OpenAI expands ChatGPT ad formats, measurement, and brand safety tools

Full analysis

OpenAI is patching the three things advertisers have complained about since ChatGPT ads launched in February 2025: formats, measurement, and brand safety. Image ads during image-generation sessions start testing in the US this month. Conversion tracking is expanding past the last click. And DoubleVerify and Integral Ad Science (the two firms that certify ads don't run next to garbage content) are piloting safety tools, but with a catch: neither gets to read actual conversations. Underneath all of it sits a supply problem that none of these upgrades touch.

What's being decided, and how hard to undo. For OpenAI this is easy to undo: these are pilots and tests. For ad-tech operators the live question is whether to treat conversational AI as a real channel now or keep it on the "test-and-learn" line. That decision is cheap to reverse too. What's actually on the table is who gets to write the rulebook for how ads get measured and verified inside a chatbot. That one is hard to undo once it sets. The deadline is set by Google, which cannot let OpenAI define the vocabulary for conversational-ad measurement unchallenged.

The Market Analyst. DoubleVerify and IAS are the clear winners here, and it has little to do with pilot fees. Being named first when an AI platform builds its ad stack is worth far more in future contract leverage than whatever the test pays. Expect both to put out press releases regardless of how thin the actual product is. For Google, the threat is not ChatGPT's 2.5 billion daily prompts against its own roughly 15 billion searches. The threat is legitimacy. Once holdcos accept conversational query as a buyable, verifiable channel, Google has to accelerate ad slots and incrementality tooling on AI Overviews to defend the same intent surface. For a general reader: the measurement vendors win by showing up, and Google now has to race on its own turf.

The Skeptic. Read the quote again. One ad placement. Prompt volume well under Google's. And a privacy rule that blocks DoubleVerify and IAS from seeing real conversations, which means they are checking placement structure, not content. Checking placement structure is not brand safety: it is a logo on a slide. For this to become a real channel, OpenAI has to multiply placements without wrecking the chat experience, solve the conversation-privacy tension so the verifiers get real signal, and close a volume gap while Google builds the same thing on the same intent. Three hard problems, zero solved. OpenAI is telling advertisers what keeps them in the conversation. The product is not there. For a general reader: the announcement promises safety tools that are switched off at the part that matters.

The Operator. A media buyer running a brand campaign has a 90-day call: throw pilot budget at this or wait. The DoubleVerify and IAS names remove the "we can't verify safety" objection from the media plan, which is real. But brand-safety teams will clock within a cycle that the vendors can't see the content, so they're reading structural signals, not context. The image format gives creative teams something to build. The reach math kills it as a core line. One placement against a fraction of Google's volume means you hit frequency concentration fast: the same users seeing the same ad over and over. Treat this as a test line item until the supply math changes. For a general reader: there's nowhere near enough ad space yet to run a real campaign.

The CFO. OpenAI gets no near-term revenue pop from any of this, and that's the part worth pricing. One ad unit against thin volume caps what you can charge per thousand impressions, because scarcity of inventory, not scarcity of demand, is the binding constraint. You can't absorb programmatic demand at scale through a single placement without a product rebuild. So the spend this attracts is small pilot money from brands who want to say they tested it. For DoubleVerify and IAS, pilot revenue is a rounding error against the positioning. The real money is a future question: can OpenAI turn multi-turn prompt intent into a targeting product without exposing conversations? That bill isn't due this year.

Where the council splits. The Strategist in the room would argue OpenAI is building credentialing on purpose, laying verification groundwork so holdcos take it seriously before the network scales, with prompt-level intent as the real prize. The Skeptic says that's a narrative outrunning the product by two years. That gap is the whole decision. Second split: the Market Analyst thinks Google has to respond fast and defensively; the Skeptic thinks the absolute numbers don't justify Google doing anything it wasn't already doing on AI Overviews.

What it hinges on. Two beliefs. First, whether OpenAI can expand inventory past one placement without degrading the chat experience that makes ChatGPT valuable. Second, whether it can give DoubleVerify and IAS access to real conversational signal without breaking the privacy promise. Both are unsolved in this announcement. Until they move, this is credentialing, not a channel. The council leans skeptical on the near term and open on the long game. What to de-risk before committing real budget: run a small image-format test to learn the creative and the labeling, but don't model reach you can't hit, and don't buy the brand-safety story until the vendors can see content.

Prediction: DoubleVerify's full-year 2026 revenue, reported in its Q4 2026 earnings release in February 2027, will show no identifiable contribution from OpenAI or conversational-AI advertising large enough for management to quantify as a revenue line.

Confidence: Medium. The pilots are real but the inventory ceiling is small.

Why: The story itself gives the reason the money stays small: one ad placement against 2.5 billion daily prompts, with the verification vendors blocked from reading conversations. That caps both how much inventory exists and how deep a product DoubleVerify can actually sell into it, so pilot fees stay in rounding-error territory against a company doing hundreds of millions in annual revenue. The positioning value is large, but positioning doesn't show up on a revenue line. The opposite outcome would require OpenAI to multiply placements and loosen the privacy constraint fast enough to generate material verified volume within a single quarter, and nothing in this announcement puts that on a 2026 clock.

Revisit by 2027-02-28: We're right if DoubleVerify's Q4 2026 results and call contain no quantified conversational-AI or OpenAI revenue contribution. We're wrong if DoubleVerify reports a specific, material dollar figure from OpenAI or conversational-AI verification.

This is the skeptical read on a partnership both vendors will market hard: the logo arrives long before the dollars do.