A colleague sent me a screenshot a while back of a conversation his daughter was having with ChatGPT about which light bulbs fit her car. She wasn’t killing time. She had a question and she was reading every word of what came back, the way you do when the answer actually matters to you.
That is a valuable thing to be able to buy against. An auto parts retailer showing up in that exchange would have been useful rather than intrusive.
Paid ads inside ChatGPT lands in that kind of moment. People are going through the response and giving it their full attention, which feels a lot like the early days of Google before anyone learned where to look away.
Media buyers seem to agree. Participation has grown quickly since self-serve buying opened, and click costs have come down from where the pilot started as more inventory entered the auction. Plenty of people ran the same math and decided the attention was worth paying for.
Whether that math works for you comes down to two things. The first is whether people ask specific enough questions about your category to make showing up worth anything. The second is whether your analytics would recognize a win if the campaign produced one. Most advertisers running this right now will pass the first test and fail the second.
How ChatGPT Ads Work
Plenty of marketers evaluating this channel have never actually seen one of these ads. They only serve to free and Go tier users, and rollout has been gradual, so anyone paying for a subscription or working from a company account has been looking at an ad-free product this whole time.
Here’s a short primer as of mid-2026:
- Ads appear inside conversations when they’re contextually relevant to the question someone asked. You are buying your way into an answer, not onto a page
- They serve to free and Go tier users
- Billing is cost per click
- Product feed campaigns reuse the structured files retailers already send to Google Shopping, so yes, there are images in there
- Buying runs through a self-serve Ads Manager, with a Conversions API and a pixel available for measurement
- Campaigns are live in seven markets with more in rollout. US targeting goes to state, designated market area and ZIP code, and ZIP is as granular as it gets
The closest familiar comparison is a retail media network or a programmatic buy, and it has less to do with the ad format than with what you get back. On a retail media network you’re buying access to a place where people are already shopping, and you take the platform’s word for what happened in there. Same trade here.
What People Actually Ask ChatGPT
Nobody opens ChatGPT and types “travel inspiration.” That’s a browsing behavior, and it happens somewhere visual, on Instagram or Pinterest or a magazine site.
What people type here is closer to: give me a three-day weekend guide to San Francisco, family-friendly stuff, and I want to stay somewhere that isn’t going to wreck the budget.
Buying against the first one in ChatGPT is close to setting money on fire, since it’s too high-funnel to produce anything you can see. Compare that to someone mid-decision, listing their own constraints. Showing up there is worth something.
The screening question for your own category is whether the competitive set is intense enough, or the product research-heavy enough, that people are genuinely comparing before they buy.
Tech is the obvious case, where someone is weighing specifications across three platforms and trying to figure out which differences matter. When our air conditioning went out, that entire process happened in ChatGPT: unit against unit, what holds up longer, what the specs mean once it’s on your roof in July.
Those conversations run long, and they’re full of specifics.
Trust Comes From the Creative
When Facebook started selling ads, the reaction was that this was strange and invasive and nobody was going to stand for it. People adjusted within about a year. Every new ad environment gets some version of that reception, and most of them survive it.
Which is why the trust question here has almost nothing to do with the placement. It’s about the creative and whether there’s any value in it for the person reading.
Yelp is the useful comparison.
You scroll past the sponsored block to get to what you think of as the real results, and then you notice the sponsored place is sitting there at 3.5 stars. That lesson got learned years ago, and people carry it everywhere.
“The copy that fails is the copy that talks over the question.”
Most of what’s running in ChatGPT right now is product advertising, which narrows the creative problem considerably. One I came across was a hot sauce showing up in a thread about what to cook for a carne asada dinner party. The ad worked because it belonged in the answer, and a product card with a price and a rating is a reasonable thing to encounter when you’ve asked what to buy.
The copy that fails is the copy that talks over the question. Someone doing a careful comparison is reading closely, and that same attention makes a sales pitch obvious in a way it wouldn’t be in a feed.
My guess is people will end up treating these like they treat every other ad, which is to say mostly fine, occasionally annoyed, and considerably more forgiving when the thing in front of them is useful.
The Signal That Tells You You’re Ready to Give It a Go
I hate being asked to make this call, because there’s no threshold I can hand you that would hold up across two different businesses. What I can tell you is where I look first.
Start with your own organic traffic. If people are already using ChatGPT to find you, or to research the topics around what you sell, the conversations are happening with or without you. Go check:
- AI referral sessions as a share of your total, month over month
- Whether direct traffic has crept up in a way nobody on your team can explain
- Whether “how did you hear about us” fields have started surfacing ChatGPT
Then ask yourself whether that number feels like a drop in the bucket or whether it feels like something. You’ll know which reaction you’re having. A brand seeing twenty percent of sessions from AI referrals sits somewhere different than one seeing twenty visits a month, and your appetite for testing counts as much as the count does.
The other half of the question is competitive.
A crowded category where everyone is fighting for the same comparison is exactly where showing up inside the comparison has value. With three players and no real research behavior, there’s less to gain even when the traffic is there.
None of this proves the ads will perform. It tells you the audience exists, which is the part you can verify before spending anything.
One thing to settle before launch: where the money comes from. Pulling budget out of paid search to fund this is the wrong trade right now, because you’d be moving dollars from a channel with clean attribution into one without it, then comparing them on terms this channel can’t meet. Fund it out of social, programmatic or whatever your experimental line looks like, and set your expectations to match.
“A crowded category where everyone is fighting for the same comparison is exactly where showing up inside the comparison has value.”
Who It Fits, and Who Should Wait
The businesses doing well here sell things people research first.
SaaS and technology, where someone is comparing platforms and trying to work out which feature differences matter. E-commerce with a product feed already built, and mature brands that have maxed out what Google will give them.
App advertisers joined that group recently, since AppsFlyer and Adjust now handle install and in-app event attribution. A team already paying for one of those contracts is looking at a setup measured in configuration screens.
My reservation applies to all of them, though. I want to know who I’m reaching, and here I mostly can’t.
Google gives me a person, their affinities, what they’re in market for. What I get from ChatGPT is geography, plus whatever the conversation implies. On the free tier, chat history gets deleted often enough that there’s no durable profile to build against anyway. Custom audiences exist, though you need 25,000 matched users to use one, which rules out plenty of advertisers on its own.
Who should wait:
- Regulated categories. Pharma, fintech, health, financial and legal services are blocked
- Hyperlocal and brick and mortar. ZIP is as tight as it gets. You can cover a big area with a DMA if you really want to, it’s just less efficient
- Impulse commerce under $20. People aren’t doing a lot of research on a phone case, and there are better ways to reach those buyers
- B2B selling into a buying committee. Decision makers on enterprise accounts may never see your ads
That last one stings, because B2B otherwise fits. Long consideration cycles, genuine comparison shopping, and an audience sitting behind a paid subscription where none of it reaches them.
Your Tracking Will Tell You This Failed
I won’t recommend this channel to a client running on basic UTM codes and standard GA4. The reporting will mislead you, and it’s the one thing I hold firm on here.
ChatGPT sessions have a habit of landing in analytics as direct traffic. Someone reads an answer, clicks through, converts two days later and last-click hands the credit to the address bar. Then you kill a campaign that was working.
OpenAI has closed most of this gap. There’s a conversions API, a pixel, advanced matching and conversion-optimized campaigns, and app advertisers now have AppsFlyer and Adjust. None of that helps until the pixel is on your site and the conversions API is passing your conversion events back.
So get that in place before launch, turn on self-reported attribution and match against your CRM where you can. Then you’re looking at real lift instead of guessing at it.
You still won’t get clean ROAS out of this, and the reporting is hit or miss enough that chasing it will waste your time. Directional lift is what I’d hold it to.
Running the Test
Start somewhere between $2,000 and $5,000, or roughly a tenth of whatever you’ve set aside for experiments. The old fifty thousand dollar entry point is gone, which is why this conversation is worth having at all. Where you land inside that range comes down to how much appetite you have for spending against a channel that takes a while to tell you anything.
A few things worth knowing before launch:
- Use daily budgets. Pacing has improved and now averages across a seven-day window, though delivery still moves around more than you’d like
- Assets that were approved can slip back into review without warning, so build in more lead time than the timeline suggests
- Buyers keep describing the platform as promising and clunky, which is about what I’d expect from something this new
If you’re weighing this against AI Max, the two are closer than they look.
Both require going broad and being okay with what you show up in. With AI Max you’re on broad match, so you might turn up on competitor terms, partner terms, who knows what, and it can pull any copy from your website into an ad. If you’ve got old content sitting on the site, that’s what might show up. Worth knowing it isn’t your only route into AI Overviews.
Give it enough runway to say something. Three weeks against aggregated reporting won’t tell you much.
Go look at how much of your traffic is already arriving from ChatGPT. Ask whether the questions people ask about your category are specific enough that showing up in the answer would mean anything. If both look promising, this is worth a small, well-instrumented test.
If the traffic is there and the tracking isn’t, that’s the part worth fixing first. Get in touch and we’ll take a look at what your setup would actually show you.
