ChatGPT Ads management

What is genuinely different about advertising in a conversation

The context is unlike search or social, and copy that ignores that tends to fail.

In search, someone types a query and scans results. In a feed, they are browsing and you interrupt. In an assistant conversation, they are mid-task, often several exchanges into describing a problem, and the assistant is helping. An advert arriving there is judged against the helpfulness of the surrounding conversation, which is a higher bar than being judged against other adverts.

The practical consequence is that promotional phrasing performs worse than plain, specific and useful phrasing. This suits businesses with something concrete to say and punishes those relying on brand adjectives. It is also, usefully, the same shift that works everywhere else in AI search.

Running it as a test, which is the only defensible way

There is no reliable benchmark data for a placement this new. Anyone quoting you an expected return is guessing, and we would rather structure around that uncertainty than paper over it.

So the engagement is built as a test. A budget you can afford to spend without a return. A written hypothesis about who this should reach and why. Success criteria agreed before launch. And stop conditions, so the decision to end it is made in advance by both of us rather than argued about later when it is emotionally harder.

That last piece is what stops a test becoming a slow drain. New channels attract sunk-cost thinking, and the defence is a number agreed while everyone is still calm.

• A budget sized to be learned from, not to be recovered
• A written hypothesis, so the result means something either way
• Success criteria set before launch, not chosen afterwards
• Stop conditions agreed in advance and honoured

Measuring a channel that will not attribute cleanly

Attribution here is harder than in established channels, and pretending otherwise produces confident nonsense.

Conversational placements sit early in decision-making, where influence is real and last-click credit is not. Someone who encounters you mid-conversation may search your name a week later and convert through what your analytics records as branded search. Judged on direct attribution alone, the channel looks worthless while actually working.

So we watch the totals as well as the platform report. Whether overall enquiries move when the test runs and when it pauses. Whether branded search volume shifts. Whether new enquirers mention having come across you. None of this is as tidy as a conversion count, and it is closer to the truth than one.

The risks worth naming before you spend

Three, and they are all real rather than defensive throat-clearing.

The platform is young, so the mechanics we set up may change during the engagement. Targeting, formats and reporting on new advertising platforms all tend to shift substantially in the first years, and some of that will land mid-campaign.

Context control is limited in a way it is not on other platforms. Your advert appears alongside generated conversation, and the range of what that conversation might be is wider than a keyword list. For most advertisers this is manageable. For regulated sectors, or brands with strict adjacency rules, it deserves a conversation before launch rather than after.

And the audience is not yet well characterised. Assistant users skew in ways that are still being understood, and that may or may not match your buyers. That is precisely what a test is for, and it is also why the budget should be one you can lose.