I would give ChatGPT ads $500 in media, 30 days, one offer, and written permission to stop. Organic ChatGPT referrals are not evidence that paid ChatGPT traffic will convert, and I would not buy an open-ended promise to keep optimizing until it does.
Barn2 disclosed what happened after spending $477 over six weeks: 19,154 impressions became 370 clicks and zero conversions. GA4 recorded 401 sessions from 374 people, so the clicks were real. The sales were not. In the same window, organic mentions in ChatGPT answers drove 552 sessions and five sales. Those visitors stayed 443 seconds on average, versus 37 seconds for paid visitors.
That gap does not prove every paid campaign will fail. It does mean the organic result cannot do the paid campaign’s homework. Here is the pilot I would commission instead: $150, then $200, then $150, with tracking checked before the first dollar leaves and a decision rule written before anyone gets attached to month two.
0. Before day one: establish the $500 ceiling ($0 in media)
This step costs $0 in media. It buys you a real constraint instead of a number on a slide.
Treat $500 as a proposed ceiling, not a promise the platform will honor. Self-serve buying runs through ads.openai.com in 89 countries with no account minimum, but each campaign has a $25 daily minimum, and OpenAI suggests opening bids around $3 to $5 per click. At $25, that is roughly five to eight clicks. One guide puts a more readable first test at $2,000 to $5,000 per month for four to six weeks. Your $500 can test whether the promise deserves more money. It cannot establish a scaling plan.
Before agreeing to the pilot, get the media ceiling, monitoring responsibility and pause authority in writing. A daily budget averages over seven days and can spend up to twice its amount in one day. If whoever runs the account cannot explain how they will keep billed media at or below $500 despite that behavior, do not call it a capped test. Resolve that problem before launch or skip it.
Price management separately. With groas, Search Ads management is $1,499 a month for Google, Bing and ChatGPT together, month to month, with no setup fee and a seven-day money-back guarantee. Spend the full $500 media cap and the combined commitment is $1,999. The guarantee is not a 30-day refund. I used to tell clients to roll management into ROAS math. I was wrong. Management is a fixed cost; media is the test cost. Write down both, and require written approval for any media dollar above $500.
There are two reasons to cut this test before it starts: you cannot trace a sale or qualified call back to its click, or the offer has never closed anywhere else. I have run tests under those conditions. They end in the same argument about bad traffic versus slow follow-up, and $500 cannot settle it. Fix tracking and prove the offer first. Do not pay to diagnose two unknowns at once.
1. Days 1–3: choose one offer and prove the tracking ($0 in media)
This stage also costs $0 in media. It buys you a test narrow enough to interpret. Pick one offer that already closes somewhere else, one landing page and one business outcome with a dollar value attached. Maybe it is a $149 plugin with a trial-to-paid path; maybe it is a booked job worth $400 on average. Write the outcome plainly: a sale or a qualified booked call counts; a click does not.
What gets cut is the multi-product split. Barn2 ran ads across four products. With $500, four offers divide the clicks and leave you arguing about which offer, page or conversation failed. One offer makes every click test the same promise.
Then run the tracking check yourself, or watch the person responsible do it:
- Fire a test purchase or lead in GA4 and confirm the conversion appears in the ad account.
- Follow the path from click to thank-you page on mobile. Check the page and the event, not just a reassuring dashboard label.
- Record organic ChatGPT referrals separately from paid visits before the pilot begins.
That last distinction matters. Barn2 recorded 552 organic ChatGPT sessions versus 210 paid sessions. Organic visitors viewed 2.18 pages and stayed 443 seconds on average, against 1.09 pages and 37 seconds for paid. Mix the streams and paid can take credit for visitors who were already arriving, and converting, without an ad. This is the tracking plan I want an agency to show before spend, the same standard I laid out in Stop Asking ChatGPT Ads Agencies for Results. Ask for Their Test Plan.
No clean test event, no day four.
2. Days 4–10: release $150 and look for disproof
Release $150, allowing for six delivery days at the $25 daily minimum within this seven-day window. At the suggested $3 to $5 per click, that buys roughly 30 to 50 clicks. Its job is to expose broken delivery, a bad destination or obviously wrong intent. It does not buy statistical confidence.
Anyone calling a 2% click-through rate an early win at this volume is reading tea leaves. Barn2 finished at 1.93% CTR with zero sales. A wider study of about 1,898 shops found paid ChatGPT visits converted at 0.37%, against 1.17% for paid Google search. Cheap clicks that leave quickly are not a result. They are a reason to inspect what those visitors thought they were clicking.
Set the stop rules now, while stopping still feels easy:
- Pause immediately if tracking breaks, the destination returns a 404 or consent error, or the conversation context is plainly irrelevant to the offer.
- Check billed spend against the $150 stage allowance during delivery. Do not assume the daily setting is a hard stop.
- Log every bid, context hint, page or pause change with its reason on the day it happens.
The context check matters because the available controls are thin: location, platform, free-text context hints and custom audiences. Inclusion audiences need 25,000 matched users, with 100,000 recommended, and ads serve Free and Go plans, not Plus, Pro, Business or Edu. You cannot wish a narrow audience into existence. You can inspect whether the traffic you do get belongs near your offer.
What gets cut is the week-three retrospective about why week one went wrong. A dated action log beats a polished explanation written after the money is gone.
3. Days 11–20: spend $200 on one consequential change
This stage costs $200, or roughly eight $25 delivery days within the ten-day window. At $3 to $5 per click, expect about 40 to 67 more clicks. It buys you one read on one change. It does not buy enough traffic to change bids, hints, creative and the landing page together, then claim to know which one helped.
Choose the variable from what days 4–10 showed. If visitors left in under 40 seconds, as Barn2’s paid visitors did on average, inspect and fix the first screen they saw. If the clicks came from conversations with the wrong intent, rewrite the context hint. Make one change, date it and say why. Leave the other variables alone for this stage. The standard agency reflex is to adjust everything and call the resulting movement learning. At this volume, it is mostly fog.
There is a reason to examine intent before celebrating price. In a 30-day B2B test, StubGroup reported a $0.76 ChatGPT CPC against $10.46 on Google, and a $92.21 cost per lead against $157.57. Yet ChatGPT converted at 0.8% against 6.64%, and Google leads closed at a higher rate. Another agency reported more than 150 qualified leads across three B2B SaaS clients at less than half Google’s cost per lead, using context hints matched to top-of-funnel ebooks, competitor comparisons and branded campaigns.
Both accounts can be true. A visitor willing to read an ebook is not necessarily ready to buy, and a low click price does not erase that difference. Your one change should address the mismatch you actually saw, not the metric that looks best in a report.
What gets cut is every other experiment and every request to raise the ceiling mid-pilot. One change, one reason, no increase above $500 without written approval.
4. Days 21–30: spend the final $150 and reconcile outcomes
The last $150 allows roughly six more delivery days at the daily minimum and about 30 to 50 clicks at the suggested bid range. Across the three stages, that is roughly 100 to 167 clicks if prices land in that range. It buys you a completed capped pilot, not a verdict on ChatGPT ads as a channel.
Now pull the CRM, not just the ads dashboard. How many paid visitors became qualified leads, booked calls or sales? Were leads worked? Is pipeline still open? Barn2 waited six weeks and recorded zero paid conversions against five sales from organic referrals. That is why this stage includes an outcome check rather than a victory lap over clicks. A lead nobody follows up is not proof the traffic failed. A worked lead that never closes tells you more.

Compare the outcome with the cost ceiling you wrote down on day one. The shop-level study found paid ChatGPT visits produced $0.89 in revenue versus $2.45 for paid Google search, implying a parity click price of $0.26 against the largest observed $1.72. You do not need to borrow that study’s ceiling. You need your own: media spent divided by closed customers is your pilot customer-acquisition cost. If nobody closes, you do not have a reassuringly small CPA. You have no closed-customer CPA to report.
At day 30, make one of three calls using the rules agreed before launch:
- Stop if paid produced no qualified pipeline, follow-up was sound and organic referrals kept producing.
- Continue if paid produced the agreed qualified outcome within your cost ceiling.
- Call the evidence insufficient if leads remain in an open sales cycle or follow-up broke the measurement. Fix the follow-up before deciding whether another capped run deserves money.
What gets cut is the automatic renewal disguised as a need for more data. Insufficient is a finding, not permission for an uncapped month two.
5. Before month two: demand the handover ($0 additional media)
The handover costs $0 additional media. It buys you the right to decide with records instead of optimism. I would require three things:
- Spend reconciliation: platform-billed media to the cent, delivery dates, the $500 ceiling and any overdelivery from daily averaging.
- A dated action log: every bid, hint, page and pause change, with the reason recorded when it happened.
- Downstream outcomes: leads worked, calls booked, sales closed, pipeline still open and pilot acquisition cost beside the ceiling set on day one.
If an agency cannot produce those, you did not run a test. You rented activity. It is the standard I hold my own work to with groas: changes logged with reasoning and a named strategist accountable for the outcome, not a retrospective story handed over after the invoice. If the $500 media cap plus management still feels heavy for what you learned, read DIY Google Ads or $1,499 Management? Price the Work, Not Just the Fee before deciding whether hands-on checking or done-for-you execution fits the next test.

No automatic budget increase. No promise that another month fixes the result. Put both in writing before day one, then honor the decision at day 30.
When the constraint lifts, the job changes from disproof to scaling. Keep the one-offer focus, day-of logging and written approval for budget increases. Raise spend only on an outcome you can trace to revenue; a more readable test may mean the $2,000 to $5,000 monthly range rather than another $500. At $4 a click, that range buys roughly 500 to 1,250 clicks instead of a handful of competing guesses. Until you have that evidence, a capped pilot that gives you a clean stop is not a failed experiment. It is the experiment working.

