Built the tracking end to end. Reconciled every platform number against real payments. Made every decision on cash.

02 Hypothesis

What each launch was actually testing

Act 1 (Ukraine)

• Does the tripwire mechanic work on cold traffic at all?
• Which placements produce paying customers — tested as two separate campaigns (FB+IG together vs Instagram-only) rather than editing one, to keep the before/after data clean and avoid resetting Meta's learning phase.
• How does the upsell sell? On Ukraine it was pitched by direct message right after purchase.

The upsell hypothesis failed: direct pitches for the big course met silence. The takeaway — a buyer has to warm up and reach out themselves, which needs a nurture sequence, not a cold pitch. That became the main change for launch two.

Act 2 (Poland)

Does the funnel carry to a new, colder, more expensive market (diaspora: higher CPM), with a setup corrected for the Act 1 lessons — Instagram-only from day one, an age focus, a clean attribution signal (strict 7-day click / 1-day view, all AI creative enhancements off, manual placements) — and a bot rebuilt around multi-step warm-up instead of direct selling?

Two further questions inside the launch, not raised on Ukraine: which placement inside Instagram (Reels / Feed / Stories) buys intent cheapest, and which of three creative angles resonates. Both were answered by campaign structure — three creatives in one ad set, letting Meta's delivery optimization decide, instead of splitting a thin budget across an A/B test.

03 Setup

A tracking stack built so campaign decisions could trust the data

Top layer: I built and owned the full stack — landing page, analytics, and conversion tracking with a CAPI safeguard (GA4, GTM, Meta Pixel + one-click Conversions API) — so decisions rested on data I could verify.

The technical layer

• Static landing on Cloudflare Pages. Page weight cut from 7.4 MB to 636 KB, LCP 2.09s (green Core Web Vitals) — ready for paid traffic.
• One GTM container, two mirrored receivers: GA4 for funnel analytics, Meta Pixel for the ad signal. Event funnel PageView → ViewContent → InitiateCheckout → Purchase on shared triggers. The checkout trigger fired on the payment-processor URL, not the button text, so copy edits couldn't break it.
• One-click CAPI as a safeguard against lost browser signal. It paid off: on Poland, Meta's Purchase events matched the payment processor 7 = 7 — closing the measurement gap diagnosed on Ukraine.
• Telegram bot (SendPulse) rebuilt: product delivery + a multi-step warm-up + engagement (chat with the author) + a buyer-source step — a third attribution layer alongside Meta and WayForPay.
• UTMs via {{ad.name}} / {{adset.name}} macros — GA4 saw traffic per creative with no manual tagging. Tagging and purchase tracking were incomplete on Ukraine; fixed on Poland, which is what opened the per-creative breakdown.

Campaign structure

Ukraine: two Sales / ABO campaigns optimizing for InitiateCheckout — FB+IG together vs IG-only — to isolate the placement variable. Landing and payments sat with the other contractor.

Poland: Sales / ABO, optimizing for InitiateCheckout (too few purchases to optimize on Purchase directly), $25/day, broad targeting with hard limits in Controls (geo, age 22+, languages RU+UA), 3 creatives on different angles in one ad set — natural delivery optimization instead of splitting the budget on an A/B test.

06 Learnings

What held up, and what I'll name as a limit

What held up

• Platform metrics can't be taken on faith. Reconciling against the payment processor isn't paranoia, it's basic hygiene — the 23% pattern replicated across both launches.
• Acquisition worked: 22 paying customers at $22.77 each, every one confirmed in the payment processor.
• Separate campaigns beat editing a live one: clean data boundaries, no learning resets. And because the campaigns ran sequentially — the mixed one Mar 26–30, IG-only from Mar 31 — auction overlap was practically absent by construction: the second launch was the implementation of a decision, not a parallel test.
• CAPI closed the tracking gap: purchases were missing in Meta on Ukraine, and matched the payment processor 7 = 7 on Poland.

What I'll call a limit, plainly

• Monetization didn't close: 0 upsells from 22. The warm-up was rebuilt between acts, but 7 Poland buyers and a short window before the project stopped aren't enough to validate the new mechanic. Monetization stayed unvalidated, and I say so directly.
• Too many variables changed between acts (market, landing, offer, stack, bot mechanic) — the Poland effect can't be isolated from the new-landing effect. With more budget I'd move one variable at a time.
• CPA rose from Ukraine ($15–20) to Poland ($35.44): the diaspora is colder and more expensive than the home market. A fact is a fact — a Poland buyer cost roughly twice as much.
• Placement-level ROAS (FB vs IG) on Ukraine can't be computed honestly: payments sat in the processor with no placement split, and there was no deduplicated CAPI. So the Facebook call rests on bot attribution, not a ROAS figure — there is no number in this case I can't defend.
• Learning Limited at these budgets is an accepted state, not a mistake: exiting it would have cost more than the monthly budget.

07 If I'd had 10× the budget

Where the next dollar would go

Not a wishlist — a direct read of where the funnel actually broke.

• The upsell step as the main object of work. The bottleneck is found; now fix it: A/B the bot warm-up sequences (length, engagement steps, chat with the author), step-by-step analytics inside the bot, and test selling the high-ticket tier via a call request instead of a payment button.
• Server-side CAPI via the payment-processor callback, deduplicated by event_id (orderReference), so Meta optimizes on real payments rather than inflated checkouts.
• Switch optimization to Purchase once volume reaches ~50 purchases/week; until then, InitiateCheckout as the proxy.
• One variable per iteration: lock the landing and offer first, then test geo, then creative angles (developing the winning "wow-result" format in Reels) — and kill criteria for every test fixed in advance, before launch.
• Systematic tripwire → upsell conversion as the core economic metric (current base 0/22 — now a baseline exists, and a rebuilt warm-up to test against it).
• Standardized UTM tagging and a weekly spend → cash cohort report — the same reconcile-to-money principle, but routine instead of one-off.