Ads inside AI assistants: what changes for agent builders
ChatGPT now carries ads; Claude promises it won't. What sponsored answers mean for trust, disclosure and agents that recommend or buy products.
For most of their short history, AI chat assistants had a business model you could explain in one line: pay for a plan, or live with a limited free tier. That changed in 2026. Advertising has arrived inside the most widely used assistant, and anyone building agents that recommend products, compare vendors or act on a user's behalf now works in a different environment.
What has actually launched?
OpenAI set out its approach in January 2026, saying it would test ads for free and low-cost users in the US. Testing began on February 9, 2026, for logged-in adults on the Free and Go plans. Plus, Pro, Business, Enterprise and Education accounts were left out, and so were conversations about health, mental health and politics. Ads sit below the response and are labelled as sponsored. OpenAI says they do not change ChatGPT's answers and that conversations are not shared with advertisers.
In May 2026 OpenAI opened a self-serve Ads Manager to US advertisers, with cost-per-click bidding and conversion tracking. For the buyer's side of the story, AffHouse keeps an explainer on how ChatGPT ads work, from formats and bidding to the gambling policy, including which markets have opened since and which categories are off limits.
Not everyone followed. On February 4, 2026, Anthropic said Claude will remain ad-free: no sponsored links beside conversations and no advertiser influence on responses. It also said that if it ever revisits that position, it will explain why.
Search got here first. Google began showing ads in AI Overviews to US mobile users in October 2024. The difference now is that the ad sits inside a conversation, where the user has often spelled out exactly what they need and why.
Is "separate from the answer" enough?
OpenAI's central commitment is architectural: the ad unit sits next to the answer, and advertisers cannot alter what the model says. That is a meaningful line. But it is a promise about one product's chat window, and it covers only the paid placement.
Agent builders work further downstream. Your agent may call a model, read web pages, query product feeds and pull results from other tools. Some of those inputs will contain sponsored listings, affiliate links or copy written specifically to persuade a model. Very little of it arrives as neatly labelled as an ad unit in a chat interface. If your agent quietly summarises a sponsored listing as a neutral recommendation, the trust problem belongs to you, not to the platform.
The principle here is not new. In December 2015 the US Federal Trade Commission issued a policy statement on deceptively formatted advertising. It says promotional messages that consumers cannot identify as advertising are deceptive if they make the content look independent or impartial, and that a misleadingly formatted ad is deceptive even when the product claims in it are true. An agent that blends paid placement into advice is the kind of format that principle was written for.
Where will agent builders feel it first?
Recommendation agents come first. If your agent compares software, plans trips or shortlists suppliers, users will increasingly ask whether its picks were paid for. Have an answer you can show rather than just assert: which sources were consulted, which results carried a commercial relationship and how the ranking was produced.
Shopping agents follow. In September 2025 OpenAI launched Instant Checkout in ChatGPT and open-sourced the Agentic Commerce Protocol it built with Stripe, a common format for agents and merchants to exchange product data and payments. Once the same agent can both recommend and buy, the line between advice and sales gets thin. Anthropic has said it is interested in agentic commerce as well, but only for actions the user initiates.
Then there are agents on the other side of the auction. Marketing agents will bid on conversational inventory, where the targeting signal is the conversation itself. That raises brand-safety questions no keyword list was designed for, such as which kinds of conversation your ad should never appear beside.
How do you keep disclosure intact through a pipeline?
The weak point in most agent pipelines is the hand-off between steps. A signal that is obvious in the source, such as a "Sponsored" tag, an affiliate parameter or a vendor-written page, is easily lost when content is fetched, cleaned and summarised. Carry commercial metadata through the pipeline as structured data, not as prose the model is free to drop.
Then show it where the decision happens. If a recommendation involves an affiliate relationship or a paid listing, say so next to that recommendation, not in a footer or terms page. The FTC's guidance on online disclosures has long made the same point: put the disclosure as close as possible to the claim it qualifies.
Finally, treat injected persuasion as an ad nobody paid you to run. Product pages increasingly carry text aimed at AI readers. Your agent should not follow it, and ideally it should flag it to the user.
So what is the real product?
The shift is not really about whether ads in assistants are good or bad. Users can choose an ad-supported assistant or one that promises none. What they cannot easily do is audit an agent's reasoning. That makes transparency about sources, incentives and who paid for what the feature that separates an agent people rely on from one they double-check.
Marketplaces that settle agent work through escrow and published service levels, Agent·Bazaar among them, already make the commercial terms of a task explicit. The next step for builders is to do the same for the recommendations inside the task.
Ad formats, markets and policies are still changing month to month. For the advertiser-side detail, the AffHouse knowledge base is a practical place to keep up.
