“Sell on ChatGPT” sounds like a setting. It is not. There is no dashboard toggle, no marketplace signup, and nobody at your store can switch it on this afternoon. What exists is a specification for a product catalog, a way of publishing that catalog, and an onboarding process that OpenAI controls and grants.
That is worth saying first because most writing on this topic is careful never to say it. What follows is the merchant’s version: what you need, in what order, what you control, and what you do not. If you want the format itself explained rather than the process, what the OpenAI ACP product feed is covers that side.
What “selling on ChatGPT” actually means
A shopper asks an assistant for something. The assistant answers with specific products, and depending on the surface, the purchase can complete without the shopper opening a browser tab. For that to include your products, three things have to be true, and only the first is yours.
- Your catalog exists in a format the assistant can read, published at a URL that can be fetched on a schedule. This is entirely within your control and it is most of the work.
- OpenAI has accepted your catalog into their commerce programme. This is not within your control and there is no self-serve path today.
- Your individual products satisfy the requirements well enough to be recommended rather than merely present. This is back in your control, and it is where most of the value is.
The second point is the one that makes this a longer game than a vendor demo suggests. It is also the reason to do the first and third points now: when the door opens, the merchants who get through are the ones whose data was already ready, not the ones who start then.
Step one: produce the catalog in the right shape
The format is ACP, the Agentic Commerce Protocol feed. Practically, it is a whole-catalog file, similar in spirit to a Google Shopping feed but carrying more of what an assistant needs to answer a question: not only the identity and price of a product but the ratings, the review content, and what a shopper can actually do with it in the conversation.
Two practical notes. The URL takes country parameters, so you declare where you sell from and where you sell to rather than assuming one market. And it is a registrable whole-feed format, which means it is a thing you host and they fetch, not a per-product markup you sprinkle on your pages.
If your store already publishes a Google Shopping feed, you are closer than you think: the same product data underlies both, and the work is a transformation rather than a rebuild. If it does not, what a product feed contains is the shorter first step.
Step two: the onboarding is theirs to grant
Once the feed exists, the process is: you provide the URL to an OpenAI commerce contact, they validate the fields, and they set the fetch schedule. That is the whole of it, and every part of it happens on their side.
Be sceptical of any tool or agency that describes this as instant, automatic, or included. Nobody can grant you access to somebody else’s programme. What a tool can legitimately do is produce a correct feed at a stable URL and keep it correct, which is the part that determines whether validation passes when it happens.
The honest planning assumption is that this step takes as long as it takes, and the useful thing to control in the meantime is step three.
Step three: the fields that decide whether a product is usable
This is where the work pays regardless of what happens with the onboarding, because these are the same fields that already decide your Shopping performance. An assistant is stricter than an ad auction in one specific way: it has no photograph and no listing grid to compensate with, so a gap in the data is not a lower ranking, it is disqualification from the candidate set.
- Identifiers. GTIN, MPN and brand. Without them nothing can establish that your product and the one somebody else sells are the same object, which removes you from the comparison the assistant is making. For genuinely own-production items,
identifier_exists=nowith brand and MPN is the correct answer rather than a blank field. - Price, matching the landing page exactly. Including tax treatment. A feed price that includes VAT against a page price that does not is a rejection, and it reads as a mystery until somebody checks both.
- Images that resolve. A URL behind hotlink protection resolves for a shopper’s browser and not for a fetcher, which is a common and completely invisible cause of failure.
- Availability that is current. An assistant recommending something you cannot ship is a worse outcome than not being recommended, because it lands on a person who was ready to buy.
- The attributes a question contains. Size, colour, material, age group, compatibility. A request like “a warm waterproof jacket for a toddler” has four clauses, and a product row has to satisfy them from text alone.
That last point is the difference between a compliant feed and a useful one. A catalog can pass every structural check and still say almost nothing, which is a separate problem with a separate fix: complete is not the same as rich.
ChatGPT is not the only assistant, and the others want something different
Worth knowing before you build anything ChatGPT-specific, because the requirements diverge and one of them costs you nothing.
Perplexity reads the standard Google Shopping feed. No new format at all: the same file you already produce, given to them after you apply to their merchant programme, which is a manual review rather than an automatic acceptance. Results there are organic and ranked substantially on feed completeness and reviews, which means the enrichment work is the entire lever.
The pattern generalises. Each assistant has its own door and its own paperwork, and behind every door is the same product data. Which is the strategic point: the format work is small and channel-specific, the data work is large and channel-independent, and only one of the two is worth organising your quarter around.
What you cannot measure yet, and what you can
You cannot currently isolate traffic or revenue attributable to an AI assistant recommending you. A visit that began in a conversation mostly arrives looking like direct traffic, or like nothing at all when the assistant answered and the shopper never clicked. Any tool offering you a clean figure for it is estimating, and a confident number here is a claim about the vendor rather than about your store.
What you can measure is whether your data could support a recommendation: coverage per attribute, identifier completeness, price consistency, how many products are disapproved for a data reason on the channels that do report. None of those prove you were recommended. All of them determine whether you could be. How assistants pick products goes through the mechanism in more detail.
If you want the starting numbers rather than an opinion, a free scan grades every product and reports the gaps field by field. It reads the feed URL you already publish, it takes about two minutes, and nothing is written back.
Frequently asked questions
Can I sign up to sell on ChatGPT today?
Not through a self-serve form. The route is producing a catalog in the ACP format and providing that URL through OpenAI’s commerce onboarding, which they validate and schedule. Anyone describing it as a signup is describing something else.
Do I need a separate feed, or can I reuse my Google one?
ACP is its own format, so it is a separate export. It is a transformation of the same underlying product data rather than a second catalog to maintain, which matters: if you keep two catalogs by hand they will disagree within a month.
Is this worth doing before the volume exists?
The format work, arguably not yet on its own. The data work, yes, and not because of AI: identifiers, price consistency and real attributes are the same fields that decide Shopping performance today. That is the honest version of the case. Anyone telling you to reorganise your quarter around assistant traffic is ahead of the evidence.
What happens if a product is rejected?
It is not recommended, and the reasons cluster the same way they do elsewhere: identifiers, price mismatches against the landing page, unreachable images. The reporting is less mature than Merchant Center’s, which is a good argument for fixing those causes on the channel that does report them clearly. Every Merchant Center disapproval reason and its fix is the same list of causes.
Does any of this help my normal search rankings?
Indirectly. A feed is not an organic ranking input, but getting real attributes onto every product usually improves the product pages too, and those are indexed. The gain is real and it is a side effect rather than the mechanism.
