For its first three years, ChatGPT had no ads, and that was part of the pitch. You asked a question, you got an answer, and nothing in the answer had paid to be there. In January 2026 OpenAI confirmed it was changing that, and by early February sponsored placements were showing up below answers for free users in the United States. By May, any business in the US could open a self-serve account and buy them.
So ChatGPT is an ad network now. The interesting part is not that it sells ads. Every large consumer surface eventually does. The interesting part is what the ad unit is. It is not a keyword you bid on. For a retailer, it is your product catalog, the same feed you already send to Google Shopping, handed to a model that decides which item to surface inside a conversation it is reading in real time. That is a different machine from the search auction, and the advertisers who treat it like one will misjudge it.
This piece works through how product-feed ads in an answer engine actually function: how OpenAI got here, how ranking and feeds differ from a search auction, the trust problem sitting underneath the whole thing, and what an early advertiser should reasonably expect.
Origin: why an answer engine needed an ad business
OpenAI resisted advertising for a long time, and said so out loud. Sam Altman had called the combination of ads and AI uniquely unsettling, on the reasoning that a paid answer is worse than no answer because the user cannot tell the difference. The reversal was not a change of philosophy. It was arithmetic.
By early 2026 ChatGPT had passed 900 million weekly active users. Subscriptions, even at that scale, do not cover the cost of running frontier models for that many people, and the great majority of those users pay nothing. Advertising is the standard way to monetize a free tier, and a free tier this large is the most valuable unsold inventory on the consumer internet. Internal projections reported in the trade press put OpenAI's 2026 ad target around 2.5 billion dollars, scaling far higher later in the decade. The wider market is moving the same direction: eMarketer forecasts US AI-search ad spending rising from about 2 billion dollars in 2026 toward nearly 26 billion by 2029.
OpenAI shipped a deliberately narrow first version. Ads appear only for logged-in adults in the US on the free and the new Go tier, the eight-dollar-a-month plan; Plus, Business, and Enterprise stay ad-free. Ads sit below the answer, labeled as sponsored. They are not eligible near sensitive subjects like health, mental health, or politics, and they do not run for users predicted to be under 18. The test began with managed campaigns from large retailers, names reported to include Target, Best Buy, Lowe's, Williams-Sonoma, and Albertsons. Then in May the self-serve Ads Manager opened the door to everyone else.
There was a separate thread running alongside the ad business: commerce. OpenAI had launched Instant Checkout, which let people buy inside the chat, then pulled back from it after finding that onboarding merchants and showing accurate product data at scale was harder than expected. It moved toward a model where ChatGPT handles discovery and the merchant keeps the checkout. That retreat matters for advertisers, because it tells you where OpenAI's confidence actually sits. Discovery and recommendation inside the conversation is the part it is betting on. Product-feed ads are that bet with a price attached.
Present: how the auction and the feed actually work
Start with the mechanics, because they are where the search-engine instinct goes wrong.
A classic search ad is bought against a query. You bid on a keyword or phrase, and when someone types something close enough, an auction runs. ChatGPT has no query in that sense. It has a conversation, often several turns long, full of context a keyword could never carry: a budget mentioned earlier, a room being decorated, a recipient the gift is for, a constraint like small kitchen or sensitive skin. OpenAI's system reads that thread and matches it to ads whose subject fits what is being discussed. There is nothing to type a keyword against. The targeting signal is the meaning of the chat.
The selection runs through what OpenAI describes as a relevance-weighted, second-price auction. Two parts there, and both count. Second-price is familiar from search: the winner pays just above the next bid, not the full bid. Relevance-weighted is the part to sit with. The bid is one input, not the input. Whether the ad and its landing page genuinely fit the conversation is weighted alongside the money, so a lower bid on a sharply relevant product can beat a higher bid on a loose one. When several advertisers are eligible, OpenAI says it shows the one most relevant to the chat. Money buys you entry to the auction. Fit decides whether you win it.
Now the feed. For retailers, OpenAI added product-feed ads: you connect a product catalog, set filters for which items are eligible, and the system generates sponsored placements from the catalog itself. The deliberate convenience is that this is the same structured feed retailers already maintain for Google Shopping, so there is no separate file to build. The platform handles up to 1 million SKUs per advertiser, while onboarding starts with a sample feed of roughly 100 items that OpenAI reviews for quality before the full catalog is accepted.
This is where feed ads diverge hardest from a keyword auction, and the divergence is the whole point of the angle. In a search auction, your keyword does the matching and your feed is mostly a display layer. In ChatGPT, the feed does the matching. The model decomposes a conversation into what the user actually needs, then matches that against the structured attributes of your products: material, dimensions, color family, finish, certifications, age group, whether an item reads as a gift. A feed where the width sits inside the title text and the use case is buried in a description paragraph forces the model to dig it out, and a thin or messy feed simply matches less often. The same structured-data quality that decides whether a product gets recommended organically also decides whether it is even eligible to be a paid placement. The catalog is the campaign.
One more piece worth getting right: the money model moved fast. The first version priced on CPM, cost per thousand impressions, with a default around 60 dollars. Impression pricing on a brand-new surface gave advertisers no clean read on value, and reported CPMs slid well below that opening number within weeks. OpenAI added cost-per-click bidding and made it the recommended way to buy, with a suggested starting maximum bid of 3 to 5 dollars per click and the minimum spend cut sharply from the managed-pilot level. CPC ties spend to an action rather than to exposure, which is the right instinct for a surface where nobody yet knows what an impression is worth. OpenAI also shipped a conversions API and pixel-based measurement so advertisers can see what happens after the click. Read the speed of that CPM-to-CPC shift as the honest signal it is: the pricing of this channel is still being discovered in public.
The trust problem nobody can design away
Every claim above rests on one promise, and the promise is the most fragile thing in the product. OpenAI states plainly that ads do not influence ChatGPT's answers: the ad systems run separately from the model that writes the response, advertisers cannot shape or reorder what ChatGPT says, and conversations are kept private from advertisers, who never receive your chats, identity, or precise location. Ads are labeled sponsored and set apart from the answer. Personalized ads can draw on past chats and memory, but only to choose the ad, not to alter the reply.
Those are good rules. The problem is structural and no rule fully closes it. A search ad sits next to ten blue links, and a user can see it is an ad and look past it to the organic results. An answer engine produces one answer. When the sponsored item appears right below a recommendation written in the same calm, authoritative voice, the line between the model's judgment and the paid placement gets thin, even with a label doing its job. The conflict of interest is not hypothetical either: OpenAI now earns more when you click the ad, and it also writes the answer above the ad. It has promised, credibly and for now, to keep those two facts apart. Whether that wall holds as the revenue target climbs into the billions is the open question of the entire category, and it is not one an advertiser controls.
It is worth knowing that this is a genuine fork in the road, not a settled one. Perplexity, another AI answer engine, ran ads earlier and then stepped back from the model, arguing that ads erode trust in the answer. Anthropic ran Super Bowl ads needling the whole idea of putting ads in an AI assistant. ChatGPT advertising is real and scaling, but the premise that every AI surface inevitably becomes an ad surface is contested by serious players. An advertiser building on ChatGPT should treat the surface as promising and unsettled at once, and should not assume the rules of engagement in 2026 are the rules of 2028.
Future and impact: what an early advertiser should do
The honest framing for a marketer is that ChatGPT ads in 2026 are an early test worth running deliberately, not a mature channel to pour budget into. A few things follow from how the machine is built.
Treat the product feed as the actual campaign asset. This is the central, practical consequence of feed-based matching. Your titles, images, and structured attributes are not back-office merchandising data anymore; they are the creative and the targeting at once. Move material, dimensions, finish, certifications, age group, and use case into clean structured fields rather than leaving them in prose. A feed built only to clear Google Shopping's minimum will under-match in a system that reads attributes to understand a conversation. Audit the feed first; nothing else you do for this channel pays off more.
Check your organic presence before you pay. OpenAI keeps organic product recommendations and paid placements on separate tracks, and the same feed quality drives both. If ChatGPT already surfaces your products well in unpaid recommendations, your feed is in good shape and paid placement amplifies a working signal. If it ignores you organically, a paid budget is propping up a feed that needs fixing first. Run the unpaid queries yourself before committing spend.
Buy on CPC and measure to an action. The pricing here is still being found in public, so do not anchor on early CPM numbers or assume today's bid guidance holds. Use cost-per-click, connect the conversions API or pixel, and judge the channel on real outcomes, purchases, leads, sign-ups, rather than on impressions or clicks alone. Treat the first campaigns as a measurement exercise: you are learning what a ChatGPT-sourced customer is worth, on a surface with no years of benchmarks behind it.
Expect the inventory to be different in kind, not just in placement. ChatGPT users arrive mid-decision, often several conversational turns into comparing options, which is a higher-intent moment than a cold keyword and a thinner-volume one. Do not expect search-scale reach. Expect a smaller pool of conversations where the buyer has already half-decided and is asking which one. Match expectations and budgets to that shape.
There is a longer arc here that connects to the rest of this category. ChatGPT pricing ads on conversational intent, Google rebuilding paid search around AI and conversational shopping, and the broader move toward AI agents that shop on a person's behalf all point the same way: structured, machine-readable product data is becoming a ranking input across discovery, not a technical afterthought. The skill that makes a feed win in ChatGPT, clean attributes a model can reason over, is the same skill that will matter wherever an AI system mediates between a shopper and a catalog. An advertiser who builds that capability now for ChatGPT is building it for the surface after ChatGPT too.
The plain summary: ChatGPT has become an ad network, but it did not copy the search-ad playbook. The unit is the feed, the match is conversational, the auction weights relevance against the bid, and the whole thing rests on a trust promise that is sound today and will be tested for years. Run it as the early experiment it is, fix the feed before you fund it, measure to outcomes, and watch the wall between the answer and the ad.
Council summary
The post argues that ChatGPT advertising is real and scaling, but it is not a search auction with a new skin: the ad unit is the product feed, matching is conversational rather than keyword-based, and the auction weights relevance against the bid. Council verified the load-bearing facts against trade and primary sources: the February rollout to free and Go tier users, the May self-serve Ads Manager, the 900 million weekly users, the 2.5 billion dollar 2026 ad target, the relevance-weighted second-price auction, the 3 to 5 dollar CPC guidance, the 1 million SKU limit, and the Instant Checkout retreat. One figure was corrected: the eMarketer US AI-search forecast for 2026 is about 2 billion dollars, not the 1 billion the draft first cited. The reader takeaway is concrete: audit the product feed before funding it, buy on CPC, measure to outcomes, and watch the wall between the answer and the ad as the revenue target climbs.
Comments