Ask an assistant which product to buy and it does not consult one ranked list. It rewrites your question into several queries. It pulls from a live search index and a merchant feed at once, then writes an answer from both. The products it names and the sources it links are chosen by two different mechanisms. Understanding how AI assistants choose products matters, because most catalogue owners are working on the wrong one.
We ran a test rather than theorising. Below is the exact question, the date, and everything we observed.
What we asked, and what came back
On 20 August 2026 we put this exact question to two assistants.
best cordless impact wrench for removing lorry wheel nuts uk
We picked real buying intent, a UK locale, and a technical constraint. It is the kind of thing a fleet workshop manager actually types.
Run one: Google AI Overview
We captured the live Google result through SE Ranking’s SERP API. Settings: United Kingdom, English, desktop. An AI Overview fired at the top of the page.
It recommended three named products. A Milwaukee M18 FUEL high torque impact wrench, an Alloyman high torque cordless impact wrench, and a DeWalt DCF900N 18V XR. Each came as a product card carrying a price, a seller and a star rating. The Milwaukee card showed 344.99 pounds with UK Planet Tools named as a seller. The Alloyman showed 129.99 pounds through an Amazon.co.uk seller. The DeWalt showed 189.99 pounds with Powertoolmate named.
The reasoning was attribute-led. It told the reader to check nut busting torque in newton metres, battery capacity in amp hours, and square drive size. It quoted a fastening torque of 1491 Nm and a nut busting torque of 2034 Nm for the Milwaukee, lifted straight from listing copy.
Now the part that matters. It cited only two editorial sources by link. One was a PistonHeads forum thread dated 17 January 2022. The other was a YouTube video published 9 August 2025. The product blurbs were attributed to Google, not to any retailer. Not one manufacturer page and not one distributor page appeared as a cited link.
Below the overview sat a discussions and forums block. It held two Facebook group posts and the same PistonHeads thread. The top of the organic results ran Car and Driver, Auto Express, Parma Group and Pro Tool Reviews. Below those came an Amazon category page, Facebook, a Screwfix category page, PistonHeads, Reddit and Machine Mart.
Run two: Claude with web search
We put the identical string through Claude Opus 5 with its web search tool. Same day. The retrieval layer returned seven results.
Auto Express came back with a 2024 group test. Then a Detailing World forum thread, a review blog called BrazeTools, and a Facebook group post. After those came the same PistonHeads thread, a United States motorcycle forum, and a syndicated round-up on AOL.
Zero product detail pages. Zero manufacturer pages. Zero distributor pages. The assistant had no retailer product data in front of it at all.
What we could not test
We had no direct access to the ChatGPT, Gemini or Perplexity consumer interfaces. So we are not reporting results for them. We planned to test five assistants and completed two properly. We are not going to invent the other three.
One supporting observation is worth adding. We track a panel of buying questions across ChatGPT, Google AI Overview and Google AI Mode for our own domain. On the same day, a services buying question in our category returned an answer citing twelve URLs. The cited set mixed supplier websites with third-party listicles and directories. The order of the citations did not match the order of the recommendations. Citation and recommendation are not the same thing, and the two lists genuinely differ.
How AI assistants choose products: the short answer
AI product recommendations draw on three separate stores of information. They behave differently and you influence them differently.
The live retrieval layer. OpenAI’s help documentation says ChatGPT may “rewrite your query into one or more targeted queries that it sends those providers”. Those providers are third-party search engines. Google describes the same shape for AI Mode as a “query fan-out” technique, “issuing multiple related searches concurrently across subtopics and multiple data sources”. Your page competes here on the same footing as a forum thread from 2022.
The merchant data layer. This is where the named products come from. Google’s Shopping Graph holds “more than 50 billion product listings, 2 billion of which are updated every hour”. That is Google’s own figure from November 2025. OpenAI now runs its own product feed for ChatGPT. Its help centre says product results “are not ads, nor influenced by any OpenAI partnerships”, and are selected independently.
The model’s priors. Whatever the model already believes about brands from training. You cannot edit this and nobody can measure it cleanly.
The Google test above shows the split perfectly. The recommendations came from the merchant layer. The citations came from the retrieval layer. A retailer with a good feed and a thin content site gets the product card. A forum gets the link.
The feed is where the product decision is made
If you sell the product, the feed is your highest-value asset. It is also the one most often maintained by whoever set up Google Shopping four years ago.
OpenAI’s product feed specification is public and it is specific. Required fields include item_id, title capped at 150 characters, and description capped at 5,000. Also required are url, brand capped at 70 characters, image_url, and price with an ISO 4217 currency code. Then availability from a fixed enumeration, plus seller_name and target_countries. Two flags, is_eligible_search and is_eligible_checkout, control what the item can be used for.
The recommended fields are the interesting ones for anyone who has built an attribute schema. gtin and mpn. product_category as a path with a greater-than separator. material. dimensions in the form LxWxH unit, plus separate length, width, height and dimensions_unit fields. weight and item_weight_unit. group_id and variant_dict for variant families. review_count and star_rating. OpenAI’s guidance is to supply a valid gtin or mpn unless you explicitly declare that no identifier exists.
Read that list next to a typical distributor catalogue and the gap is obvious. Torque sits in a description string. Weight is a number with the unit typed into the same cell. Variants are separate unrelated items. Half the range has no MPN because the ERP never needed one.
This is ordinary product content enrichment work. It is now the difference between being a named product and being invisible. Our applied AI work almost always starts with a feed diff rather than a content project. The feed shows you exactly which attributes an assistant can filter on today.
Structured data on the product page
There is a lot of noise about schema markup as a route into AI answers. Google’s AI features page addresses it directly.
You don’t need to create new machine readable files, AI text files, or markup to appear in these features. There’s also no special schema.org structured data that you need to add.
That kills the llms.txt argument and the idea of AI-specific markup. It does not kill Product structured data, because that serves a different purpose.
For merchant listing experiences, Google requires name, image and an offers object holding price and priceCurrency. It recommends brand.name, sku, the applicable gtin variants, mpn, description and availability. It also recommends shippingDetails and hasMerchantReturnPolicy. Google says that supplying both page markup and a merchant feed maximises your eligibility to these experiences. The two together help Google understand and verify your data.
So the honest position is this. Markup does not buy you a place in an AI Overview. Markup and a feed together make your product eligible for the merchant surfaces an AI Overview draws its product cards from. Those are different claims and vendors routinely blur them.
OpenAI describes its inputs as “structured metadata from first-party and third-party providers” plus “other third-party content”. It says it weighs “availability, price, quality, and whether they are the maker or primary seller of that item”. Being identifiable as the primary seller of an item is a data problem, not a marketing one.
How AI assistants choose products when the question carries constraints
The test question had a constraint buried in it. Lorry wheel nuts imply a torque threshold. Watch what the assistant did with that.
It converted the constraint into an attribute range. It told the reader to look for roughly 1300 to 2000 Nm of breakaway torque. Then it matched products against that range. That is a filter operating on a numeric attribute with a unit.
A product whose torque figure lives only in a paragraph of marketing copy cannot be matched by that filter reliably. A product recorded in foot pounds, while the query reasons in newton metres, is a coin toss. One of the three recommended products carried both units in its listing title. That is presumably why it survived.
This is the part most teams underestimate. The assistant is not reading your brand story. It is looking for a number, a unit and an identifier. We covered the same pattern in industrial product data and technical specs. AI retrieval has not changed the underlying requirement. It has raised the penalty for getting it wrong.
Three fixes cover most of it. Get every filterable spec out of prose and into a typed attribute with a unit. Normalise units within an attribute rather than storing whatever the supplier sent. Link variant families, using group_id in a feed or isVariantOf in markup. The assistant then sees one product with options, not nine unrelated items. Our product attributes piece covers the schema design side.
What nobody knows about how AI assistants choose products
We would rather say this plainly than pretend to a certainty nobody has.
No assistant publishes its ranking function. Perplexity’s own help documentation says only that it searches the web, identifies trusted sources, and returns numbered citations. It does not explain how trust is scored or how sources are ranked. Google says AI features surface “a wider and more diverse set of helpful links”. It does not say how those links are chosen. OpenAI names its inputs but not their weights.
Answers are also not stable. Run the same question twice and you can get different products and different citations. Any single observation, including ours, is a sample and not a measurement.
The surfaces are moving fast too. OpenAI published its product discovery programme in March 2026. Google’s generative AI reporting for site owners only reached Search Console in June 2026. Anything written about this in 2024 describes a different system.
What is not in doubt is the direction. Both major assistants now take product facts from a structured feed. Both take narrative from the open web. That split is stable even while the ranking is not.
Key takeaways
- Recommendations and citations come from different pipelines. In our 20 August 2026 test, Google recommended three products from merchant data. Its links went to a 2022 forum thread and a YouTube video.
- Claude’s retrieval layer returned seven sources for the same question. Not one was a product, manufacturer or distributor page.
- The merchant feed is where the product decision happens. OpenAI’s feed specification asks for identifiers, category paths, materials, dimensions with units, weights and variant groups.
- Google states directly that no special schema.org markup is needed for AI features. Markup plus a feed is what buys merchant eligibility, which is a different claim.
- Constraints in a question become attribute filters. Specifications trapped in prose, or stored in inconsistent units, do not survive that step.
- Nobody outside these companies knows the ranking weights. Treat any confident claim about them, including a consultancy’s, as a hypothesis to test.
If you want to know how your own catalogue behaves here, the test is cheap to run and uncomfortable to read. We run a fixed panel of buying questions for your category. We capture what gets recommended and what gets cited. Then we trace each gap back to the field that caused it. Thirty minutes is enough to tell whether it is worth doing. Get in touch to talk it through, or start with our applied AI expertise page. Our AI in product data piece is the wider view.