ChatGPT Ads Is Growing Up. The Results Still Need Proving.

The Canary Gets a Bidding Upgrade

ChatGPT is the first large AI assistant to sell ads at scale, which makes it the canary in the coalmine for every assistant that follows. How it bids, what it reports and whether it can prove it works will shape what advertisers expect from AI ad inventory for years.

From today, 9 November, ChatGPT Ads bidding options are changing. According to OpenAI's email to advertisers, as reported by Search Engine Roundtable:

• Target CPA replaces Max Bid for new conversion campaigns, and Target CPC replaces Max CPC for new click campaigns.
• Targets are averages, not ceilings. OpenAI says the system can bid above or below your target in any individual auction. The old strategies capped each bid.
• Maximize results stays, renamed intelligent bidding. It has been the preselected strategy for eligible new ad groups since August.
• Max Bid and Max CPC retire on 1 January 2027. Remaining campaigns will be moved to targets automatically, and the email does not say what target value they will be given.
• No minimum conversion volume is stated for Target CPA. OpenAI does warn that restrictive targets or weak conversion signals can reduce delivery and leave budget unspent.

This is what a maturing ad platform looks like, and it is broadly good news. Target-based bidding, a pixel and Conversions API, an Event Quality Score and a growing list of measurement partners are the building blocks every serious performance channel has. But it also removes the one hard control advertisers had, at a point when the evidence on performance is thin and the reporting is hard to trust. Back in September we advised capping ChatGPT tests with Manual Max bid. From January, that option goes.

What the Public Results Actually Show

We reviewed the public performance write-ups we could find. Almost all come from agencies or vendors that sell ChatGPT Ads services or tools, most are single accounts or small budgets, and very few use the same definitions. With that health warning, the picture is consistent.

Attention is low

Click-through rates sit far below search. PPC ONE's compilation of international results puts the median across 24 campaigns at 1.25%, with a range from 0.35% to 7.9%. In Workshop Digital's same-advertiser test, ChatGPT delivered 1.06% against 5.2% on Google Search.

Conversion is weaker still

This is the number that matters, and it is the most consistent finding. In all four public same-advertiser comparisons we found, Google Search converted far better: 0.46% against 5.3%, 0.21% against 3.71%, 0.9% against 5.95%, and 0.8% against 6.64%. Several advertisers reported spend with no conversions at all.

Clicks can be cheaper, but that is not the point

Cost per click varies wildly, from under $1 to over $13 across PPC ONE's US examples. Some advertisers paid less per click than on Google, with Workshop Digital at $3.14 against $6.76. But a cheaper click that converts at a fraction of the rate is not a cheaper customer.

ROAS depends on who is counting

The most transparent single-account result we found, from Opascope, was 1.49x blended ROAS on about $60,000 of spend over 15 days, with daily ROAS swinging between roughly 0.2x and 2.9x. Other published figures run from 3x to 6.8x, but every one of them is attribution-based, and in one case ROAS moved from 3.3x to 6.8x depending purely on the attribution model chosen.

There are positive results. Some lead generation advertisers report lower cost per lead than Google, and OpenAI's own partner case studies show lower attributed CPA and high shares of new customers. But taken together, the public evidence does not yet support the idea that ChatGPT Ads is an efficient performance channel.

The Reporting Problem

Weak performance would be manageable if you could see clearly why. You largely cannot.

• No prompt-level data. Advertisers cannot see what was asked when their ad appeared. Buyers have been pushing for prompt-level reporting and stronger placement insights, and it is the single biggest visibility complaint.
• Thin breakdowns. Device and country are the only documented breakdowns. There is no demographic, audience, reach or frequency reporting.
• Modelled conversions. OpenAI notes that reported totals may include modelled conversions, so part of what you see is an estimate, not a count.
• Discrepancies in both directions. One US B2B advertiser saw 36 platform conversion events against 2 real inquiries. Another saw zero conversions in the dashboard against three actual sales, after a tracking parameter was lost at checkout. Others report platform clicks far exceeding the visits recorded in their own analytics.

When a platform both decides what to bid and grades its own results, and you cannot see the inputs, the honest position is to treat every number it reports as a claim to be checked, not a result.

And the Uplift Question

The deeper question is whether ChatGPT Ads creates sales that would not otherwise have happened. On that, the public evidence is very thin.

In October OpenAI opened up to third-party measurement partners, including attribution platforms, data connectors and incrementality specialists such as Haus, Measured and WorkMagic, who are exploring geo experiments. That is the right move. But by OpenAI's own account incrementality measurement is at an early stage.

The only published lift-style result we found is a geo test for one brand, Dose, run with WorkMagic and released alongside OpenAI's own case studies. It reported significant lift, mostly from new customers, and more than twice what last-click attribution captured. It is encouraging, but it is a single partner-published case. None of the independent write-ups we reviewed ran a holdout.

OpenAI has also shared that 52.7% of eligible one-day view-through conversions happened within an hour of the ad being seen. That tells you about timing. It does not tell you whether the ad caused the sale.

How to Test ChatGPT Ads Without Taking It on Trust

There is a real audience on ChatGPT, and advertisers are spending regardless, so testing is reasonable. Test it like a channel that has to prove itself:

1. Set targets deliberately before January. Do not let the 1 January migration pick your Target CPA for you. Review existing Max Bid and Max CPC campaigns now, decide your targets, and keep tests small while hard caps still exist.
2. Reconcile weekly, not quarterly. Tag everything, check the click identifier and UTMs survive to checkout, and compare platform clicks and conversions with your own analytics every week. Investigate any gap above a few percent.
3. Judge on conversion and quality, not CPC. Compare cost per qualified lead or customer against search for the same offer. A cheap click is irrelevant if the lead never closes.
4. Run a holdout. Use a geo or audience holdout and judge the result on incremental revenue. If the uplift is not measurable, the spend stays at test scale.
5. Look for the halo. Some ChatGPT influence will convert later through brand search or direct visits and be credited elsewhere. Measure full customer journeys rather than last click, which is exactly what independent attribution such as Corvidae AI is built to do.

The Canary's Lesson

ChatGPT Ads is maturing quickly, and target bidding, conversion APIs and measurement partners are what any serious ad platform needs. But maturity in the tooling is not the same as proof in the results. Right now the public evidence points to lower attention, much lower conversion, reporting that advertisers struggle to trust, and very little independent proof of uplift.

Every AI assistant that sells ads will face the same three questions: does it convert, can we trust the numbers, and does it add anything we would not have had anyway? ChatGPT is answering them first, in public. Advertisers who insist on independent answers now will be far better placed when the rest of the AI ad market arrives.

If you are planning a ChatGPT Ads test, or want to know what your existing tests are really delivering, our paid search and analytics and attribution teams can help you set it up so the results stand up.