I run marketing at GravityKit, a B2B WordPress plugin company. When OpenAI opened its ChatGPT ads beta to advertisers, we ran a proper test: three weeks in late July and August, a capped budget, two ad groups pointed at our best landing pages, and full conversion tracking wired up before launch.
Clicks were cheap and plentiful, engagement was near zero, and we measured no conversions or key events at all. We paused the campaign early.
Almost everything written about ChatGPT ads so far is agency content marketing or secondhand news coverage, so this post is the account I wanted to read before we spent the money: how the platform works as of August 2026, what it costs, the results we saw, the rough edges you only find by spending, and who should and shouldn’t test it.
How ChatGPT ads work
The structure is familiar from Google or Meta: campaign, ad groups, ads. You pick an objective per campaign (Views billed on CPM, Clicks billed on CPC, or, more recently, conversion-optimized campaigns), set a daily budget and max bid, choose countries, and submit ads for review. OpenAI’s advertiser documentation describes a relevance-weighted, second-price auction. Ads render as clearly labeled sponsored units beneath the assistant’s answer.
Targeting is the new part. There are no keywords. Each ad group gets a single context hint: one full sentence describing the conversations you want to appear in, like “a user is asking how to display form submissions on their WordPress website.” ChatGPT matches your ads to conversations using that hint, your ad copy, and your landing page. Measuring anything beyond clicks means installing OpenAI’s measurement pixel on your site first.
Only free-tier users see the ads. OpenAI’s own FAQ is explicit: ads may appear for Free and Go plans, while Plus, Pro, Business, Enterprise, and Edu accounts never see them. Keep that in mind whenever someone quotes ChatGPT’s total user numbers at you.
The platform is also moving fast. Features have been shipping monthly since the beta opened in early 2026, so some of the limitations below may be fixed by the time you read this. I’ve dated everything for that reason.
There’s also an Advertiser API covering campaigns, ad groups, ads, and reporting. We ended up needing it for reasons we didn’t expect (more below), but it has quirks: historical per-ad numbers can restate between pulls, insights requests silently cap at 20 daily buckets (ask for a longer window and your earliest days just vanish from the response), and you’ll want to script against it with curl, since the default clients of some HTTP libraries get blocked at the edge.
What ChatGPT ads cost
Our setup: one campaign, Clicks objective, $25 a day (currently the minimum daily budget), a fixed $3.00 max CPC (the bottom of OpenAI’s recommended $3 to $5 starting range), targeting the US, UK, Canada, Australia, and New Zealand. Two ad groups with three ads each, one context hint per group, pointing at two different use-case landing pages.
Over three weeks we paid about $550 for roughly 12,800 impressions, a 1.7 percent CTR, and an average CPC of $2.57. That CTR is healthy for a placement that sits underneath an answer the user already got.
On pricing, expect fixed bids to pin near your max while budget delivers: for most of the test our effective CPC hugged the $3.00 cap before drifting down, and under the Clicks objective there’s no smart bidding to do the work for you. The floor is low, though. At the $25 a day minimum, a real test costs a few hundred dollars, not a few thousand, which makes it cheap to buy your own data instead of borrowing someone else’s conclusions, including ours.
The results
Front-end metrics were fine, and the back end was zero across the board. Roughly 150 tracked users from paid ChatGPT traffic produced no product page views, no checkout starts, and no purchases. Average engagement for most ads sat far below what the same landing pages see from search traffic. We killed the campaign on day 20 of a planned 34.
The same window (July 28 to August 17) in GA4, alongside our other channels:
| Traffic source | Avg engaged time per session | Pages per session | Key events per 100 sessions |
|---|---|---|---|
| ChatGPT ads (paid) | 5 s | 1 | 0 |
| ChatGPT organic referrals | 40 s | 3.5 | 13 |
| Google Ads (search) | 65 s | 4.8 | 7 |
| Organic search | 51 s | 3.6 | 5 |
The tracking wasn’t broken. During the test, one organic (unpaid) ChatGPT referral converted into a purchase, flowing cleanly through the same measurement pipeline. The paid traffic just didn’t have intent.
Our results rhyme with the few other first-person accounts out there. SE Ranking ran a bigger test and reported plenty of clicks but very few signups, and Omni Lab’s write-up on ChatGPT ads for B2B found the same structural problems we did, including a poll where 74 percent of their SaaS-marketer audience was on a paid, ad-free ChatGPT plan.
You also can’t control devices. The only targeting lever the platform offers is geography (country, region, or DMA), so campaigns serve wherever ChatGPT runs, mobile included, with no way to turn it off. About 70 percent of our paid sessions came from mobile. My hunch going in was that the cheap clicks were mostly mobile fat-fingers, and the mis-tap signature is real (lots of three-to-seven-second sessions), but desktop sessions were just as shallow. Intent was the problem, not loose thumbs.
The targeting black box
Conversation-level reporting doesn’t exist. The platform has no equivalent of a search terms report, so you can’t check whether your ad surfaced in a relevant conversation or something loosely adjacent. You write your hint, the system matches, and you read aggregate numbers. You’re trusting the targeting entirely.
The closest we got to visibility came from a bug. Early in the test, one ad group was soaking up 89 percent of impressions while its twin starved. The UI had split a comma-containing context hint into one-word fragments, so an ad group was effectively targeting the word “show.” We caught it by pulling the ad group config through the API, since the reporting surfaced nothing. The bug came back once after we fixed it, so we ended up re-verifying hints via API after every UI edit.
That incident proved hints steer delivery, but we only learned it because something broke. In normal operation you have no idea what contexts you’re buying, and there’s no negative targeting to exclude the ones you’d never want.
Other rough edges
All as of August 2026:
- Brand-safety review produces false positives. Our best-performing ad got pulled into review mid-flight for a “landing page mismatch” and went dark for two days, even though the landing page contained the ad’s exact phrases. Re-saving the ad got it re-approved the same day. One gotcha: ads in review still show status “active” in the API; the field that gates serving is the review status.
- The conversions objective wouldn’t commit to delivering. We built a parallel conversion-optimized campaign, and the delivery estimator warned “may not deliver” at every bid we tried, so we never launched it. That may just be cold-start pessimism, but it didn’t inspire confidence.
- Measurement undercounts. GA4 captured only about two-thirds of billed clicks as sessions. Some of the loss looks like consent and webview attrition, and some looks like the accidental taps mentioned above. If you’re judging this channel, instrument beyond the click.
Is advertising on ChatGPT worth it in 2026?
For us it wasn’t, and we paused the campaign early. That’s a verdict on the channel today rather than forever: the platform is shipping fast, and the reasons it didn’t work for us are structural rather than creative.
- Our buyers mostly can’t see the ads. Technical and professional audiences skew heavily toward paid ChatGPT plans, which are ad-free.
- Free-tier conversations in a narrow B2B niche are thin, so even a modest budget struggles to find relevant inventory.
- You can’t verify context relevance or build on what works, because nothing is visible at the conversation level. Retargeting wasn’t viable for us either: audience list minimums sit far beyond what a niche product’s traffic can feed.
The channel has a better chance of working for consumer products with broad appeal, impulse-friendly price points, and audiences that use free ChatGPT. If someone asks ChatGPT for dinner ideas and you sell meal kits, the logic is sound.
The math also changes if OpenAI ships some combination of paid-tier inventory, conversation-level reporting, negative targeting, working conversion optimization, and lower retargeting thresholds. Given the current shipping pace, we’d re-evaluate in six months rather than write the channel off.
If you test it anyway
What we’d do differently, or did and would repeat:
- Install the measurement pixel and verify it’s firing correctly (fire a test event, confirm it lands in Ads Manager) before spending a dollar.
- Write context hints as single full sentences, never comma-separated lists, and verify them through the API after every edit.
- UTM-tag everything and reconcile platform clicks against your analytics from day one. Expect a gap.
- Judge the test on engagement and pipeline rather than clicks, which will look cheap without being good.
- Cap the budget and set an end date before you start.
The placement competes with the answer
The bigger lesson goes past this one platform. A paid unit under an AI answer competes with the answer itself, and the answer already satisfied the user. The durable win is being in the answer, cited and linked as a source. That’s earned through the kind of work I’ve written about in the 7 pillars of GEO, and you can measure it today by tracking AI referral traffic. Our one conversion from ChatGPT during this whole test came from exactly that: an organic recommendation rather than a paid unit.
Ads on ChatGPT will mature. The mid-funnel shift I described in AI ate the mid-funnel guarantees the ad dollars will follow the attention. But right now, if you have budget to spend on AI channels and a technical audience, I’d put it into becoming the answer instead of renting the space below it.
Have you run your own test? I’d like to compare notes, especially if your results diverge from mine.