ChatGPT Ads conversion tracking crossed a line in September. In a single month, three vendors that sell to agencies stopped treating the newest ad surface as an experiment and started treating it as production inventory: Supermetrics added a Conversions metric to its ChatGPT Ads connector, Windsor added write actions that let an AI assistant enable campaigns and set their budgets, and TracerGram wired OpenAI Ads into the conversion events it syncs back to the platforms. Every one of those moves is shipped, dated and checkable.
In the same month, the buy side told AdExchanger the surface still feels “very rudimentary.” Both things are true at once. And the gap between them is the whole story for a regulated advertiser: this surface learned to count conversions before it learned to produce a record.
Three vendors moved the newest surface into production
Start with what actually shipped, because the speed is the point. In its September product update, Supermetrics wrote that the ChatGPT Ads connector now supports three new segment-based report types — Products, Country and Device — and adds Conversions as a metric on the Ad report. Then it wrote the sentence that matters most: Previously, only ad-level aggregation without conversion data was available. That is a vendor telling you, in its own release notes, how young the surface’s measurement is.
Windsor went further. Its changelog entry for 18 September reads: Windsor.ai MCP write actions expanded: ChatGPT Ads (OpenAI Ads): enable and pause campaigns and ad groups, and set campaign budgets. Windsor has carried OpenAI Ads as a data connector since 5 June; the write actions landed in September. Its write-actions documentation now reads 137 write actions across 19 connectors, and growing — up from 131 when this pipeline last checked, and 123 before that. And it kept opening the surface after that: the 26 September entry added Meta Ads and LinkedIn Ads write actions, and the 2 October entry added an upload app that pushes a local file into a connector action directly from the assistant.
TracerGram took the third route: not reporting and not writing, but feeding the funnel. Its homepage now promises to sync qualified conversion events to Meta, Snapchat, TikTok, OpenAI Ads, and Whop — adding the newest ad surface to a conversion-sync list that already covered four others.
Three different products, three different layers, one pattern: on this surface, the capability to act arrived quickly and in volume. What arrived with it is a count, a write and an event. What did not arrive is the thing that says who decided.
What the surface can now tell you — and what it cannot
A conversions metric is a count. It standardises how many conversions the platform attributes to a campaign, and it is genuinely useful — until September, by Supermetrics’ own account, an advertiser working in ChatGPT Ads could see ad-level aggregation with no conversion data attached to it at all. Adding the metric is progress. It is not a record.
The buy side is blunt about where the measurement still sits. In its September report on the surface’s $1 billion run rate, AdExchanger quoted Rajeev Nair of Lifesight saying the lack of competitor data and very limited conversion-tracking capabilities and audience options makes the platform feel very rudimentary. Scott Ensign of Butler/Till called it particularly challenging for agencies and large enterprises. And OpenAI declined to comment on its road map for introducing new measurement solutions.
Read that alongside the shipped features and you get the true shape of the surface: it is a place where an agent can now pause a campaign and set a budget, on a channel whose own conversion tracking the buy side still describes as limited. That is not a reason to avoid it. It is a reason to be precise about which half of the problem a new feature solves.
There is a quieter problem underneath the count, too. A conversion is attributed to the platform that reports it, by that platform’s own rules. When three vendors each sync the same funnel outcome back to a different ad network, the same conversion can be claimed three times. More measurement does not settle that argument. A record of which change produced which outcome does — because it is written once, by the party accountable for it.
A metric describes an outcome. A record describes a decision.
The distinction matters because the surface that accepts a change is the surface that stores the evidence — and the newest surface has the thinnest evidence layer of any. When an agent pauses a campaign on a channel you adopted last quarter, that change is written into a log the platform chose, in the fields the platform chose, readable for as long as the platform chooses to keep it.
| Capability | Supermetrics | Windsor | TracerGram | Ott |
|---|---|---|---|---|
| ChatGPT Ads reporting | ✓ Conversions metric, Sep 2026 | ✓ Data connector since Jun 2026 | ✓ Conversion-event sync | Not connected |
| Writes to ChatGPT Ads | — | ✓ Enable/pause, set budgets | — (sync only) | — |
| Names an AI agent as the actor | ✗ | ✗ | ✗ | ✓ |
| Scopes the change to a client engagement | ✗ | ✗ | ✗ | ✓ |
| Record stored independently of the ad platform | ✗ | ✗ | ✗ | ✓ |
None of the three vendors above is doing anything careless. They are shipping what the surface allows, on the surface’s terms. The point is that the platform’s terms are count the outcome, not explain the decision — and a regulated agency is asked for the second one. That is what an operations activity log is for: the actor, the value before and after, the timestamp and the linked object, written once and stored outside the tool that made the change.
The answer a platform gives is a state, not a decision
When a platform reports what happened, it reports the state it ended in. The campaign is paused. The conversion count is 412. The budget is £2,000. Those are states, and they are reliable and cheap to produce. A decision is a different kind of object: who asked, who approved it, against what limit, for which client, and when. Platforms are very good at states and structurally uninterested in decisions, because a decision implies an accountable person — and the platform is not that person, does not want to be, and will not accept the liability of being mistaken for them.
Ott does not connect to ChatGPT Ads. No official API or MCP exists for the surface, and Ott is Meta-ecosystem-first by design. The argument here is not that Ott instruments the newest channel. It is that the record a regulated agency is asked for — the actor, the value before and after, the timestamp, the linked object — has to live off the ad platform precisely because the platform’s job is to accept the change, not to explain it. The discipline is surface-agnostic; it matters most where the surface ships the least.
The number you quote has a shelf life
There is a small, useful lesson buried in Windsor’s write matrix. In six weeks the published figure has moved three times: 123, then 131, then 137 as of 2 October. Every one of those numbers was accurate on the day it was published. None of them was accurate a fortnight later.
That is not a criticism of a vendor shipping fast. It is the reason a count is a bad thing to buy on. A count changes without a date attached to the change. A dated record — this change, by this person, at this time, on this account — cannot move without leaving a trace, which is the entire property you are paying for.
Why this is a finance-agency problem before it is a general one
A general advertiser that loses a conversion metric loses some optimisation efficiency. A regulated agency that cannot reconstruct a decision loses its answer when a client or a supervisor asks what happened. Since August 2026, the EU AI Act has treated an agency operating an AI agent on a client’s ad account as a deployer, and ESMA’s new supervisory priority puts how supervised entities use AI squarely in scope from 2027.
The newest ad surfaces are exactly where agents are pointed first, because that is where the tooling is newest and the guardrails thinnest. Which is why a channel’s measurement maturity and its record maturity are two different questions — and why an agency should ask both before it lets an assistant write.
What to ask before an agent writes to a new surface
- Who approved the change, and where is that written down?
- Which client engagement does this change belong to?
- Under which rules was it permitted?
- What was the value before and after, and at what timestamp?
- Can you export the answer without asking the platform?
If the only answer to the first question is “the platform accepted it,” there is no record — only a state. A named actor and a before-and-after value are what turn an activity log into an answer for a supervisor. Paused is not a record; budget moved from £800 to £2,000 at 14:32 is. And evidence you can only read inside the tool that produced it is a feature, not a record — which is why the price of the evidence should never depend on how often you look at it.
The half nobody shipped
Give the shipped half its due. Supermetrics made the surface measurable, Windsor made it writable, and TracerGram made it feed the funnel. Those are real products doing real work, and the pace is impressive. What none of them shipped — because none of them can, on a surface this young — is the operations record an agency gets asked for months later, after the campaign, the budget line and the person who signed it off are the only things anyone cares about.
That record is not an ad-platform feature and never will be. It belongs one layer below the ad account, to the agency that made the decision. The newest surface has made that gap visible again.