A $750 ChatGPT Ads test can produce a cheap-looking CPA and still tell you almost nothing. Before I approve one, I want three things settled: what counts as a conversion, who can approve more spend, and what makes the test stop.

This glossary covers twelve terms I would put in the proposal before a dollar goes live. Each starts with plain language, then the distinction that matters when your AI pulls reports and runs the work through groas. The point is not to learn advertising vocabulary. It is to give a small budget one question it can actually answer.

The starting math is simple. OpenAI lists a minimum campaign budget of $25 USD per day, with equivalent floors in other currencies. Thirty days costs $750 before management fees. At a $4 click, that buys about 190 clicks. At a 2–6% conversion rate, that is roughly 4–11 conversions. A small test can mean low evidence, not just low risk.

A small test-budget coin disappearing into a large chat window

Your AI reads the reports; groas supplies the connection that lets it act on ChatGPT Ads, Google Ads, and the rest from the same chat, with every change logged and undoable. That connection only helps if the words mean the same thing to you, your agency, and the machine.

Platform access

Platform access means your billed business is eligible to run ChatGPT Ads in its country, with the audience coverage you expect.

An agency saying it can test ChatGPT Ads is not proof that your account can launch. The distinction is between access to a service and access to the inventory you are paying to reach.

The draft proposal needs to name three things:

  • The billing country of the business running the ads.
  • The ChatGPT plans whose users can see them.
  • The date the account can actually start spending.

Current availability and plan coverage matter more than a slide about the size of ChatGPT’s audience. Ads shown to logged-in adults on Free and Go do not reach people using ad-free Plus, Pro, Business, or Enterprise plans.

My rule: have your AI confirm account eligibility before you approve spend. A proposal is not a launch confirmation.

Management fee

A management fee is what you pay someone to run the test, separate from the money spent on ads.

One agency makes the split explicit: LeadGulls charges $1,500 per month, with ad spend billed directly to the client’s OpenAI account. That separation belongs in every proposal. Otherwise, “a $750 test” can quietly mean something much more expensive.

For running the work through your own AI, groas Unlimited is $19 a month, flat, for a whole team. The tool cost stays separate from platform spend, so the media math stays readable.

I want two costs shown, not blended:

  • Platform CPA: media spend divided by recorded conversions.
  • All-in cost per qualified lead: media spend plus management or tool fees, divided by qualified leads.

Tell your AI to report both. The cheaper-looking number does not get to stand in for the cost to the business.

Media spend

Media spend is the money paid to the advertising platform for ad delivery.

Not management. Not creative. Not the landing page build. Those costs matter, but they do not belong in the media column.

Say you spend $20k a month on Google and test $750 on ChatGPT Ads. The test is 3.75% of that Google spend. It cannot rescue the quarter. It can answer one question about one audience if the tracking holds.

My instruction to the AI is deliberately boring: keep the media envelope visible, report spend against it, and never pull fee dollars into that column. For a separate comparison with Google costs, use the 2026 CPC and budget benchmarks by industry. Do not let the comparison quietly expand the test.

Test budget

A test budget is the maximum media spend approved to answer one question within named dates.

No dates and no cap means no test. Just spending.

At $25 per day, a 30-day test has a $750 media envelope. A practitioner cost roundup argues for $50 per day over 30 days, or $1,500, as a more readable starting point. Neither number makes the result useful by itself. The question does that.

Here is the shape I want in writing:

Spend no more than $750 over 30 days to test whether Free and Go traffic requests quotes at under $90 per qualified lead.

That gives the AI an envelope, a duration, and a business standard. It also gives you something to reject when someone proposes three new audiences halfway through.

Freeze new ideas until the test closes. A test that can grow mid-flight is a retainer wearing a costume.

Budget guardrail

A budget guardrail names who must approve an increase and how that approval is recorded.

It is separate from the spending controls inside the platform. I want both: a platform setting and a written authorization rule.

That distinction is not paperwork for its own sake. A roundup of practitioner threads includes complaints about lifetime budgets pacing poorly compared with daily budgets. A written ceiling tells everyone what to do even when platform behavior gets untidy.

For a $750 test at $25 per day, the instructions can be short:

  • Alert at 50% and 80% of the approved envelope.
  • Do not raise the daily budget or total envelope without written approval.
  • Name the person who can approve it and where that approval must appear.

Tell your AI those rules before it runs the work through groas. An alert is not permission to spend more. Neither is “performance looked promising.”

Conversion

A conversion is the action your tracking tells the platform to count.

Nothing more. The dashboard does not know whether the event deserves applause.

ChatGPT Ads measurement uses an oaiq pixel, an oppref click ID, and a server-side Conversions API. Pixel and server events need a shared event ID for deduplication. Those are implementation details, but the business question is simpler: does the event fire when someone submits a quote request, or merely when they view the quote page?

Get that wrong and your AI can faithfully report success that never happened.

Before launch, I want the reporting instructions to name:

  1. The exact action counted as a conversion.
  2. The pixel and Conversions API setup used to record it.
  3. The tracking failure that requires spending to pause.

Name the action, not just the event label. “Lead” is a label. “Completed quote submission” tells me what happened.

Qualified lead

A qualified lead is a recorded conversion that passes your business’s sales filter.

That might mean the right service, the right area, valid contact information, and intent to buy. Platform conversion counts do not apply that filter for you.

Early practitioner reports describe cheaper clicks for some local services alongside complaints about low-quality leads and clicks without conversions. Cheap traffic and useful demand are not synonyms. I would rather settle that distinction in the proposal than argue about it after the invoice.

A qualification rule might be: service within the coverage ZIP codes, valid phone number, project over $500. The exact filter belongs to the business, not the advertising dashboard.

Tell your AI through groas to show raw conversions and qualified leads side by side. Then show their respective costs. Do not relabel the platform’s raw conversion count as qualified just because both fit under the word lead.

Attribution window

An attribution window is the period after an ad click or view during which the platform can credit a conversion to that ad.

ChatGPT Ads offers click windows of 7, 14, or 30 days, with 30 as the default, and view windows of 0 or 1 day, with 1 as the default. Click attribution takes precedence when both qualify.

This is where two reports can appear to disagree without either containing a calculation error. General Insights counts click-through conversions; the dedicated conversion report includes 1-day view attribution by default. Change the window or report, and the same spend can show a different CPA.

For this test, my instruction is:

Use 30-day click and 0-day view for the decision-making report. Show 1-day view separately, not folded into the headline result.

Lock the window before comparing CPAs. Otherwise, you may be comparing reporting settings rather than performance.

Conversion lag

Conversion lag is the delay between an ad click and the recorded conversion.

It is not the same as the attribution window. The window sets the eligibility rule; lag describes how long the result takes to arrive.

There can also be reporting delay. Practitioner reports describe attributed conversions taking 24–48 hours to appear. Judge today’s spend against only today’s visible conversions and you can make a working test look broken.

My rule for the first assessment is to give every dollar seven full days to convert, then show spend by click date beside conversions by conversion date. That waiting period does not change the 30-day attribution window; later conversions can still change the report.

Stopping spend and judging the result are separate decisions. The budget can close on schedule while the evidence catches up. A daily CPA chart without that distinction is a panic generator.

Cost per acquisition

Cost per acquisition, or CPA, is spend divided by the number of acquisitions you have explicitly defined.

The denominator is the argument. Is an acquisition a recorded form submission, a qualified lead, or a customer? Name it before calculating anything.

ChatGPT Ads can report CPA without billing per conversion. A reported outcome metric is not a promise that each outcome costs that amount. Likewise, a Target CPA is an average-cost goal, not a per-auction cap.

I keep two clearly labeled calculations in the report:

  • Platform CPA = media spend ÷ raw conversions.
  • All-in cost per qualified lead = media spend plus fees ÷ qualified leads.

On a $750 test with a $1,500 management fee, 10 recorded conversions produce a $75 platform CPA. If all 10 qualify, the all-in cost per qualified lead is $225. Same test. Very different number to take into a budget meeting.

Tell your AI which measure decides the next budget. Do not let it substitute the cheaper one.

Stop rule

A stop rule is a written if-then instruction that pauses spending before hope consumes the remaining budget.

I want four in a small test, with the conversion-lag rule applied before judging lead quality:

  • Budget spent: pause at the approved $750 or $1,500 envelope. No rollover without new approval.
  • Tracking broken: pause if the pixel or Conversions API stops firing, or required oppref IDs go missing.
  • Lead quality fails: pause if 100 clicks produce zero qualified leads after the agreed waiting period, even if raw conversions look fine.
  • Evidence too thin: stop spending at the end date. If conversions remain under five after the waiting period, call the result unreadable, not promising.

These are instructions for this test, not universal performance benchmarks. Their job is to settle the decision before everyone becomes emotionally attached to the campaign.

Tracking deserves its own stop condition: practitioner threads describe conversions appearing as direct traffic without oppref or pixel data. If the measurement breaks, spending more does not repair the evidence.

Give your AI explicit commands through groas. “Keep an eye on it” is not a stop rule.

A kitchen timer wired to a pause button above a small stack of coins

Incremental lift

Incremental lift is the extra revenue or qualified leads the ads create that would not have happened without them.

Attribution counts credited touches. Lift asks what changed because the ads ran. Those are different questions, and a small dashboard does not magically answer the second.

This is the term I would retire from small-test decks unless someone names the comparison. Four attributed conversions do not tell me how many customers would have arrived anyway. Neither does a tidy CPA.

My reporting instruction is blunt:

Report attributed conversions by window, then state what the test did not measure. No holdout or geographic comparison means no claim that it proved new demand.

Before approval, put the other definitions beside that instruction: access, fees, media envelope, dates, approval owner, conversion event, qualification filter, attribution window, lag allowance, decision-making CPA, and stop rules. Have your AI use those definitions when it pulls reports through groas.

Then choose one of the 7 small-budget ChatGPT ad tests, each with one question and a kill rule.

A small budget can answer one question well. These definitions decide whether you hear the answer. If an agency calls attribution lift, ask what would have converted anyway.