and keep products in stock. None of these are revolutionary - but the combination, executed consistently, creates the signal profile that AI agents need to recommend with confidence.

The brands that act on these findings in the next 6-12 months will build the compounding advantage we described in Finding #4. Those that wait will find themselves fighting an increasingly uphill battle against established recommendation incumbents. In AI agent optimization, as in so many competitive domains, the best time to start was six months ago. The second best time is today.