Your catalog is talking to robots now. Here's what they hear

Every product page you publish has two audiences now. One is the shopper you designed it for: the photography, the layout, the persuasive copy. The other is a machine that never sees any of that. When an AI assistant decides whether to recommend your product, it is not looking at your page. It is listening to your data. And what it hears is often very different from what you think you are saying.

So what does the machine actually hear? Roughly five things.

1. Structured product data. The cleanest signal is structured markup and feeds, the machine-readable version of your catalog (name, price, availability, brand, materials, sizes, GTINs). This is the difference between the assistant knowing a fact about your product and inferring it. Assistants increasingly read product feeds and structured data directly; if the attribute a shopper asked about is not in there, the assistant cannot confirm the match, and is unlikely to recommend on a guess.

2. Descriptions, for meaning, not style. The assistant reads your descriptions, but not the way a copywriter wrote them. It extracts meaning and differentiation. Two products with near-identical boilerplate sound like the same product to a machine, which makes it hard to recommend one over the other. Specific, distinct descriptions are how the assistant tells your products apart.

3. Claims, checked against each other. When your description says "waterproof" and your spec sheet says "water-resistant," a human shrugs. A machine notices the conflict and gets cautious. Assistants tend to avoid asserting claims they cannot reconcile, because shoppers punish errors and verify what they are told (Yext, 2026). Internally consistent claims get repeated confidently; contradictory ones get dropped.

4. Reviews and the wider web. The assistant does not only hear you. It hears what reviews, forums, and authoritative sources say about your brand and products, and it weighs whether the outside world backs up your catalog's story. Shoppers do the same: after a recommendation, a majority go verify on Google and on the brand's own site (Yext, 2026).

5. The front door: can the bots even get in? Before any of this, the assistant's crawler has to be allowed to read your pages, and your product content has to render without a browser. Plenty of demand is being left on the table here. Adobe found AI-driven retail traffic surging, up 693% over the 2025 holiday, while concluding that many U.S. retail sites are not entirely readable by machines, with product pages averaging just 66% readability (Adobe, April 2026; traffic figure via Digital Commerce 360).

The gap between what you say and what it hears

The reason this matters: the machine's version of your catalog is the one that gets recommended. In a CatalogSignal CEI Benchmark of 100 brands across 10 verticals and four AI providers (more than 100,000 AI shopping queries), mean funnel accuracy was 47.7%, mean hallucination was 9.3%, and mean commercial harm was 10.6%. The inferred harm taxonomy included wrong product claims, blocked purchase paths, bad competitor substitutions, invented products, and wrong return policies. (Directional panel metrics, not shopper-level rates; harm taxonomy inferred; no individual brands reported.) Most of that is not malice; it is the machine filling gaps your data left open, or repeating a conflict your catalog never resolved.

The fix is not to write better copy for humans. It is to make sure the data layer underneath says what you mean: complete attributes, consistent terminology, claims that reconcile, content the crawler can actually reach. When the machine's version of your catalog matches your intended version, recommendations get accurate, and accurate is what converts.

Your catalog is already talking to robots. The only question is whether you know what it is telling them.

See what assistants can read. A baseline Commerce Eligibility Index™ assessment shows exactly how AI reads your catalog, and where what it hears diverges from what you meant. Request one at catalogsignal.com.

Sources