The click is not the conversion. The agent is.

The IAB is writing a framework for AI advertising measurement, due November 12. Digiday filed it as an industry standards story. It is also an operator story about a control surface that just stopped working.

Here is the problem Caroline Giegerich, the IAB’s VP of AI, put on the record: if an agent reads a product page, weighs competitors, and buys on a user’s behalf, who gets credit? The publisher whose content shaped the answer? The platform that ran the agent? Nobody, because the evidence trail does not exist yet? Traditional signals — UTMs, referrer headers, the landing-page click — do not reliably survive that journey. The working group is arguing about what should count as proof when the click never happened.

That is not a future tense problem. Ads are already being served to agents. Time has started serving them. Perplexity has blocked some of them as deceptive. The marketplace is being upended while the receipts are still written for a human who clicked a link.

What shipped this summer, and what it does not fix

OpenAI plugged AppsFlyer and Adjust into ChatGPT ads so brands can attribute installs, in-app purchases, and subscriptions. That is real progress. App marketers buying ChatGPT inventory since the February 2026 pilot finally have a third-party MMP path instead of faith. Click-based attribution links, Conversions API forwarding, deduplication against the pixel — the plumbing looks familiar on purpose.

It is also the Facebook pattern, restated. Measurement vendors integrate into a platform that still owns the black box. Bishop, co-founder of OpenAds, said it cleanly in the Digiday piece: Facebook coded the integration and measurement vendors rubber-stamped it. The black box stayed. ChatGPT ads now have a trust layer. They do not give you an independent picture of what the agent did before the click.

Meanwhile, on the organic side of the same channel, the referrer is dying. Attrifast’s aggregate across dozens of sites puts the median share of ChatGPT visits filed as Direct in GA4 around 71%. Native apps and in-app browsers strip the Referer header. GA4 launched a native AI Assistant channel on May 13, 2026, which is a floor, not a ceiling: it only sees sessions that still arrive with a usable referrer. AI Overviews still look like Organic Search. Most of the high-intent traffic from the assistant sits next to typed-in URLs in Direct, converting better than the Direct next to it, and your channel report cannot tell the difference.

So in one season you get a standards body trying to invent credit for agent-mediated buys, a platform shipping MMP integrations for the click it can see, and an analytics default that misfiles most of the organic assistant traffic that still leaves a human on your site. Different layers. Same failure mode: the control surface assumes a click on a landing page.

The operator check

I build attribution systems for a living. When the lake and the ad platform disagree on a Facebook account, last-click is theater. When a scheduled report dies at fifteen minutes, the dashboard still looks calm until someone asks for the receipt. The mechanism is older than ChatGPT ads: you cannot manage what you cannot reconcile at a grain the seller did not choose for you.

Agents buy without a click. That does not make measurement optional. It makes UTM-on-the-landing-page the wrong primary control surface, the same way permissions were the wrong primary control surface for agents that narrate instead of send. The IAB framework will try to name what evidence should exist. Until that evidence exists, the operator move is not to trust the channel’s self-report. It is to keep an independent spine — your own event stream, your own identity keys, your own reconciliation against a grain you did not aggregate — and treat every platform’s agent scorecard as a claim to audit.

The work is the same whether the buyer is a person or an agent: prove the touch, reconcile the total, refuse a black box that grades its own homework.