ChatGPT Will Show Product Ads as It Generates an Image

Someone asks ChatGPT to draw a birthday card for their sister. For a few seconds nothing happens, because generating a picture takes longer than writing a paragraph. From later this month, in the United States, that waiting space fills with products.

That is the whole story, and it is smaller and bigger than it sounds. Smaller, because it is one country, one surface and an invited group of advertisers. Bigger, because the slot sits in the one moment of a chat where the user has nothing to read and nowhere else to look.

What exactly did OpenAI announce?

A new visual display format inside ChatGPT, appearing alongside the results of image generation. TechCrunch reports the ads will be clearly labelled, will not influence the model's answers, and begin showing later in October 2026 in the United States only, with an initial test group of advertisers supplying the products.

The Decoder describes the shape of it: the unit renders as a product carousel or a row of product images with a Learn more button, placed below the generated image while the user waits. Both outlets put the audience on the free and lower cost subscription tiers, against a platform reaching roughly 1.2 billion people a week, of whom about 5 percent pay for a subscription.

So the people who will see this are, almost by definition, the people who have not paid to avoid it. That is a large pool of ordinary shoppers rather than a narrow professional audience.

Why does the image generation moment matter more than another banner?

Because of what the user is doing. A shopper who types a question is mid thought and scanning for an answer. A user waiting on an image has finished asking and has idle attention, which is the condition every advertiser pays for everywhere else.

There is a second reason, less obvious. Image prompts leak intent in a way text questions often do not. A request to draw a nursery mural, a logo for a dog grooming round, a poster for a village fete: each one names a life event, a trade or an occasion. The commercial read on that is strong without anyone needing to look at the conversation itself.

We wrote in August about the smaller text unit that reached European users, an advertiser name, a short title, a line of copy and an image. This is a visual upgrade on that, and it arrives in the United States first, which is the same ordering as last time.

Can a small shop actually buy one of these slots?

Almost certainly not yet, and the reason is the feed rather than the budget. The ads are generated from a product catalogue, and the catalogue programme has a size floor.

According to GeekSeller's walkthrough of the Ads Manager beta, a feed must carry at least 1,000 products, uploaded over SFTP, in Parquet, JSONL.gz, CSV.gz or TSV.gz, UTF-8 encoded. The required fields include item_id, title, description, url, image_url, price, availability, brand, seller_name, seller_url, return_policy, target_countries and store_country, plus links to your privacy policy, terms of service and return policy.

Read that floor again. A thousand products is not a big catalogue for a marketplace reseller, and it is an impossible one for a ceramicist with forty glazes or a deli with a hundred and twenty lines. The gate here is not the money. It is the shape of the business.

What the feed programme asks for, line by line

The requirements are worth laying out together, because no single source lists them in one place and the pattern tells you who the programme was designed for.

Parquet as the preferred format is the detail that gives the game away. No shop owner has ever asked for Parquet. It is what you choose when the party on the other side expects a data warehouse export, and it tells you the programme was built around retailers with a feed team rather than around sellers with a product page.

What should a seller with eighty products do instead?

Work on the listing you are not paying for. PPC Land's walkthrough notes that feed setup includes a toggle for participation in organic product listings, which is the unpaid surface where ChatGPT shows products it has not been paid to show. Placement there cannot be bought, so it is decided by whether your product data is accurate, complete and corroborated.

That is a slower job than buying a slot and it is the one available to you. Three things carry most of the weight, and all three live on your own site rather than in anyone's ad platform.

1. Stock and price truth. An assistant that recommends a sold out item in the wrong currency has been embarrassed by you once and will not be keen to repeat it. Accuracy is the price of entry, not an optimisation.
2. Attributes a machine can read. The required field list above is a useful specification even if you never upload a feed anywhere, because it is a plain statement of what the reader on the other side expects to find. Our guide to getting a product feed into shape for AI shopping surfaces goes through those fields in order.
3. A page that survives being read aloud. If the only way to learn your return window is a PDF in the footer, the model will guess, and a guess about your returns policy is a liability with your name on it.

If your product pages are the weak link rather than your range, that is a build problem rather than a marketing one, and it is the thing an AI store builder that writes your product pages and schema is for. The listing cannot be accurate if the page it is drawn from is not.

How much of this is a real channel and how much is a press cycle?

The money says real. The Decoder reports an annualised run rate of 1 billion dollars from ads alone, a stated goal of 100 billion dollars by 2030, and ads already live since February across more than 40 countries. Those are not the numbers of a pilot somebody will forget about.

The measurement build out says the same. The attribution list now runs to AppsFlyer, Triple Whale, Adjust, Northbeam, Branch, Singular, Kochava, Airbridge, Tenjin and DV Rockerbox, with Fospha, Measured and INCRMNTAL for full funnel work and Haus, Measured and WorkMagic for geo experiments. Brand suitability pilots with DoubleVerify and Integral Ad Science evaluate the environment without reading user conversations, and advertisers can set negative keywords to keep out of contexts they do not want.

Nobody wires thirteen attribution vendors into a format they intend to drop. Equally, nobody needs thirteen attribution vendors unless the buyers are large enough to demand their own measurement stack. The plumbing confirms both that the channel is serious and that it is not being built for you.

There is a quieter signal in the tier choice. Ads land on the free and low cost plans, and only about 5 percent of weekly users pay for anything. A seller reading that as a downgrade in audience quality has it backwards. The free tier is where price sensitive shoppers live, and price sensitive shoppers are the ones who compare before they buy. For a small seller competing on something other than price, that is the audience most likely to be told about a cheaper alternative while it waits for its picture.

What does clearly labelled actually protect?

It protects the answer, not your position in it. Both reports state the same two things: the unit carries an advertising label, and it does not influence what ChatGPT says. So a competitor cannot pay to be recommended to a shopper who asked for advice.

What a competitor can do is appear beside the recommendation. That is a weaker form of influence than buying the answer and it is not nothing, because the eye does not keep a careful ledger of which box was paid for. Anyone who has watched a shopper read a search results page knows how little the label changes.

The controls on that sit entirely on the advertiser side. Negative keywords let a buyer keep its products away from contexts it does not want, and the brand suitability pilots give large advertisers a read on the environment without anyone reading user conversations. None of those levers point outward. There is no seller side control that keeps a competitor out of a carousel next to a conversation about your category, and there is no reason to expect one.

Does this change what a shopper sees when they ask about your product?

Not directly, and that distinction is worth holding onto. Both reports state the ads are labelled and do not influence the model's responses. The recommendation and the sponsored card are earned separately, which is the same separation we described when ads first reached European users.

What changes is the surrounding furniture. A user who has just watched a carousel of competing products while their image rendered arrives at your page in a different frame of mind than one who came from a plain recommendation. You cannot buy into that carousel, and you also cannot stop a competitor appearing there.

The honest summary is that this is a competitor advantage you cannot match on equal terms, in a place where your absence is invisible. That is uncomfortable and it is not fatal. The unpaid listing sits in the same conversation, and the shopper who clicks through to a page that answers their question will not be counting who paid for what.

The part of the announcement nobody is talking about

PPC Land's piece carries a detail that reframes the whole feed story. The product feed infrastructure did not start life as an ads system. It was built for instant checkout, launched in September 2025 through a payments partnership, and by March 2026 Walmart had disclosed that instant checkout converted at roughly a third of the rate of clicking out to the retailer's own site. The approach was dropped, and the feeds migrated to the Ads Manager in June 2026.

So the catalogue pipe you are being invited to fill was originally plumbed for a checkout that did not work. That explains the shape of the requirements, which read like a commerce integration rather than an advertising one, and it explains the thousand product floor: a checkout integration is worth the engineering for a large retailer and absurd for a small one.

It also carries a lesson that outlasts this announcement. When a large platform retires a commerce feature, the pipework usually survives and gets pointed at the next revenue idea. We have written about what changes for sellers when buying moves inside an assistant, and the direction of travel in that piece has not reversed. The checkout attempt failed and the data requirements it created are now the price of admission to something else.

What to actually do this month

Nothing expensive, and three things that are free.

Count your range honestly against the thousand product floor so you stop reading advertiser coverage that does not apply to you. If you clear it, the sample of 100 products is your real first test, and failing it means your catalogue data is the blocker rather than your budget.

Then check the unglamorous fields: stock status, currency, brand, return policy, a seller name that matches your legal entity. Those are the ones the required field list asks for and the ones a hand maintained catalogue gets wrong, because they are boring and nothing on your own site breaks when they drift.

Last, decide whether you want to be in unpaid listings at all. Some sellers do not, and that is a legitimate position rather than an oversight. What is not legitimate is being in them with wrong prices, because the shopper who finds out blames you and not the assistant.

What is worth watching next?

The sequence from last time gives a usable expectation. The text unit ran in the United States from February and reached European users in late August, roughly six months apart, and the self serve buying tool opened to United States advertisers well before any European buying date was confirmed. If the visual format follows the same path, European users see it before European advertisers can buy it.

The other thing to watch is the floor itself. A thousand products is a business decision rather than a technical limit, and the limit on a feed programme tends to come down as the inventory of unsold ad slots goes up. Nothing has been announced on that and nobody should plan a quarter around it. It is simply the number to check again in six months, because it is the single figure standing between a small catalogue and this channel.

Meanwhile the useful work is unglamorous and portable. Accurate stock, honest prices, readable attributes and a return policy on the page rather than in a PDF pay off in unpaid listings today, in a paid slot if the floor ever drops, and in ordinary search either way. None of it is wasted if this particular carousel turns out to be a footnote.

The carousel in the loading screen is new. The work it rewards is the work that was already sitting there.