OpenAI did not announce a new ad format. It did not announce a holding company partnership or a measurement standard. It announced that ChatGPT advertising is now running at a billion dollars a year, six months after the first test, and then released a model three days later that lets people stop looking at the screen where those ads sit. Amazon took nine years to admit in public that it had an advertising business. OpenAI took six months to build one and three days to complicate it. What follows is what the number actually measures, what the Amazon versus Perplexity ruling already settled, and what a creative organisation should change before the autumn plans are signed.

Key takeaways

ChatGPT advertising reached a one billion dollar annualised run rate in under 200 days, against three to four years for Amazon, around three for Facebook and roughly one for TikTok.
The advertising exists only on the tiers where a person still reads an answer on a screen. GPT-6 Astra, released 3 September, completes the same tasks without that step.
Amazon sued Perplexity in November 2025 for letting an agent bypass its advertising, and won in March 2026. The conflict is a decided precedent, not a forecast.

A billion dollars in under 200 days, and what a run rate does not tell you

On 31 August, OpenAI said its advertising business had reached a one billion dollar annualised run rate. The company also opened self-service buying through Ads Manager in India, Europe, the Middle East and North Africa, after reaching 31 European markets earlier in the month. In April the same business was worth around 100 million dollars.

A run rate is a pace. It takes the revenue of the most recent period and multiplies it out across a year. Roughly 83 million dollars in a month becomes a billion on paper. It is not a year of booked revenue, and the distinction matters when the comparison set is made of companies reporting audited annual figures.

The pace is still remarkable. It is also short of OpenAI's own stated target of 2.5 billion dollars in advertising revenue for 2026, which the current trajectory does not reach.

Nothing in advertising has ever grown this fast

Amazon launched its advertising arm in October 2012. It took three to four years to reach a billion dollars, and nine years before the company would break the line out in its published results, disclosing 31.2 billion dollars for 2021 after a decade inside an unhelpful category called Other. Facebook launched self-serve advertising in November 2007 and passed a billion in 2010. TikTok, the fastest of the previous generation, went from around 340 million dollars in 2019 to 1.14 billion in 2020.

Six months against nine years is not a marginal improvement. It is a different category of event.

Two things explain it. OpenAI started with an audience it did not have to acquire, close to a billion weekly users already inside the product. And it skipped the phase every previous platform went through, where the ad business is built quietly while attribution is solved.

The advertising lives where a human is still reading

Ads appear underneath answers, on the Free and Go tiers. That is where the majority of those weekly users sit. Paying tiers do not carry them.

The composition of the spend says more than the total. Sensor Tower measures advertising density per user up 163 percent between April and August. Shopping fell from 37 percent of US ad spend on the platform to 21 percent. Financial services went from a single advertiser in the top 100 to four in the top ten.

Read that again in plain terms. Each user is being shown considerably more advertising than four months ago, and the mix is tilting toward a category with very large budgets and very patient measurement expectations. Several agencies said publicly during the same week that performance does not yet justify the spend.

That is a channel filling up. It is not yet a channel proving itself.

Amazon already took this question to court

In November 2025, Amazon sued Perplexity. The complaint was that Comet, Perplexity's agent, was walking into the Amazon store on a shopper's behalf, retrieving what it needed and completing the purchase without passing through the sponsored placements that fund the storefront. Amazon won in March 2026.

The world's largest retail advertising business established, in front of a court, that an agent bypassing advertising is a problem worth litigating.

Now apply that to 3 September. GPT-6 Astra is built to operate software through screens the way a person does. It opens the interface, fills in the fields, compares the options and finishes the task. No human has to watch the screen while it happens.

OpenAI is selling attention on the free tier and building the thing that removes the need for it on the paid one. Both businesses are growing. They are not growing in the same direction.

What the agent reads instead of your creative

An agent does not watch a fifteen second pre-roll. It does not register a hero image, a colour, a tagline or a piece of art direction. It reads price, specification, availability and whatever structured attributes it can retrieve, then returns a shortlist.

OpenAI's own shopping research tool reached 52 percent product accuracy on queries with multiple constraints, against 37 percent for standard ChatGPT search. Accuracy here means how well the result matched stated requirements on price, material and specification. The gap between being in the shortlist and being absent from it comes down to whether the product data is complete and machine readable.

So a creative organisation now serves two readers from one brief. A distracted human on a free tier, who responds to an image and a line. And a system with no eyes at all, which will never see either.

Those two readers require different artefacts, produced by the same team, from the same asset pipeline, on the same deadline.

Where this argument does not hold

Agent-mediated sessions remain a minority of commercial journeys. Most people still type a question and read the answer themselves. A channel serving that behaviour has years of usefulness ahead of it, and treating ChatGPT advertising as already obsolete would be wrong.

Nothing prevents OpenAI from placing advertising inside the agent layer itself. Sponsored positions in an agent's shortlist are technically straightforward, and the Amazon ruling suggests platforms will defend their right to monetise their own surface rather than surrender it. If that happens, the question stops being whether ads survive and becomes who controls the shortlist.

There is also a simpler possibility. A billion dollar run rate reached this quickly may reflect the novelty of new inventory more than durable performance. OpenAI has published no advertiser retention figures and no repeat spend data. If the channel disappoints, the agent question will not be what decided it. Measurement will.

Three moves before the autumn plans are signed

First, for media directors. Buy this as a test line with a defined end date and a defined success threshold, not as an always on channel. The inventory is real, the audience is real, and the attribution is not settled. Set the kill criteria before the first invoice, because the density figures suggest the cost per impression will not stay where it is.

Second, for heads of creative ops. Treat product data as a creative asset with an owner. Specification, availability, dimensions, materials and compatibility now determine whether a product appears in a machine generated shortlist. In most organisations that information sits in a spreadsheet nobody has updated since the launch, maintained by nobody in particular, while the hero image goes through four rounds of approval.

Third, for brand directors. Decide who signs off on what the agent says about the brand. When a system summarises a product without ever loading the page, the summary becomes the brand message, and no one in the current approval chain has authority over it.

This is the part that lands on creative operations rather than on media planning. Two readers, one brief, one asset pipeline, and a production chain that was never asked to output structured attributes alongside a key visual. MTM was built for that pipeline, and teams at L'Oréal Paris, Lancôme and Helena Rubinstein use it to keep creative and the data attached to it moving through the same process rather than two disconnected ones.

See how MTM keeps creative assets and their metadata in one pipeline →

FAQ

What is an annualised run rate, and why is it not the same as revenue? A run rate takes the revenue of the most recent period and projects it across twelve months. OpenAI's billion dollar figure means the business earned roughly 83 million dollars in its latest month. It does not mean a billion dollars has been received. The distinction matters because the companies used for comparison, Amazon and TikTok among them, reported audited annual revenue. Comparing a pace with a booked year flatters the newer number, which is why both are worth stating.

Should we put budget into ChatGPT Ads this quarter? Yes, as a test with limits. The audience is close to a billion weekly users and self-service buying is now open across Europe, India and the Middle East, so the channel is accessible. Set a fixed budget, a fixed end date and a measurable threshold before starting. Several agencies have said publicly that returns do not yet justify large commitments, and advertising density per user rose 163 percent between April and August, which usually precedes a change in pricing.

Do AI agents see advertising at all? Not in the form most advertising takes. An agent retrieves structured information, price, specification, availability, and returns a result. It does not load a banner, watch a pre-roll or register art direction. This is what Amazon argued in its case against Perplexity, and Amazon won in March 2026. Some platforms are now building sponsored placements into agent responses, which would be a different product from display advertising rather than a continuation of it.

What did the Amazon versus Perplexity ruling actually decide? Amazon sued Perplexity in November 2025 over its Comet agent accessing the Amazon store on a user's behalf and completing purchases without passing through sponsored placements. Amazon won in March 2026. The practical effect is that a large platform successfully defended its right to control how automated systems interact with its commercial surface. It establishes that the conflict between agents and advertising is a live legal question, not a theoretical one.

We are a B2B brand with no e-commerce. Does any of this apply to us? It applies to the data, not the checkout. Whenever a system answers a question about your category without loading your website, whatever it can retrieve about you becomes the description a prospect receives. For a B2B organisation that means specification sheets, pricing structures, integration lists and documentation are now brand-facing content. The practical starting point is the same as for retail: find out who owns that information internally, and whether anyone has updated it this year.