What We Took Away From OpenAI's ChatGPT Ads Webinar

A few of us sat in on OpenAI's ChatGPT Ads webinar last week, and it was more substantial than most vendor sessions tend to be, although the thing we came away arguing about hasn't appeared in any of the coverage since. It was a single slide roughly halfway through, and it changed how we think about the channel more than any of the product demos did.

The slide nobody is talking about

OpenAI laid out the customer journey as six steps, running from discover through research, compare, evaluate and decide to purchase, which is standard enough and not in itself worth writing about. What caught our attention was that only four of those steps were marked as places where brands get to contribute. The highlighted section covered research, compare, evaluate and decide, leaving out discovery at one end and the purchase itself at the other.

We think that's a more honest description of this channel than most of the write-ups have managed, and it points at something more useful than a new place to spend media budget.

They illustrated the point with somebody shopping for hiking boots, which works well precisely because it's so ordinary that everyone has done a version of it. The questions ran roughly as what kind of boot do I need, which ones work for a beginner, is cushioning or stability more important, will these survive long hikes, and is this particular pair worth the price. That's five questions inside one conversation, and what has actually changed isn't the questions themselves, since people have always asked those, but the fact that they used to get asked in five different places and now they get asked in one.

Why the middle of the journey matters more than it used to

Those four steps are where consideration happens, and until fairly recently that process was scattered across the whole internet: review sites, forums, a thread from 2019 that somehow still outranks everything, three video comparisons of wildly varying quality, the manufacturer's spec page, a retailer's filter menu, and a dozen tabs that nobody ever closes.

Brands never really controlled any of that. What you could do was turn up in enough of those places, described consistently enough, that the overall picture tended to come out in your favour, which was slow and indirect and close to impossible to attribute properly. It was also quite forgiving, and that's the part we think gets forgotten. If one source described your product wrongly, several others would contradict it, and the buyer sorted out the difference themselves without especially noticing they were doing it.

So you never controlled that stage, but mistakes in any one place didn't cost you much, because there were always other sources correcting them. Move the whole process into a single conversation and those other sources are no longer present. Whatever the assistant can find out about your product stops being one input among many and becomes the only account the buyer sees.

We'd stop short of calling that a crisis, and it isn't a reason to be nervous about the channel, but it does change where the risk sits and most media plans haven't caught up with that yet.

The numbers worth writing down

The session opened with survey data rather than product detail, which told us something about who OpenAI thinks still needs convincing. Across fifteen markets, 55% of people say they use generative AI platforms, and the age breakdown is the part we'd pay attention to, running at 74% among 18 to 28 year olds and 66% among 29 to 44 year olds before falling away to 48% for 45 to 60 and 28% above 61.

That reads less like a description of today than a forecast, since the people who will be making the largest household and business purchases over the next decade are already doing their research this way.

OpenAI also showed where commercial activity sits inside ChatGPT, and product discovery came out as the largest single category at 31% against 13% for purchase, which means most of what happens in there isn't buying at all but people working something out.

What's actually available today

It's worth separating the vision from the live product, because the coverage tends to blur them together.

Targeting currently covers custom audiences built from your own customer or prospect lists, geographic targeting at country level with states, metro areas and postcodes available in the US, and platform targeting so you can split iOS, Android and web. Reporting in the Ads Manager beta gives you impressions, clicks, spend, click-through rate, average cost per click, average CPM and conversions, and there's a pixel and a conversions API for tracking alongside an advertiser API if you want to manage campaigns programmatically.

Anyone who has run search or social campaigns will find all of that familiar. What's new is where the ad appears, not how you buy it.

The part we're still unsure about

There's one complication we haven't resolved, and it matters most if you're planning UK or European spend.

Most of the commentary so far has assumed ChatGPT Ads simply match whatever conversation is currently on screen, which would make this contextual advertising, where the ad is chosen to suit the content. The alternative is behavioural targeting, where the ad is chosen to suit the person based on what they've done previously. The session muddied that distinction for us, because there was a slide showing a single user with three unrelated things on the go at once, and the example ad served was for a minivan appearing inside a conversation about where to stay near Lake Tahoe. It was labelled as suggested based on interests and positioned as high personal relevance rather than high contextual relevance, which is a meaningfully different claim.

The targeting controls point the same way, since custom audiences are built from uploaded customer lists rather than conversational context, and OpenAI's own description of its signals mentions broader ones once ad personalisation is switched on.

Matching the conversation somebody is currently in and tracking their interests across separate conversations aren't the same thing, and in the UK and the EEA they don't sit under the same legal basis, so we don't know yet how this resolves in practice. What we'd suggest is watching which disclosure line turns up in your market, because the two shown during that session weren't describing the same product.

Where it's heading, and what we'd do now

The roadmap moves consistently in one direction, away from an ad that sits beside an answer and towards one that takes part in the task itself, with automatic campaign creation and sponsored agents both named as what comes next. As a format that sounds like a minor addition, but structurally we don't think it is. An agent doesn't show somebody an ad and leave them to decide. It compares the options, narrows them and recommends one, inside a task the user has asked it to complete. At that point you aren't writing the words any more, and what you're really doing is making sure there's enough accurate information about your product available for a system that reads text to select it.

The early advertiser results shown in the session looked encouraging, though they should be read as early, and the question we find more useful isn't whether a new channel can manage a decent click-through rate in its first six months but what happens once the assistant stops answering questions and starts making the choice.

Which brings us to the practical part. The instinct with any new channel is to work out the bid, and we think that's the wrong place to start, because the better question is whether the assistant doing the comparing has enough to work with. Are your specifications actually complete or only roughly right, do your claims say the same thing on the product page, the datasheet, the partner listing and the retailer's description or have they drifted apart over the years, do the comparisons buyers ask for exist anywhere in a usable form, and has anyone credible written about the product in language that survives being summarised?

None of that is a media buy, but all of it decides how the media buy performs, and it takes considerably longer to sort out than this channel will take to arrive, which is awkward given that the shortlist gets built whether or not you're ready for it.