Learn / Assistants
How ChatGPT shopping chooses products
An assistant choosing a product works through the constraints a shopper stated, filters candidates that fail them, and chooses among the survivors. Most of what decides the outcome is therefore arithmetic on published facts rather than judgement about quality, which is why a store with a genuinely better offer can lose to one that simply published its shipping cost.
Selection is constraint checking before it is judgement. A shopper says what they want, each clause becomes a filter, and candidates that fail a filter are removed before anything resembling taste is applied.
That ordering is the single most useful thing for a merchant to understand, because nearly all losses happen in the filtering stage rather than in the final choice.
What the filters are made of
A request like "a waterproof jacket under a budget, in stock, delivered this week, returnable" is four filters and one preference. Each filter is answered from published data or not answered at all.
- Budget, which needs a delivered total rather than an item price.
- Stock, at the variant the shopper asked for rather than the product in general.
- Delivery, which needs a stated expectation rather than a calculation at checkout.
- Returns, which needs terms in text rather than in a graphic.
Silence is not treated as a pass
The asymmetry that costs merchants most is between "does not meet this" and "does not say". The first is a commercial answer you can live with. The second removes you, because an agent acting for somebody does not resolve an unknown optimistically.
A store with genuinely cheap delivery and no published rule loses to a competitor with worse delivery and a stated number. See price transparency and AI agents.
Identity decides whether you are compared at all
Two listings of the same product differ in title, photography and description. Deciding they are the same item is a guess unless an identifier says otherwise, and an unmatched listing is not a competing offer, it is a separate unknown.
See how to add GTINs to your catalogue.
What to do with this
Audit a product against the five ways your customers describe what they want, and check each one is answerable from what you publish. That is a more useful exercise than any amount of copy rewriting. See how to appear in ChatGPT shopping results.
Questions
Can I influence the choice with better copy?
Where the copy states a checkable fact, yes. Where it makes an unverifiable claim, no. "Best in class" answers no constraint; "waterproof to a stated rating, ships in two days, free returns for thirty" answers three.
Does brand reputation count?
Less directly than you would expect at the filtering stage, which is where most candidates are lost. Reputation matters more among survivors, and being a survivor is decided by data.
Why did it pick a more expensive competitor?
Most often because the competitor’s total was computable and yours was not, or because a constraint you meet was not stated anywhere it could be read. An unknown gets resolved against you rather than in your favour.