Running ChatGPT Ads for the First Time

After nearly a decade managing campaigns across Google, Microsoft, Meta and Amazon, it's genuinely rare that a brand-new advertising platform comes along. Platforms evolve constantly, with Performance Max, Demand Gen and the ever-expanding range of Meta campaign types, but always within existing ecosystems. So when ChatGPT Ads launched, we decided to be early. 

But not with a client's budget, with our own.

How do ChatGPT Ads actually work?

If you're coming from a Google or Meta background, the first thing to get your head around is targeting. ChatGPT Ads don't use keywords in the traditional sense. Instead, they rely on what OpenAI calls contextual hints. Rather than bidding on search terms, you're describing the type of conversation that should trigger your ad; what's the nature of the discussion someone is having with ChatGPT when your ad becomes relevant?

When I first set up our campaigns, I fed the R3 website to both Claude and ChatGPT and asked a simple question: "If you were having a conversation with somebody, how would you present R3?" That gave me a starting point for phrasing our contextual hints, and guidance on how to present the agency succinctly in the form of a conversation rather than a search result.

The biggest surprise? It’s pretty basic

Honestly? The thing that surprised me most was the simplicity of the platform. Google and Meta are incredibly mature, with menus and sub-menus, endless reports and granular controls. ChatGPT Ads is far more minimal by comparison. When you first log in, you’re taken straight to campaign creation, and the structure is simply: 

Campaign → Ad Group → Ads. 

The bare bones are there, but the insight and reporting layer isn't (yet).

To give you an idea of the gap, with Google I can see what hour my ad was shown, on what device, in what location, right down to postcode level. With ChatGPT Ads you get top-level performance only which covers clicks, impressions, click-through rate, and conversions if you've set up conversion tracking (which we have). For everything else, you're flying blind, unless you bring your own measurement, which I'll come back to.

I also expected targeting to be more specific, especially as I'm used to being able to serve ads to 45-year-old males on smartphones at a particular time of day. But with ChatGPT Ads, you get a contextual hint, geographic targeting and audience layering via first-party data, and that's about it. My mental model going in and the reality were definitely a little different.

It’s one account per client, and the client owns it

Another structural difference worth knowing is that OpenAI doesn't allow agencies to create ad accounts on behalf of clients. There's no equivalent of Google's MCC or Meta Business Manager where an agency can house a portfolio of client accounts. It's strictly one account per client, set up and owned by the client themselves (and there's a fair amount of business and billing information required during setup), with access then handed over to the agency.

It’s this that shaped how we approached testing ChatGPT Ads. Rather than pitching it cold to a client, we decided to run it on R3's own marketing.

So why did we test it on ourselves first?

There were two reasons. The first is straightforward; we want leads too. We'd been running Google Ads and Microsoft Ads for R3, so ChatGPT Ads was a natural extension of that. A new source to test.

The second reason is about risk. It would be contradictory for R3 to recommend a platform to clients that we'd never run ourselves. In digital marketing it's very easy to spend budget without producing a return, and there's no risk to an agency recommending a shiny new channel with someone else's money. 

Instead, we took the burden of risk onto our own ad spend. So when we offer this service to our clients, we'll be confident with the interface, the campaign setup, and, crucially, the pitfalls.

One of R3's strengths as an agency is that we run our own e-commerce businesses. So we understand the day-to-day pressures clients face from budgets and ROAS targets to the need to be profitable. Testing ChatGPT Ads with our own money and budgets is the same philosophy applied to a new channel.

What went wrong first (and how did we fix it?)

My first campaign was a lesson in itself. Because R3 has so many arms to the business, including RevOps, development and marketing, that initial iteration was a catch-all. Essentially: "We're R3, we do absolutely everything, come give us a lead."

The click-through rate was incredibly poor. But when I thought about it, the reason was obvious; somebody looking for marketing services is not the same person looking for a Laravel developer. You don't walk into a clothing shop wanting a coat, trousers, a shirt, jeans and boots all at once. You go in looking for one specific thing.

So I went back and rebuilt the account with service-specific ad groups, each with contextual hints focused on what a particular user is actually looking for and where we want to direct them, rather than "we can do everything, give us a shout."

What the results are like so far

The restructure made a real difference. With the original catch-all setup, average time on site from ChatGPT Ads traffic was less than 10 seconds, and not a single user reached our contact page.

Since moving to service-specific contextual hints:

• Average time on site has tripled to 30 seconds
• Four users have reached the contact page
• We've had our first genuine enquiry come through

Early days, and we're yet to speak to that lead, but it's a directional signal that the channel can drive quality traffic when the targeting is structured properly.

You'll need GA4 alongside it to measure more effectively 

This is my single biggest recommendation for anyone considering the platform. ChatGPT Ads will tell you that you had 30 users on a given day, your click-through rate, and your average CPC, and not much more.

To judge the quality of that traffic, we set up UTM tracking and cross-reference everything in GA4. Of those 30 users, what was the average time on site? What was the engagement rate? Which pages did they view? That's how you build a picture of whether the traffic is any good, and right now, it's the only way.

What KPIs are realistic?

With so little benchmark data available, I'd keep it simple initially: how many people are you getting to the site, and what percentage of them are good quality? It'd be naive to expect 100% quality traffic from any channel, but if 30 to 40% of your ChatGPT Ads visitors are engaging properly, that's a worthwhile early KPI.

You can set up conversion tracking for contact forms and lead events, and optimise bidding towards those. But bear in mind that the benchmarks we rely on elsewhere don't exist yet. On Google, the golden rule for maximise-conversions bidding is 30 leads in 30 days. For ChatGPT Ads, nobody knows what the equivalent is.

One more note of caution: be sceptical of the platform's own recommendations. When I first created our campaign, ChatGPT suggested a £15/day budget and a £1.80 bid, before it knew our industry, our campaigns, or (I'm fairly sure) before I'd even added a contextual hint. Next time, I'd trial a much lower bid rather than taking those numbers at face value. Could a 50p bid deliver the same traffic?

Should you run ChatGPT Ads now, or wait six months?

Truthfully, it depends on the business.

If you have an experimental marketing budget with no hard CPA or ROAS target attached, go ahead and run with it. Six months of hands-on experience on a brand-new channel will be invaluable, and you'll be ahead of the curve when it matures.

If your budgets come with strict ROAS or CPA targets, keep away for now. Let others run the tests, then work with an agency that has been doing exactly that, so when you do launch, you can set clearly defined targets backed by real experience rather than guesswork.

Will ChatGPT Ads be a core channel 12 months from now?

Whether ChatGPT Ads becomes a core channel or stays niche depends on the trajectory of AI itself. We're right in the thick of it. People are writing shopping lists in ChatGPT before they'll reach for a pen and paper, and user uptake has been enormous. The question is whether that level of interaction increases, plateaus or falls away over the next year. Ultimately, ads only make sense where the users are.

Either way, we'd rather have a year's experience under our belt when the answer arrives than be starting from scratch. That's why we test new platforms on our own money first, and it's why, when ChatGPT Ads is ready for clients, we'll be ready too.