Why a product called Summer Dress is invisible to AI
Last reviewed
Short answer
Shoppers ask assistants constrained questions: linen, under $80, ships to Canada, machine washable. An assistant answers by matching constraints against product data, and a product titled Summer Dress with a two-line description offers nothing to match. Attribute completeness, in titles, descriptions and structured fields, is what makes a product answerable, and the published feed specs reward exactly that.
The question changed shape
A Google shopper typed summer dress and did their own filtering on the results page. A ChatGPT shopper says something like a midi dress for a June wedding in Portugal, linen or cotton, under a hundred dollars, and expects the filtering done for them. OpenAI's own description of selection is matching against the shopper's intent using structured metadata such as price and product description.
Every constraint in that sentence is an attribute. Fabric, occasion, length, budget, availability. A product that states its attributes can be matched to five constraints. A product that states none can be matched to none, and no amount of brand charm fixes an empty field.
Where attributes live, in order of leverage
The title is first because it travels everywhere: the feed caps it at 150 characters, and the useful shape is brand, product type, then the two or three attributes a buyer would actually filter on. The description is second, with 5,000 plain-text characters to state material, dimensions, care, fit and use cases in sentences an assistant can lift. Structured fields are third: vendor, product type, variant options, and identifiers like GTIN where you have them.
OpenAI's guidance says recommended attributes improve ranking, relevance, and user trust, and names reviews and rich media among them. So the ladder is: title attributes for matching, description substance for answering, structured fields for certainty, evidence for trust. The title rung has its own page.
The five attributes most catalogs are missing
Graded catalogs fail in the same places. These five cover most of the gap between unanswerable and answerable, and none of them requires a copywriter.
- Material or composition, the single most common constraint in product prompts
- Dimensions or size with units, stated in the description, not only in a variant dropdown
- Compatibility or fit: what it works with, what it does not
- Care and practical facts: washable, dishwasher safe, weatherproof
- The identifier, GTIN or MPN, when the product has one, so the assistant knows which product this is at all
No GTIN? State what you are
Handmade and private-label products often have no barcode, and the feed spec treats identifiers as optional, so this is a gap, not a wall. What replaces certainty is description: brand, materials, origin and precise product type, stated plainly, so the assistant can place the product without an identifier.
The trap to avoid is inventing identifiers or borrowing a similar product's barcode, which makes the assistant confidently wrong about what you sell. An honest none beats a wrong number.
The test that takes one minute
Pick your best-selling product. Write down the three questions a buyer would ask before purchasing it. Then read only your product title and description and try to answer them. If the answers are not on the page, they are not in the feed, and the assistant does not have them either.
Scaling that test across the catalog is the tedious part, and it is what the free score exists for: one grade per product against the published rules, weakest products first.
Frequently asked questions
Which attributes matter most?
The ones your buyers filter on, which varies by category: material and size for apparel, dimensions and compatibility for home and electronics, ingredients and quantity for consumables. Write the three questions buyers ask before purchase, then make the answers literal text in the title and description.
Should I stuff attributes into the title?
No. The spec caps titles at 150 characters and warns against all-caps, and ChatGPT may rewrite titles it receives. Two or three high-value attributes in the title, everything else in the description as sentences.
Do variant options count as attributes?
Yes, variants carry size and color as structured data. But an assistant answering a constraint question also reads the description, so key attributes should appear there in words too, not only inside a dropdown.
Is this different from normal SEO advice?
It overlaps and diverges. Google rewarded keywords and let the shopper filter. Assistants do the filtering themselves, so completeness beats repetition: stating linen once matters, stating dress nine times does not.
Sources
Keep reading
Score your own catalog
Entitled grades every product against the rules on this page, explains each failure, and fixes them with AI you approve. The score is free forever.