Creator Attribution: Codes, Links and Holdout Tests

Codes and links tell you who claimed credit. They do not tell you what the program caused. So every setup I build carries both: per-placement codes and links for day-to-day operations, and at least one geo holdout or staggered start for the number you put in front of finance. When those two numbers are compared, last-click typically overstates creator contribution by somewhere between 1.4x and 3x, and the gap widens the more branded search defence you are running. The operational layer: A unique code per placement, not per creator, because a creator running three videos across two quarters produces one number you cannot decompose otherwise. A locked UTM taxonomy with a fixed vocabulary, enforced by a generator rather than a shared spreadsheet, since free-text sources are how six months of data becomes uncountable. Short links chosen because they pass parameters through rather than stripping them. A landing page that carries the code forward so the code applied at checkout ties back to the session that produced it. And a post-purchase survey question, one line, because self-reported source is a weak signal that happens to correlate with the strong one and catches whole channels the deterministic layer cannot see. What is knowable on mobile: For apps the setup runs through your existing MMP, whether that is AppsFlyer, Adjust or Branch, with deferred deep linking so a code survives the trip through the App Store. On iOS I will tell you plainly what SKAdNetwork and AdAttributionKit permit at creator level, which is coarse, delayed and privacy-thresholded. Android gives you more. Any vendor promising per-creator install attribution on iOS with the fidelity you had in 2019 is selling you modelled output described as measurement, and you should ask them which of the two it is. The causal layer: One of two designs, chosen against your business rather than my preference. Geo holdout, where creator activity is suppressed in matched provinces or metros for a defined window, which suits national local-service and retail businesses. Or staggered start, where markets enter in sequence, which suits app launches and anything where withholding a market is commercially unacceptable. Both need a pre-registered read date, a pre-registered metric and a minimum spend level to produce a detectable effect, and I will tell you before we start if your volume cannot support either. That conversation is the service. What I will not do: Build a dashboard that reconciles to nothing, which is most creator dashboards. Attribute on last click while you also bid on your own brand terms, because you are then paying two parties for one purchase and calling it growth. Report reach, impressions or engagement rate as outcomes. Or forecast a return figure before a single placement has run, since that number depends on your landing page, your pricing and your product, none of which I control. Not for you if: You need the reported number to be large. Roughly half this work reduces a figure somebody has already presented internally, and that is politically expensive. Also not for you if creator activity is too small to move a measurable line, in which case keep the codes for operations and skip the causal work until volume justifies it. Delivery: The naming and taxonomy standard with its generator. Code and link provisioning through your commerce or MMP stack. GA4 and warehouse event mapping so the code survives into reporting. The test design, pre-registered, with its read date and stopping rules. The read itself, written as a decision rather than a chart pack: what to keep funding, what to stop, and what the operational numbers should be discounted by from now on.