Steal my $750 ChatGPT Ads test plan

Every week, someone replies to this newsletter with the same question: "OK, but do you think ChatGPT Ads will actually work for MY business?"

Sure there are guidelines of what will probably work…

…but the only way to know is to test it for yourself.

For $25/day, you can stop guessing and get a real answer in 30 days.

Why most "tests" don't actually tell you anything:

• No baseline written down, so "$4 clicks" doesn't mean anything without a number to compare it to.
• Too many context hints or landing pages changed at once, so nothing gets isolated.
• Touching the campaign mid-test, which resets what the matching system has learned.
• Watching CTR instead of conversions, which is the wrong signal on this platform entirely.
• Giving up around day 10, before the system's had time to learn anything.

Simple test: Set a daily budget cap in OpenAI Ad Manager at $25/day.

3 steps, 30 days, $750 total.

Doing it right = buying your answer.

Step 1: Set the test up right (days 1-3)

The common mistake here is skipping the boring setup and just turning campaigns on. Don't.

• Write down your baseline first. Open your Google / Meta dashboard and record 3 numbers per channel: average CPC, conversion rate, CPA. Skip this and "$40 conversions" means nothing without "$190 on Google" next to it.
• Install tracking on day 0: the pixel plus the Conversions API, documented at developers.openai.com/ads. Testing without tracking is just donating $$$ to OpenAI. Get AI to help you (connect Web Browser access + MCPs… if you need help then email me)
• Build one campaign, 3 ad groups, one context hint each:
  • The buying conversation: "when users compare [your category] for [your niche] and have mentioned budget or timeline"
  • The problem conversation: "when users describe [the pain you solve] and ask what their options are"
  • The wildcard: written in your customers' own words, stolen from support tickets and reviews
• Write 5 killer ads. You have far less real estate here to make an impact so make it count. Those should map directly to the context hints for each ad group
• Use one landing page, your best-converting existing one. Building new pages for an unproven channel is how tests die of scope creep.

Quick note on budgets: as of June 9, 2026, OpenAI treats daily budgets as averages. They can spend up to 2x your daily limit on a single day, as long as weekly spend stays within 7x. Plan accordingly.

Why 3 hints? OpenAI is explicit that context hints "are not exact-match keywords and do not guarantee delivery."

3 hypotheses give the model 3 doors to walk through, and give you a real comparison at the end.

Step 2: Let it run and touch nothing (days 4-20)

This is the part that wrecks most tests. The matching system needs data to learn which conversations fit your ads, and every edit resets that learning. Set it, then sit on your hands:

• Expect impressions to trickle, not flood. Your traffic here is long-tail by design, so day 5 will look quiet next to a Meta launch. That's normal.
• Ignore CTR. Early industry reporting puts ChatGPT ad CTRs well below search benchmarks for everyone, and conversions are what matter here.
• Compare everything against your baseline sheet, not against vibes.
• Check twice a week, no more, or you'll start "fixing" things that were busy working.

Quiet and converting beats loud and broke.

Step 3: Read the verdict and make one move (days 21-30)

By day 21, the data is STARTING to tell you something, but you’re not fully there yet.

3 possible outcomes, with a next step:

• Converting at a CPA below your current channels → scale it. Kill the other 2 groups, write 3 variations of the winner, raise budget slowly while CPA holds.
• Clicks but no conversions → your landing page is the suspect. The ad promised a conversation and the page is shouting like a billboard. Fix the page before blaming the platform.
• Barely any impressions → your hints are too narrow, your budget is too low or your category sits in restricted territory (more on that in an upcoming issue). Widen the hints one notch and rerun - at least another 15 days.

Run the test for yourself (and stop asking people who haven't)

Whichever result you get, you bought a real answer for $750. Most business owners spend $25,000 asking an agency to guess, and the agency hasn't run this platform either.

The window for cheap tests closes as the auction fills up, so run yours while a wrong guess only costs lunch money.

Run the test, then reply with your numbers. I read everything, and I'll feature the most interesting results (anonymized) in a future issue.