Signal Engineering: Strategic Data Filtering for Better Ad Performance — Thomas Petit, Independent Consultant

Audit starts in Meta Events Manager / Google Ads Goals, not Ads Manager. For the last 7 or 30 days, compare total installs, free trials and paid subscriptions the platform received (not attributed ones) against RevenueCat, Amplitude or Mixpanel before sending. ~5% gap is fine; above 10% must be fixed (change source, parameters or the event itself). Watch for pitfalls: double sources that Meta deduplicates but Events Manager shows twice; missing parameters so only opted-in (IDFA) users are attributed (one change left only 20% of conversions reported). Once fixed, it can run 6-12 months but must be rechecked.

Default Meta SDK auto-mapping often reports a direct purchase, a trial converting to paid, and a renewal all as 'Purchase'. Map each individually, verify each passes correctly, then decide which combination to optimize toward.

When trial-to-paid from paid traffic trails other sources or its own past, run a regression on non-converting trials. Hard factors: device, OS version, language. Soft factors: onboarding answers (stated goal, intent, e.g. 'just browsing'). Keep the raw trial event unchanged; engineers add a new event (e.g. mapped as 'subscribe' or 'purchase') fired via the MMP (Adjust, AppsFlyer, Singular) or platform SDK only when a trial meets the criteria, and campaigns optimize to that. Most predictive factors across apps: age, goals, and consumer vs prosumer/solopreneur vs business. Advanced: add onboarding questions designed to create this variance. Exclusion is a last resort; skip it if trial quality is already fine.

Meta's value rules (formerly bid multiplier, now out of beta) let you weight segments: e.g. +20% for Google Pixel or newest Samsung Galaxy (5-10% of Android, convert like iPhones), don't show ads to very old Android versions, value under-25s ~30% lower instead of excluding everyone under 25. Guest combines this with qualified-trial filtering ('hammer from both sides').

Default SDK revenue for subscriptions is wrong: trial conversions arrive too late and yearly plans look better than weekly ones that renew for months. Instead send an engineered value per user: predicted value at a fixed day (guest likes month 13, which includes the first yearly renewal and compares monthly vs yearly fairly), set by plan, device, onboarding answers, refund-risk behavior and so on. Find the value drivers with a regression over high/low/zero-value users. Skip this if revenue has little variance across users; go straight to it if variance is extreme (e.g. $1 IAP vs $20/month vs $200/month users).

Rather than hardcoding 'plan X on device Y with onboarding answer Z = 56', the app/SDK calls an internal API with the user's parameters and gets the current value (56 this week, 68 next). Needed when plans, promos and currencies change often, especially volatile currencies like Turkey. Adjust per destination (Firebase, Meta SDK, MMP, RevenueCat). Engineers found it simple once the request was clearly specified; the delay was communication.

Optimize toward something that happens on day one, not a 7- or 14-day outcome (shortening the trial to 3 days doesn't solve this). Optionally hold the event back 1-2 hours and fire it only for trial starters who show engagement, e.g. three meditations/workouts within 2 hours. Don't wait too long.

Networks need roughly 50-100 events per week per campaign; guest uses 10/day/campaign as the floor. Below it, don't filter trials; instead add earlier high-intent events (strong onboarding answers, first activity) to raise volume. Well above it, sending 30 filtered events beats 300 unfiltered. Revisit as budget grows (e.g. $100K to $500K allows much tighter filtering).

E.g. report $5 users as $3 and $50 users as $100 to steer the network harder toward valuable users. Trade-offs: in-platform ROAS becomes unreadable, so read ROAS internally; pushing too far makes delivery swing to extremes. Guest has made mistakes doing it.

Early on, send 100% of each event to every network to validate data. Then map filtered events under separate event names; what works differs by network (e.g. revenue on Meta, qualified trials on Google, direct purchase + trial conversion on TikTok; filter young-audience trials for TikTok but not Meta). Sources may differ (Firebase for Google, Meta SDK, MMP). Keep alternative signals ready and run e.g. a trial-optimized campaign alongside a value-optimized one for audience expansion and as a fallback.

Track a curve of value sent to networks vs realized value; a widening gap signals a problem. Don't trust the total (2% overall deviation can hide 10% of users causing 80% of it); check at country and platform level. Revisit value assumptions about every 6 months as plans, channel mix and budgets change. In one client the value window is the last 2 weeks if the sample is big enough, else 3 months, else drop one criterion (country, onboarding answer) to avoid noise from small samples.

In 2017, Google Ads trials converted to paid at ~15% vs ~40% on organic and Meta. Identified reasons Google-sourced trials weren't converting and sent Google a different, filtered trial signal.

Sent weighted signals by plan and user type: weekly subscription from a normal user = 1, yearly from a business user = 5. Multipliers were hardcoded, then pricing was iterated for a year and a half, so they no longer matched real user value; fixing them to track real value became a priority.

A fairly young app changed the events it sent to ad networks while also working on creatives. Details not given.

Scaled on a couple of platforms, relying on one setup. When performance fell, creative and event fixes didn't restore it and there was no alternative ready. Guest's lesson: keep other properly mapped signals available and test them on the side.

Client already had a fairly sophisticated in-house signal system; handed optimization to Voyantis, which filters events and sends them back to the networks itself (including event-timing tricks). Running two tests.

In a mental health app, users who report a high anxiety level in onboarding are much more valuable, so that answer is factored into the signal sent to networks. Mildly lower-value answers are not excluded (handled via value optimization or value rules instead); only near-zero intent answers like 'just browsing' are filtered out.

At one client, users who finished a 30-screen onboarding in under a minute (clicking next, starting a trial, cancelling) had little value, so completion time was used as a criterion. Checked at other apps: there, the fastest finishers were the best users because they knew what they wanted. Not universal.

Client optimized toward predicted LTV, then realized cash flow mattered more and that different users maximize 3-month vs 5-year returns; switched the value sent to networks to maximize short-term payback.

AI app with usage costs offers tiered subscriptions plus credits bought on top, so users who hit a limit can add $5-10 of credits instead of upgrading a full tier. A team of several people spent months designing it to avoid cannibalizing subscriptions and keep pricing fair.

A not-big app that had just crossed the App Store Small Business Program threshold (fee rising from 15% to 30%) tested app-to-web checkout, against the guest's general advice.

Users who don't convert within 24 hours but show high likelihood of retaining as freemium users or converting later get a small value (~$1, vs the $10/month or $90/year plan values) instead of zero, so networks don't learn to avoid them.

A gaming company reports whale players (worth $500-1,000) as medium-value users, because reporting the full value makes Meta 'done for the day' after one big user. Guest notes it's rarer in subscriptions but could apply to very high tiers (e.g. reporting a $200 plan as $100).