Six Months of ChatGPT Ads and Still No Benchmark Worth Trusting

ChatGPT ads benchmarks are everywhere right now, and none of them agree. A $3-5 CPC gets repeated as gospel in every deck. One ad-penetration tracker swung from 25% to 0.05% back to 51% within a few weeks. OpenAI has published no rate card at all.

Six months after ads went live, a media buyer sitting on real spend data still can't tell a client whether the number in front of them is good — and that's not a research gap you close by finding the right tracker.

There is no trustworthy ChatGPT ads benchmark because OpenAI has published none, and the third-party figures that exist disagree by orders of magnitude — one study found ads in 25.94% of 50,006 commercial prompts, while a separate tracker's ad-penetration reading swung between 0.05% and 51% in the same stretch. Buyers who need a defensible number should build one from their own account data instead of averaging conflicting external figures into a false consensus.

Six months in, chatgpt ads benchmarks still don't exist

The timeline moves fast enough that "still don't exist" needs a date stamp. OpenAI announced plans to test ads on January 16, 2026. Ads went live February 9 for logged-in US adults on the Free and ChatGPT Go ($8/month) tiers only — Plus, Pro, Business, Enterprise and Edu stay ad-free. The format is a labeled Sponsored block that renders below a completed answer, run by a separate system that can't shape or rank the chat response itself.

Self-serve Ads Manager opened to all US businesses on May 5, introducing CPC bidding in the $3-5 range and removing the managed-pilot's spend minimum (reported as both $50K and $200K in different outlets, never reconciled). Shopping ads followed May 12, letting retailers connect catalogs up to 1M SKUs for auto-generated product cards.

Europe got 31 markets on August 24. By late August, ChatGPT Ads had reportedly reached a $1B annualized run rate in under 200 days. Four days before this was written, on September 16, OpenAI launched Sponsored Agents: post-click conversational agents piloted with select advertisers including HubSpot and Shopify integrations. No aftermarket tracker claims to cover that yet.

Against that pace, Search Engine Journal reported on September 8 that advertisers six months in still don't know what good looks like. When we ran ChatGPT ads ourselves for two weeks, the number we'd have handed a client depended entirely on which week we asked.

That's the honest starting point for anyone chasing a chatgpt ads benchmark: the platform is still being built while you're trying to measure it. The pace itself is already reshaping how attribution and media buying get planned around this channel, well before any stable number exists to plan against.

What OpenAI actually publishes about its ad platform

OpenAI publishes its ad policy — what's disallowed (adult and dating, gambling, alcohol, tobacco, illegal drugs, crypto, credit-repair and debt-settlement), plus an outright ban on political ads and no ads near sensitive health topics or to under-18 accounts. It also documents, at a policy level, that the ad system runs separately from the chat model, that memory can inform targeting only if a user has memory enabled, and that advertisers never receive chat content or history.

What it doesn't publish is anything you'd need to answer "is this a good CPC": no rate card, no performance benchmark, no category-level average, no case studies with real numbers attached. That absence isn't an oversight.

It's the same posture Perplexity landed on from the other direction — Perplexity launched Sponsored Questions in late 2024, stopped onboarding new advertisers in October 2025, and fully discontinued ads in February 2026, pivoting to a subscription-only, ad-free positioning. Google took the opposite architecture entirely: its AI Mode and AI Overviews formats inject ads into the answer itself, tied to Performance Max and Shopping, rather than separating them into a labeled block.

Two platforms, two philosophies, and neither one gives a buyer a benchmark to hold onto. Placement architecture explains more of that gap than targeting does, a split we broke down when comparing ChatGPT ads against Meta's feed interrupt.

Where the reported $3-5 CPC number actually comes from

Every deck citing "$3-5 CPC" is citing the same origin: Axios and Digiday reporting on the May 5 self-serve launch, not a figure OpenAI confirmed as a rate card. That distinction matters every time the number reappears, because it keeps getting cited as if OpenAI stated it.

OpenAI didn't. The CPC range is a reported observation of early auction behavior, and auction behavior six months into a cold-start ad market moves fast — advertiser count, format mix, and geo footprint have all changed since May.

The CPM figures floating around are on worse footing. $25-65 CPM shows up on agency blogs with no sourcing chain back to anything OpenAI said or any study that measured it. Treat any chatgpt ads benchmark that quotes a CPM range as a claim to verify, not a number to plan a budget around.

The same pattern shows up in every chatgpt ads CPC and chatgpt ads CPM figure floating around right now. Compare what LinkedIn charges per click, what Meta charges, and why CPM keeps rising everywhere else — platforms with years of published auction data still produce ranges that vary by vertical and season. ChatGPT has neither the years nor the published data, so a single point estimate is doing more work than the evidence supports.

Why ad-penetration trackers disagree by 50 points

"What percentage of ChatGPT replies carry an ad" sounds like a fact you could just go look up. It isn't, yet. SE Ranking's study of 50,006 commercial prompts found a paid placement in 25.94% of them.

Similarweb reported roughly 26% on August 17. A separate tracker's reading on the same underlying question swung from 25.5% to 0.05% and back up to 51% within a matter of weeks. Same platform, same rough time window, numbers that don't even round to the same order of magnitude.

The loudest chatgpt ads benchmarks in circulation right now are ad-penetration percentages, and they're also the ones that disagree hardest. The disagreement is methodological, not mysterious. Every tracker samples with a synthetic prompt panel or, in Similarweb's case, a real consumer panel, and none of them discloses panel size or geo skew.

A synthetic-prompt panel structurally misses personalization variance and mid-conversation ads, because it fires a fixed prompt list and reads whatever comes back once. Ads are targeted per account and per session. Two prompts that look identical to a tracker can trigger completely different ad logic depending on who's asking. This table lays out what each aftermarket tracker actually measures, since that's the piece missing from most coverage that just quotes the headline percentage:

None of these trackers is lying, exactly. They're measuring different populations at different moments with different instruments and reporting the output as if it were one number. That's the actual source of the 50-point spread.

What the SE Ranking study proves and what it doesn't

The SE Ranking study is the most rigorous public data point on ChatGPT ads behavior, and it's worth being precise about what it establishes. Across 50,006 commercial prompts, 25.94% showed a paid placement.

Of the ads served, 14.35% were semantically unrelated to the prompt that triggered them — a meaningful gap between what someone asked and what got advertised next to it. The advertiser's own brand was cited organically above its own paid ad in only 3.63% of placements, and the advertised URL turned up in the model's citations just 0.09% of the time.

What that proves: paid and organic function as largely separate layers inside ChatGPT, with almost no crossover between what the model cites and what gets advertised. What it doesn't prove: a stable market-wide ad-penetration rate, a benchmark CTR, or anything about attribution once a user clicks through.

It's one study's finding on one panel at one point in time, not a settled rate you can plug into a forecast. Competitive research on this platform behaves closer to tracking share of voice in paid social than to reading a stable Nielsen number — the panel and the methodology are the whole story, and different panels will keep producing different numbers until one becomes the industry default.

The measurement fields Ads Manager won't give you

This is where "just benchmark it yourself against the market" runs into a wall Meta and Google buyers don't hit. OpenAI's Ads Manager withholds the triggering prompt from advertisers as a stated privacy decision. It gives no demographic breakdown, no reach or frequency reporting, no log-level export, and no third-party verification. Per Improvado's and Taggrs' breakdowns of the platform, it also gives no incrementality measurement of any kind.

That last gap is the one that actually blocks normal benchmarking. Every mature ad platform's benchmark culture rests on being able to compare a campaign against a holdout or a lift test, not just against last quarter's average.

Without reach, frequency, or incrementality, a buyer can't separate "this CPC is good" from "this CPC is good relative to what would have happened anyway" — and that second question is the one a CFO actually cares about. None of the chatgpt ads benchmarks passed around right now account for that gap. A benchmark built on impressions and clicks alone, with no incrementality layer under it, tells you less than it looks like it tells you.

How to build chatgpt ads benchmarks you can actually trust

Given all of that, the workable move isn't finding a better tracker. It's the same discipline that comes before any creative test: find the angle first, before you chase a number. Here, the angle is deciding what "good" would even mean for your account before you go looking for one. If you haven't set up the account yet, the advertising on ChatGPT guide covers that groundwork; this is about what comes after you've already spent the money.

1. Pull your own account history on a fixed cadence. Export every campaign's spend, clicks and reported CTR from Ads Manager weekly, through API access if you're centralizing it alongside data from other platforms rather than screenshotting a dashboard. For deeper programmatic pulls, wiring ChatGPT into ad research directly is worth a look.
2. Segment by format as the platform adds them. Standard Sponsored blocks and shopping/product-feed ads are different auctions with different competition. A blended average across both hides more than it shows.
3. Compute your own CPC trend instead of borrowing an average. A CPC calculator turns raw spend and clicks into the same ratio the $3-5 figure claims to describe, except this one is actually yours.
4. Structure what you save the way you'd structure a competitor ad. If you're logging Sponsored block screenshots for reference, AI ad enrichment tagging keeps them searchable the same way you'd tag a saved competitor creative.
5. Track share of voice against named competitors you chose, not an undisclosed panel. You know who you're actually bidding against. A tracker's synthetic prompt list doesn't.
6. Re-baseline after every platform change. Sponsored Agents, the EU rollout, new ad formats — each one resets the auction dynamics under your numbers, so a baseline from April tells you little about October.
7. Report a range and a method, not a single confident number. A defensible "our CPC ran $2.80-4.10 over six weeks, computed from account exports" beats a borrowed $3-5 every time someone asks how you got it.

Watching that trend over months is where ad timeline analysis earns its keep — the value isn't a single snapshot, it's the shape of your own curve. This is close to what our campaign benchmarking use case documents for platforms with public ad libraries.

On ChatGPT you're building the same discipline without a public library to lean on, a harder version of the same problem. If you're advising a client rather than running your own account, how agencies should prepare for this channel covers the conversation that comes next.

What real chatgpt ads benchmarks would require

Two things, realistically. First, OpenAI would have to publish structured platform data the way other platforms' ad libraries already do, even partially. Second, a regulatory mechanism would have to force transparency OpenAI hasn't volunteered.

The second path already has a name: Article 39 of the EU's Digital Services Act requires a public ad repository covering who paid, on whose behalf, the targeting parameters, and total recipients by Member State.

The European Commission designated ChatGPT a VLOSE on August 31, 2026 — the first AI chatbot to get that classification, triggered by roughly 159.1M average monthly EU users against a 45M threshold. The Article 39 repository compliance deadline lands in January 2027. No such repository exists for ChatGPT yet.

Even the platforms that have had years to comply aren't there. X remains the worst case, missing content, payer and targeting data, with a €120M DSA fine from December 2025 under appeal. TikTok and Meta each have 7 of 11 sub-requirements unmet. No VLOP has ever been fully examined against Article 39 specifically.

Meta's own Ad Library, launched in 2019, is the closest precedent for what a working repository looks like, and even it omits spend, CTR, cost-per-result and audience size — third-party tools built an entire market layering that missing metadata back on top, a pattern that's directional, not proof that the same market will exist for ChatGPT.

The structural problem runs deeper than politics, too. Meta assigns every ad a stable, persistent ID that a permalink can point to forever. ChatGPT has no equivalent object: no static placement, no per-ad permalink, heavy personalization that breaks the assumption an ad is a fixed, verifiable thing two different people could both look up.

OpenAI's usage policies also prohibit automated scraping outside an official API, and no official transparency API exists. That's the actual reason a Meta-style ad transparency archive can't just get rebuilt for ChatGPT by a motivated developer. The object it would need to archive doesn't reliably exist in the first place.

Common questions about chatgpt ads benchmarks

Does OpenAI publish an official ChatGPT ads benchmark? No. OpenAI publishes ad policy and placement rules but no rate card, no CPC or CPM benchmark, and no performance case studies with attached numbers.

What is the average CPC for ChatGPT ads? There's no official average, and no openai ads benchmark to check it against. The $3-5 figure that circulates comes from Axios and Digiday reporting on the May 2026 self-serve Ads Manager launch, describing observed auction behavior rather than a stated OpenAI rate.

What percentage of ChatGPT responses show ads? It depends which study you ask. SE Ranking found ads in 25.94% of 50,006 commercial prompts, Similarweb reported roughly 26% around the same time, and a separate tracker swung between 0.05% and 51% within weeks. Treat any single figure as one study's finding, not a settled rate.

Can you measure ROI on ChatGPT ads the way you measure Meta ads? Not directly. Ads Manager withholds the triggering prompt, demographic data, reach, frequency and incrementality reporting, which removes most of what a normal ROI or lift analysis needs.

Will a real ChatGPT ads benchmark ever exist? Only if OpenAI publishes structured platform data on its own, or a regulatory mechanism like the DSA's Article 39 ad repository forces disclosure. Neither exists as of this writing, and the January 2027 compliance deadline is the nearest forcing function on the calendar.

A CFO doesn't want a citation. They want a number they can defend next quarter, and "there isn't one yet, so here's ours" is a legitimate answer if the number behind it is real. Until OpenAI publishes something structured or a regulator pries one loose, the only trustworthy figure among all the chatgpt ads benchmarks in circulation is the one built from an account's own spend, computed the same way every time.

Six months of borrowed numbers is enough — start a free trial and start tracking your own ChatGPT ads baseline against your own history, or check what it costs on pricing.