ChatGPT now sells ads. The question is whether anyone can measure them

This week, Branch announced support for ChatGPT Ads attribution across mobile apps, websites, and desktop. Advertisers can attribute installs, reengagements, and post-install events to ChatGPT campaigns, and pass selected conversion events back to OpenAI for optimization.

Read that announcement the way a media buyer should read it. Branch did not announce a new ad format or a new targeting capability. It announced measurement. Attribution across mobile apps, websites, and desktop, covering installs, reengagements, and post-install events, plus the reverse flow: selected conversion events passed back to OpenAI so the ad system can learn and optimize. That is the full measurement loop in one press release: credit where credit is due, and signal flowing back to the machine that bids the budget.

For an app marketer, the three event types named are the whole vocabulary of mobile performance. Installs are acquisition. Reengagements are win-back and retention. Post-install events are the revenue and engagement milestones that happen after the download: the purchase, the subscription, the level completed. Saying all three are attributable to ChatGPT campaigns means an app advertiser can now evaluate this channel the way they evaluate every other channel, on a real funnel, not on vibes and click counts.

Why this announcement matters

That matters because of what OpenAI's own help center now says. The named measurement partners for ChatGPT Ads are AppsFlyer, Adjust, Branch, and Singular, with click-based attribution and attributed events appearing in the Conversions metric after 24 to 48 hours. Four named partners is a real measurement story, not a pixel and a prayer. AI discovery is becoming a media channel, not a science project.

Count what those four names represent. AppsFlyer, Adjust, Branch, and Singular are not four random startups. They are the four mobile measurement providers that the performance marketing industry actually uses. When a new ad platform names one of them, it is a courtesy. When it names all four, it is a strategy. OpenAI is signaling that ChatGPT Ads will plug into the measurement stack advertisers already run, rather than asking advertisers to adopt OpenAI's own black-box reporting as the sole source of truth. That choice is the difference between a channel you can test and a channel you have to take on faith.

The mechanics OpenAI describes are deliberately conventional: click-based attribution, with attributed events landing in the Conversions metric after 24 to 48 hours. Conventional is a compliment here. It means the crediting logic works the way every other channel's click-based attribution works, and it means the reporting delay is stated up front instead of discovered the hard way. A new channel that grades its own homework generously is a known industry pattern. A new channel that ships with third-party measurement and documented lag is a channel that has read the room.

And the framing line stands: AI discovery is becoming a media channel, not a science project. Science projects get innovation budgets and loose accountability. Media channels get tested against CAC targets, incrementality reads, and the merciless question every other channel answers: what did it actually drive. The Branch announcement is the moment ChatGPT Ads enters that regime.

A new channel, same old rule

But a channel without measurement is a gamble with extra steps. The advertisers who win here will be the ones who wire up attribution before they scale spend, not after. That means getting your MMP integrated now, deciding which conversion events you will pass back for optimization, and setting expectations internally before the first invoice arrives.

The sequencing point is the whole argument. Most advertisers integrate measurement after the channel is already spending, which means the first weeks of data are untrustworthy, the first optimization decisions are guesses, and the first internal report comes with asterisks. The advertisers who wire up their MMP before the first dollar clears get clean reads from day one. Their learning curve is shorter because their data is better. In a new channel, where every early data point is already noisy, the advertisers with disciplined measurement have a compounding advantage over the ones flying blind.

Deciding which conversion events to pass back deserves more thought than it usually gets. The instinct is to pass everything, on the theory that more signal is better. The better discipline is to pass back the events that represent real business value: the purchase, the subscription start, the qualified lead. Passing back shallow events, like app opens or page views, trains the optimization algorithm to buy you more app opens and page views. The algorithm will get you exactly what you tell it to get. Choose the events that match your actual P and L, not the ones that make the dashboard look busy.

Your setup checklist

1. Confirm your MMP supports ChatGPT Ads attribution. If you run AppsFlyer, Adjust, Branch, or Singular, you are inside the named partner set. Get the integration configured before you brief creative, not after.
2. Map the full event set. Installs, reengagements, and post-install events should all be attributable to ChatGPT campaigns across mobile apps, websites, and desktop. Verify each surface you care about, because attribution that works on the app and breaks on the web is half a measurement story.
3. Choose your feedback events deliberately. Pick the conversion events you will pass back to OpenAI for optimization, and make them the events that define business value. Document the choice so it can be revisited with data, not with vibes.
4. Build the reporting delay into your plans. Attributed events appear in the Conversions metric after 24 to 48 hours. Set that expectation with stakeholders now, before anyone asks why the numbers moved two days after launch.
5. Establish your reconciliation rule. Decide up front how you will compare platform-reported conversions against your MMP numbers, at what discrepancy level you investigate, and who owns the call. Write it down before the first invoice arrives.

Two things to watch as you test

First, the 24 to 48 hour reporting lag affects optimization loops, so early reads will be noisy and patience is part of the strategy. Second, compare what the platform reports against your own MMP numbers before you trust either one completely. New channels always grade their own homework generously at first.

On the lag, be specific about what it breaks. Optimization loops run on feedback, and feedback that arrives 24 to 48 hours late means the algorithm is always bidding on a picture of the world that is one to two days old. Early in a test, when volume is low and every conversion moves the numbers, that lag makes day-to-day reads almost meaningless. The practical answer is pacing: give the test enough days of budget for several full lag cycles to complete before you judge performance, and resist the urge to kill or scale a campaign on same-day data. Patience is not a soft virtue here. It is a measurement requirement.

On the reconciliation point, the rule is simple and non-negotiable. Your MMP is your source of truth for what happened. The platform is the source of truth for what it charged you. Where those two disagree, and on a new channel they will disagree, you trust the MMP and you investigate the gap. This is not anti-platform paranoia. It is the same discipline that made every mature channel trustworthy: independent measurement, compared against the platform's numbers, as a standing habit.

The caveats

Click-based attribution is the stated mechanism, and it has known limits. It credits the click, which means it systematically undercredits channels that influence without being clicked, the same decoupling problem that AI answers created on the organic side. If ChatGPT exposure drives branded search or direct traffic that converts elsewhere, click-based attribution will not see it. Plan for an incrementality read once the channel has enough spend to be worth measuring properly.

And the honest framing on "four named partners": it is a real measurement story, but it is still day one. Integrations have edge cases. Reporting has bugs. The first advertisers through will find them. That is not a reason to wait. It is a reason to test with real money, real measurement, and realistic expectations.

Would you spend real budget on a channel you cannot measure?

Data over opinions.