The short answer
Create an OpenAI Ads account in the advertiser's own name, complete verification and billing, install the OpenAI Pixel and a real conversion event, tag every destination URL with UTMs and preserve the oppref click parameter. Then build one campaign per objective and geography, two to four ad groups where each one describes a single conversation state, and at least three genuinely different ads in each. Start on CPC or Maximize results, not conversion bidding. UK minimum is £15 a day, so a thirty day test starts at roughly £450.
Everything below is the detail behind those four tiles. If you want the honest verdict on whether the channel works, that lives in our comparison with Google Ads. This page is the build.
Decide whether you should test at all
Most of the money wasted on a new channel goes before anyone even opens the ads manager. Score yourself honestly against these eight. Two points for a clear yes, one for a maybe, nothing for a no.
| Factor | Why it decides the test |
|---|---|
| High-consideration purchase | People describe context before they choose. That description is the only targeting signal you get. |
| Buyers research this in ChatGPT | If your category is a quick price check rather than a considered problem, there's nothing to match against. |
| Customer value tolerates £3 to £8 clicks | Observed CPCs so far sit around the two to five dollar mark. Low-value offers die at that price. |
| You already see ChatGPT referral traffic | Existing organic referrals are the strongest single signal that paid placements are worth buying. |
| A specific landing page exists | The page is part of the relevance model, not just the destination. |
| Conversion tracking is healthy | Without it you're buying clicks and guessing. |
| You've a benchmark channel | You need something to compare cost per qualified lead against, or the result means nothing. |
| Free and Go users plausibly include your buyer | Plus, Pro and Business accounts never see ads. In some B2B markets that removes the decision maker. |
Twelve or more out of sixteen and you have a strong test candidate. Eight to eleven, test carefully with money you can afford to write off. Below eight, spend it on the channel you already know converts.
The reachable audience is Free and Go users. In plenty of B2B markets the person signing the invoice is on Business, and no amount of clever setup will reach them.
Account, access and verification
The client owns the ad account and the billing relationship. You get invited to it. This is the opposite of the habit Google Ads Manager Accounts and Meta Business Manager have trained agencies into, and it matters here because OpenAI's account roles are newer and messier to unpick later.
- Legal advertiser name, exactly as it should appear on the ad.
- Verified domain and the advertiser favicon, which is what shows next to the ad.
- Billing details and the account ID.
- Who at the client is admin, and which of your people hold access.
Verification and payment setup are the two most common reasons a brand new account sits at zero impressions on day one. Do both before you build anything, because you can't tell the difference between "not verified" and "context hints too narrow" from the delivery screen.
Measurement before media
Don't launch and promise to sort tracking next week. Early practitioners keep reporting gaps between clicks reported in the ads manager and sessions landing in GA4, so you want both numbers from day one rather than an argument in week three.
- OpenAI Pixel sitewide, firing on every page.
- One meaningful standard conversion event. A submitted lead or a booked meeting, never a scroll depth or a button click.
- Conversions API where the stack allows it, because browser-only matching is weak on long B2B journeys.
- Automatic Advanced Matching where it's appropriate and your privacy notice covers it.
- UTMs on every destination URL, plus
opprefpreserved through redirects. - CRM source and campaign fields populated from those UTMs, so the deal that closes in March still says ChatGPT.
The URL pattern we use:
https://yoursite.co.uk/your-landing-page
?utm_source=chatgpt
&utm_medium=paid
&utm_campaign={CAMPAIGN_NAME}
&utm_content={ad_id}
&utm_term={ad_group_id}
&campaign_id={campaign_id}
&oppref={oppref}Then run a live test conversion yourself before spend starts. Click a real ad, submit a real form, and confirm the event appears in the ads manager, in GA4 and on the CRM record. Three places, one action.
Campaign, objective and bidding
Three objectives exist. CPM buys reach, CPC buys clicks, and oCPC optimises toward a conversion event while still charging you per click. There's a trap in the third one: you can't convert an existing CPC campaign into oCPC later. You build a new campaign. So decide deliberately.
| Objective | Use it when |
|---|---|
| CPC / Clicks | New account, unknown traffic quality, tracking just installed. This is where almost every first B2B test should start. |
| Maximize results | You want the system pacing bids inside the budget and you're comfortable handing over that control early. |
| oCPC / Conversions | Tracking is proven, a supported standard event fires reliably, and there's enough volume for the model to learn from. |
| CPM / Reach | Rarely, for B2B. Visibility inside a chat window isn't a business outcome you can bank. |
OpenAI suggests a starting maximum CPC around three to five dollars, roughly £2.20 to £3.70. Treat that as opening bid guidance and nothing else. It isn't a benchmark, and hitting it isn't a result.
The auction is relevance-weighted and second price, which means creative and landing page quality change what you pay, not just what you win. Geography goes beyond country in supported markets, down to regions, cities and postcodes, and you can pick platform: iOS, Android or web.
Want us to build and run your first ChatGPT Ads test?
Book a 30 minute call. We'll tell you straight whether your buyers are reachable on this channel before anyone spends anything.
Ad groups and context hints
This is the part that looks like keyword targeting and isn't. At ad group level you write context hints: a description of the conversations, topics and situations where your offer is a sensible next step. OpenAI is explicit that these aren't exact match, not hard audience rules, and not a promise your ad appears beside particular wording.
One theme per ad group. The structure that works:
One business objective, one geography, one broad offer. For example: UK, demo demand, ChatGPT.
One conversation state. Existing solution failing. Comparing vendors. A specific technical problem. Looking for a specialist.
Three or more genuinely different angles, not three rewrites of the same sentence.
A hint that works reads like a description of a person mid-problem. Ours, for the LinkedIn side of the business, runs close to this:
What to avoid: thousand word audience manifestos, job title stuffing, unrelated services bundled in, and every conceivable pain point crammed into one box. One advertiser reported a roughly thousand word hint setup barely delivering at all, then delivery climbing once they cut it back to a few hundred words of business type, intent and use case.
Creative that suits a conversation
The static format is advertiser name and favicon, a title, a description, a destination URL and a square image. That's it. The person seeing it is halfway through working out a problem, so the ad has to read as a useful next step rather than an interruption.
- Problem. Name the thing they're currently describing to the model.
- Specialist. Who you're for, narrowly. Vague breadth reads as nobody.
- Outcome. The commercial result, in their language.
- Diagnostic. A second opinion before they make an expensive decision.
- Proof. One real, quantified result you can stand behind.
On images, test a founder or expert photo against a branded graphic and a product screenshot. One practitioner reported photos lifting click-through from around half a percent to several percent against logos and service graphics. That's a single anecdote and I wouldn't plan a budget on it, but the gap is big enough to be worth an hour of your time.
Budget and what a real test costs
UK minimum is £15 a day. Run continuously for thirty days and you're at £450 before you've learned anything, so plan the whole test up front rather than topping it up nervously.
Below £450 you're not running a test, you're buying an anecdote. If control matters more to you than pacing, set a total campaign budget instead of a daily one and let it run out.
For our own retained clients we cover the setup and management of the first test at no extra fee and the client funds the media. Partly because it's a fair way to share the risk on a channel this young, and partly because every test adds to a benchmark set nobody else has yet.
Launch checks and troubleshooting
Reporting can lag several hours, so resist the urge to diagnose anything on launch afternoon. When something is genuinely wrong, work the symptom rather than the vibe.
| Symptom | Check in this order |
|---|---|
| No impressions | Verification, billing, ad review status, campaign dates, budget, bid competitiveness, overly narrow context hints, location and platform availability. |
| Impressions, no clicks | Creative is weak or the hint is too broad. Sharpen the title, get more specific about who it's for, change the image, point at a more relevant page. |
| Clicks, no engaged sessions | Curiosity traffic or a mismatch between ad and page. Compare platform clicks against GA4 sessions before you blame the page. |
| Engaged traffic, no leads | Offer or page. Test a lower friction, diagnostic style next step for people who are researching rather than buying. |
| Leads, poor quality | Tighten the ICP language in the hint, the ad and the page. Qualify harder on the form. Don't chase lead volume. |
Rejected creative and rejected landing pages are common enough to be worth a mention. The fix is usually a claim you can't substantiate on the page itself.
How to judge it after 30 days
Not on click-through rate. Not on cost per click. Cheap clicks from a semantically relevant conversation can still be worthless traffic, and that's the single most repeated finding in the public tests so far.
- Primary: qualified leads, booked meetings, pipeline created.
- Secondary: lead rate, engaged session rate, cost per click, click-through rate.
- Against: your existing Google or LinkedIn cost per qualified lead, over the same window.
Then pick one of three, without sentiment:
Qualified economics are competitive with your other channels, and the demand looks incremental rather than stolen from search.
The signal is promising but the sample is too thin. One more month, same hypothesis, more budget.
Traffic quality is weak after a fair test of intent and creative. Move the money. No sunk cost theatre.
Read next
Common questions
Want ChatGPT Ads set up and tested properly?
- Client-owned OpenAI Ads account, verification, billing and access configured.
- OpenAI Pixel, a real conversion event, CAPI where feasible, UTMs and CRM source capture.
- One campaign, two to four ad groups built as distinct conversation states, three plus ads in each.
- A context hint test with one variable changed, so the result means something.
- A thirty day read against your existing cost per qualified lead, and a straight scale, extend or stop recommendation.
Starter £1,250/mo, Growth £1,725/mo, Scale £4,025+/mo. For current clients we absorb the setup and management of the first test and you fund the media. We're learning this channel alongside everyone else and don't think you should pay a premium for that.