- The assistant selects rather than lists: being "roughly relevant" is no longer enough.
- It compares attributes: what is not written does not exist for it.
- It cross-checks sources: a contradiction creates doubt.
:::
Cause 1: descriptions that say nothing verifiable
"Comfortable, elegant, designed for you": these words allow no comparison. A page that helps an assistant answers the buyer's questions: material, dimensions, weight, intended use, compatibility, care, limits.
Fix: for each category, list the questions customers actually ask (customer service, reviews, on-site searches) and check that every page answers them with a fact. Brand messaging stays, but after the facts.
Cause 2: attributes missing or buried in free text
Many catalogues store colour, size or composition in a paragraph rather than in fields. A human copes; a system comparing products copes much less well.
Fix: structure attributes in your catalogue system (PIM or e-commerce back office) with normalised values: "100% cotton" everywhere, not "cotton", "pure cotton" and "coton" depending on the page.
Cause 3: information that contradicts itself
The price differs between the page and the Google Merchant Center feed, the page says "in stock" while the feed says "sold out", the marketplace shows another composition. Faced with contradictions, an assistant has good reason to prefer a competitor whose data agrees.
Fix: a single source of truth for each product data point, feeding the site, the feeds and the marketplaces. Among its best practices for its AI features, Google itself recommends keeping Merchant Center information up to date and structured data consistent with the visible text of the page.
Cause 4: structured data missing or incomplete
Schema.org markup (Product, Offer, AggregateRating…) describes the page in a machine-readable format. It is not a magic formula: Google states that no special markup is needed to appear in its AI answers. But complete, accurate markup (brand, reference, GTIN, price, availability, attributes) reduces ambiguity.
Fix: audit the markup of a sample of pages per category, fill in missing fields and check they match exactly what the page displays.
Cause 5: missing from the feeds assistants read
Assistants do not just crawl pages. Google relies on Merchant Center for its shopping experiences. OpenAI has published a product feed specification for ChatGPT, with required fields such as title, description, brand, image, availability and price. A catalogue missing from these feeds, or present with empty fields, starts at a disadvantage.
Fix: check which feeds you supply, which fields are filled in, and how often prices and stock levels are updated there.
Cause 6: a page robots read poorly
Content loaded only through JavaScript, tabs hiding the specifications, pages blocked by mistake in robots.txt: if a robot cannot see the information, it cannot use it.
Fix: view the page as a robot receives it (URL inspection in Search Console, for instance) and check that the text, attributes and prices are there.
Where to start
:::etapes
- Measure: check whether ChatGPT and Gemini mention your brand on your key queries, for example with our free GEO Radar tool.
- Sample: audit around twenty pages per category against the six causes above.
- Fix at the source: normalise attributes in the catalogue, then push them to the site and feeds.
- Scale: when the catalogue holds thousands of references, AI can enrich pages from supplier data, under human review.
