ChatGPT Ads: what we learned in the first £10,000

Every new advertising channel goes through the same three phases. First, nobody can buy it. Then a small number of people can buy it and nobody knows what it is worth. Then the benchmarks arrive, the agencies standardise, and it becomes a line item like any other.

Paid placement inside AI assistants is squarely in phase two. There is inventory, there are buyers, and there is almost no reliable public information about what performance looks like. That combination is dangerous, because it means the first campaign a company runs will define its opinion of the whole channel — and the first campaign is usually structured badly.

This is an account of how we think about the first meaningful tranche of spend: what to buy, what to expect, and — the part almost everyone gets wrong — how to know whether it worked.

A note on numbers: we deliberately do not publish invented performance figures. Where this piece refers to results, it refers to the shape of the results and how to read them, not to a fabricated case study. Anything we quote as a client outcome will be quoted with permission and attribution or not at all.

Why the first campaign usually fails

The most common structure we see is a straight port: someone takes the search campaign that works, duplicates the ad groups, rewrites the copy slightly, and pushes it into the new surface. Spend goes out, some clicks come back, the cost per acquisition looks worse than search, and the channel gets written off within six weeks.

The reasoning error is treating the assistant as a search engine with a different skin. It is not. The context in which the placement appears is fundamentally different, and that difference changes both what should be said and what should be expected.

• In search, the user has already compressed their need into a query and is scanning a page of options. They are in selection mode.
• In an assistant, the user is mid-conversation. They are describing a situation, being asked clarifying questions, and receiving a considered answer. They are in orientation mode.
• Someone in orientation mode is further from a transaction and much more responsive to being educated than to being sold to.

You are not interrupting a search. You are appearing inside somebody's thinking. Copy that works in the first situation reads as an intrusion in the second.

Structuring the first tranche

We treat an initial budget as a research budget with a commercial upside, not as a performance campaign with a research side-effect. That framing changes how it is split.

Roughly, we want the money to answer four questions in order: does this surface reach our buyer at all; which conversational contexts produce engaged traffic; what does that traffic do once it lands; and is any of it incremental to what we would have got anyway. Each question needs enough spend behind it to produce a signal, which is precisely why splitting a small budget across twenty variations produces nothing but noise.

In practice that means far fewer, far broader groupings than a mature search account would use. Concentration beats granularity when you have no priors. You can slice later, once you know which slices exist.

What the traffic actually behaves like

The consistent pattern — and the thing that most changes how you should judge the channel — is that assistant-sourced visitors arrive better informed and less ready.

Better informed, because they have just had a conversation that explained the category to them. They land knowing the vocabulary, the trade-offs and roughly what they are looking for. On-site behaviour reflects it: they skip the explainer content and go to specifics.

Less ready, because that conversation was often the beginning of the process rather than the end. They are orienting. Judging them against a bottom-of-funnel search cohort — someone who typed your brand name plus "pricing" — will make the channel look bad for a reason that has nothing to do with the channel.

The correct comparison is not brand search. It is the paid social or display spend you use to reach people who do not yet know they need you. Against that benchmark the picture usually looks quite different.

The measurement problem is the real problem

Here is the part nobody wants to hear: a significant share of the value this channel creates will not be attributable to it.

Conversations happen inside an application. Someone hears about you in an assistant, does nothing, and searches your brand name two days later on a different device. That conversion lands in organic or direct. Your paid assistant spend gets no credit and looks inefficient. Meanwhile your brand search volume — which you are probably not looking at closely — goes up.

If your only measure is last-click, this channel will always look worse than it is, and you will always under-invest in it for reasons that are an artefact of your analytics.

What we do instead, in rough order of usefulness:

• Watch branded search volume and direct sessions as a leading indicator alongside the campaign, not separately from it.
• Ask on the enquiry form. A single optional free-text field — how did you first hear about us — is unfashionable and unreasonably informative. Assistants get named in it more often than most marketing teams expect.
• Run geographic or temporal holdouts where volume allows. Turning spend off in a region for a fixed period is blunt, but it is the only cheap way to see incrementality without a proper test infrastructure.
• Track assisted paths, not just last touch, and accept that the model will still under-count.
• Watch on-site behaviour rather than immediate conversion. Depth of visit, pricing page views and return rate tell you whether the traffic is real long before the conversion data has enough volume to be trusted.

What we would tell someone before they start

Three things, and they are all about expectations rather than tactics.

First, commit to a duration, not just a budget. A campaign killed at four weeks produces an anecdote. The learning curve here is longer than in a mature channel because there are no priors to start from, and the inventory itself is still changing underneath you.

Second, fix the destination before you buy the traffic. The most common cause of poor performance we see is not the campaign at all. It is an informed, oriented visitor landing on a page written for someone who has never heard of the category. The conversation has already done the educating; the page needs to do the deciding.

Third, do the organic work in parallel. Paid placement puts you in the conversation. Being cited organically puts you in it as a recommendation rather than an advertisement, and those two things are not received the same way. Companies that run only the paid side are renting a presence in a surface where the earned presence is worth more.

Is it worth it yet?

For most companies, at most budgets, this is not the channel that will transform the quarter. Volumes are still modest relative to search, targeting controls are still coarse, and the reporting is thin.

It is worth it if you meet at least one of three conditions: your category is one where buyers genuinely research through assistants; your competitors are visibly present in those conversations and you are not; or you can afford to buy the learning now while it is cheap and unfashionable, before the benchmarks arrive and the price follows them.

If none of those is true, the honest recommendation is to spend the money on being retrievable organically and revisit the paid side in six months. We would rather tell you that than take the retainer.