What a $1,500 ChatGPT Ads Test Can Prove—and What Takes $5,000

At the $3.14 CPC recorded in one B2B ChatGPT Ads test, $1,500 buys about 478 clicks. If those clicks convert at 0.5% to 2%, you get roughly 2 to 10 conversions. That is not a ROAS test. It is a small sample with a media budget attached.

The same B2B test spent $2,319 for 739 clicks and got zero MQLs, while Google Search on the same offer produced 2 MQLs at $6.76 CPC. Cheap clicks did not mean cheap customers. I have run this budget math for clients for years, and I used to get the order wrong: size the budget, run the campaign, then ask what we learned. The useful order is the reverse. Choose the question first.

OpenAI publishes no cross-advertiser benchmarks and warns that delivery and spend fluctuate in beta. The scenarios below are worked examples, not ChatGPT Ads market averages. They show where $1,500 can produce a decision and where $5,000 or more may be needed just to get a readable conversion count.

Set the click and conversion assumptions before you spend

ChatGPT Ads pricing has changed quickly. It launched at $60 CPM with a $250,000 minimum in February 2026, then moved to $25 CPM with a $50,000 minimum by April 2026. It now uses CPC billing, with a $25 daily minimum and a recommended starting max bid of $3 to $5 per click in a relevance-weighted second-price auction. Early spenders report average CPCs of $2 to $5 across industries; the platform warns when a max CPC sits below $3.

For the budget math, I use a cheap $2 click, the $3.14 click from the documented B2B test, and a $5 click for a more competitive case. These are scenarios, not promises of what an auction will charge.

The low end is possible, not guaranteed. One agency reported $0.76 average CPC on ChatGPT versus $10.46 on Google over the same 30 days, alongside a client median of $2.12 CPC, 1.39% CTR and $29.61 CPM. That $0.76 example would buy far more clicks than my table shows. Your CPC determines the sample size; your conversion rate determines what that sample can tell you.

Why I model lead forms, demos and purchases separately

Google conversion rates do not transfer neatly to ChatGPT. In one early same-site comparison involving about 1,500 users, Google converted at 3.71% with 41-second engagement; ChatGPT converted at 0.21% with 17-second engagement. Another B2B account recorded 0.8% on ChatGPT versus 6.64% on Google over the same 30 days. Those are individual observations, not a platform-wide rate. They are enough to make copying a Google forecast look reckless.

The click can arrive in the middle of a conversation and leave quickly. Published landing-page baselines run higher: around 6.6% all-industry median, 3% to 7% for typical lead generation, 2% to 10% for booking a call, and 1% to 5% for a demo. Those baselines describe different mixes of traffic and intent. For this exercise, I model:

• 4%: a strong, low-friction lead form. This is the optimistic case.
• 2%: a lead offer with some friction. This is the middle case I plan around for a new channel.
• 0.5%: a demo, booked call or purchase. This is where a small budget breaks first.

The mechanism is plain division: CPC divided by conversion rate equals cost per conversion. As an illustration, $5.42 CPC at 8.18% conversion is about $66 per lead, while a $3 ChatGPT click at 0.21% conversion is about $1,429 per conversion. Lower CPC cannot rescue a much lower conversion rate. Do not choose a budget from the click price alone.

How many clicks buy 10, 30 or 50 conversions?

Clicks needed equal target conversions divided by conversion rate. Spend equals those clicks multiplied by CPC. The table uses an illustrative $3 CPC, near the $3.14 documented B2B click and within the $2 to $5 range early spenders report. At $5 CPC, multiply the spend column by about 1.67; at $2, multiply it by about 0.67. Rounded conversion counts are expectations, not a guarantee that the next click will complete the row.

Read the last row twice. At a 0.5% conversion rate, $1,500 buys an expectation of 2 or 3 conversions and a story you will be tempted to tell about them. Even the 4% case produces only 20. Practitioners often look for 15 to 30 conversions before calling a test readable, and roughly 30 conversions in 30 days is a familiar Google Smart Bidding learning reference. Neither is a published ChatGPT Ads requirement or a statistical guarantee. Thirty is a useful planning yardstick, not a magic switch. At 0.5%, even $5,000 does not buy that yardstick.

Ten conversions leave a wide error bar

Suppose you pay $3 per click, get 500 clicks and record 10 conversions. The observed conversion rate is 2%. A rough normal-approximation 95% interval runs from 0.8% to 3.2%. At the same CPC, those endpoints imply a CPA from about $94 to $375. The observed CPA is $150, but the estimate is nowhere near precise enough to defend a narrow CPA claim. With 20 conversions from 500 clicks, the approximate conversion-rate interval tightens to 2.3% to 5.7%. Better. Still wide.

Now split the original 500 clicks between two variants. Variant A gets 8 leads from 250 clicks; variant B gets 12 from 250. B appears to win by 50%. At counts that small, a swing of 4 leads can be ordinary noise. I used to tell clients we could read creative signals at low spend. I was wrong. What I was reading was randomness with a headline attached.

This does not mean small samples tell you nothing. It means below 10 conversions, read the result as a range, not a verdict. Even above that, ask whether the difference is large enough to change your decision before you call a winner.

The two questions $1,500 can answer

A small test is useful when it asks a small question. I would use $1,500 to check repeatable volume and break-even fit, in that order.

1. Does the category produce enough ad opportunities? At the $25 daily minimum, $1,500 could stretch across about 60 days at the floor. Watch impressions and whether relevant prompts keep producing opportunities or go quiet. The documented B2B test recorded 69,630 impressions for 739 clicks. If your own campaign produces only 2,000 impressions after 30 days and $750 at the floor, limited scale is worth knowing before you commit another month. It does not prove the category will never grow; it gives you a reason not to scale this test yet.
2. Can the click price plausibly fit your margin? Take an illustrative customer worth $600 in gross margin. If 1 in 5 qualified leads closes, break-even cost per lead is $120. At $3 CPC, you need a 2.5% click-to-lead conversion rate to hit it. After 500 clicks, an observed rate near 0.5% looks different from one near 4%, even allowing for wide error bars. A narrow miss around 2.5% needs more data, not a triumphant slide.

Those are decisions a small test can support: pause for lack of volume, pause for a clear margin mismatch, or earn the right to keep measuring. Use $1,500 to test feasibility, not to announce ROAS.

What $1,500 cannot settle: creative, audiences or ROAS

Split 500 clicks across three headlines and each gets about 165 clicks. At a 2% conversion rate, that is roughly 3 conversions per headline. You cannot pick a creative winner from that. The same denominator problem applies to audience splits. Custom-audience targeting requires 25,000 matched users, which may rule it out before a small advertiser spends a dollar. Broader prompt-level splits still need far more clicks to distinguish a 1.8% converter from a 2.2% converter.

ROAS is harder. It needs closed revenue divided by spend, not merely form submissions. If 10 leads close at 20%, you have 2 sales; lose one deal and the observed return changes sharply. Early paired tests have also raised a lead-quality question alongside lower ChatGPT conversion rates than Google. A $1,500 report that leads with clicks and CTR while declaring ROAS proven has measured activity, then labeled it certainty. Do not make the claim the denominator cannot carry.

Run one offer, one page and one decision rule

With $1,500, the structure should be boring: one tight campaign, one offer, one landing page, no creative or audience splits. Practitioners warn against spreading tiny budgets across variants; concentrate spend and judge cost per lead against margin rather than CTR. Write the decision rule before launch, while nobody is emotionally attached to the results.

My rule would be to review at 500 clicks or $1,500, whichever comes first, and keep spending only if cost per qualified lead sits within 30% of break-even and impression volume shows the category repeats weekly. Anything else is a pause, not permission to scale. That is the difference between doing the conversion math most agencies skip and holding a meeting about promising engagement.

• One offer, one conversion event. Choose a short lead form or a booked call, so every click feeds the same denominator.
• One landing page already used for paid search. A page with an established search conversion rate gives the new-channel read more context. It does not guarantee the same result on ChatGPT.
• One written kill rule. For example, pause at $750 with zero qualified leads, or if CPC stays above $5 when break-even requires it below $3. The seven small-budget tests with one question and a kill rule each use the same discipline: decide what would stop the test before you see a flattering CTR.

Before an agency asks for $5,000, ask for the conversion count

Make the agency show the break-even sheet before it raises the budget. What CPC did you pay? What qualified-lead rate did you observe? What cost per qualified lead does that imply against margin? Then ask what the next $3,500 is expected to buy in conversions, not clicks.

At an assumed $3 CPC and 2% conversion rate, $5,000 buys about 1,667 clicks and 33 conversions. That reaches the roughly 30-conversion planning yardstick that a $1,500 test at the same rate does not. At 0.5%, the earlier table shows the limit: $5,000 still falls short. A bigger budget is not automatically a better test. It needs to buy enough outcomes for the question on the table.

I run ChatGPT tests inside groas Search Ads work because execution is continuous and changes are logged with reasoning. The budget stays gated until volume and break-even checks pass. That is the approval-before-scale control I want from a small test: one question, one kill rule, no increase on noise.

Ask for three answers in writing:

1. What qualified-lead rate are we assuming, and what paired evidence supports it?
2. What is the pause trigger, in numbers, and who calls it?
3. If the test earns more spend, what conversion count will that spend buy, and what result would count as a winner?

An agency that can answer in conversions and kill rules may have a defensible case for $5,000. An agency that answers only in impressions and CTR is selling spend, not a test.

The decision: buy the answer your budget can support

A $1,500 ChatGPT Ads test that promises ROAS proof will fail, and the failure may look remarkably like data. Ten conversions leave a wide CPA range. Two conversions leave a story. Size the question before you size the budget.

Run the $1,500 test if you need to know whether your category produces repeat volume and whether its click cost can fit your break-even CPA. Keep one offer on one page, write down the kill rule, and do not approve the next tranche until its expected conversion count is large enough for the decision you want to make. At a 2% rate and $3 CPC, $5,000 can plausibly buy that next read. At 0.5%, it cannot. Spend for the answer, not for the report.