Advertising inside ChatGPT is getting off the ground, and nobody has figured out yet whether it actually works.
The official numbers (suspicious)
From what I’ve read on Adweek, a leak from StackAdapt — OpenAI’s DSP partner — confirmed that ChatGPT ad inventory is currently priced at a CPM somewhere between $15 and $60.
OpenAI, for its part, keeps insisting the CPM is $60.
If memory serves, that’s roughly three times Meta’s average. More or less where Netflix Ads started.
The $1,000 experiment
With those numbers, if you ran a $1,000 test across the main platforms, here’s what you’d get in reach:
- Google Search: half a million impressions
- Meta: just over 100,000
- TikTok: around 100,000
- ChatGPT: 16,000, on a good day
Half the story: intent
And yet those numbers only tell half the story.
For one thing, on ChatGPT — much like early Google Ads — the user isn’t killing time. They’re asking specific questions because they have a problem and want it solved.
That clearly changes the value of each impression compared to platforms where users are browsing with a very different mindset. A smaller reach could still translate into a more interesting number of conversions — likely sales.
Criteo reports that users arriving from LLMs convert at roughly 1.5x the rate of other channels. That tracks: the intent is cleaner.
But everything else still doesn’t hold up.
The real problems
Low volume: the CTR measured by Adthena is under 1%. Google Search, in the same context, clears 6%.
I came across a case of a company that burned through just 3% of a $250,000 budget over several weeks — simply because there wasn’t enough inventory to spend against.
Tracking: from what I can tell, OpenAI sends you weekly CSV files. In 2026, that should be an insult. Apparently it isn’t.
Google Ads circa 2000
My read is that we’re still firmly in test-and-learn territory.
This feels more like Google Ads in 2000: high prices, few benchmarks, a first-mover advantage for whoever jumps in now — but almost nothing to decode yet. My friends Marco Loguercio and Duccio Lunari might remember what that was like.
The difference is that here everyone assumes intent is worth a premium. But without volume, there’s precious little to actually measure.
The question
I keep wondering whether it’s worth paying over the odds to run tests right now, or whether it makes more sense to wait for the technology to mature — and, above all, to figure out which sectors can actually justify paying this much for this little inventory.
What do you think?