OpenAI must label a sponsored agent as sponsored. The label protects the answer by making the ad weak.

OpenAI began testing Sponsored Agents on 16 September. A buyer clicks an ad inside ChatGPT, opens a separate labeled conversation with the advertiser's own AI agent, asks follow-up questions, and follows a link out when ready. The same day, HubSpot became the first CRM to sell ChatGPT Ads from inside the CRM, with attribution back to closed deals.

The mechanics matter less than the word. OpenAI has to stamp that conversation "sponsored", and every buyer who sees it decides in about a second what the stamp means. Three separate accusations live inside that one word: this is worth less, this is less true, and I am not getting a straight answer. They have three different answers, the data on all three now exists, and two of them are worse news for the advertiser than for the buyer.

Less valuable: 44% of buyers discount a post the second they see who paid

Yes, and it keeps getting confirmed. IAB Australia's Creator Connect report, out today, found 44% of shoppers are less likely to act once they know a post is paid, rising to 54% among 50 to 70 year olds. The same report puts creator-driven brand discovery at 84% to 85% among under 40s. The reach is enormous and the persuasion drains out the moment the label goes on.

We found the same effect in AI creative two weeks ago: the work loses its clicks at the moment the audience is told, whatever made it. The label is a discount, applied by the reader, instantly, every time. Nobody badges inventory to make it worth more.

Less true: 14.35% of the ads have nothing to do with your question

Wrong question, and the numbers show why. SE Ranking ran 50,006 commercial prompts across 20 US niches in August. Ads appeared on 25.94% of them, close to Google's AI Mode at 29.45%. Of those, 14.35% were semantically unrelated to the prompt they sat beside, above 50% in Relationships and News and Politics. Concentration is heavier still: one advertiser, BestMoney, held 13.56% of all ads observed.

That is not a biased recommendation. That is adjacency, sold by auction, and in one case in seven it is not even adjacent. Asking whether it is true is a category error. It rented a wall next to the road you were already driving down.

Objective: 3.63% of advertisers are cited in the answer above their own ad

This is the question that matters, and the honest answer annoys the marketers more than the skeptics. In the same study, only 3.63% of advertisers were cited as a source in the answer directly above their own ad. Just 0.09% of advertised URLs appeared in citations at all. OpenAI says ads do not influence ChatGPT's answers, and the independent measurement of its ad product is consistent with that: the recommendation and the advertisement are strangers to each other.

So the buyer does still get an objective recommendation, near enough. The money is fenced off from the answer, and the fence appears to hold.

Read that as a valuation rather than a compliance note. OpenAI calculated, correctly, that the unbought answer is the entire asset, and that letting advertisers into it would burn the thing generating $83 million a month. The paid conversation gets a badge and a separate room, and the label is where they wrote the number down.

So the company that sells the answers now sells ads beside the answers, and is obliged to put a sticker on the ad explaining that it is not the answer.

What this actually costs a GTM team

Run the three answers together. The sponsored conversation is discounted on arrival by nearly half your audience, carries no authority by design, and does not buy you into the recommendation that decides the shortlist, because that lane is fenced and the fence is the product.

HubSpot's side is the honest part. "Customers are asking on ChatGPT. They live in LLMs," VP of product Angela De Franco told B&T. True, and worth spending against. Notice what the integration measures, though: it ties a closed deal back to a campaign through a CRM record, and we reported on 8 September that 37.6% of revenue teams call updating those records their single biggest admin bottleneck.

"Sponsored" has stopped meaning biased and started meaning weak, and the marketing industry has not caught up to how bad that is for it. A biased recommendation was at least a recommendation. This is a paid slot beside the decision, sold by auction, that the decision has been engineered to ignore. OpenAI built that fence on purpose and deserves credit for it, because the unbought answer is the only asset it has. Nobody should expect it to sell the thing it is protecting.

Which leaves the uncomfortable half. Every GEO and AEO budget approved this quarter is money spent trying to climb that fence, and the vendor who built it publishes numbers saying it holds. The route into the recommendation is the route it has always been: be the company people independently cite and vouch for when nobody is being paid to say it. Some of that ends up written down where a crawler can reach it. Most of it happens in rooms, over dinner, on a panel, in the conversations that were never for sale in the first place. Twelve months from now the sponsored lane will be bigger, cheaper per impression, and still fenced off from the answer. Anyone still calling that a discovery strategy will be explaining a flat pipeline to a board that can read the same study we just did.