OpenAI began testing ads in ChatGPT in February 2026 and opened its self-serve Ads Manager in May, so by autumn advertisers have had the better part of two quarters of live data. What they still do not have is a benchmark. OpenAI publishes no cross-advertiser figures — its own position is that a single average CTR or CPA cannot represent the channel — and the third-party numbers floating around are small, early, and contradictory. Reported click-through rates cluster around 0.6% to 1.3%, well below Google Search and Meta; cost per click has ranged from under $2 to around $18 depending on vertical; and one 15-day real-spend test posted a 2.35% conversion rate and a 1.49x blended ROAS with daily swings from 0.2x to 2.9x. Meanwhile the channel itself is real: ChatGPT ads crossed a $1 billion annualized run rate in late August 2026, fewer than 200 days after launch, with tens of thousands of advertisers. Two problems muddy every number — dashboard-reported clicks that analytics tools cannot verify, and a meaningful share of placements landing off-topic. This guide reads the available data honestly, explains why no public benchmark exists yet, defines what "effective" actually means on a surface you cannot copy a playbook onto, and shows why performance here is decided by how fast you can build your own baseline.
If you have been trying to answer "is my ChatGPT ad performing well," you have run into the real problem: there is nothing to compare it to. OpenAI does not publish cross-advertiser performance benchmarks. Its own stated position is that it does not yet have benchmarks across advertisers, industries, objectives, or campaign types, and that a single published claim — an average CTR, an average CPA — cannot represent a channel this new and this varied. That is an unusually candid thing for an ad platform to say, and it is the honest frame for everything below. Six months into a live ad product, advertisers still lack the one thing every mature channel gives them: a number that tells them whether they are winning.
This guide is the performance companion to two others. If you want the mechanics of running a campaign, see the guide on how to advertise in ChatGPT; if you want the strategic read on where the ad formats are heading, see the guide on image and video ads in ChatGPT. Here the question is narrower and harder: given that no authoritative benchmark exists, what does the early data actually show, and how should you decide what "effective" means for your own account?
The vacuum of official data has been filled by independent tests, agency case studies, and third-party estimates. They are worth reading — but read them the way OpenAI suggests, as directional signals from small and early samples, not as targets. Almost every credible source that publishes a figure attaches the same caveat: this is one account, over one window, in the channel's first months, when early-mover economics do not hold forever. With that stated, here is the shape of what has been reported.
The most-cited number is CTR, and it looks alarming at first. Third-party estimates from the channel's first months cluster around 0.6% to 1.3% — one independent analysis pegged a single brand at roughly 0.9%, others report an overall figure closer to 0.68%. That is well below the click-through rates advertisers are used to on Google Search and Meta. But a low CTR on this surface is not automatically a failure. The ad sits under an answer to a question someone deliberately asked, so the click that does happen carries more intent than a feed impression. The mistake is optimizing CTR as if it were the goal; on a high-intent, low-volume surface, a smaller number of better clicks can outperform a larger number of idle ones.
Cost has been just as variable. Reported CPCs have ranged from under $2 in one documented real-spend test to around $18 in other agency-reported estimates for software and finance verticals, well above the roughly $3 to $6 cited for ecommerce — those verticals running higher than ecommerce, exactly as they do elsewhere. The Ads Manager supports three objectives: CPM for reach (a launch-era bid around $60 per thousand impressions, with real clearing prices reported lower as inventory has scaled), CPC for clicks (OpenAI's own recommended starting bid is $3 to $5), and a conversion objective that charges per valid click rather than per conversion. Because clearing prices are set by an auction on a thin, fast-moving surface, treat any published CPC as a starting sanity-check, not a rate you can plan a budget around.
The more encouraging signal is downstream. Because the ad answers a stated question, conversion rates tend to run higher than social advertising — when the click lands, it lands warm. But the samples are small and the day-to-day swing is large. One documented 15-day test at roughly $60,000 of real spend posted a 2.35% conversion rate and a 1.49x blended return on ad spend, with daily ROAS swinging between 0.2x and 2.9x. That volatility is the real lesson, more than the averages: a single day tells you almost nothing here, and any figure like 1.49x is one advertiser's window, not a rate to expect.
It would be easy to read the messy numbers as "the channel does not work." The scale says otherwise. ChatGPT ads crossed a $1 billion annualized revenue run rate in late August 2026, fewer than 200 days after launch, with tens of thousands of advertisers already buying. A surface does not reach that run rate on advertisers who see nothing back. What the gap between "growing fast" and "no benchmark" tells you is that the channel is genuinely working for some advertisers and genuinely not for others, and the difference between the two groups is not yet legible from the outside — which is precisely why a borrowed benchmark would mislead you.
Before you trust any performance figure — yours or a third party's — account for the two issues that currently distort the data at the source.
The most documented complaint is a measurement gap. Multiple advertisers have reported discrepancies between the click counts in OpenAI's reporting and the sessions their own analytics tools record — clicks that the dashboard shows but Google Analytics never registers. Some of this is the normal difficulty of cross-domain attribution inside an AI assistant, some is reporting that is still maturing, and some clicks simply never resolve into a trackable session. The practical response is to stop trusting raw dashboard click counts as truth and lean on server-side measurement instead: OpenAI's Conversions API and pixel, UTM tagging, and — above all — downstream conversions you can verify in your own system.
The second problem is match quality. One study of over 50,000 prompts found that roughly one in seven ads was off-topic relative to the user's query, rising sharply in categories like news and politics, and that in the large majority of placements the advertised brand went unmentioned in the answer text above it. That matters for performance two ways: an off-topic placement wastes an impression against the wrong intent, and an ad running beneath an answer that does not mention your brand is working harder than one that does. The relevance of the placement, which you steer but do not fully control, is a hidden variable inside every CTR and conversion figure you see.
Put the missing benchmark, the volatility, and the measurement problems together and a different definition of "good" emerges — one that does not depend on a number you can borrow. Four principles hold up.
First, your benchmark is your own account, on a rolling window. Since daily results swing wildly and no external baseline exists, effectiveness is a trend you establish for yourself over weeks, not a target you hit against the field. Second, optimize on downstream conversions, not clicks. Given the attribution gap and the low, intent-heavy CTR, the click count is the least trustworthy and least meaningful number in the report; the verified conversion is the one that survives all the noise. Third, budget to get a read. Because you are generating your own baseline from scratch, plan for enough spend — practitioners suggest roughly $5,000 to $10,000 — to reach statistical honesty rather than reacting to a bad Tuesday. Fourth, test enough creative to actually learn. On a surface with no shared playbook, the account that runs the most on-brand variants against the same intent is the one that finds what works before the budget is gone.
That last point is where performance is actually won right now, and it is easy to miss because it is not an ad-account setting. In a channel with no benchmark, thin data, and volatile days, the advertiser who establishes a reliable baseline first has a decisive edge — they know what "good" looks like for their offer while everyone else is still guessing. Building that baseline is a testing problem, and testing at speed is a content-supply problem. You cannot find your winning creative-and-intent combinations if you can only afford to make one image and one headline; you need a set to run against each intent, refresh the losers, and scale the winners, week after week, while the auction is still cheap.
The same supply drives the two levers that quietly move your numbers most. The destination the click lands on decides whether a low, expensive click still converts — a landing page written for the exact question the ad answered continues the conversation, while a generic homepage wastes the intent. And the organic answer above your ad is a second, free placement: ChatGPT already names brands in its responses, and being one of them lowers your effective cost by converting a click that a cold brand would waste. Earning that mention is generative engine optimization, and it runs on the same fuel as a search ranking — a broad, on-brand content footprint the model can cite, covered in the guide on AI search visibility. Creative velocity, intent-matched destinations, and a citable footprint are the three things that turn an un-benchmarkable channel into a measurable one for you specifically — and all three are supply, not settings.
Be precise about the boundary first. Kompozy does not bid in the auction, does not run OpenAI's Ads Manager, and does not fix OpenAI's attribution — you buy, target, and measure inside ChatGPT's own platform. What Kompozy does is remove the one constraint that keeps most advertisers from ever building a trustworthy baseline: the rate at which they can produce on-brand creative and content to test. It is an AI content generation and multi-platform publishing engine, and in a benchmark-less channel that rewards iteration, generation speed is the performance lever.
For the test velocity, point Kompozy at a single source and it produces a whole set of variants to run against the same intent instead of a lone asset: Photo Posts, Persona Photos, Infographic Photos, and Quote Graphics for the native unit today, and Persona Shorts, Marketing Shorts, and Clipped Shorts as richer formats open up. Every variant is governed by one Persona Brief and a face-locked persona pool, so a dozen creatives stay recognizably the same brand while you kill losers and scale winners — which is exactly the cadence a no-benchmark auction demands. For the destination, Kompozy generates Blog Articles and text landing content tuned to the specific question a campaign targets, so the click lands somewhere that matches the intent that earned it rather than a generic page that squanders it.
For the free placement above the ad, the same engine builds the citation footprint: Text Posts, Carousels, Blog Articles, and Email Newsletters fanned across eight social platforms plus blog and email from one scheduling queue, with autopilot and a per-post review gate so the volume never ships off-brand. That breadth — and the consistency of it — is what an AI model draws on when it decides which brands to name, and keeping that identity uniform across every surface is its own discipline, covered in the guide on AI search brand consistency. The result is that one content operation feeds all three levers at once: the variants you test in the paid slot, the pages your clicks convert on, and the footprint that gets you cited above the ad — which is how you build a reliable baseline while the channel is still cheap and everyone else is still guessing.
Six months in, the reason advertisers still lack clear benchmarks for ChatGPT ads is not an oversight — it is the honest state of a channel too new and too varied for a single number to describe, which OpenAI itself acknowledges. The early data shows low click-through rates, wide cost swings, promising but volatile conversion, real attribution gaps, and a channel nonetheless growing fast enough to cross a billion-dollar run rate in under 200 days. The takeaway is not "wait for a benchmark." It is that "effective" is something you define for your own account by iterating faster than the uncertainty — running enough on-brand creative, against sharp intent, into destinations built to convert, to establish a baseline before the auction stops being cheap. The advertisers who win the early ChatGPT ad economy will be the ones who could produce that supply, not the ones who found the perfect benchmark.
There is no official figure, and OpenAI does not publish one. Third-party estimates from the channel's first months cluster around 0.6% to 1.3%, which is well below the click-through rates advertisers are used to on Google Search and Meta. Read those numbers as directional, not as a target — they come from small, early samples, and a low CTR on a high-intent surface is not automatically bad if the clicks it does earn convert. Judge your own account against its own rolling trend rather than a borrowed benchmark.
No. OpenAI's stated position is that it does not yet have performance benchmarks across advertisers, industries, objectives, or campaign types, and that a single published claim such as an average CTR or CPA cannot represent the channel. That is why advertisers six months in still feel like they are flying blind — there is no authoritative baseline to plan against. The Ads Manager reports your own impressions, clicks, CTR, average CPC and CPM, spend, and conversions, but nothing about how you compare to anyone else.
It varies too much to state a single number honestly. Conversion rates tend to run higher than social because the ad answers a question the user just asked, often one with buying intent — but the samples are small and volatile. One documented 15-day, roughly $60,000 real-spend test posted a 2.35% conversion rate and a 1.49x blended ROAS, with daily ROAS swinging from 0.2x to 2.9x. Treat any figure like that as one account over one window, not a rate you should expect to hit.
This is one of the most common complaints about the channel. Multiple advertisers have documented discrepancies between the click counts in OpenAI's reporting and the sessions their analytics tools record. The reporting is still maturing, cross-domain attribution inside an AI assistant is genuinely hard, and some clicks never resolve into a trackable session. The practical response is to lean on server-side measurement — OpenAI's Conversions API and pixel with UTM tagging — and to trust downstream conversions over dashboard click counts.
Because there is no benchmark to borrow, you have to generate your own read, and that takes both spend and creative volume. Practitioners running the channel suggest budgeting roughly $5,000 to $10,000 and judging results on rolling windows rather than single days, since daily performance is volatile. Just as important is testing enough creative variants against the same intent to find what works before the budget runs out — the account that iterates fastest builds a usable baseline first.
Six months in, there is no official ChatGPT ads benchmark — OpenAI publishes no cross-advertiser figures, and third-party samples are small and early. Reported click-through rates cluster around 0.6% to 1.3%, below Google and Meta, while conversion rates run higher because the ad answers a question with buying intent. What "effective" means is best defined per account: budget enough to get a read, judge on rolling windows, and optimize on downstream conversions, not clicks.
Get started → · ← All guides · Compare Kompozy vs other tools