ChatGPT Product Feed Ads: What Decides If You Show Up

Google Shopping ads have run on product feeds since the format existed. Keyword ads bid on words a shopper types; Shopping ads bid on the feed itself, because the feed is what gets matched to the query. In 2026, OpenAI brought that same mechanic into ChatGPT with what it calls product feed ads.

Since testing began in February 2026, OpenAI has rolled sponsored product listings into ChatGPT conversations across the US, then into Canada, Australia, New Zealand, the UK, Mexico, Brazil, Japan and South Korea. An Ads Manager Beta opened in June to a much wider set of merchants, and by mid-2026 the mechanism looked structurally familiar to anyone who has set up a Shopping campaign: upload a product feed, and ChatGPT generates the ad from what’s in it.

That changes what “advertising on ChatGPT” actually means for a merchant. It isn’t a slogan you write or a headline you A/B test. It’s a feed you submit, and the feed’s quality decides whether the ad shows at all, and how well it performs once it does.

Worth being precise about what this is and isn’t. OpenAI has said plainly that ChatGPT’s organic shopping results, the products the model surfaces when someone just asks a question, are not ads and not influenced by any partnership. We covered that side of the system in ChatGPT Shopping Visibility. ChatGPT ads are a separate, paid surface sitting next to those organic results, with their own product feed spec and their own eligibility gate. The two systems draw on similar data. They are not the same door.

What ChatGPT product feed ads actually run on

OpenAI’s commerce documentation lays out a product feed spec that reads like a close cousin of Google’s: JSONL as the primary format, CSV and TSV accepted in Google-compatible shape, UTF-8 encoded, gzip-compressed at scale. Nine fields are required on every item: a stable item_id, title, description, product url, brand, seller_name, image_url, availability, and price. Past that, a long list of optional fields does most of the work that separates a generic listing from one an AI system can actually reason about: gtin and mpn for identity, product_category, material, color, size, gender and age_group for attributes, plus review counts and star ratings for social proof.

Eligibility is explicit here, not implied. A product carries an is_ads_eligible flag before it can appear as a sponsored listing at all, on top of the search-eligibility flag every product needs just to be considered. The Ads Manager Beta adds its own gate on top of that: a minimum catalog of 1,000 products, feeds delivered by SFTP, a recommended daily refresh, and a review of the submitted products by OpenAI before any campaign goes live. It’s the same idea Google Merchant Center has enforced for years, applied to a newer surface. Your ad is only as eligible as your feed.

Where feed quality decides an ad’s fate

A few things in this system reward or punish data quality, and none of them are the bid.

Eligibility comes first. A product missing required fields, carrying a price string that isn’t in the “amount CURRENCY” format the spec expects, or pointing at an image URL that isn’t a publicly reachable HTTPS link doesn’t get reviewed down. It doesn’t run. The spec is specific about this in a way that reads almost like a warning: omit an optional field you don’t have data for rather than fill it with a placeholder, because a fabricated value is worse than an empty one.

Relevance follows close behind. ChatGPT ads are matched to a conversation, not typed into a search box. A recipe question can surface a meal-kit or grocery-delivery ad, which means the system needs enough structured signal in a product’s attributes to know when it belongs in an answer. A feed that carries only a title and a price gives the matching system almost nothing beyond the most literal keyword overlap. A feed with category, material, size and a real description gives it something closer to what a shopper actually asked for.

And then there’s trust, which OpenAI has been more explicit about than either of the above. “Protecting the trust people place in ChatGPT” is the company’s stated guiding principle for how ads get shown at all. A platform built on conversational trust has less patience for a sponsored product that turns out to be inaccurate, out of stock, or misrepresented than a search engine has for a misleading blue link. Feed accuracy isn’t just a performance lever here. It’s closer to a license to participate.

The same data, increasingly the same job

The commentary on ChatGPT’s product feed ads keeps landing on the same observation: the setup “mirrors how shopping campaigns work on platforms like Google,” and structured product data now “matters far beyond syndication.” Titles, descriptions, attributes, taxonomy and imagery all feed both the organic AI shopping answers a product might appear in and the paid ones a merchant chooses to run.

That’s good news if you’ve already been doing feed work for Google. It’s bad news if you’ve been treating your Merchant Center feed as a one-time setup task. A feed built to clear Google’s disapproval checks is a strong starting point for ChatGPT’s eligibility flags, since both systems are asking a version of the same question: does this record describe a real, purchasable product accurately enough to show someone deciding whether to buy it? A strong starting point isn’t the same as an identical requirement, though. OpenAI’s spec has its own field names, its own eligibility flags, and its own review step, so a feed engineered purely for Google still needs mapping and checking before it’s ready for a ChatGPT campaign. The field-level requirements for showing up in ChatGPT organically are a useful reference here too, since the same structured attributes that help an assistant describe your product are what an ads reviewer is checking for.

Keyword ads bid on the words a shopper types. Product feed ads bid on how well your data describes what you’re selling, and a thin feed loses that bid before any spend is placed.

A feed checklist before you turn on ChatGPT ads

Before submitting a feed to OpenAI’s Ads Manager, a few checks catch most of the rejections and underperformance merchants run into early.

Confirm every item carries the nine required fields, with price formatted exactly as amount plus currency code and no placeholder text standing in for data you don’t have. Check that is_ads_eligible and is_eligible_search are set on purpose, not left at whatever default you never verified. Fill in the optional attribute fields you can support with real data: brand, category, material, color, size, GTIN or MPN where one exists. These are exactly the fields that separate a literal keyword match from a system that understands what the product is. Treat identifiers the way any responsible feed does: source a real GTIN from the manufacturer or GS1 when one exists, and leave the field empty rather than inventing one. A fabricated identifier is a policy problem waiting to surface, not a shortcut around a data gap.

Catalog size matters too. The beta’s 1,000-product minimum means a small or highly seasonal catalog may not clear the bar yet, which makes feed completeness on the products you do carry even more important. There’s less room for weak records to hide in volume.

Budget doesn’t buy your way around a thin feed, either. Reported minimums for the beta sit around a $25 daily budget with bids starting near $3, numbers small enough that a merchant could reasonably treat the whole thing as a low-risk pilot. But a low bid on a poorly attributed product just means you lose the match slower. The spend is not what determines whether ChatGPT can describe your product to someone asking a real question. Structured product data is what does that, in ads and outside them.

Once a campaign is live, tracking where the resulting traffic actually comes from is its own problem. Referral data from AI platforms undercounts by nature, so treat any early click or conversion numbers as a floor, not the full picture.

Conclusion

ChatGPT ads didn’t invent a new discipline. They imported the one Google Shopping merchants already know and gave it a second surface, with its own field names and its own review gate. A feed that’s complete, accurately attributed and honestly identified clears both bars at once. A feed that’s been getting by on the bare minimum will find both harder in 2026 than it used to. Checking a feed against nine required fields and a handful of eligibility flags is a lighter lift than it sounds, and it’s the same kind of work UCP Radar was built to keep current automatically, product by product, as feeds keep getting asked to do more.