ChatGPT Ads launched as a US test in February 2026 and is now buyable in more than fifty markets. Seven months of real spend has produced something the platform itself still does not provide: published results from named advertisers. This post reviews what they reported.
The short version: the channel works, but in a narrow band. Profitable campaigns and zero-conversion campaigns both exist, and what separates them is usually not the channel but how the campaign was built. The most consistent complaint is not performance — it is measurement. And the same campaign can cost four times as much in one market as another.
What the Platform Reports About Itself
Per OpenAI’s August 31, 2026 announcement, ChatGPT Ads passed $1 billion in annualized revenue run rate in under 200 days. Tens of thousands of advertisers use it, the ecosystem includes more than 50 technology and measurement partners, and CPC plus outcome-optimized bidding now account for the majority of campaigns.
Those figures describe demand, not performance. The only two outcome examples in that announcement are anonymous: “an ecommerce advertiser” achieving 3x ROAS over 28 days, and “a technology partner” reporting that more than 80% of ad-driven traffic came from new customers. No brand, no vertical, no budget.
OpenAI does not hide the gap. Its help center states that there are no performance benchmarks across advertisers, industries or campaign types. There is no official yardstick for whether a CPC is good. Everything below therefore comes from what advertisers published themselves.
The Published Numbers
Search Engine Journal’s September 8, 2026 roundup collected the tests with meaningful spend behind them; MediaPost’s August 28 report added the advertisers who got nothing. Together:
| Advertiser | Spend | CPC | CTR | Outcome |
|---|---|---|---|---|
| Hostinger (hosting) | ~$70,000 | Not above Google Search | The bigger worry | CPM above $65; narrow use cases worked, broad messaging did not |
| Common Thread Collective (high-AOV ecommerce) | $9,620 | $4.41 | 0.94% | 3.3x–6.8x ROAS depending on attribution model |
| Floyd Blaikie (B2B) | ~CAD 7,000 | $9.29 | 0.7% | Of 146 companies behind 336 clicks, only 5 matched ICP |
| Synter (multi-market) | $4,428 | $9.89 avg | — | 22.89 depending on market |
| Choice OMG (Canada, agency) | $415 | ~$7 | 0.6% | 60 clicks, no interested buyers |
| Cleverly (B2B lead gen) | — | — | “Decent” | Zero conversions; campaign paused |
| PerfectTablePlan (software) | £289 | — | — | 2,988 clicks → 14 installs; £20.68 per install |
A CPC spread of 22.89, CTRs between 0.6% and 0.94%, and 6.8x ROAS sitting beside zero conversions. That distribution is the finding: “is ChatGPT Ads good or bad” is the wrong question.
The Recurring Problem Is Measurement, Not Cost
Independent advertisers in different countries reporting the same defect makes it hard to dismiss as individual misconfiguration.
Nicholas Verity, CEO of B2B lead generation agency Cleverly, added UTM parameters to his campaign and found OpenAI reporting 57 clicks while Google Analytics showed under 20 sessions. His reasoning was blunt: if the platform’s own numbers do not match reality, either the metrics are unreliable or clicks from the same user are being counted twice. They stopped spending — while still arguing the channel could become a major one in time.
Others reported the same thing. We Scale Startups founder Daniel Johnson said click counts never lined up with Analytics, and not always in the same direction; until the platform’s own reporting is trustworthy, he treats the spend as a research budget rather than a channel to scale. Camino5 co-founder Ryan Edwards called it a common problem, describing a roughly $1,000 test that produced clicks but no Analytics data at all.
The operational consequence: measurement has to exist before the campaign, and it cannot rely on a single source.
Clicks Arrive, Interest Does Not
The second recurring theme is traffic that lands and does nothing.
The most detailed public record comes from Andy Brice, developer of event seating software PerfectTablePlan. He spent £289.52, received 2,988 clicks, and got 14 installs — a 0.46% conversion rate at £20.68 per install. Average time on page was 7 seconds and bounce rate 69%. He ruled out bot fraud, concluding the clicks were largely human, and attributed the result to targeting. His own comparison is the striking part: on the same product, Google Ads converted at 5.3% and ChatGPT’s free organic referrals at 4.2%, while ChatGPT’s paid clicks managed 0.46%. Organic traffic from the same assistant worked roughly nine times harder than paid.
Peter Jaffray, managing director at Canadian agency Choice OMG, spent $415 across three campaigns targeting Ontario and Alberta. He received more than twice the impressions he expected, but CTR stayed at 0.6% against the roughly 2% he sees on search. The 60 clicks came from legitimate IP addresses, yet nobody engaged with the site. His summary: it behaved more like a display channel than an engaged one.
The most instructive B2B finding came from Floyd Blaikie’s team. The $9.29 CPC visible in Ads Manager was not the real problem. Using visitor deanonymization, the team identified 146 companies behind 336 paid clicks and found only five matched the client’s ideal customer profile. Ads Manager surfaced none of that. In B2B, evaluating this channel requires a layer the platform does not provide.
Geography Changes the Price by 4x
Synter’s test is the clearest warning against importing anyone else’s cost assumptions. Across 9.89 — and the average conceals the actual picture:
| Market | CPC |
|---|---|
| United Kingdom | $5.10 |
| United States | $10.62 |
| Australia | $17.59 |
| New Zealand | $22.89 |
Australia and New Zealand had considerably less volume, so those high costs should be read alongside thin inventory. Either way, the same campaign selling the same thing paid four times more in one market than another.
This is also why OpenAI’s 5 figure cannot serve as a benchmark: it is a recommended starting maximum bid, not a platform average. Since August 2026, “Maximize results” has been the default bid strategy for eligible new ad groups, setting and adjusting bids automatically. Advertisers who want a hard ceiling must opt back into manual bidding.
What Actually Worked
The positive results follow a pattern rather than arriving at random.
Hüseyin Ograk, head of PPC at Hostinger, published an assessment that matured as spend grew. Early on he reported CPCs no higher than Google Search and purchases already coming through. After nearly 65, but CTR was the bigger concern. The core finding was that specific use cases performed while broad messaging consistently struggled. Traffic quality was inconsistent, and judging ROAS on direct conversions alone was difficult.
Common Thread Collective managed to scale a high-AOV ecommerce client from 1,000 per day inside a month. Across 4.41 CPC and 0.94% CTR, returning between 38,000 in attributed revenue depending on the model — 3.3x to 6.8x ROAS. Notably, they did not leave measurement to the platform; they evaluated downstream performance with Triple Whale alongside Ads Manager.
Three patterns emerge. Narrow messaging beats broad. High order value makes a low CTR tolerable. Measurement outside the panel is the only way to see what the channel contributed.
Prices Fell, but Inventory Was the Real Constraint
Costs moved substantially over seven months. Per Digiday in April 2026, CPMs fell from 25 within nine weeks. Jai Amin, chief of media activation at agency Jellyfish, said the 45 depending on inventory composition. One executive buying through Criteo reported a 35 band.
For context, Gupta Media puts Facebook at 7.63, Google Display at 3.02 and LinkedIn at 250,000 to $50,000, then disappeared entirely with May’s self-service launch.
The less-discussed side of that story is supply. Per Digiday’s May 2026 follow-up, one pilot advertiser managed to spend just 250,000 commitment over four weeks, with OpenAI initially able to push roughly $100 per client per week. High minimums were being asked for against inventory that did not exist yet. Fill rates later improved, and four of seven ad executives reported increased delivery over a six-week stretch.
Anyone starting now enters after that phase — self-service, lower prices, deeper inventory.
Where the Ads Actually Appear
On brand safety there is measured data. SE Ranking analyzed 50,006 commercial prompts across 20 niches, reported by MediaPost on August 11, 2026, and found 14.35% of ads carried no topical connection to the conversation. In categories such as news and politics, the disconnect rose above 50%. The same study found advertisers were cited as a source in only 3.63% of placements, against 11.53% in Google’s AI Mode.
This is not something an advertiser controls. In a system that matches on conversational context, placement decisions sit largely with the platform, and Ads Manager has no report showing which contexts you appeared in. That is precisely why Brice could not optimize his way out of the problem.
On the user side, the only available data is OpenAI’s own. Chief revenue officer Denise Dresser said at Cannes Lions in June 2026 that the rate at which users close the ad unit had dropped 50% since the February launch, reading that as improved relevance. Worth treating carefully: a falling dismissal rate can indicate habituation as much as relevance, and the figure is not independently verified.
OpenAI’s Own Reversal: Buying Did Not Move Into the Chat
The most instructive experience may be OpenAI’s own. The company launched Instant Checkout in September 2025, letting users buy inside the conversation, and billed it as the next step in AI commerce. Per CNBC on March 24, 2026, it failed to take off and OpenAI changed direction: merchants keep their own checkout experiences while the company focuses on product discovery.
Analysts told CNBC that OpenAI had underestimated how hard enabling transactions would be, struggling to onboard merchants, show accurate product data, and support multi-item carts or loyalty memberships. The current model has merchants share product feeds and promotions so their catalog is represented inside ChatGPT, with the purchase completed on the merchant’s own site.
The practical implication for media planning: the landing page is still where conversion happens. Planning around the assumption that users will stop visiting sites is, for now, contradicted by the platform’s own experience.
What to Do Differently
- Instrument before you spend, and never trust one source. UTMs, the measurement pixel and the server-side Conversions API belong in place before the first dollar. Assume a gap between panel and analytics.
- Do not import another market’s CPC. A 4x spread was measured inside one campaign. Your own test is the only valid reference for your market.
- Reset the CTR expectation. Published tests sit at 0.6% to 0.94%. What carries a low CTR here is high order value or high lifetime value.
- Narrow the message. Multiple tests reported specific use cases working where broad messaging failed.
- Look behind the click in B2B. The panel does not show the firmographics of the traffic. Click quality needs an external layer.
- Do not blame the channel for zero conversions too quickly. With no competitive reporting, query breakdown or context data, separating a genuine performance problem from delivery or matching variance is hard. Early results warrant more scrutiny than they would on a mature platform.
- Treat the landing page as a variable. In a channel reporting 7-second dwell times, page speed and message match may matter more than bid settings.
Quick Take
The most defensible conclusion from seven months of data is that this is not yet a performance channel, but it earns a test. OpenAI’s own documentation calls it beta, says it has no benchmarks and acknowledges that delivery can fluctuate. At the same time, a $1 billion revenue run rate shows the demand is real.
The advantage now belongs to advertisers entering after the fact. Most early testers spent before instrumenting, and then could not interpret what happened. Not repeating that is the most concrete gain available today.
Related News
- ChatGPT Ads Opens in Türkiye: Date, Eligibility and Minimum Budget
- What Is ChatGPT Ads and How Does It Work?
- How to Set Up Conversion Measurement on ChatGPT Ads
- New ChatGPT Ads Campaign Tools
- GA4 AI Assistant Traffic Channel
Sources
- Search Engine Journal: 6 Months Into ChatGPT Ads, Advertisers Still Don’t Know What ‘Good’ Looks Like (September 8, 2026)
- MediaPost: Data Reveals ChatGPT Ads Not Performing (August 28, 2026)
- MediaPost: ChatGPT Serves Ads In Results On Irrelevant Topics (August 11, 2026, reporting SE Ranking’s study)
- Successful Software: ChatGPT ad targeting is garbage (September 2, 2026)
- Digiday: ‘Everything is coming down’: ChatGPT ads are getting cheaper (April 17, 2026)
- Digiday: As OpenAI’s ChatGPT ad delivery improves, the doubts it created aren’t so easily fixed (May 26, 2026)
- Adweek: OpenAI Sees Fewer Users Dismissing ChatGPT Ads as It Scales (June 23, 2026)
- CNBC: OpenAI revamps shopping experience in ChatGPT after Instant Checkout (March 24, 2026)
- OpenAI: A milestone in expanding access to AI (August 31, 2026)
- OpenAI Help Center: ChatGPT Ads frequently asked questions
- OpenAI Help Center: Daily Budgets
Frequently asked questions
- Do ChatGPT Ads work?
- Published results split sharply. High-AOV ecommerce and products with specific, narrow use cases have produced profitable campaigns; tests launched with broad messaging and no conversion tracking mostly ended at zero conversions. Campaign construction appears to matter more than the channel itself.
- What is a typical ChatGPT Ads CPC?
- Published tests range from $4.41 to $22.89. In a single Synter campaign the cost was $5.10 in the UK, $10.62 in the US, $17.59 in Australia and $22.89 in New Zealand. OpenAI's $3 to $5 figure is a recommended starting bid, not an average.
- Why is CTR on ChatGPT Ads so low?
- Published tests cluster between 0.6% and 0.94%. The ad sits below the end of the response in a separate labeled box, so the user has to choose clicking over continuing to read. Expecting search-style rates of 2% to 6% is the wrong reference.
- Why don't platform-reported clicks match Google Analytics?
- This is the single most reported problem. In Cleverly's test OpenAI reported 57 clicks while Analytics showed under 20 sessions. Other advertisers reported the same mismatch, sometimes in the opposite direction. Set up UTMs and server-side measurement before spending.
- Which verticals are seeing results?
- Published wins concentrate in high-AOV ecommerce and considered-purchase products. On the B2B side, one agency deanonymized the 146 companies behind 336 paid clicks and found only five matched the ideal customer profile, so click volume alone is not a useful signal there.
- How much budget does a meaningful test need?
- Published tests range from roughly $300 to $70,000. The $415 tests produced no usable direction; campaigns that generated real signal generally spent several thousand dollars. Having measurement in place matters more than budget size.
- Do ChatGPT ads appear next to irrelevant content?
- SE Ranking analyzed 50,006 commercial prompts across 20 niches and found 14.35% of ads had no topical connection to the conversation. In categories such as news and politics, that disconnect rose above 50%.
- Has OpenAI published performance benchmarks?
- No. OpenAI states plainly that it has no performance benchmarks across advertisers, industries or campaign types. Ads Manager also has no Auction Insights or Impression Share equivalent, so there is no external way to judge whether a given CPC is high or low.