Advertising inside ChatGPT went from announcement to something you can actually buy faster than most people noticed, and the mechanics are different enough from Google and Meta that the usual instincts will get you into trouble. The most consequential difference: several campaign settings cannot be changed after you create the campaign.

Here’s what’s established as of late September 2026, what it means operationally, and what I’d do before spending real money on it.

A caveat worth stating up front, because it applies to everything below: this is a young platform, details are shifting week to week, and some features are limited to selected advertisers or markets. Verify eligibility in your own account rather than planning around a feature you read about.

The rule that will cost someone money

Billing model, objective and conversion event are fixed at creation. You cannot change them afterwards. If you pick the wrong conversion event, the fix is a new campaign — and a new campaign means starting the learning over.

On Google or Meta you can adjust an objective, swap a conversion action, or change how you’re billed, and the campaign carries on. Bringing that habit here produces a graveyard of half-learned campaigns. Decide the three settings deliberately, write them in the brief, and have someone check them before the campaign goes live.

Conversion campaigns bill on impressions

This is the part that catches performance marketers.

Conversion-optimised campaigns are billed on impressions, not on conversions. You’re optimising toward a conversion while paying for reach. That’s not unusual in itself — it’s how most social platforms work — but combined with the next point it changes how you should read early results.

The bid cap is not a guaranteed CPA. It’s an input to the auction, not a promise. Treating it as a target cost per acquisition and budgeting accordingly is the most likely way to be surprised by the first invoice. Set budgets you can afford at zero conversions, and judge the campaign on outcome after the fact.

Conversion events: standard only

Only standard conversion events are supported. Custom events are not.

Practically, this means the measurement work happens before the campaign exists:

  • Map what you care about onto the standard event set. If your business’s real outcome is a custom event, you need a standard event that sits close enough to it to be a useful proxy.
  • Implement and verify the event before creating the campaign, because you can’t change it later.
  • Check the event isn’t double-firing. A conversion event that counts twice is worse here than elsewhere, since you can’t swap it without rebuilding.

Measurement runs through an OpenAI Ads pixel, distributed as a Google Tag Manager template from a public GitHub repository, imported manually. Two things to check before publishing the container: that the template’s permissions are what you expect, and that the pixel ID and event fields are right. GTM’s template status page will show you the imported commits and any errors attached to them — some of those error messages are still undocumented, so budget a little time for the first setup.

Flexible attribution windows — and what they don’t do

You can now choose click windows of 7, 14 or 30 days and view windows of 0 or 1 day.

One thing to be very clear about internally: this changes reporting, not optimisation and not billing. Widening the window increases reported attribution. It does not increase sales.

If you change the window mid-flight, restate the prior period on the new basis before comparing, and say so in the report. Otherwise you’ve manufactured an improvement out of a settings change, which is the kind of thing that gets discovered in a quarterly review. This is the same discipline that reporting on white-label accounts requires, and the same underlying issue as last-touch attribution.

Targeting and creative

Platform targeting is available across Android app, Android web, desktop web, iOS app and iOS web. Worth splitting early: behaviour and conversion rates between app and web surfaces are rarely similar, and you want the data separated before you start drawing conclusions.

Creative assistance will suggest copy and imagery, and optionally adapt or translate text. Useful for volume. Review every output against your brand claims before it ships — automated adaptation is confidently wrong about regulated language, pricing and product specifics, and a translated claim can be legally different from the original.

Announced 16 September, currently a US test with selected advertisers: a user who clicks an ad can enter a clearly labelled, business-sponsored conversation, separate from ChatGPT’s independent answers and from the conversation they were having.

This is the genuinely new format, and it’s worth thinking about seriously even if you can’t access it yet. A sponsored conversation is closer to a landing page than to an ad unit — which means the skills that matter are the ones that make a good conversational funnel: knowing what the buyer’s actual questions are, having accurate product information available, and knowing what the conversation is supposed to conclude with.

The obvious failure mode is treating it as a chat-shaped brochure. If your sponsored agent can’t answer the third question a real buyer would ask, it’s worse than a static page.

Integrations

HubSpot connects account setup, ad creation, reporting and lead follow-up. For lead gen teams this matters more than it sounds: it’s the shortest path to getting outcome data back out, which is the difference between optimising toward form fills and optimising toward revenue. Same principle as lead quality in Google Ads.

Shopify is available to US merchants, with international expansion announced for 23 September. Check eligibility and tracking before switching anything on.

What I’d do before spending money here

  1. Implement and verify the pixel first. Check the GTM template’s permissions, confirm the standard event fires once, and confirm it fires on the real outcome.
  2. Write down the three locked settings — billing model, objective, conversion event — and have someone else check them before creation.
  3. Set a budget you can lose. Bid cap isn’t a CPA, so size the test on spend, not on expected acquisitions.
  4. Split app and web from day one.
  5. Fix the attribution window and leave it alone for the duration of the test. Change it later, deliberately, with a restated baseline.
  6. Judge it on blended efficiency, not on platform-reported conversions. A new channel’s own numbers are the least reliable thing about it. See how to calculate ROAS.

The short version

It’s an impression-billed conversion channel with immutable core settings, standard events only, and a bid cap that isn’t a target CPA. The measurement setup has to be right before the campaign exists, because you can’t repair it inside the campaign afterwards. Sponsored agents are the part worth paying attention to strategically, and the part most likely to be done badly by people treating a conversation as a banner.

Test it like a new channel: small, measured against a number the platform doesn’t control, and with the tracking verified before the first impression.

Most of the detail here is tracked week to week in Yoann Ferrand’s Paid Media News newsletter, which is where I follow platform changes at this pace.

If you’re deciding whether a new channel deserves budget this quarter, that’s the kind of call I help with.