MarketMirror 以 120 名模擬買家測試 8 則 Facebook 廣告，排名幾乎倒置卻找出註冊頁阻礙

Simulation tools love to show the run where they were right. This is a run where one was wrong, and why it was still worth doing.

The setup

csca.ai sells prep for the CSCA exam (the test for studying in China) to students and parents in South and Southeast Asia. They ran 11 Facebook ads for a free mock test across Nepal, Indonesia, Bangladesh, Vietnam and Sri Lanka, and let us use the real results.

Real results (Meta Ads Manager): $56.05 spent, 274,557 impressions, 4,466 clicks, 1,500 landing page views, 16 sign-ups.

Then we gave MarketMirror the 8 that got most of the budget, plus a brief, without the results. It built 120 simulated students and parents (8 audiences × 15 people, each with a short life interview), put the ads in realistic Facebook feeds three times over with different random seeds, and ran a survey on the sign-up offer.

Where it was wrong: clicks

We compared the simulated click ranking with the real one. Rank correlation: -0.84. That's not "a bit off", that's almost backwards. The simulation's favourite ad was the Nepal broad audience; the real winner was the Indonesia parents ad at 2.32% CTR.

If you'd used it to pick which ad gets the budget, it would have pointed you the wrong way. We'd rather say that out loud than hide the run.

Where it was right: the leak

Look at the real funnel again: 1,500 people reached the page and 16 signed up. About 1 in 100.

The simulated buyers had already said why, in their own words and their own languages:

"Gratis tapi minta password lagi? Jangan." (Free, but asking for a password again? No way.) (Indonesia)

"Phone number dhele spam asbhe." (Give them my number and the spam starts.) (Bangladesh)

"Free sign-up? They want my phone and email. Harvesting data for sure." (Nepal)

In the survey, trust ("looks like a scam") was the top blocker for 56% of the simulated people, and privacy for 28%. The real sign-up form asks for full name, email, phone number, country and a password, for a free test.

The ad wasn't the main problem. The page after it was.

What it suggested next

From the real ad data plus the run, MarketMirror's ideas step wrote five next tests. Two of them:

Zero-friction entry: a 3-question quiz first, no email, no personal data.
Privacy-first page: say plainly how the data is used, cut the form to 3 fields.

These are proposals, not results. The point is the order: fix the page before buying more clicks.

What we take from it

Don't use a simulation to rank ads by CTR. At least not until a backtest on your own past ads says it can. (MarketMirror has backtest and calibrate for exactly this.)
Do use it to find the "why". Objections, in buyers' own words, by audience, before you spend.
Real data in, next test out. The useful loop is: export real results → compare → test the fix on simulated buyers → run the small real test.

Try it

MarketMirror is open source (AGPL-3.0) and runs from Claude Code, Codex, Cursor, Gemini or any MCP agent.

👉 https://github.com/DarkWorldCoder/market-mirror: if it saves you one bad launch, a ⭐ helps other people find it.

Real ad results shared with csca.ai's permission. Simulated buyers are synthetic; their quotes are labelled. Written with help from an AI assistant; every number was checked against the saved data.