How Do You Know Your Growth Is Real and Not Just Stolen Credit?
You hired a UA agency, spend went up, installs went up, the dashboard said ROAS was positive. Then you cut the agency and installs dropped by exactly the budget you removed. That isn't growth — it's a receipt for money you spent. Last-click attribution is why most "profitable" channels are quietly unprofitable. Here is how to tell real from credited.
Q: Why does last-click attribution lie to you?
Because everyone claims the last click. Self-attributing networks — the big ad platforms — take credit by default for any conversion that happens near their impression, even when the user would have installed anyway from a friend, a search, or a catalog they already trusted. In a multi-touch world, last-click just hands the win to whoever was loudest at the finish line. The result: you overpay the channel that closes and starve the one that opened.
Q: What is incrementality, and how do you actually run it?
Incrementality measures the one thing that matters: would this conversion have happened without the spend? You find out with a holdout — a randomized slice of your audience, or a matched geo, that gets no ads while the rest does. The difference in conversion between the two groups is your true lift. Geo holdouts are the cleanest: pick two similar regions, spend in one, watch both. If the control region converts the same, your "campaign" added zero. A PSA creative as a placebo control is the gold standard for creative incrementality.
Q: Why do cross-promo networks double-count?
Because the network sees the user everywhere and credits itself everywhere. If one player moves across three games in your catalog and the cross-promo engine touches two of them, it may claim both as "acquisitions" when one was just a hand-off. Internal attribution is blind to itself — it can't see that the user was already yours. The fix is a single source of truth that deduplicates by user, not by touch. Without it, your "cross-promo ROAS" is arithmetic, not measurement.
Q: How do you read a cohort without fooling yourself?
Decide deterministic versus probabilistic before you look at the number. Deterministic attribution — a logged-in user ID — tells you what happened. Probabilistic — device fingerprint, probabilistic matching — tells you a guess, and guesses drift toward "credit everyone." The trap is mixing the two and calling it precise. Report them separately, label the probabilistic one as an estimate, and never let an unvalidated model set your budget. The cohort that looks best is usually the one your attribution flatters most.
| Method | Answers | Blind spot |
|---|---|---|
| Last-click | who closed | who opened, organic |
| Self-attributing | the platform's view | double-counts, bias |
| Geo holdout | true lift | slow, needs scale |
| Deterministic cohort | what happened | cross-device gaps |
