Why AI shows a competitor instead of your product

Usually not because a competitor won a ranking. The model could not confirm which product yours is, or found a more quotable description elsewhere, or had a retailer's page as the only linkable destination. Identity, reachable facts and an official destination are addressable. Ranking is not.

The question is not "why did they rank higher"

There is no ranking. A model is asked "which trekking backpack around 45 litres should I buy", and it composes an answer from what it can retrieve and is willing to state. Your product can be absent from that answer for reasons that have nothing to do with quality or with anyone outbidding you.

Four causes we can distinguish from our own measurements, and they need different fixes.

1. The model cannot tell which product is yours

Product identity is the most common break. A brand publishes "Trail 45" with no GTIN, no MPN, and three variants that differ only by a colour name in the title. A retailer publishes the same product with a different name, a different capacity — last season's — and a barcode. The model has two candidate products and one is better identified.

What to check: open your product page with JavaScript disabled. If the identifiers are not in that HTML, they are not in what the crawler read.

2. Your facts are not reachable without a browser

Storefronts render specifications in the browser. Crawlers do not run them. The page that a model reads is often a headline, a price and a marketing paragraph — everything a buyer asks about lives in a tab that never executed.

3. Your page is not a destination the model can link

A model answers with somewhere to go. If the only linkable thing is a retailer's listing, that is what appears, and the retailer becomes the authority about your product. We keep the official destination as a separate, explicitly recorded field for this reason: "product shown" and "official URL shown" are two different outcomes, and only the second one is worth calling a success.

4. Someone else's description is simply more quotable

A page that states a fact plainly — capacity, compatibility, what is in the box, what the product does not do — gives a model something to reproduce. A page of adjectives does not. Between 26.07.2026 and 16.09.2026 we stored 4,078 model answers about products with 22,789 extracted citations; the pages that get cited are consistently the ones that answer the question in sentences.

How to check your own products in an hour

Pick five products and one buyer question each — the question a customer

actually asks, not your category name.

Ask each question in ChatGPT, Perplexity, Gemini and Claude. Record for

each: the provider, the exact model name and the date. A model's behaviour changes under a stable name, and an observation without the model and date is not checkable six months later.

Write down three things per answer: was your product named, whose URL was

shown, and was any stated fact wrong.

Repeat in two weeks without changing anything. That is your baseline. Any

change you make afterwards is measured against it.

This is exactly the procedure we run, and the reason our pages report ranges of days rather than a single result: gpt-4o-mini has been measured for 39 days, sonar 30, grok-4-latest 18, gemini-2.5-flash 15, claude-haiku-4-5-20251001 4. One answer is an anecdote; a series is evidence.

What we will not tell you

That fixing these four things guarantees a recommendation. It does not, and anyone promising it is selling a ranking that does not exist. What is provable is narrower and more useful: whether your product is identifiable, whether its facts are reachable, whether your official page is the destination, and whether the answer changed after you fixed something. The ladder we use to keep those apart is at /evidence.