The Conversion Count Behind a ChatGPT Ads Result

A ChatGPT ads case study can show an $82 CPA on 22 conversions. At that count, ordinary counting noise is large enough that a roughly 40% swing is plausible. The first number I want from an agency is the conversion count behind its CPA.

The shortcut is one divided by the square root of the conversion count. At 25 conversions, that gives about 20% relative counting noise at one standard error, or roughly 40% at two. It is not a precise confidence interval for CPA: spend, conversion quality and attribution can add uncertainty, and a CPA range is not perfectly symmetric. But the shortcut exposes the problem. A reported $82 CPA on 22 conversions is nowhere near precise enough to settle a close channel comparison. I would treat something like $49 to $115 as a rough illustration of the spread, not as certified endpoints.

That is the filter for any ChatGPT ads ‘result’ an agency can show you today. Ask for results alone and you may get a convincing screenshot. Ask how many conversions produced the number and you can judge how much the screenshot tells you.

Start with assumptions, not a borrowed benchmark

OpenAI has not published performance benchmarks by advertiser, industry or campaign type, according to reporting six months after the February 2026 launch. Reported CPCs range from below $3 to $10–$13 and higher. A single average CPC will not tell you what your account can buy.

I used to tell clients to judge a new channel by its average CPC. I was wrong. CPC sets the price of a chance; conversion rate determines how often that chance becomes evidence. The table uses stated guesses so you can inspect the arithmetic, not mistake it for a forecast.

Delivery comes before efficiency. Reported auction guidance says bids under about $3 win little delivery, and valid clicks are billed even on conversion-bidding campaigns. In this model, $1,000 divided by a $3.50 CPC buys about 286 clicks before any practical delivery constraint. A $2.50 bid that rarely enters the auction will not buy those clicks just because the spreadsheet says it could.

Write down the CPC and conversion rate you are assuming before anyone quotes a prospective CPA. If the manager changes either assumption in the pitch, the conversion count changes with it.

What three monthly budgets buy on paper

Clicks equal budget divided by CPC. Expected conversions equal clicks multiplied by conversion rate. These are arithmetic outputs, not observations or guarantees; fractional conversions are averages you might expect across repeated tests, not something a campaign can record.

The table holds CPC at the illustrative $6 middle case when calculating conversions. Brackets show how clicks change at $3.50 and $9.00 CPC, and how conversions change at 1% and 5% conversion rates while CPC stays at $6. The error column uses the middle-case conversion count and the 1/√n counting shortcut.

At the middle assumptions, $6 divided by 2.5% gives a model CPA of $240. A $3,000 month therefore produces an expected 12.5 conversions, not a proven $240 CPA. Applying the table’s ±28% counting shortcut gives a rough $173–$307 span around that model figure. It is a useful warning about precision, not a formal interval or a promise that the actual CPA will land there.

After 90 days and $9,000 of spend, the same assumptions yield 37.5 expected conversions. The counting shortcut narrows to about ±16% at one standard error. More spend buys more observations; it does not remove questions about attribution or whether the conversions were qualified.

A one-month small-budget test can establish delivery and check whether tracking fires. It cannot establish a precise CPA. That distinction should shape the claim you allow at the end of the month.

What 10, 30 and 100 conversions let you say

The 1/√n shortcut gives a quick scale for counting noise. These figures are approximate relative errors at one standard error, not pass marks or formal channel-comparison thresholds.

To halve that error, you need four times the conversions. I keep those three counts handy because they end arguments that a good-looking early CPA ought to end by itself. Ten conversions are not a verdict. Thirty make for a more serious conversation. Even 100 do not rescue a comparison built on different conversion definitions.

Published Google Ads bidding guidance points to similar volume constraints: it cites 30 conversions in the last 30 days per campaign for Target CPA, 50 or more for Target ROAS, and roughly 15–20 as a workable starting range for Maximize Conversions, with stronger learning beyond 30. Those are bidding-guidance figures, not a statistical certification stamp for ChatGPT ads. They do show why 18 conversions and a winning-looking CPA should not settle a budget decision.

Use the count to decide how strong a claim you can make, not as permission to ignore what counted as a conversion.

Why a 20% channel gap can be noise

Consider an illustrative month in one account. ChatGPT ads spend $3,000 for 12 conversions: $250 CPA. Google Search spends $3,000 for 15 conversions: $200 CPA. The deck calls the difference a 20% win for Google. I call it two small counts with overlapping uncertainty.

The mechanism is simple. Each channel has its own counting noise, so a comparison must clear uncertainty on both sides. A discussion of incrementality testing contrasts an advertiser with 1,000 weekly conversions, which may detect about a 10% lift, with one at 100 weekly conversions, which may detect only changes above 25%. That is a different test from comparing two reported CPAs, so those detection figures do not transfer directly. They make the same volume problem hard to miss: small changes are difficult to distinguish when counts are small.

Channel behaviour complicates the comparison further. In a study of about 10 billion sessions, paid ChatGPT visits converted at 0.37% versus 1.17% for paid Google Search and generated $0.89 versus $2.45 in revenue per session. That report also said 97% of 87 advertisers using both earned less per ChatGPT visit and put a breakeven CPC near $0.26. Those are the study’s aggregate observations, not a forecast for your offer.

A direct-response CPA can also miss assisted or delayed B2B conversions. It should not be declared a win merely because someone expects assists to appear later; those have to be measured. Keep the conversion definition fixed, then inspect direct and delayed outcomes before moving budget on a small CPA gap.

Read the count before the case-study headline

When I open a ChatGPT ads deck, I look for n: the number of conversions behind each CPA. If it is missing, the headline result has not earned a meeting. If it is 14, 31 or 47, I read the CPA alongside the counting-error table and ask what happened after the click.

One published B2B test reported $2,319 spent on 739 clicks at $3.14 CPC and zero MQLs; Google produced two MQLs in the same comparison. Clicks arrived. Downstream proof did not. That does not prove ChatGPT ads cannot work for B2B, and two Google MQLs do not prove Google won the test. It shows why a lively click report cannot stand in for pipeline.

This is why asking for results is the wrong filter when the result comes without its denominator and test conditions. A screenshot might show a $41 CPA on 19 conversions. Using the same shortcut, 19 conversions carry about ±23% relative counting noise at one standard error, before any attribution problem enters the room. The number looks tidy because the dashboard rounds it.

Ask for the count, dates, spend, CPC and conversion definition beside every CPA. Ask whether the reported conversions were qualified and whether the attribution window could capture delayed outcomes. No n, no useful result.

Put one test protocol on the table before spending

Here is the question I would ask a manager to write down: Can ChatGPT ads produce qualified conversions at a CPA within 30% of Google Search after 60–90 days, with delayed outcomes measured? This is a prospective test, not a claim that I have run one. My expectation is weaker direct CPA and a possible role for assists, because a person reading an answer may still be researching rather than ready to buy. The measurements, not that expectation, get the last word.

1. Setup: Use one offer, one landing page and one definition of a qualified conversion. Keep ChatGPT ads separately identifiable in tracking. Cap daily spend against the planned $3,000 monthly budget rather than letting the first week consume the test.
2. Controls: Keep the offer, landing page and conversion definition stable during the comparison. Record any unavoidable change. A mid-test landing-page swap can leave you comparing different funnels under one campaign name.
3. Measure: Report conversions as n; spend, CPC, CTR, conversion rate and CPA; and assisted or delayed conversions at 30 and 60 days. Show the 1/√n counting shortcut beside CPA as a warning about precision, not a confidence interval.
4. Stop and approval gate: Stop at the spend cap or 90 days, whichever comes first. Do not increase budget solely because a low-count CPA looks good. Require a review after n passes 30, including whether the remaining uncertainty fits the CPA decision you need to make. For narrower tests at this spend, use one question and one kill rule per variant.

The point of pre-registering the rules is not ceremony. Without a fixed conversion definition, cap and stop rule, a manager can keep changing what ‘worked’ until something looks presentable. The limits of a $1,500 test deserve the same honesty. Decide what evidence would justify a scale-up before the first click arrives.

The calendar behind 30 and 100 conversions

Time is the number that often disappears from the pitch. At the illustrative $6 CPC and 2.5% conversion rate, a $3,000 month yields 12.5 expected conversions. Divide a target count by that monthly pace and convert months to weeks. The result is a planning estimate, not a date on which evidence automatically becomes conclusive.

At $3,000 a month, reaching 30 conversions takes roughly 10 weeks on these assumptions; reaching 100 takes roughly 35. At $5,000, those estimates fall to about six and 21 weeks. At $1,000, the model does not reach 30 within a quarter. The actual calendar moves if CPC or conversion rate moves, which is precisely why the assumptions belong next to the result.

That is also why I would test ChatGPT ads under the same budget discipline used for Google, rather than buy a separate miracle. groas lists Search Ads coverage for Google Ads, Bing Ads and ChatGPT Ads at $1,499 a month, with a flat fee rather than a percentage of spend. Its always-on execution and logged changes make the work inspectable; the human strategist provides direction and accountability. None of that manufactures conversions or shrinks an error bar without observations. It gives you a way to run the test while keeping the spend, changes and decision in view.

So stop asking which agency has the best ChatGPT ads result. Ask for the n behind its CPA, the CPC and conversion-rate assumptions, and the time your budget would need to reach 30 conversions. If the manager cannot put that math on one page, do not buy the screenshot. If they can, cap the test and let the count determine what you are entitled to believe.