The short verdict
ChatGPT Ads is worth a properly designed test if you sell a considered purchase, you already see ChatGPT showing up in your referral data, and you can stomach clicks in the three to eight pound range while you learn. It's not worth moving money out of a working search account for, and it isn't a channel you can run on autopilot next to your Google campaigns.
The targeting is genuinely clever. The advertiser controls and the feedback loop are primitive. Those two facts sit together uncomfortably, and everything else in this guide falls out of them.
Interesting enough to test. Not mature enough to trust. Both halves of that sentence matter.
What ChatGPT Ads actually is
Ads appear inside ChatGPT conversations. You don't bid on keywords. You describe, in plain language, the situations where your business is a sensible next step, and the system decides when a live conversation matches that description well enough to be worth showing your ad.
The four things that matter most before you spend anything:
- Ads only serve to Free and Go users. Plus, Pro and Business accounts never see them. In a lot of B2B categories that quietly removes a chunk of the buying committee.
- The minimum spend is about £15 a day. Low enough that a test is affordable, high enough that a badly structured account burns a month's budget learning nothing.
- You get metrics, not diagnostics. Impressions, clicks, CTR, CPC, conversions and performance by ad. No prompts, no themes, no search terms equivalent, no negatives.
- The user already has their answer. Your ad has to persuade someone to leave a conversation that's already helping them. That's a harder job than appearing above a search result.
Who should test it, and who shouldn't
- Considered purchase with a real research phase
- ChatGPT already appears in your referral or self-reported source data
- Buyers likely to be on free plans, or researching before anyone buys them a licence
- Deal values that survive an eight pound click
- You have a specific landing page, not a homepage
- Tiny target market where you can name every account
- Enterprise buyers almost certainly on Business plans
- No conversion tracking or CRM source capture in place
- Budget that only exists by cannibalising a profitable channel
- You need an answer inside two weeks
If you sell into a market of a few hundred nameable companies, this probably isn't your channel yet. That's a tiny-TAM problem and it gets solved with account penetration, not conversational reach.
Not sure whether your buyers are even reachable on this channel?
Book a 30 minute call. We'll look at your organic ChatGPT referrals and your buyer's likely plan tier before anyone spends anything.
Setting it up
Build it in your own account, not an agency one. You'll want the history, the audiences and the pixel data if anyone ever hands the account over. The build order that works:
- Account and business verification first. This takes longer than you expect and blocks everything behind it.
- Pixel and conversion events before the first campaign. Launching without measurement wastes the only expensive part of the test.
- One campaign, one geography, one offer. Blended campaigns are unreadable on a channel with no diagnostics.
- Two to four ad groups, each one conversation state. The ad group is your hypothesis container. Treat it that way.
- Context hints written per group, kept concise. Over-engineered hints have strangled delivery in public tests.
- Specific creative and a matching landing page. The page is part of the relevance model, not just the destination.
Measuring it properly
On a channel that won't show you demand, your measurement stack has to do the job the platform's reporting doesn't. Three layers, all of them needed:
- Platform layer. OpenAI Pixel plus the Conversions API for server-side events, with Advanced Matching where you're allowed it.
- Analytics layer. Clean UTMs and the click parameter so GA4 doesn't quietly file the traffic as direct or referral. Expect discrepancies and know which number you're managing to.
- CRM layer. Source written onto the record at creation. Without this you'll be arguing about attribution in month three instead of making a decision.
Add the human layer too. A "what were you trying to solve" field on the form and a note from every sales call is the closest thing to a search terms report you'll get here.
The blind spot you have to design around
There's no search terms report. No prompts, no themes, no negatives, nothing that tells you which part of your context hint did the work. You're paying for a targeting decision and you're not allowed to look at it.
That's why published results are all over the place. One agency reports over 150 qualified leads across three B2B SaaS clients at a cost per lead around 60% below Google. Another reports roughly $415 spent, 60 clicks and not one interested buyer. Both are believable, and neither can explain the gap, because neither was shown the demand they bought.
The practical consequence: change one variable at a time, and accept that every lesson costs a test cycle rather than an afternoon reading a report.
How it differs from Google
The important difference isn't the interface. It's the order of events. On Google the user asks and has to click to get an answer, so your ad is standing on the route to the thing they want. In ChatGPT the answer arrives first, and your ad has to be worth interrupting it for.
More contextual intent does not automatically mean more click intent. The conversation is already solving the problem your ad is trying to sell.
Which means the creative bar is higher, not lower. A generic ad served into a relevant conversation will lose to the conversation every single time.
Context hints and the new optimisation craft
Context hints look like keywords and behave nothing like them. They're a description of a situation, not a match on a string, and you get no feedback on which part of the description was used.
The public evidence is thin and contradictory. Search Engine Land found simple keyword-style hints outperformed conversational ones in a controlled test. Another advertiser found very long hints throttled delivery almost to nothing. That contradiction is the reason to run your own test rather than copy anyone's template.
A 30 day test plan
This is the shape of the test we run. It's deliberately small, deliberately slow, and designed to produce a decision rather than an opinion.
Verification, Pixel and CAPI, CRM source capture, one landing page per conversation state. Nothing goes live until an event fires end to end in a test.
One campaign, two to four ad groups, £15-30 a day. Concise hints. Two creatives per group, both specific, neither clever for the sake of it.
Which groups actually spend, which creatives earn clicks, what engaged session rate looks like. Conversion volume won't be readable yet and pretending otherwise is how people kill working tests early.
Kill the group that didn't deliver, or change hint style, or change creative. One of those, not all three.
Scale, extend or stop, judged against your existing cost per qualified lead. Qualitative notes from sales calls carry as much weight as the dashboard.
Budget roughly £450 to £900 in media for that month. If your average deal can't justify that as a learning cost, the honest answer is to wait.
The rest of the cluster
Account, verification, Pixel, bidding, ad groups, the £15/day minimum and launch checks.
How conversions really get attributed, the oppref parameter and getting source onto the CRM record.
You pay for a targeting decision the platform will not show you. What that costs B2B.
Contextual intent is not click intent. Why the answer arriving first changes everything.
Not keywords, but the closest thing to a targeting lever you have. What the tests show so far.
Common questions
We run ChatGPT Ads and publish what happens
- A suitability read before spend: organic ChatGPT referrals, buyer plan tier, click price tolerance.
- Account, Pixel, CAPI and CRM source capture built in your own account, not ours.
- A structured hypothesis test across two to four conversation states, one variable at a time.
- Qualitative capture on every lead, so you learn what people were actually researching.
- A thirty day scale, extend or stop call measured against your existing cost per qualified lead.
Starter £1,250/mo, Growth £1,725/mo, Scale £4,025+/mo. For current clients we cover setup and management of the first test and you fund the media. This channel is young and we don't think anyone should pay a premium while the whole industry is still learning it.