How AI Agents Choose Which Products to Recommend

The AI Agent Decision Pipeline

Before diving into individual signals, it helps to understand the four-stage pipeline that most AI agents follow when responding to a shopping query.

Stage 1: Discovery

The agent identifies candidate products that could satisfy the query. This stage draws from product feeds (Google Merchant Center, Microsoft Merchant Center), crawled web data (schema markup on your product pages), and indexed content (your site's LLMs.txt file and structured content). If your products are not in any of these data sources, the agent cannot discover you at all. You are invisible.

Stage 2: Evaluation

Once discovered, each product is evaluated against the query intent. Does it match what the shopper asked for? Are the specifications correct? Is it in stock? Is the price within the shopper's stated or implied budget? This is where structured data quality becomes decisive. An agent with complete Product schema, accurate pricing, and detailed attributes can confidently evaluate your product. Without that data, the agent moves on to a competitor it can evaluate with certainty.

Stage 3: Ranking

Products that pass evaluation are ranked against each other. This is where trust signals, reviews, pricing competitiveness, and protocol readiness create separation. Two products that both match the query will be ranked based on which merchant provides more reliable data, better value, and stronger trust indicators. This stage is where most of the competitive battle happens.

Stage 4: Presentation

The top-ranked products are presented to the shopper. How they appear depends on the agent. Google AI Mode might show a rich product card with images, pricing, and a buy button. ChatGPT might describe the product in conversational text with a purchase link. Perplexity might present a comparison table. The presentation quality depends on how much data the agent was able to extract from your structured markup, and richer data means richer presentation.

Signal 1: Structured Data Quality

Structured data is the single most influential factor in AI product recommendations. Princeton researchers found that GPT-4's accuracy at recommending the right product jumped from 16% to 54% when structured data was available. That is a 3.4x improvement from a single signal category.

AI agents rely on schema.org markup to understand your products. Without it, they are parsing unstructured HTML and guessing at product attributes. With it, they have machine-readable data they can evaluate with confidence.

What Agents Look For in Your Schema

• Product schema completeness: Name, description, brand, SKU, GTIN, images, and category. The more fields populated, the higher the agent's confidence in your product data.
• Offer schema accuracy: Price, currency, availability status, seller information, and valid-through dates. Agents cross-reference this with feed data and landing page content. Mismatches between your schema and your visible page content erode trust immediately.
• AggregateRating schema: Star rating, review count, and best/worst rating. Agents use this for social proof evaluation without having to crawl third-party review sites separately.
• BreadcrumbList schema: Helps agents understand your site hierarchy and product categorization. This aids in matching products to category-level queries like "best trail running shoes."

Signal 2: Product Feed Completeness

Your product feed is the primary data pipeline between your store and AI agents. Google AI Mode, for example, draws recommendations directly from Merchant Center feed data rather than crawling your website in real time. If your feed is incomplete, you are invisible to the largest AI shopping platform on the web.

Required Attributes Are Just the Starting Line

Meeting the minimum required attributes (title, description, price, availability, image, link) gets your products into the feed. But AI agents evaluate far beyond the minimum. The optional attributes are where competitive advantage lives, because most merchants do not bother filling them out.

High-Impact Optional Attributes

• product_highlight: Up to 10 short bullet points summarizing key features. AI agents use these as the primary feature summary when comparing products against a query. Think of these as the talking points an agent uses to pitch your product.
• product_detail: Structured attribute-value pairs (e.g., "Material: Gore-Tex" or "Battery Life: 40 hours"). These give agents precise specification data for technical queries like "waterproof jacket with pit zips" or "noise-canceling headphones with 30+ hour battery."
• GTIN / UPC / EAN: Global Trade Item Numbers allow agents to match your product against the same item from other merchants. This enables price comparison and competitive evaluation. Without a GTIN, your product exists in isolation and cannot be compared on value.
• shipping and return_policy_label: Agents factor total cost (product price plus shipping) and return flexibility into their value assessment. Free returns can tip a recommendation in your favor over a slightly cheaper competitor.

Feed Freshness Matters

If an AI agent recommends your product at $79.99 and the shopper lands on a page showing $89.99, that agent's credibility suffers. AI systems penalize merchants with frequent data mismatches by reducing recommendation frequency. Target a feed refresh interval of 15 minutes or less, using event-driven updates via the Content API for Shopping rather than relying on scheduled file uploads.

Signal 3: Merchant Trust Signals

AI agents do not just evaluate products in isolation. They evaluate the merchant behind the product. An agent recommending a product is implicitly vouching for the merchant, and agents protect their own credibility by favoring trustworthy sellers.

What Agents Assess Algorithmically

• Business age and domain history: Established domains with consistent product data histories are weighted more favorably than brand-new stores with no track record. Domain age is not decisive on its own, but it contributes to a composite trust score.
• SSL and security: HTTPS is a baseline requirement. Missing SSL means your products will not appear in AI recommendations at all. This is table stakes, not a differentiator.
• Contact information accessibility: Agents check for visible phone numbers, email addresses, and physical addresses. Hidden or missing contact information is a red flag that suggests a merchant is not confident enough to be reachable.
• Policy pages: Return policies, shipping policies, and terms of service pages signal operational maturity. Agents verify these pages exist and contain substantive content. A return policy page that says "Contact us for returns" is far weaker than one with clear timelines, conditions, and process steps.
• Customer service accessibility: Live chat, visible support hours, and multiple contact channels all contribute to the merchant trust score. Agents evaluate this algorithmically by checking for the presence and quality of support infrastructure on your site.

Signal 4: Reviews and Reputation

Reviews are the social proof layer that AI agents weigh heavily, particularly for non-branded queries. And here is why that matters so much: 95% of AI shopping searches do not include a brand name. Shoppers ask for "best wireless earbuds for running" not "best Sony earbuds." When brand is not specified, reviews become the primary trust differentiator between competing products.

Review Signals AI Agents Evaluate

• Aggregate rating: The average star rating across all reviews. Agents treat this as a confidence-weighted score, meaning a 4.3 with 500 reviews carries more weight than a 4.8 with 12 reviews. Volume calibrates confidence.
• Review count: Higher review counts signal a more established product with broader market validation. For two products with similar ratings, review count breaks the tie.
• Review recency: Recent reviews are weighted more heavily than older ones. A product with 200 reviews all from 2023 is less trustworthy than one with 150 reviews where 40 are from the last three months. Recency signals ongoing quality and relevance.
• Sentiment analysis: AI agents do not just look at star ratings. They parse review text to identify recurring complaints (shipping delays, quality degradation over time, misleading product photos) or consistent praise (durability, customer service responsiveness, value for money). Negative sentiment patterns reduce recommendation likelihood even when the aggregate star rating looks decent.
• Third-party review signals: Agents cross-reference reviews from Trustpilot, Google Business Profile, BBB ratings, and other third-party sources. Consistent ratings across platforms increase trust. A 4.5 on your own site combined with a 2.1 on Trustpilot is a red flag that suggests review manipulation.

Signal 5: Price and Value Positioning

AI agents evaluate price not in isolation but in the context of total value. They are optimizing for the best outcome for the shopper, which means the total cost relative to perceived quality and reliability.

How Agents Assess Value

• Total cost calculation: Agents compute the all-in price: product price plus shipping, plus estimated tax. Free shipping at $79.99 can beat a $69.99 product with $14.99 shipping, because the total cost is lower and the shopping experience feels more transparent.
• GTIN-based price comparison: When a product has a GTIN, agents can compare your price against every other merchant selling the identical item. If you are significantly above market price with no differentiator (faster shipping, better return policy, bundle offer), you lose the recommendation to the lower-priced competitor.
• Promotional offers: Active sale prices, coupon codes surfaced via structured data, and limited-time offers can boost recommendation priority. Agents recognize sale_price and sale_price_effective_date from your feed data and factor them into value calculations.
• Price consistency across channels: If your price differs between your feed, your schema markup, and your landing page, agents lose confidence in your data integrity. Price consistency across all channels is a trust signal in its own right, separate from the actual price level.

Signal 6: Technical Performance

The technical foundation of your site determines whether AI agents can access and trust your data. A beautifully optimized product feed is undermined if your site is down when an agent tries to verify the information or if your pages load too slowly for crawlers to process efficiently.

Performance Signals That Agents Evaluate

• Site speed: Slow-loading product pages increase the risk that an agent times out during verification crawls. Target under 3 seconds for Largest Contentful Paint. Pages that load slowly also correlate with poor user experience, which agents factor in indirectly.
• Uptime reliability: Frequent downtime means agents encounter errors when verifying your product data, which degrades your trust score over time. Each failed verification is a mark against you. Aim for 99.9% uptime or better.
• Crawlability: Your robots.txt and meta robots tags should allow AI crawlers access to product pages, images, and structured data. Blocking CSS or JS resources prevents agents from rendering and verifying your pages properly. Implement an LLMs.txt file to give AI crawlers a structured map of your most important content.
• Mobile readiness: Google uses mobile-first indexing, and AI agents that rely on Google's index inherit this preference. If your mobile product pages are missing schema, have layout issues, or load slowly compared to desktop, you are invisible to the largest mobile-first AI systems.
• API response time: For protocol-based interactions (UCP, ACP), your checkout API endpoints need to respond quickly. Slow APIs degrade the agent-mediated shopping experience and can cause transaction timeouts that reflect poorly on your store.

Signal 7: Protocol Readiness

The newest and fastest-growing signal category is protocol support. UCP (Universal Commerce Protocol) from Google and ACP (Agentic Commerce Protocol) from OpenAI and Stripe enable AI agents to go beyond recommending products to actually facilitating purchases within the agent experience itself.

Why Protocol Support Boosts Recommendations

An agent that can recommend a product AND complete the purchase in one interaction delivers a dramatically better user experience than one that can only link to your site and hope the shopper follows through. Agents actively prioritize merchants that enable this full-loop experience because it increases conversion rates and user satisfaction. The agent looks better when it can close the loop.

• BuyAction schema: Adding BuyAction to your Product schema signals that your product supports programmatic purchase. Semrush research indicates that BuyAction combined with comprehensive structured data increases AI citation rates by 3.2x compared to products without it.
• UCP integration: Google's Universal Commerce Protocol allows agents to query inventory, get real-time pricing, and initiate checkout through your Merchant Center data. This is the path for merchants already invested in the Google ecosystem. Learn more in our agentic commerce guide.
• ACP integration: OpenAI's Agentic Commerce Protocol, powered by Stripe, enables ChatGPT and other agents to facilitate transactions directly from your store. This is particularly relevant for merchants already using Stripe for payment processing.

Protocol readiness is still in its early stages, which means merchants who adopt now gain a significant first-mover advantage. By the time these protocols become table stakes, early adopters will have months or years of trust-building data with AI agents that latecomers will need to earn from scratch.

The Ranking Formula: Putting It All Together

No AI agent publishes its exact ranking algorithm. But based on observed behavior, published research, and the technical architecture of these systems, the seven signal categories combine with roughly the following weighting.

What You Can Do Today

Here are the five highest-impact actions you can take right now to improve your standing with AI shopping agents, ordered by expected return on effort.

Frequently Asked Questions