Ad inventory has a new address. It sits inside the conversation your customers have come to trust more than any feed, any inbox, any results page they have used before. Brands are already asking how to buy it, what it costs, who to call. 

Far fewer are asking what happens to the trust that made the surface worth buying in the first place, and that is the question that should be running the meeting.

What Actually Shipped in February

The facts themselves are not seriously in dispute. OpenAI announced the move in January and began testing ads inside ChatGPT in the United States on 9 February 2026, limited to logged-in adults on the free and Go tiers. Paid plans stayed clean. 

The rules were published before a single placement ran, and they are stricter than almost anything the open web tolerates. Ads appear below the end of a response and carry a sponsored label. Health, mental health and political queries are walled off. Advertisers cannot shape or reorder what the assistant says, because the advertising system runs on separate rails from the model that writes the answer, and personalisation, while switched on by default, can be turned off. 

The company has also set out which accounts are eligible to see ads at all, including a commitment that nobody identified as under 18 will be shown them.

Every Promise Here Is a Design Decision

Read that list again as a marketer rather than as a user. Nearly every promise on it is a design decision, and design decisions tend to get revisited when a revenue line starts asking questions. The wall between the ad and the answer exists because somebody drew it, which means somebody can move it. There is nothing sinister in saying so. It is broadly how ad-supported products have matured, from print to broadcast to search, and pretending otherwise is how media buyers end up startled by their own channel three years in. Academic commentary has already put the point more sharply, asking whether a set of voluntary safeguards can hold once advertising becomes central to the business rather than peripheral to it. The promises look real today. Treating them as permanent is a planning error rather than a moral one.

The Counter-Force Nobody Budgeted For

Users, meanwhile, are not waiting for the industry to finish arguing. Ad blockers began filtering the new placements almost as soon as they appeared, and the AI Block-o-Meter keeps a public running count of how many have been caught across ChatGPT and other assistants, alongside survey work that should give any campaign planner pause. When Harris Poll put the question to 2,066 US adults in May, 72 per cent said they were concerned that AI chatbots might use their conversation data to target them with ads. Forty-four per cent called themselves very or extremely concerned. Among people who already run a blocker, the figure climbed to 77 per cent.

The Trust Gap Underneath the Opportunity

That finding would be easier to dismiss as an interested party’s data if independent research pointed somewhere else. It does not. Pew Research Center’s survey of 5,119 US adults, fielded in February, found 71 per cent expecting that wider use of AI will make their personal information less secure, against 3 per cent who expect the opposite. Two separate pieces of research, different sponsors, different questions, landing within a point of each other.

The same survey is worth reading for reach as well as for anxiety, because the two sit oddly together. Around half of US adults now use chatbots, a quarter of them daily, and six in ten say they read the AI summaries at the top of search results. Adoption at that scale usually arrives with enthusiasm attached. Here it has arrived with roughly two-thirds of the public saying the technology is moving too quickly. You are being offered access to an audience that is large, engaged and unusually suspicious of the thing it keeps opening.

The Company That Got There First, Then Left

Then there is the part of the story the trade press has underplayed. Perplexity got to this market first and then walked out of it. Sponsored follow-up questions launched in late 2024 with real brands attached, new advertisers stopped being taken on through 2025, and by February the company had stepped away from advertising altogether on the grounds that it made users suspicious of everything. The reasoning its executives gave the Financial Times was blunt enough to be worth sitting with. A person has to believe the answer in front of them is the best one available. Put a paid unit anywhere on that page and the belief gets harder to hold, and the doubt does not politely confine itself to the ad.

Two serious players have now examined this business model up close and declined it.

Why Merit Still Beats Purchase

Which exposes the uncomfortable arithmetic underneath the whole opportunity. A sponsored card beneath an answer that never mentions your brand looks like a weak position by construction, because the reader has just been told what the assistant thinks and is now being shown what somebody paid to add. Being named inside the answer on merit tends to be the stronger outcome, and it rarely gets cheaper to buy your way around that. 

If anything, the arrival of paid units makes the unglamorous business of organic search visibility more valuable in an assistant-led market rather than less. That work is slower and harder to report on. It also cannot be switched off by a policy change at a company you do not control.

The Refusal, and the Case Against It

Anthropic took the argument further and said never. Its February essay held that even a clearly separated ad would change what the assistant is for, because advertising money creates a pull towards engagement, and engagement is not the same thing as being useful. The most helpful exchange a person has with an assistant might be a short one that ends. The essay also made the point that ad incentives, once introduced, have a habit of widening as they get folded into revenue targets. Whether that reads as principle or as very well-timed positioning probably depends on your cynicism, and it is worth remembering the same company bought a Super Bowl slot to say it.

And yet the purist case can be pushed too far. A labelled ad sitting below a straight answer is arguably more honest than a great deal of what currently passes for content marketing, including some of the sponsored coverage brands buy without a second thought. Nobody writes essays about the ethics of a paid listicle. Search results carried paid links for two decades and the world adjusted to it. Reviewers accept gifted product. 

Publications run advertorial under a small grey word that almost nobody reads. Set against that history, a labelled box beneath an answer looks close to exemplary, and a brand refusing the channel on principle may simply be handing the placement to a competitor with fewer scruples and better timing. The assistant is being held to a standard the rest of the industry has never once met, which tells you something about how much people have already invested in it emotionally, and possibly something about how quickly that investment could turn.

Google Took a Third Road

Google went a different way again, and it is the route that should occupy planners most. Ads have stayed out of the Gemini app while going deeper into search itself, where formats built with Gemini now generate advertising copy inside AI Mode answers and recommendation lists. Conversational Discovery units. Highlighted Answers that can surface a brand inside a list of suggestions. 

An explainer written by the model sitting next to the paid creative to justify its presence. The careful separation drawn elsewhere does not exist here in the same form. In these formats the ad no longer sits beside the answer so much as inside the shape of one, which is a meaningfully different proposition to buy and a meaningfully harder one to disclose.

Questions to Settle Before the Budget Moves

So the media plan needs a harder question than it usually gets asked. Before any budget moves, work out what share of your audience will ever see the unit, because only the free tiers carry ads and blockers are already trimming what remains of that. Then work out what a click from a conversational surface is genuinely worth when set against the same money in managed paid search, where the behaviour is understood and the benchmarks have years behind them. For a lot of categories in 2026, the honest answer is probably a small test and a patient read of the numbers rather than a confident reallocation.

Measurement deserves its own warning. A conversational channel sends people who behave nothing like search traffic, arriving on pages built for a different sort of visitor entirely, and the temptation will be to judge those early results against benchmarks that were never designed to hold them. Give the test its own reporting line. Give it longer than a quarter before deciding what it proved.

The Pharmacist Problem

A brief digression, because the analogy people reach for tends to be the wrong one. Nobody resents a billboard. A billboard has never once pretended to be advice, and nobody has ever turned to one and asked it a follow-up question. The closer comparison is a pharmacist taking money from a single manufacturer, or a sommelier with a stake in one vineyard. The wine is still good. The recommendation is probably still sound. The relationship has changed anyway, and the customer tends to feel that before they can articulate it.

Where Professionals Should Start This Quarter

Some practical ground, then, for teams who have to make decisions before the theory settles. Start by checking whether your category is eligible at all, because the exclusions are wider than they first appear. Health, mental health and political advertising are blocked at the platform level, and Pew found that a fifth of chatbot users go to these tools for medical advice while one in ten use them for emotional support. A substantial share of the most emotionally weighted conversations happening inside assistants is closed to advertisers by policy, which removes several sectors from the opportunity before anyone opens a spreadsheet.

Next, audit your organic position before buying anything. Ask the assistant the ten questions your best customers actually ask and record whether your brand appears unprompted. A sponsored placement supporting an answer that already names you works differently from one attached to an answer that names three competitors instead.

Then write your disclosure position down now, while it is still a choice. Formats that generate copy dynamically will make it harder to say afterwards exactly what was published in your name, and the teams that thought about this in advance will be the ones able to answer a journalist, a regulator or a client without improvising. The self-serve tools where campaigns are built and budgets set are open and straightforward enough that the operational side is not the hard part. 

Judgement is the hard part, which argues for putting the team through proper training before the first campaign rather than after the first mistake.

The Advertising You Cannot See

If you take one thing into your next planning session, make it this: treat assistant advertising as a small, well-measured experiment with its own reporting line and its own disclosure policy, and treat your organic presence inside those answers as the asset actually worth funding. The placement can be bought this quarter. The credibility cannot.

There is one more thing, though, and it is the reason none of this settles. Everything written above assumes the advertising announces itself. Labelled units, separate boxes, a wall you can point at. Filters catch those, and so do readers, and the argument stays legible to everyone involved. Advertising written into the phrasing of an answer is a different animal. A product named a shade too warmly inside a sentence that reads like judgement leaves no box to catch and no label to read. The people counting blocked ads say plainly that their totals are a floor rather than a ceiling for exactly this reason, and that they are still working out how to detect the rest. Nobody has yet explained how a brand would know which kind it had bought. Nobody has explained how a reader would know which kind they were being handed.