One counterintuitive detail

One detail first, and it may strike you as a little strange.

When you click an ad in ChatGPT, it does not open the brand’s website.

It starts a conversation — with an AI agent sponsored by the brand.

The feature is called Sponsored Agents. OpenAI announced it on its own blog on September 16 and is testing it in the US with a set of advertisers.

The official description:

Sponsored Agents give users the option to go deeper. After seeing a relevant ad, users can choose to start a clearly labeled conversation in ChatGPT with an agent sponsored by the brand. They can say what they care about, ask follow-up questions, and when they are ready for the next step, click through to the brand’s site.

Watch the wording here. It carries a lot:

  • “Clearly labeled” — OpenAI knows the sensitive part is whether users can tell an ad from their assistant talking
  • “the conversation is kept separate from your original conversation, and separate from ChatGPT’s independent answers” — an ad conversation must not pollute your existing context
  • “when ready for the next step, click through to the site” — the link is still there, but it became an option after the conversation, not the destination of the ad

This is a substantive change in ad format.

The old ad logic was send you away: click, redirect, leave. Every platform is anxious about click-through rate, because a click means the user left.

Sponsored Agents invert it: keep you here and talk.

This is not a side experiment. It is version one of AI-native advertising

Digiday quoted OpenAI CFO Sarah Friar, and I think that quote is the key to the whole thing.

At Goldman Sachs’ Communacopia tech conference on September 8, she said the ads currently sitting under ChatGPT answers are just a “basic starting point.”

Then this:

This is all before we’ve really launched a format that feels truly endemic to AI.

(This is all before we’ve really launched a format that feels truly endemic to AI.)

Translation: what you are looking at is not even the real version yet.

What does AI-native mean?

Traditional internet advertising is designed for reading web pages — banners, feeds, auction-ranked results, all built so you bump into them while scanning.

AI-native ads are designed for an assistant in conversation with a person. In a conversation, the user is asking questions. So the ad should not be a banner. It should be something you can ask questions of.

The ad goes from display item to salesperson.

One brand is already testing it: Wayfair. Their paid media director said something refreshingly plain:

Wayfair is taking part in OpenAI’s Sponsored Agents pilot, exploring how conversational commerce can help customers make smarter, more confident decisions. We are testing at limited scale, with guardrails around product accuracy, transparency and service handoff.

The word guardrails matters. Brands know that having an AI speak for them is a completely different order of risk from buying a banner.

One number, and one line that got overturned

There is something almost theatrical in the details here.

The Wall Street Journal reported in early September that ChatGPT ads have hit a $1 billion annualized revenue run rate.

The same piece notes that Sam Altman once called ads a last resort.

“Ads are a last resort for our business model.” “I hate ads.” He said both. And now OpenAI is testing Sponsored Agents, letting advertisers buy directly from HubSpot and Shopify dashboards, and has brought in Amazon Ads for a conversational ad pilot.

36Kr’s headline nailed it: "$1 billion is just the starting point."

From “I hate ads” to “$1 billion annualized and still scaling” is roughly the distance of a compute bill.

Meanwhile, another company took a completely different road

In the same week OpenAI was laying down an ad network, Anthropic’s news read differently:

  • Profitable two quarters running (multiple outlets, the 14th)
  • Annualized revenue of $65 billion in July, projected to top $100 billion this year (sources, the 18th)
  • IPO in motion, possibly on Nasdaq as early as October

One more item from the same window is more telling: 21st Century Business Herald ran “Anthropic races to Nasdaq, OpenAI is not coming.”

The two most similar companies publicly split on business model this month:

One trades attention for money — ChatGPT has 800 million users, and turning attention into ad revenue is the road Google and Meta have walked for twenty years.

One trades capability for money — Claude runs on subscriptions and enterprise API, sold to people willing to pay for a stronger model.

What is interesting is that both work:

  • OpenAI’s user base is too large for ads not to be the obvious choice
  • Anthropic turned “we don’t do ads” into a selling point — for enterprise buyers, a model that is not bent by ad incentives is valuable on its own

“No ads” went from a technical choice to a market position.

So what does this mean for people building AI products

I am writing an AI chat client myself, so my angle on this is a bit different.

First, conversational ads will redefine what counts as good product design.

In the past, client design meant figuring out how to show an ad without getting in the user’s way.

Sponsored Agents make the ad a participant in the session. That creates new problems the product has to handle:

  • How do you make it unmistakable that I am now talking to a brand, not my assistant (OpenAI uses labeling plus conversation separation)
  • How do you hand the user from my assistant to the brand agent, and then hand them back, without it feeling clumsy
  • What is the right way to isolate context

None of these have answers in current product design practice.

Second, not selling ads becomes a durable differentiator.

Anthropic’s road shows one thing: when your competitor starts selling ads, your refusal becomes pricing power.

The same logic applies to anyone building a vertical AI product — put ads in your product and users will reprice how much they trust what it tells them.

Third, and most practical: advertisers will start demanding AI-readable assets.

When an ad goes from banner to agent, brands need product knowledge built for an AI — not campaign assets, but a knowledge base that can be interrogated, answer details, and say “I am not sure about that” when it isn’t.

That may be the next real gap in demand.

One hard engineering problem: keeping the ad in bounds

Sponsored Agents sound like a product design question. In implementation it is all engineering.

Problem one: isolating context.

OpenAI specifically stresses that the sponsored agent’s conversation is kept separate from your original one. That sentence hides a hard architectural constraint:

The brand agent must not be able to read what you told ChatGPT earlier.

Think about it. A user mentioned three hours ago that they are paying off a mortgage and money is tight, then an ad for a sofa shows up. If the brand agent can read that, it is not advertising anymore, it is harvesting.

So there has to be a session-level isolation boundary, and it has to be verifiable and auditable.

Problem two: making the label actually work.

“Clearly labeled” sounds simple, but product history has almost no successes here — user attention to ad labels is basically zero.

So what works is not a label but changing the whole interaction shape: a different avatar, a different bubble color, plain text saying this is the brand’s AI, not ChatGPT. The goal is that users cannot confuse the two visually or interactively.

Problem three: handoff without accidents.

The user finishes with the brand agent and moves to the next step: view the product, place the order. If the agent got a parameter wrong — who is liable?

That Wayfair line about guardrails around product accuracy, transparency and service handoff is exactly this. Brands have realized: let an AI speak for you and every sentence it says has legal consequences.

None of these three has a standard answer yet. Whoever works it out first defines the engineering spec for the next generation of AI ads.

But let’s stay sober for a minute

First, the results are nowhere near proven.

The very title of that 36Kr piece — A Complete Guide to Why ChatGPT Ads Underperform and How to Optimize Them — tells you the story: ads underperforming is already an industry consensus that needs a dedicated guide.

Advertisers care about conversion and ROI. Turning “chatted with the brand AI for three turns and left” into money is, for a CFO, still an unsolved sum.

Second, trust is the most dangerous thing here.

ChatGPT’s product strength rests on the default assumption that it is on your side.

Once ads enter the conversation surface, that assumption erodes. OpenAI clearly knows — hence the repeated emphasis on clearly labeled and kept separate. You can see them walking carefully along that line.

Third, even OpenAI is putting limits on the ads.

A detail: on September 10 word came that OpenAI bans ads for competing AI products in ChatGPT (Adobe and others were affected).

It sells the inventory and screens who may buy it. What that tells you: the rules of this ad slot are not set by the market, they are set by OpenAI.

Last

Lined up together, these are worth a look:

  • OpenAI: $1 billion annualized in ads, Sponsored Agents live, from I hate ads to building an ad empire
  • Anthropic: two profitable quarters, $100 billion annualized in sight, IPO in October, no ads as a selling point
  • Amazon, Shopify, HubSpot: racing to plug into the ChatGPT ad pipeline
  • Wayfair: testing carefully, guardrails first

Together they say one thing: the free lunch of the AI assistant is being priced.

Only this time the ad does not look like an ad — it is not something stuck in front of your eyes, it is the character sitting across from you, ready to answer your question.

I want to ask you:

If ChatGPT popped up a “chat with the brand AI” button, would you click it?

Or put another way: what would an acceptable ad in an AI assistant look like to you? Comments are open.

Sources: OpenAI’s blog (9/16), PYMNTS, Digiday, The Wall Street Journal, Reuters, 36Kr, Cailian Press and other public reporting.