A $1,500 ChatGPT ads test should answer one question, not buy you “early presence.” If you run a small business and want an agency to handle the test, these nine rules help you find out whether the ads produce a real inquiry, booking, or sale at a cost you can live with. No conversion signal, no reason to scale.

1. Cap the entire test before you buy a click

Put the total media budget, daily limit, start date, end date, and approval owner in one written plan. A daily limit alone is not enough: ChatGPT delivery can be bursty, and a day can spend up to twice that limit under seven-day average pacing. Reporting can lag behind the spend. You do not want to discover a spike by reading yesterday’s dashboard.

The $1,500 scenario is workable as a capped experiment, not a promise of a decisive result. Self-serve buying opened May 5, 2026 without the old pilot’s roughly $200,000 lifetime minimum. The USD floor is $25 per day per campaign. At a suggested $3 to $5 max bid per ad group, two weeks at $25 a day amounts to $350, or roughly 70 to 115 clicks if clicks come in at those prices. Clicks are not the answer you are paying for.

I used to tell clients a verbal cap was enough. I was wrong. If the cap is not written down, it moves.

First step: email the agency your $1,500 total, your chosen daily limit within the $25 to $50 range, and your four-week test window. Ask: “Will you confirm both caps in writing and make no increase without my written approval?” If they hedge, end the call.

2. Choose one result that would make the spend worthwhile

Pick one downstream measure before the agency shows you a dashboard: a booked job, paid signup, or qualified inquiry. Not CTR. Not impressions. Not what the deck calls engagement.

Here is the mechanism: an ad click does not necessarily become a site session, and a platform conversion does not necessarily become a customer. In one B2B comparison, an account recorded 36 Ads Manager conversions against two real inquiries. SE Ranking’s first-hand check covered three campaigns, eight ad groups, 48 ads, and 1,263 clicks, with almost no sign-ups. Those figures do not predict your result. They show why an attention metric cannot settle a revenue question.

First step: write a pass condition the agency can read without interpreting it. For example: “This test wins if we get 10 qualified bookings at or below $150 each.” Define qualified before launch, too. One success metric means one answer.

Ask: “What single result are you willing to be judged on?” If they offer three metrics and a story about brand lift, end the call.

3. Write the stop rule while everyone is still optimistic

A cap limits the money you can lose. A kill criterion limits how long you can explain away a weak result. Decide both before the first click.

At $3 to $5 per click, $1,500 buys roughly 300 to 500 clicks if it all spends at those prices. At a 1% site conversion rate, 1,000 clicks would produce about 10 conversions and cost roughly $3,000 to $5,000. That is why a smaller test may give you a useful warning without giving you statistical certainty. A cleaner read can require far more conversions than this budget is likely to buy.

Your stop rule should use the result you chose in Rule 2 and allow for the reporting lag. For example: stop if 300 clicks produce zero qualified inquiries with tracking verified, or if cost per qualified inquiry remains above $200 after 400 clicks. These are example thresholds, not universal benchmarks. Choose the numbers your business can tolerate.

First step: put the kill number in the same email as the budget cap. Ask: “What result would make you tell me to stop spending?” If they cannot name one, they can always sell you another week of testing. End the call.

4. Send the click to a page that answers the buyer’s question

Build one test page around one buyer situation. Give the visitor the price range, timeline, what is included, who the offer fits, and how to book. Do not make them hunt through a homepage for the answer.

ChatGPT targeting uses conversational context hints about situations, not a Google-style keyword list. A hint might describe someone looking for cushioned shoes for a first 5K rather than simply say “running shoes.” Those hints do not guarantee delivery, and you do not get a query or conversation report to mine afterward. The page therefore has to do more work without the usual search-term feedback loop.

The ad appears below the answer, labeled sponsored; it does not change the answer itself. The reader has already seen a summary and chosen to click. Make that choice pay off. For an emergency plumber, the first screen should show response time, callout fee, service area, and a tap-to-call button. A generic homepage makes the visitor start over.

First step: name the buyer question and draft the page’s opening around it. Ask: “Will you build a dedicated page for this test?” If their plan is to send every paid click to the homepage, end the call.

5. Test alongside Google and Bing, not at their expense

Give ChatGPT its own budget and traffic tag. Leave your existing Google and Bing search running so you can compare cost per qualified inquiry on the same sheet.

ChatGPT ad reach remains a slice of its audience. Ads began as a US test for logged-in adults on Free and Go plans and now run in more than 60 countries; several account types, under-18 users, and Temporary Chats do not see them. Pausing proven search to fund that slice turns a $1,500 experiment into a possible revenue hole.

Early same-advertiser comparisons also show why a baseline matters: one compilation reports ChatGPT conversion rates of 0.21% to 0.9%, against 3.71% to 6.64% for Google. Your offer may behave differently. That is an argument for measuring it beside the search channels you already know, not abandoning them on the strength of a launch deck.

First step: ask for one weekly comparison of spend and qualified inquiries across all three channels. Do not let an unproven test replace a working baseline. If the agency wants to pause search to pay for ChatGPT, end the call.

6. Give the test four weeks, not four promising days

Set the dates up front and give each week a job:

  1. Week one: confirm delivery and tracking.
  2. Week two: compare two or three variants.
  3. Week three: refine the better performer without raising the budget.
  4. Week four: judge the result against your Google and Bing baseline.

That four-week shape protects you from making a budget decision on a screenshot. Impressions, clicks, and CTR can refresh in about 15 minutes, while spend can lag seven to eight hours and conversions 24 to 48 hours. Add uneven delivery and a small sample, and day two can look much more certain than it is.

I have killed small Google tests too early because CPC looked ugly on day five. A few expensive clicks are information. They are not a verdict. The same goes for a few cheap ones.

First step: lock the four-week window and agree that no scaling decision happens before the final read. You can still stop for a verified tracking failure or a pre-agreed kill condition. Ask: “Will you put that decision schedule in writing?” If they want to raise spend after a few good days, end the call.

7. Require a fresh written yes before any budget increase

The original cap is not permission to raise it later. Make approval a separate rule: no higher daily limit, no larger total, and no extra campaign budget without your explicit written yes before the additional spend.

This matters even when the agency is actively managing the account. Continuous adjustments to bids, creative, or allocation inside an approved cap are different from making the test bigger. The former may improve the read. The latter changes the amount you agreed to risk.

ChatGPT account setup has its own sharp edges: the billing entity must be in a listed country, some account settings are permanent, and new accounts can face delivery limits until verification and home-market spend requirements are met. Those are reasons to plan carefully, not reasons to authorize a rushed increase when delivery disappoints.

groas manages Search Ads across Google, Bing, and ChatGPT on a flat fee, with around-the-clock work on bids, budgets, and landing pages. That is the operating model I want here: active management inside the boundary you approved, with a human accountable for changing that boundary.

First step: add one sentence to the agreement: “Any budget increase requires my written approval before it takes effect.” Ask whether the agency agrees. If the answer is some version of “trust us,” end the call.

8. Ask for a change log, not a prettier dashboard

Have the agency send a short weekly email with spend, real inquiries, and each change made with its reason. For every edit, record the date, what changed, why it changed, and its effect on spend. Include bids, context hints, creative, pages, and caps. If nothing changed, say so.

This is not paperwork for its own sake. ChatGPT offers aggregated reporting rather than a query or conversation report, so you cannot inspect targeting the way you would inspect Google search terms. Delivery can also concentrate unevenly: in SE Ranking’s test, two ad groups took about half the impressions. Without an action log, “we optimized performance” can mean almost anything. Usually it means you need to ask another question.

I treat a missing change log much like missing conversion tracking: I cannot judge work I cannot inspect. A weekly PDF of impressions and CTR does not fix that. If you want a fuller reference for management responsibilities, use this managed Google Ads checklist to pressure-test the agency’s answer.

First step: request the log format before launch, not after the first report disappoints you. Ask: “Will you send it every week, including weeks when nothing changes?” If they will only report outcomes and not decisions, end the call.

9. Prove the conversion before spending the first dollar

This is the rule I would start with, even though it comes last here. Install the OpenAI Pixel and Conversions API, make sure they are deduplicated with a shared event ID, and verify that the oppref click ID survives redirects. Otherwise, the ad click may not tie back to the lead you care about.

Then separate paid ChatGPT traffic from organic chatgpt.com referrals with its own utm_source. Expect Ads Manager and GA4 not to match perfectly: Ads Manager can include modeled conversions, and the systems report on different timelines. The point of the check is not to force identical totals. It is to prove that a real test conversion appears in both places before your budget becomes the tracking test.

The risk is not theoretical. One small-business account spent about $500 in just over an hour for 140 reported clicks and zero leads; only about 13% of those clicks were verifiably seen reaching the site, while the dashboard lagged. That account alone cannot tell you why the numbers diverged. It can tell you how little comfort a click total offers when you cannot trace the visit and the inquiry.

First step: have the agency show you a live test conversion in Ads Manager and GA4 before launch. Ask: “Will you spend nothing until I have seen it in both?” If the answer is “we can fix tracking afterward,” end the call. A $1,500 test with clean attribution can tell you what to try next. Without it, all you know is that clicks happened.