A B2B ChatGPT ads test matched 336 clicks to 146 companies, and only five fit the ICP

Six months into ChatGPT Ads, enough advertisers have published their spend data to compare notes, and the notes do not agree. Across the four detailed tests collected by Search Engine Journal on September 8, costs per click run from $4.41 to $22.89, with other advertisers reporting figures below $3 — and OpenAI has published no benchmarks by advertiser, industry or campaign type to explain the distance between them.

The platform gives you little with which to close that gap. Ads Manager offers spend, impressions and clicks, the averages that fall out of them — CTR, CPC, CPM — and conversions. There is no Auction Insights equivalent, no impression share, no competitive reporting. Selection runs on a relevance-weighted second-price auction driven by the context of the conversation, the landing page, the creative, and advertiser-supplied "context hints" that OpenAI is careful to say are not keywords. When your CPC moves, nothing in the reporting separates competition from relevance, inventory, your own bidding, or a change in which conversations you were matched to.

The spread is not noise, it is the finding

Four published tests, four different pictures. Hostinger's is among the largest shared publicly: after close to $70,000, head of PPC Hüseyin Ograk reported CPMs above $65 and said click-through rate concerned him more than cost, with narrow use cases outperforming broad messaging and traffic quality inconsistent. Common Thread Collective scaled a high-order-value ecommerce advertiser from $7 a day to more than $1,000 a day; after $9,620 it reported a $4.41 CPC and 0.94% CTR, and between $19,000 and $38,000 in attributed revenue depending on the model — a ROAS anywhere from 3.3x to 6.8x, measured in Triple Whale rather than in Ads Manager. Synter spent $4,428.84 for a $9.89 average that concealed $5.10 in the UK, $10.62 in the US, $17.59 in Australia and $22.89 in New Zealand.

The test that should stop a B2B software marketer is Floyd Blaikie's. Her team spent roughly $7,000 CAD at a $9.29 CPC and 0.7% CTR, then took the traffic to a visitor deanonymization tool. Behind 336 paid clicks sat 146 organisations. Five matched the ideal customer profile.

Five ICP matches out of 146 identified companies is not a cost-per-click problem. It is a question Ads Manager was never built to answer.

Every number in that campaign looked ordinary. The CPC sat between Synter's UK and US figures; the CTR was a rounding error from Common Thread Collective's. Judged on the metrics the platform hands you, it was a normal test. The only reason anyone knows otherwise is that Blaikie's team asked a different system who the clicks belonged to. For software companies with long cycles and small buying committees, that question is not a refinement — it is the entire evaluation, and none of it happens inside OpenAI's reporting.

OpenAI is shipping matching, not diagnostics

Four days before that round-up, Search Engine Land — owned, worth noting, by Semrush — reported OpenAI's broadest batch of advertising updates so far. Ads are rolling out across Europe, India, the Middle East and North Africa. Custom audiences can now be edited rather than rebuilt and exceed five million members. The Measurement Pixel now takes more hashed identifiers off your customer records — names and phone numbers, postal codes, regions. A planned optimization model will weigh view-through alongside click-through conversions and bill by impression.

Read the two together and the pattern is clear. Almost every one of those updates improves OpenAI's ability to match a conversion you already got back to a click. Not one of them explains a price you paid. Identity resolution and auction transparency are different products, and only the first is being built.

Which leaves the $3-to-$5 figure now circulating as a rate card. It is OpenAI's recommended starting maximum bid, not an average. It is also increasingly decorative: since August, eligible new ad groups default to "Maximize results", which sets and adjusts the bid for you and promises no particular CPA, CPC or ROAS. An advertiser who wants a hard ceiling has to opt back into manual bidding.

What a software marketer should do differently

Treat ChatGPT Ads as a channel you are measuring, not one you are optimising. The audience limit alone reframes the buy for B2B: ads appear only to Free and Go users, while Plus, Pro, Business, Enterprise and Edu accounts are ad-free. Go is itself paid, so this is not simply the non-paying audience — but the seats a company buys for its staff sit in the excluded tiers.

• Write down what would make the channel worth keeping before the first dollar goes out, in your numbers, not someone else's CPC
• Stand up firmographic identification on paid ChatGPT traffic from day one — the click-to-company question is the test
• Budget for a market-by-market read; a blended CPC hid a 4.5x range in Synter's data
• Set manual bidding if you need a ceiling, and log which strategy each test used
• Hold early results loosely; with no diagnostic signals, a bad week and a bad channel look identical

As Osmundson puts it, "there is no universal number that determines whether a ChatGPT Ads campaign is performing well." That is not a reason to sit out. It is a reason to arrive with your own instrument.