Stop asking ChatGPT ads providers to show you real results. Right now, that question rewards whoever can tell the cleanest story about a small, short test.

I used to give the opposite advice. For Google Ads, I told prospects to demand proof: show me an account like mine, show me what you changed, show me CPA before and after. Years of account history made that a useful filter. ChatGPT ads do not have that history. Budgets are small, measurement is shaky, and a polished PDF can make a selected win look like a repeatable method. Ask for the test design instead. A provider’s control, budget cap, kill criteria, and plan to read ChatGPT beside Google and Bing tell you more than their best slide.

Why a ChatGPT ads case study is the wrong filter

It sounds responsible to ask which agencies have run ChatGPT campaigns and can show results. Best agency for a small business. Recommended provider for a small-budget test. Who can actually manage ads inside ChatGPT? Same question in five wordings.

For an established channel, that filter can work. Here, it mostly tests presentation skills. ChatGPT ads launched on Feb 9, 2026, for Free and Go users. Self-serve Ads Manager opened May 5, 2026, and the October 2026 expansion brought the total to 89 countries, 1.2B weekly users, and 100,000+ advertisers. No provider you interview has a year of results from this channel. What they can show you is a handful of short tests, usually led by the one that photographs well.

A good result is not the same as a good process. The process is what you will be paying for when the next test is yours.

The numbers are too scattered to be a playbook

OpenAI publishes no performance benchmarks by industry or campaign type. A 49-case collation reports a median CTR under 1.25%, with results ranging from 0.35% to 7.9%, and a median $4.41 CPC across 15 dollar-billed campaigns. That spread is not a planning benchmark. It is a warning against borrowing someone else’s forecast.

All four same-advertiser comparisons in that set showed worse conversion rates than Google Ads. The scaled wins OpenAI publishes are partner-reported picks: WeightWatchers reported CPA 15.3% below blended paid search, Dose reported that 67% of incremental purchases came from new customers, and Portland Leather reported that 93% of visitors were new. OpenAI calls incrementality testing still in its early stages.

I would not throw out those wins. I would throw out the leap from a selected win happened to this provider can repeat it for my business. The question is whether ChatGPT brought in customers Google and Bing would not have brought in. A case study without a comparison group cannot answer it, however handsome the charts.

A click can be real and still be the wrong click

On Google, I could point to a keyword and say: this term produced this CPA last month; now watch what happens when I change the account. ChatGPT does not give you the same fence around a test. One niche advertiser counted 1,370 US searches over two months across its 14 core keywords. Its ChatGPT test nevertheless served 19,154 impressions, about 14 times that total search demand. Ads followed context hints across conversations, without the keyword negatives an operator would use in Search.

That test spent $477 over six weeks for 19,154 impressions, 370 clicks, and zero conversions. The same pages converted organic search at 2.43%, which predicted about five sales from 210 ad sessions; the stated probability of zero by luck was about 1 in 176. Paid visitors averaged 37 seconds and 1.09 pages, versus 443 seconds and 2.18 pages for organic ChatGPT referrals.

I read that as an intent mismatch, not a creative brief begging for another round. The clicks arrived curious. The buyers arrived convinced. If a case study reports clicks but hides what happened after the click, it hides the part you need to buy.

A polished deck can hide the messy part

The collation authors note that positive comparisons come mostly from sellers of ChatGPT ad management or tools. That suggests selection, not necessarily dishonesty. Providers show prospects the tests that worked. Four of the five highest CTRs in the set came from markets outside English: France at 4.83%, India at 4.25%, Taiwan at 3.85%, and Brazil at 3.30%. The authors warn against using someone else’s CTR as a planning benchmark. I agree. Language, category, and the Free and Go users who happened to see an ad all matter.

The measurement is uneven, too. One UK lead-generation shop saw form fills arrive while the platform still showed zero conversions two weeks later. It saw about 100 reported clicks against about 20 in its own tracking. Another partner reported a 24-to-36-hour conversion delay, versus hours on Google. AdRoll and Jellyfish reported no click gap. That is not proof that tracking always fails; it is a reason to check your own before declaring victory or defeat.

And the scale headlines? $1B in annualized run-rate in under 200 days and 100,000 advertisers tell you the channel is attracting spend. They do not tell you its CPA, ROAS, repeat-spend, or incrementality against a control. Measurement firm Measured puts equivalent ChatGPT CPC at about six times Google’s and describes early tests as generally pretty bad. The deck that skips the cost and the tracking gaps is not proof. It is editing.

Ask for a plan that can survive a disappointing result

Hand every shortlisted provider the same one-page brief. Judge the reply on four things: the control, the budget cap, the kill criteria, and how ChatGPT spend will be read beside Google and Bing. I want to know what they will do if the numbers look flat, not just how they will describe a win.

The channel gives providers more ways to sound certain than to be certain. OpenAI’s October 2026 engineering post says a separate LLM orchestrator handles ad retrieval and ranking while answer generation runs elsewhere. Winning an auction does not guarantee an impression: a later relevance check and delivery window still apply. OpenAI tracks funnel stages internally that it does not fully expose. Its policy says advertising does not influence answers, and ads face limits in regulated verticals and sensitive conversations.

An operator cannot steer every reason an impression appears or disappears. So I do not want a promise to optimize CTR until a good-looking number turns up. I want a plan to find out whether the spend added purchases or leads. Without a control, the provider cannot tell you what ChatGPT added.

The test brief I would send every provider

Copy the following into your email. Give each provider the same question and ask them to return a proposed setup, measurement plan, and exit rules. This is not a request for a perfect experiment. It is a way to see who recognizes what a test must rule out.

Question: Does ChatGPT add conversions Google and Bing would not have produced? Count purchases or leads in your own system against a control. Do not substitute impressions, CTR, or platform-reported clicks for the answer.

Setup: Propose a capped budget you can afford to lose, a four-to-six-week window, one conversion definition shared across channels, and a holdout agreed before launch. Ask how the provider would use a holdout or geo-lift design and what they would compare at the end. Mid-market buyers have been advised to keep a microscopic learning budget and define a holdout or geo-lift before launch. For a small business, the point is the same: decide how you will judge the spend before you spend it.

Make the provider name the reach constraints. Custom-audience inclusion has a 25,000 minimum and 100,000 recommended, so many small businesses will be limited to exclusions. Free and Go users see ads; Plus, Pro, and Business users do not. A plan that ignores who can see the campaign is not a plan, even if the PDF has a lovely funnel diagram.

Controls and weekly measurements

Have the provider write down the holdout, the shared conversion definition, and the Google and Bing comparison before launch. Then ask for the same lines each week. Measure money and added buyers first; use platform ratios to diagnose the result.

  • Spend against the cap, in dollars. A percentage without its base tells you little.
  • Conversions in your analytics, using the agreed definition. Put the platform count beside it, not in its place.
  • Platform clicks against your sessions. Log the gap weekly so a tracking break does not masquerade as performance.
  • Time, pages, and new-customer share. These help distinguish browsing from buying. The early partner wins providers quote include 67% new customers for Dose and 93% new visitors for Portland Leather.
  • CPA beside Google and Bing, plus the holdout result. A cheaper click does not earn more budget if it adds no customers.

Allow for the reported 24-to-36-hour conversion delay before judging a week’s conversions. If platform and analytics counts disagree, show both. Do not let the cleaner number quietly replace the messier one.

Budget cap and kill criteria

Write the exit rules before the first dollar goes out. Say you spend $20,000 a month on Google at a $60 CPA. Give ChatGPT $1,500 over four weeks. That is a learning budget, not a growth bet. A proposed decision rule might read:

  1. Extend: If CPA lands within 20% of Google after the full reporting delay, and the control supports added conversions, extend four more weeks and raise the cap.
  2. Kill: If spend hits the cap with zero added conversions in your system, stop.
  3. Pause: If tracking gaps exceed 30% for two straight weeks, fix measurement before judging performance.

Those numbers are an illustrative rule for that budget, not a benchmark a provider gets to impose on every account. What matters is that the provider commits to a rule you can check. Large brands can treat early spend as an innovation tax. A small business needs to know when the experiment ends.

What I expect the result to change

I expect many small-business tests to read flat for now. The mechanism is intent: the shopper who types a keyword into Google may be ready to buy; the person chatting with ChatGPT may still be thinking. Broad context matching can bring browsing clicks at premium prices. That is my expectation, not a result I can promise you in advance.

A flat result means clean tracking but zero or near-zero added conversions at this budget. Do not scale it. Wait six months rather than paying to repeat the same question immediately. A bad read is different: clicks your analytics never saw, conversions the platform missed, or post-click behavior that looks nothing like your buyers. Fix measurement or targeting before you claim the channel failed; then decide whether one more capped test is worth it.

Either result changes your next spend decision. A pretty CTR does not.

Four answers that disqualify a provider

I would reject a proposal that does any of these:

  • Leads with CTR or impressions instead of added conversions counted in my system.
  • Promises a CPA or ROAS before seeing my tracking, despite the absent benchmarks and wide variation in early results.
  • Has no holdout or kill rule, only a plan to keep optimizing platform clicks.
  • Charges a percentage of ChatGPT spend while asking me to treat a capped experiment as an open-ended campaign. The deck calls that alignment. I call it paying more when the test gets expensive.

A serious answer is less exciting: here is the cap, here is the control, here is when we decide, and here is the CPA beside Google and Bing that would earn more budget. If they cannot write those sentences before launch, do not pay them to narrate the result afterward.

Some businesses should not run the test yet

If ChatGPT ads cannot reach your buyers, waiting is the right design. Health advertisers outside the US, finance and legal teams with tight review rules, and businesses focused on Plus, Pro, and Business users should sit out; those users do not see the ads. So should anyone who cannot spare $1,000 to $2,000 and learn nothing. Not running a weak test is a decision, not a failure.

For everyone else, this is why I run ChatGPT tests inside groas Search Ads, with Google, Bing, and ChatGPT managed together for one flat $1,499 a month. Bids, budgets, and landing pages are worked around the clock, with changes logged. The point is to read ChatGPT CPA beside Google CPA rather than buy a separate ChatGPT story that hides the comparison.

Keep asking for case studies and you will hire the best storyteller in the newest channel. Ask for the test design and you can find out whether ChatGPT added a customer you would not have won anyway. Ignore that distinction and you may keep funding the story long after the test should have stopped.