Your first week in ChatGPT Ads goes on looking for buttons that aren't there. Here is the map: every gap, checked in our own account in September 2026, with what you do instead. Almost all of them have a replacement, and hardly any replacement looks like the original.
There is no conversation or search terms report
The ad is chosen by the context of the conversation, and nobody shows you that conversation: you get aggregated data (impressions, clicks, spend, CTR, CPC, CPM and conversions) and not a word of what the person typed. It's a privacy decision by OpenAI, not a report on its way.
What you do instead: one ad group per intent. If all your context hints live in the same group, performance comes back mixed. With one group per intent, the group report becomes a poor man's search terms report: you don't know what people asked, but you do know which family of conversations brings you cheap clicks. And give each ad its own UTM so your analytics tells you the rest.
There are no negatives of any kind
No campaign negatives, no ad group negatives, no shared lists, no topic exclusions. You cannot exclude anything. For anyone whose craft is built on mining search terms, this is the hardest change in the module.
What you do instead: filter up front, not afterwards.
- Narrow with your hints, which are free text: describe the specific situation you fit, not the whole category.
- Qualify in the card itself. The title and the copy are your cheapest filter: saying who it's for ("for campaign managers"), the price or the format saves you clicks you never wanted to pay for.
- And if a whole group brings bad traffic, the only real "negative" is rewriting its hints or pausing it.
There is no manager account
Every advertiser has their own account and an agency can't create one on behalf of a client. One person can have access to several accounts and switch between them, but there is no combined view.
What you do instead: a naming convention from day one (market, intent, device), exporting the reports and consolidating them yourself —in a spreadsheet, if need be— and granting access person by person.
No scheduling, and no bid adjustments by device or country
There is no scheduling by hour or by day of the week, and no bid modifiers by device, location or audience: a campaign has a start date and an end date, and that's that.
What you do instead: what can't be adjusted gets separated.
| What you'd want to adjust | The real way out |
|---|---|
| Business hours only | Turn the campaign on and off (by hand, or with a tool that does it for you) |
| Bid lower on mobile | One campaign per device: the platform does let you pick where each campaign runs |
| Bid higher in one country | One campaign per country |
⚠️ Pitfall: when you duplicate a campaign to split devices, remove that device from the original or they will compete with each other, and remember that budgets add up: two campaigns at 25 a day are 50 a day.
There is no value-based bidding
The strategies are impressions, clicks and conversions (OpenAI's documentation). There is no equivalent of tROAS: the campaign optimises towards the number of conversions, never their amount. And the reports carry no conversion value: the revenue report is yours.
What you do instead: bid for quality by choosing which event counts as the campaign's conversion. If the campaign counts raw leads, it will chase volume; if it counts qualified leads, it will chase quality. It's the most powerful lever on the platform and almost nobody uses it. With one condition: volume, because a conversion strategy needs enough signal to learn and qualified leads are always far fewer than leads.
The amount, on the other hand, you measure in your own analytics.
And in Europe, no custom audiences either
They exist (with a minimum of around 25,000 matches to include), but they cannot be used in the European Economic Area or Switzerland. There is no personalisation there either: the ad is chosen using the current conversation, the language and the approximate location.
What you do instead: what you'd do in Google Ads by excluding customers, here you do with the message: a card that speaks to someone who doesn't know you yet filters out existing customers on its own.
The mindset table
| In Google Ads | In ChatGPT Ads | What you do |
|---|---|---|
| Search terms | Doesn't exist | One group per intent + a UTM per ad |
| Negatives | Don't exist | Narrow hints and a card that qualifies |
| Manager account | Doesn't exist | Consistent names, consolidate yourself |
| Ad scheduling | Doesn't exist | Turn on and off |
| Bids by device or country | Don't exist | One campaign per device or country |
| tROAS | Doesn't exist | Pick the conversion event (quality, not volume) |
| Value in the report | Doesn't exist | Your analytics |
| Audiences (in the EEA) | Not available | Qualify with the message |
What it does have, and often gets overlooked
- Performance by hour of day. You can't schedule, but you can see when you spend. A warning from our account: confirm which time zone the platform reports in before deciding anything with those hours.
- A breakdown by surface (iOS app, Android app, desktop web, mobile web), with one breakdown at a time and no crossing of surface with ad.
- Thirteen conversion event types, one of them custom, and deduplication by your own identifier: if the same conversion arrives twice, the first one stands.
- View-through conversions, reported separately and not billed.
- A product feed (from 100 to 1,000,000 items, using the schema of a Google Shopping feed).
- New since 16/09/2026: "Sponsored Agents" —after the ad, a separate conversation opens with the advertiser's own agent, in pilot— and integrations with Shopify and HubSpot.
What to remember
- No search terms and no negatives: filter up front —with the hints, the card and the page— and split by intent so the report can be read.
- No scheduling and no bid adjustments: what can't be adjusted gets split into campaigns, watching that budgets add up.
- No value-based bidding: you bid for quality by choosing which event counts as the conversion, and you need volume.
- No audiences in Europe: qualify with the message.
- Almost everything has a replacement; the only thing that never comes back is knowing what the person asked.