How ChatGPT picks products and merchants, and why that has nothing to do with ads

Two mistakes are being made in online retail at the same time. The first mistake: whoever puts media budget into ChatGPT ads assumes this also improves their organic standing. The second mistake: whoever spends no media budget on ChatGPT assumes ChatGPT is not their concern. Both are wrong, and OpenAI's own help centre says why in both cases.

1. Two systems, explicitly separated

The sentence sits verbatim in the help article on shopping in ChatGPT Search: product results are selected independently, are not ads, and are not influenced by any partnership. The help article on shopping research states the counterpart from the other direction: ads are separate from shopping research.

For you that means ad budget does not buy organic visibility. And conversely, organic visibility does not disappear just because you do not advertise. These are two pieces of work, and only one of them costs media money.

An observation from 2 September 2026, viewed while logged out. Asked "my espresso does not taste good to me any more, I currently have Tchibo beans, can you recommend something?", ChatGPT shows four product cards: Kimbo Espresso Napoletano, Caffè Borbone Miscela Rossa, Lavazza Crema e Gusto and an espresso from Murnauer Kaffeerösterei. The listed sources are Amazon, Jassas, EDEKA24 and the maker itself. None of the cards is marked as an ad.

The notable part is not the selection but what the cards lead with: 4.8 from 1,237 reviews. 4.9 from 1,172. 4.7 from 986. The review count comes before price and merchant.

At the same time, a paid campaign was running for the coffee merchant measured in section 7, delivering impressions and clicks. It does not appear in this answer. That is the separation from the paragraph above, this time observed rather than quoted. And it shows the currency the organic side counts in: not budget, but reviews.

2. What drives the product selection

A product appears when ChatGPT considers it relevant to the intent of the query. According to OpenAI it draws on:

• structured metadata from first-party and third-party providers, such as price and product description, plus other third-party content

• the model response generated before it considers any new search results

• OpenAI's safety standards and product policies

On top of that comes the person's context: query, memory, custom instructions. State a budget and price gets weighted more heavily. Say you dislike clowns and clown costumes drop out.

Two sentences from the same source temper the expectation and belong in the picture: not all available products will necessarily be shown, and the model can misread intent.

3. What merchants are ranked by

When someone clicks a product, ChatGPT may show a list of merchants offering it. Ranking is based on availability, price, quality, and whether the merchant is the maker or primary seller of the item.

That is the most concrete statement in the whole documentation, and it is uncomfortable. Carry the same item as twenty others and you compete on price and availability. Be the maker or primary seller and you hold an attribute that cannot be discounted away. Be that and fail to make it visible in your data, and you give it away.

4. What ChatGPT does with your copy

Precision pays here, because many shops maintain product copy that never reaches a reader.

ChatGPT generates simplified product titles and descriptions from what it receives from third-party providers, because merchants name the same item differently. Your copy is raw material, not output.

The labels on product images, for example "Budget-friendly" or "Most popular", are generated by the model from available information. OpenAI states explicitly that they are not guarantees and not verified statements. The same holds for review summaries: they come from public sites and are not verified by OpenAI.

And prices: they come from third-party providers, the first price shown is not necessarily the lowest, and changes to price or shipping terms may arrive with a delay. Feed your price data only through third parties and you have no influence on that delay.

5. The point that is not in the marketing but in your firewall

Shopping research, the comparative research inside ChatGPT, works on publicly available retail sites: it reads product pages directly, cites sources, and avoids low-quality sites.

And then comes the sentence this is all about:

"In addition, some retailers block automated access to their sites. When this happens, shopping research will either skip those sources or rely on other sites with similar products."

Block automated access and you get skipped. Not ranked lower, skipped, with competitors carrying similar products taking your place.

The awkward part is that nobody usually made that decision on purpose. It rarely sits in robots.txt, where you would see it. It sits in the WAF, the CDN, a bot mitigation layer or a JavaScript challenge, and those answer with a 403 without anyone noticing. OpenAI lists exactly these layers in its guidance and names what to do for the major providers: a rule on the detection ID for Cloudflare, a bot category for Akamai, nothing at all for Vercel, where it works automatically.

The mechanism behind it is Web Bot Auth: the agent signs its requests per RFC 9421 and the public keys are published for retrieval. That is why genuine traffic can be told apart from claimed traffic, if you verify it.

6. The Shopify case, and who owns your access to ChatGPT

For merchants on Shopify, OpenAI writes that product data is already integrated through Shopify Catalog and no additional work is required.

That sounds like relief, and for the moment it is. But the second thought is worth having. Being in that way means being in through a third party's channel. If Shopify changes its terms, if OpenAI changes its sources, or if you switch shop systems, your visibility changes without you doing anything.

Today this route costs nothing extra. When Agentic Storefronts launched in March 2026, Shopify wrote: "Products become discoverable in ChatGPT by default via Agentic Storefronts, with no separate integrations, no apps, and no transaction fees beyond standard processing rates." That is a statement about today, not a promise about tomorrow. Whether it stays free, and which plan it ends up tied to, is Shopify's decision, not yours.

Owning your access looks different: your own product pages that a fetcher can genuinely read, your own rules about who may read, and your own product feed, for which OpenAI grants access. That is more work, and the honest reason for it is not that OpenAI requires it, because for Shopify merchants it explicitly does not. The reason is that you then own your access yourself.

7. What the standard route leaves of your data, measured

So far this has been about rules. This section is a measurement on a real shop, so that the ownership question from section 6 does not stay abstract.

What the catalogue can take in at all. Shopify's own documentation on the Global Catalog names exactly three filterable attributes: Color, Size and Target gender. Alongside them come category, rating, price tier, availability, condition and shipping country. It states verbatim that unsupported attribute names are ignored. A hint like "suits beginners" or "works offroad" has no field in that grid.

Two further sentences from the same source belong here. First, some fields are inferred by Shopify's AI and are "not always be present or have varying accuracy". Part of what stands about you in the catalogue is therefore not maintained by you. Second, results are clustered by a Universal Product ID, with several merchants' offers on the same node.

What a well-kept shop actually holds. We measured it on our pilot customer, a coffee retailer. A single product carries six fields in Shopify's standard namespace: flavour, roast level, caffeine content, country, dietary form, product form. Exactly the class you would expect.

In that same product's own fields there are more than thirty. And that is where the buying decision sits: use cases per machine, altitude, varietal, processing, harvest, the cooperative with its number of farmers, brewing recipes with dose and time, plus two lists that are precisely what a customer asks about in conversation: who this coffee suits and who it does not.

Two real sentences from those fields:

"drinks a lot with milk and wants an espresso that does not vanish in a cappuccino"

"comes from Brazilian espresso and is used to plenty of body, cocoa and hardly any acidity"

That is the information a buying conversation runs on. It exists, maintained, in the merchant's own words. And the route by which Shopify merchants land in ChatGPT by default does not carry it.

The point is not that Shopify is doing something wrong. A catalogue spanning millions of shops has to be comparable, and comparable means narrow. The point is that this narrowness is a decision somebody else made for you. Anyone who wants to curate what is transmitted, and how much, needs a route of their own.

8. What we built from this

Our free check has measured both sides separately since this week, because OpenAI keeps them separate. The ads side asks whether OAI-AdsBot may read, whether a conversion measurement is present, and whether product data is structured. The organic side asks whether OAI-SearchBot may read, whether a plain fetch gets through at all, and whether structured product data exists.

The middle signal on the organic side is why we built it: it shows the blocking that is invisible in robots.txt.

What the check does not say: that you will appear afterwards. It measures preconditions. The selection is made by ChatGPT on the criteria from section 3, and none of us influences those.

Try it yourself: check.klariton.com, enter a domain, read the result.

For the content work: the buying decision questions a good product page has to answer are what Buying Answers draws from your material. The depth of product detail is what Owned Intelligence builds. Whether you appear in AI answers is what Prompt Monitoring measures.

9. Where things stand today, briefly

• Ads and organic product results are separate. Budget does not buy selection.

• Merchants are ranked by availability, price, quality, and maker or primary seller status.

• Your product copy is raw material, not output. ChatGPT rewrites it.

• Labels and review summaries are model-generated and not verified by OpenAI.

• Block automated access and shopping research skips you.

• Shopify merchants are already integrated through the catalog. Convenient, but not your own access.

Sources

• OpenAI Help: Shopping with ChatGPT Search (retrieved 30 Aug 2026)

• OpenAI Help: Using shopping research in ChatGPT (retrieved 30 Aug 2026)

• OpenAI Help: Cloud browser allowlisting (retrieved 30 Aug 2026)

• Shopify Dev: Global Catalog MCP (retrieved 31 Aug 2026)

• Shopify Dev: About Catalogs (retrieved 31 Aug 2026)

• Shopify News: Millions of merchants can sell in AI chats (retrieved 01.09.2026)