ChatGPT Ads opens to all US advertisers, and the search measurement stack does not come with it

In a search account the keyword does two jobs: it buys the impression and it labels the row in the report. ChatGPT Ads keeps the first job and drops the second. Advertiser context hints sit at the ad-group level ^[3], and matching runs off conversation context rather than exact-match strings ^[2].

Ads Manager returns delivery metrics and accepts conversions through a pixel or the Conversions API ^[9]. What that set does not contain is any breakdown by the context an impression was matched against ^[14]. So the smallest thing an advertiser can steer is the ad group, and the only way to learn that one situation performs better than another is to buy the two situations separately and compare totals. The bid pivot that used to take an afternoon of keyword pruning now takes a campaign restructure and a waiting period.

The attribution problem follows from the same gap. Conversions arrive through the advertiser's own pixel or API ^[9], so the platform reports the media and the advertiser's stack decides the value, with nothing in between naming which conversation produced the click. The write-up, published on dev.to, advises testing whether the context an ad appears in produces meaningful downstream intent rather than assuming a click carries the same meaning as a search click ^[16]. That is the correct instinct, and it is also an admission that no transferable benchmark exists yet.

Pricing shows the same immaturity from the other side. Third-party observation puts clicks as low as $2, while OpenAI's Help Center tells advertisers to open at $3 to $5 ^[7][8]. The recommended floor is 50 percent above the cheapest clearing price anyone has reported ^[12]. CPMs, meanwhile, have come down from launch levels as inventory expanded ^[10]. Both readings are what a filling auction looks like: guidance runs conservative because variance is high, and prices move because supply moves. Neither figure is a planning input.

Custom audiences and conversion-optimized campaigns have been added as the product matured ^[11], and that is where advertiser-side identity and outcome signals start re-entering a system that declines demographic targeting ^[4]. About three months separated the US pilot from open access ^[13], and the source characterises the whole thing as an actively evolving ecosystem rather than a finished replacement for search or social platforms ^[15]. Its practical advice reduces to two points: set your own baselines with controlled test budgets, and keep on-platform delivery numbers separate from what you can actually attribute ^[17]. For anyone whose reporting cadence assumes a keyword report exists, that is a change to the operating model, not a change to a media plan.