📗 This is the user manual for the Suite for ChatGPT Ads, the tool in the Ninja Scripts panel. Here you can search it, listen to it, highlight it and save it to favourites. It is updated with every release of the tool.
Without measurement, «maximise conversions» has nothing to maximise and cost per conversion is a decorative number. This is the part that takes the most work and the part that changes results the most.
The two keys, again
They aren't the same and they always get confused:
| Key | What for | Where it's created |
|---|---|---|
| Ads API | Reading the account and applying changes | Settings → API Keys |
| Conversions | Sending conversions from the server | Tools → Conversions → key icon 🔑 (top right, no label) |
The 🎯 Conversions screen carries a diagram pointing at exactly where that icon is, with a direct link to that screen in your account. You'll also need the pixel ID, which sits just under the data source name, on the Data Source tab. Don't go by how it starts: some begin with Fp… and some don't.
🚩 Sending isn't measuring: the actions must be created by hand
This is the most expensive thing to learn the hard way, and we learned it in our own account:
OpenAI does not create conversion actions for you. You can be sending events for days, get a
200 OKon every one, and still see zero in the report because no action exists with that event type. And there is no warning telling you.It happened to us: 46 events accepted over several days and not one conversion counted. The moment the action was created, 22 conversions appeared at once.
Actions are created under Tools → Conversions → Conversion Events → + Create. You need: the data source, the base event (a closed list: that's what matches what you send), a name (yours, it's only a label) and the attribution windows.
A trick to save time: Event Stream tells you in ten seconds whether it's «not arriving» or «arriving and nobody counting it».
The tool can't create those actions for you because OpenAI's API doesn't allow it. What it does do is show you, on the screen itself, the exact name you have to use for each one, with a copy button, and each one's real status based on what has already been sent: «already arrived (n)», «pending», «couldn't be sent» or «none yet».
One event per campaign
The most important rule in the whole measurement setup, and the one that messes up the most money when broken:
- Two actions describing the same fact add up: you'd count every conversion twice.
- Also attaching the stages (qualified, opportunity, sale) would make a single lead count four times.
One campaign, one event. The other actions keep being measured on their own in OpenAI's own report; what they must not be is attached to the campaign.
And since there's no value-based bidding on this platform, choosing which event the campaign counts is the only quality lever there is: if it counts the raw lead, «maximise conversions» chases volume; if it counts the qualified one, it chases quality. That said: qualified leads are always far fewer, and with low volume the campaign is left without a signal to learn from. The sensible minimum is 30 in 30 days, and 💡 Recommendations tells you when you reach it.
⚠️ Changing the event is not changing the objective, and it's worth not mixing them up. The event a campaign counts is changed in the ads manager (the tool can't: the API won't even list the events). The objective —clicks or conversions— can never be changed once the campaign exists: that's what 🧬 Clone is for, in section 14. If your campaign was created for clicks, attaching the qualified event to it will not make it bid for conversions.
🎯 Sending the Ninja Leads stages
If you have Ninja Leads (Lead Rating), the tool can send OpenAI the lead's four stages — raw, qualified, opportunity and sale with its amount — so the campaign learns from quality rather than volume.
How it works, every 30 minutes:
- The leads arriving from your site are read and matched against your Lead Rating sheet by the lead identifier.
- One event per lead and stage is sent, with a deterministic identifier: if it's re-sent, OpenAI deduplicates it.
- The raw lead goes with the lead's real date. Each stage goes with its milestone's date when your CRM sends it; if it doesn't, with the day the panel sees it ticked.
- The sale amount is sent in the currency you pick in the settings (EUR by default): check it if you don't invoice in euros.
- ✖ Bad stops the raw lead from being sent. Careful: if that same lead has a stage ticked (qualified, opportunity or sale), that stage is still sent — the bad-lead filter only applies to the raw one.
- There's a switch per stage and a Test button that validates without really sending anything.
- And two checkboxes that widen what gets sent, both off by default. They are explained in detail below, under «The two checkboxes: what qualifies and what travels».
Two platform limits worth knowing: an event older than 7 days is rejected (OpenAI's window), and what's been sent cannot be withdrawn — unlike Google Ads, there are no adjustments or retractions here.
💰 The 7 days apply to the stages too. If you tick the month's sales all at once at month end, the ones older than a week never get sent and the campaign never learns from them. Tick milestones as they happen: that's the difference between «maximise conversions» chasing quality or never finding out.
🩺 Why each lead is sent — or isn't
Below the send figures there's a block answering exactly that question, with figures taken from the database (not from the last pass). It exists because the most common case is the hardest to diagnose: a table of zeros that looks like «nobody came» when it's actually blocked.
If none of your leads can be sent, it says so at the top and in red, with the cause in plain words. The two causes of being blocked:
- Your site's tracking doesn't declare advertising consent: without it, neither the click identifier nor the email is stored, so there's nothing to attribute with. Fixed by updating the tracking on your site.
- The lead carries neither click identifier nor email: it didn't come from a ChatGPT ad, or its form never captured the email.
And the ones that were attempted but didn't make it are broken down by their real reason —expired past the 7 days, the source page was missing, conversions key rejected— instead of being a lone number.
🔀 The two checkboxes: what qualifies and what travels
There are two checkboxes that look alike and do different things. Neither changes the normal case: a lead that carries the click identifier and accepted advertising is always sent, whatever you do with them. What they do is widen the edges, each on its own side.
| What it adds | What that lead needs | |
|---|---|---|
| «Leads without the click identifier too» | leads that did not come from a ChatGPT ad | to have an email |
| «Also send leads who did NOT accept advertising» | leads that did come from an ad but gave no consent | to carry the click identifier |
Which leads qualify
| The lead has… | Neither box | Only the 1st | Only the 2nd |
|---|---|---|---|
| identifier + email, accepted | sent | sent | sent |
| identifier, no email, accepted | sent | sent | sent |
| no identifier, with email, accepted | not taken | sent | not taken |
| neither identifier nor email | not taken | not taken | not taken |
| identifier, did not accept | not taken | not taken | sent |
What travels in each case
| Case | Identifier | Hashed email | Phone | Out of audiences |
|---|---|---|---|---|
| Accepted · identifier + email | ✅ | ✅ | — | no |
| Accepted · identifier only | ✅ | — | — | no |
| Accepted · email only (1st box) | — | ✅ | — | no |
| Did not accept · identifier (2nd box) | ✅ | — | — | yes |
Two things worth fixing in your mind: the phone is never sent, and for someone who did not accept advertising the email never travels even if the lead had one — for that lead only the click identifier is stored, the one the platform itself put in the URL, and its event is flagged to stay out of audiences and personalisation: it counts as a conversion and nothing else.
When to switch each one on
The second is the easy one: with it you measure 100 % of the leads ChatGPT brought you instead of only those who accepted, and you hand over no personal data of anyone who said no. Its only requirement is that your privacy policy covers it, because the click identifier is still visitor data.
The first is a trade-off, and it is worth understanding before ticking it. It sends the hashed email of all your leads — including those from your other campaigns and organic ones — so the platform can try to recognise them. That rescues the person who clicked your ad, did not fill the form, and came back days later typing your address: that lead no longer carries an identifier and without this box you lose it. But it opens the door to the opposite: the platform claiming leads another channel brought you.
📊 A real case, with figures. In an account we manage, out of 906 leads in one week only 4 came from ChatGPT Ads; 772 came from another ad platform. With this box on, 318 events were sent, of which only 3 carried the click identifier. And the only conversion the platform claimed was dated on a day with not a single ChatGPT lead: it came from the email match of a lead another channel had brought — and that lead was already counted on the other platform.
How to read the figures with the first one on
It is not that switching it on is wrong: it is that the conversions the platform counts then cannot be used to judge the channel. They serve another purpose, which also matters: giving «maximise conversions» signal to learn from.
So it helps to separate the two questions, and the panel gives you a different answer for each:
| What you want to know | Where to look |
|---|---|
| Whether the bidding has enough signal to learn | the platform's conversions |
| What this channel really brings you | 📈 Report → real cost per lead, which counts by each ad's tag and only adds up the leads that arrived from there |
That way you can have both: the generous figure feeding the algorithm and the clean figure for deciding where the budget goes.
🛡️ Visits · each ad's traffic quality
If you have Ninja Shield, this screen tells you how many of each ad's visits are people: measured visits against the clicks OpenAI counts, % bots, % with no interaction, duration and forms submitted.
It's the counterpart to cost per lead: a cheap ad with junk traffic stops looking cheap. Ads pointing at pages without the tracking are flagged as «not measured», with the reason, rather than showing up as zero.
Which network each visit comes from (since 16/10/2026). For each paid visit the network of its IP is looked up, and only that is stored, never the IP. Two more figures appear, per ad and in total:
| Figure | What it means |
|---|---|
| % data center or VPN | the visit comes from a hosting or VPN provider, which is where bots usually live. iCloud Private Relay networks are left out on purpose: those are people on iPhones |
| % outside the area | the IP is in a country none of the account's campaigns target. It is compared by country, because an IP doesn't reliably tell the postal code |
Both are calculated over the visits with that data. Visits before 16/10 only have it if they were still in the queue (7 days); otherwise the figure shows «—», never 0. A VPN is not always a robot: read it together with the % of bots and the forms.
Crawlers are set aside. OpenAI and others open the ad URL, with its parameters, to review the page. That is not a click. If the browser declares itself a crawler (OAI-SearchBot, ChatGPT-User, Googlebot…), the visit is set aside: it counts neither as a visit nor as a bot, and the «Where the visits come from» block says how many there were and whose.
This screen's thresholds (when a visit counts as a bot) are calibrated with Google Ads traffic and are provisional for ChatGPT: the ChatGPT app opens websites in its in-app browser, and that produces quick bounces that aren't robots. They'll be recalibrated with our own data.