How AI shopping systems choose which products to recommend

For retailers, the practical question is no longer only “can this product be found?” but “does the available product information make its suitability clear?” A shopper may not search for a category name at all. They might ask for:

• a smart dress that will not crease in a suitcase;
• a lightweight waterproof coat for dog walking;
• art supplies suitable for a university student starting illustration;
• a gift for someone who loves coffee but does not own a grinder.

Those requests combine product type with use case, behaviour, audience and constraints. They are the requests SO-ME, the consumer shopping platform Sighted powers, receives from shoppers, and they differ from a keyword search in a way that decides what product data has to hold. Matching a request in the shopper's own words to the product information retailers already publish is difficult, because conversational discovery needs more depth about a product than a keyword search.

There is no single universal pipeline across ChatGPT, Google, Gemini, Perplexity and retailer-specific agents. Depending on the platform, product information can come from combinations of:

• merchant or commerce feeds;
• platform catalogues and integrations;
• Merchant Center and other product databases;
• crawlable product pages;
• structured product data;
• merchant availability, price and fulfilment information;
• reviews and other web sources;
• contextual information supplied by the shopper.

OpenAI says ChatGPT shopping results are selected based on relevance to the user's intent and may consider factors such as price, reviews and ease of use. It also takes product feeds from approved partners, which puts a retailer's own product data in front of ChatGPT directly. The two decisions ChatGPT makes, and what it weighs in each, are taken apart in the ChatGPT Shopping ranking factors guide.

Google has moved further in the same direction by adding optional conversational attributes to Merchant Center specifically to help AI systems understand product nuances; the Merchant Center conversational attributes guide lists them.

No one field “ranks” a product. Machines now have more structured ways to evaluate whether a product satisfies a natural-language shopping need.

Sighted uses a six-stage model: eligible, retrieved and understood, surfaced, recommended, clicked, purchased. A retailer can lose at several of those stages:

• The product is absent from a source the platform uses.
• It is present but has incomplete or contradictory data.
• It surfaces for a broad category but lacks enough information to satisfy a more specific request.
• It is recommended but loses the click because of price, availability, delivery or merchant confidence.

That is why a single “AI visibility score” is useful only if it leads to a diagnosis.

In Sighted's pilot with The Game Collection, 200 products were enriched and 17 collections created. AI citation coverage went from 0% to 32% of the priority products tested, and the store's organic conversion rate rose 64%. The coverage figure is a share of a defined set of products, which is the kind of measure the six stages call for, where a single rank position is not.

Traditional catalogue data is often very good at answering what a product is. AI-led shopping creates more demand for data that also helps answer:

• Who is it suitable for?
• What problem does it solve?
• What occasion does it fit?
• How does the material behave?
• What is compatible with it?
• Which variant is right for this particular constraint?

Some of that information belongs in standard product attributes, some on the product page, some in channel-specific attributes, and some should not be invented at all. The job is to make useful, supportable product facts available in the places shopping systems can use them, rather than to generate more words; the feed, schema and product page guide separates those places.

Do AI assistants simply use Google rankings?

No. Different systems use different combinations of web information, product catalogues, feeds and platform-specific systems. Conventional search visibility can still matter because crawlable web content is one source, but it is not accurate to reduce AI shopping selection to a Google rank position.

Does schema make a product more likely to be recommended?

Structured data helps systems understand explicitly marked-up information and remains good ecommerce and SEO practice, but there is no credible basis for claiming that adding schema alone causes a fixed uplift in AI recommendations.

Can retailers pay for organic product recommendations?

Paid ads and organic product recommendations should be treated separately. On ChatGPT, OpenAI says product results are selected independently and ads do not influence the assistant's answers.

1. Establish which AI shopping surfaces matter for your category and market.
2. Define a realistic prompt set based on actual shopper intent.
3. Check whether your catalogue is present and correctly represented. For ChatGPT, the troubleshooting guide runs the eligibility checks in order.
4. Identify the specific information missing from products that lose relevant recommendations.
5. Publish validated enrichment into the appropriate source: feed, Merchant Center, product page, structured data or commerce integration.
6. Rerun the same tests and track downstream commercial signals.

The free AI visibility audit is the first measurement: it reads up to 50 of your product pages and reports what an assistant can and cannot read on each. Inside the product, Sighted enriches the catalogue at scale, writing the attributes each product's category needs from your own descriptions, options and photos for your approval, and exports the result to Merchant Center and the assistants' feeds. The same enrichment technology powers SO-ME, where the requests in the first section arrive from the shopper's side.