Why ChatGPT Ads Are Underperforming According to Early Advertiser Data

ChatGPT Ads are underperforming because the placement sells attention from a user who is in the middle of a conversation, not at the end of a decision, and the measurement stack behind the auction has not caught up with the price. Rand Fishkin's recent crowdsourced thread on live advertiser results put numbers to what operators running CPC campaigns had already noticed: click-through rates sit near 0.7% while Google Search clears 6.4%, reported CPCs cluster between $3 and $22 with no stable benchmark, and attribution leans on exports, pixels installed weeks ago, and manual reconciliation. The signal in the data is not that the channel is broken. It is that early advertisers are paying search-grade prices for research-grade intent, and the distance between those two things is where ROAS disappears.

What the Fishkin Dataset Actually Says

The Fishkin post pulled responses from practitioners running live ChatGPT Ads spend across retail, SaaS, and B2B services. The pattern in those self-reported numbers lines up with the few third-party studies that exist. ChatGPT ads produced a 0.91% CTR against Google's 6.4% Search benchmark in Adthena's early pilot tracking, and six months in, CPCs range from $3 to $22 across advertisers with OpenAI publishing no cross-advertiser benchmark to explain the spread.

Conversion data is thinner but points the same way. A new working paper from WU Vienna and the Frankfurt School, which analyzed sales data from almost 1,900 online shops, found that conversion rates ran at 0.37% for ChatGPT ads compared with 1.17% for Google search ads, and that a paid ChatGPT visit generated $0.89 in revenue per session versus $2.45 for a paid Google search visit. The researchers calculated that a ChatGPT click would need to cost roughly $0.26 to match Google's return on ad spend. Reported CPCs sit above $1.70.

Three things stand out when you line up Fishkin's crowdsourced inputs with the academic and vendor data. Early advertisers are heavily concentrated in B2B SaaS and developer tools, which skews the sample toward research-intent queries. Nearly every respondent reported difficulty tying clicks back to downstream revenue. And the dispersion of CPCs is wide enough that the median hides the story.

Intent Mismatch Is the Primary Failure Mode

A ChatGPT ad fires after the model has already answered the user's question. That is a structurally different moment than a Google results page, where the user has stated an intent and is scanning for the nearest match. Search Engine Land's reporting on practitioner results makes the structural point directly: CTR is lower because the ad appears after ChatGPT finishes answering, and conversion measurement tools only became usable in May 2026.

The implication is practical. Queries that route into ChatGPT skew toward comparison, synthesis, and "help me decide" moments. That is the top of the funnel, not the bottom. If you are buying ChatGPT inventory with Google-style expectations — immediate form fills, same-session purchases, CPL under $50 — the numbers will not clear. The channel reads more like a paid-social discovery placement than a search bid, and the shift AI search is driving in paid intent is the same shift that is reshaping how to budget for assisted conversions across the whole stack.

There is also a composition problem. The academic study found that among the 102 financially tracked advertisers, visitors converted far less often, placed orders $79 smaller and generated $2.86 less revenue per session. Smaller baskets plus lower conversion means the revenue shortfall compounds. Even brands that treat ChatGPT as a brand-lift play should expect the per-session revenue math to look weak for the first several quarters.

Placement Opacity Compounds the Cost

You can run a Google Search campaign and know, down to the query, where your spend went. ChatGPT inventory is the opposite. Independent analysis of commercial prompts found that ads showed on 25.94% of the commercial prompts tested, and in 14.35% of cases the ad matched the query at best as well as a randomly picked ad would. That is placement without targeting control, priced like placement with targeting control.

The reporting surface is similarly thin. Early coverage documented that there was no tracking pixel, no post-click attribution, and no standard conversion tracking, with performance data delivered through weekly CSV files during the pilot. Conversion tracking shipped in May, but the historical record is still too short to build a reliable CPA baseline — and the kinds of things a paid-media team would normally use to catch waste, from search-term reports to placement exclusions, do not exist yet in the form operators are used to from Google Ads. The audit discipline that catches six-figure waste on mature platforms has nothing to grab onto here.

Attribution Shortfalls and the Fishkin ROAS Problem

The crowdsourced thread's most common complaint was not price. It was that respondents could not confidently say what their spend had produced. The Hostinger case Search Engine Journal covered is illustrative: the team relied on Triple Whale alongside the native reporting to evaluate downstream performance, because platform numbers alone did not resolve the question. OpenAI itself explicitly states that ChatGPT Ads does not yet have performance benchmarks across advertisers, industries or campaign types, which leaves the burden of interpretation on the buyer.

That has two downstream effects on how you should budget. First, any claim about ChatGPT ROAS that is not backed by first-party conversion data and a holdout test is noise. Second, early-mover auctions produce outlier wins that will not survive scale. Opascope's widely cited 15-day run returned a 1.49x blended ROAS at a ~$1.72 CPC, with daily ROAS swinging from 0.2x to 2.9x. That is a honest result, not a benchmark. It tells you the channel can work on a specific offer in a specific window, and it tells you nothing about what your campaign will do.

A Budget Allocation Framework for ChatGPT Ads

Given the data, three allocation decisions make sense depending on account profile. These are the choices performance teams should actually be weighing, not whether to "test AI" in the abstract.

• Pause. If your account is under $50K monthly paid spend, is dependent on last-click attribution for CFO reporting, or sells a product where the buying moment is decisively transactional (emergency HVAC, legal intake, local services), there is no budget case for ChatGPT Ads in the current beta. The measurement overhead alone will burn margin.
• Test small. Accounts spending $50K to $500K monthly with a working first-party conversion stack, server-side tagging, and a product that benefits from research-phase exposure (SaaS, D2C considered purchases, financial products) should allocate 2% to 5% of paid budget to a 90-day ChatGPT test with pre-defined guardrails: CPC ceiling, daily budget cap, pixel plus incrementality holdout. Expect noisy daily ROAS and judge on rolling 14-day windows.
• Wait and watch. Accounts whose internal reporting cannot yet handle a channel with partial attribution should skip the test and read the data others publish. The inventory is not going anywhere. A 2027 entry with real benchmarks beats a 2026 entry burning budget to learn what OpenAI will eventually publish for free.

Pair that decision with a cap on what you count. ChatGPT-sourced revenue should sit in its own line in your dashboard, not blended into Google reporting through smart bidding. The teams doing this cleanly are treating it the same way they treated ChatGPT Ads against Google Ads benchmarks when the first comparison data arrived: as a parallel channel under test, with its own CAC and payback math.

What Operators Should Actually Do Now

If you are running ChatGPT Ads today, three operational moves protect the budget while the channel matures. Set a hard CPC ceiling at or below your blended Google Search CPC for the same offer. The parity price research from WU Vienna suggests even that is generous. Instrument every click with first-party tracking and a UTM scheme separate from your paid search taxonomy, so the data stays clean for the comparison that matters. And run a geographic holdout for at least six weeks before letting the channel into any bidding-model signal mix.

Agency-side, the AI PPC work that is holding up at the moment is not creative prompting tricks. It is the boring attribution plumbing — server-side tagging, offline conversion imports, incrementality tests — that tells you whether a new placement is producing pipeline or scraping credit from channels that already were. A PPC Audit that includes ChatGPT spend alongside Google and Meta should start with whether that plumbing exists before it touches a single bid.

Return to the account that opened this piece: search-grade prices, research-grade intent, attribution that does not close the loop. Apply the framework, and the ChatGPT line item either sits at 2% to 5% of paid budget inside a holdout structure, or it comes out entirely until the measurement stack catches up. Either way, the reporting stops lying to the CFO. The upside of the channel is real but it is a brand and discovery asset priced like a search asset, and Fishkin's data is the clearest signal yet that paying full freight for the placement without accounting for the mismatch between search-grade price and research-grade intent is how early advertisers end up with negative ROAS and a story about "testing AI" instead of a number.