A shopper types a sentence into ChatGPT and gets three to five products with reasons. No results page, no ten blue links, no "sponsored" row at the top of the list. For a brand, the question is what put those products on the shortlist and kept yours off it.
This post explains the mechanism: the inputs ChatGPT uses, how each one shapes the recommendation, why answers vary, and what a brand can and cannot influence. For the practical to-do list, see the ChatGPT shopping optimization checklist.
What happens between the question and the shortlist?
Think of a recommendation as four steps. OpenAI does not publish its internals, so this is a model of the process built from what it does document, not a description of its code.
- Read the request. ChatGPT works out what the shopper wants: the product type, the constraints (budget, size, use), and the priorities between them. OpenAI says it assesses intent from the question and other context it has, such as memories or custom instructions.
- Gather candidates. It looks for products that fit, using structured product data from third-party providers and web content it can retrieve.
- Weigh them. It compares candidates against the request, using what the data and third-party content say: reviews, guides, discussions.
- Write the answer. It picks a few, explains why each fits, and shows where to buy.
Every step is a place a product can drop out.
How does the shopper's wording change the result?
More than most brands expect. "Best coffee grinder" and "quiet coffee grinder for a small kitchen under a modest budget" are different questions with different winners. The second has three constraints, and only products whose descriptions and reviews mention all three survive step 3.
This cuts both ways. A big brand that is well described in general can lose a specific question to a small brand whose page and reviews state exactly that fit. If you sell into a niche, the questions that describe the niche are where you can win.
Context adds a second layer. A shopper whose memories say they are vegan, or who mentioned earlier that they live in a small flat, will get answers shaped by that. Two people typing identical words can get different shortlists.
Where does ChatGPT get product facts?
From structured product data and from the web. OpenAI says it uses structured metadata from third-party providers, such as price and product description, alongside other third-party content such as reviews.
In practice that means the facts ChatGPT states about your product, such as its price, availability and key specifications, may come from a source other than your own site. When the sources agree, that is fine. When a comparison site still shows last season's price, or a marketplace listing has a different description, the engine has to choose.
The practical rule: your product facts should be identical on your product page, in your structured data and in every listing you control.
Why do reviews matter so much?
Because the answer needs reasons, and reviews are where reasons live. "Quiet", "easy to clean", "runs small", "good value": these are the words that let ChatGPT say why a product fits a request. Specifications say what a product is; reviews say what it is like.
OpenAI gives an example worth noticing: a label such as "budget-friendly" can mean that reviewers often mention good value, not that the product is the cheapest available. In other words, the way a product is described in an answer can come from what reviewers say, not from your data.
This is also where sentiment enters. A product with many reviews mentioning a flaw will be recommended with that caveat, or not at all for questions where the flaw matters. See brand sentiment tracking for how to measure the stance answers take.
What about buying guides and roundups?
They act as pre-made shortlists. When a guide titled "the best quiet grinders" is among the pages ChatGPT reads, its list is a strong signal for which candidates belong in the answer. A product that is in the guides for its category starts with an advantage; a product in none of them has to win on data and reviews alone.
That makes buying-guide coverage one of the highest-value pieces of work in AI commerce, and also one of the slowest. It is classic PR: a product worth writing about, sent to the right editor, with the facts they need.
Why does the same question return different products?
Three reasons, in order of size:
- Generation varies. Answers are written fresh each time. Even with the same inputs, the shortlist can change between samples.
- Context varies. Memories, earlier messages and settings change what the shopper is taken to want.
- Sources change. Prices, stock, reviews and guides update, and the engine's view updates with them.
The consequence for measurement is simple: one answer proves nothing. Ask each question several times on several days, and count how often each product appears. A product named in 9 of 20 answers is in a different position from one named in 1 of 20, and you will not see the difference from a single screenshot.
How do you test what ChatGPT recommends in your category?
Run a small, repeatable test rather than a single search.
- Write ten to thirty shopping questions the way your customers describe their needs. Mix broad ("best coffee grinder") with specific ("quiet grinder for a small kitchen").
- Use a fresh session without memories or custom instructions, so the answer is closer to what a new shopper sees.
- Ask each question three times, and repeat on three different days.
- Record every product named, the merchants shown, the reasons given and the sources cited.
- Count per product. "Named in 4 of 9 answers to the specific question, 0 of 9 to the broad one" tells you where you are competitive.
- Read the reasons. The phrases ChatGPT uses to justify a competitor are the phrases your own description and reviews are missing.
Repeat the same test monthly, with the same questions. The trend matters more than any single round, and a fixed question set is the only way to see one.