The challenges of marketing attribution

1. You only ever observe one version of history

This is the root problem and everything else descends from it. You can see that someone saw an ad and then bought. You cannot see whether they would have bought anyway, because that version of the world did not happen. Every attribution method is an attempt to estimate that missing half. Rule-based models (last click, first click, linear, time decay) do not estimate it at all: they apply a convention and call it credit. That is not a measurement, it is a bookkeeping choice, and it is why two tools using different rules produce different numbers from identical data.

2. Every platform grades its own homework

Meta, Google, and TikTok each report on their own performance while selling the inventory being evaluated. Each counts a conversion it touched inside its own attribution window, so the same order is claimed several times. Add up what every platform reports and you will usually exceed your real order count. This is not fraud and not a bug; it is what happens when each party sees only its own slice and none of them are neutral. The practical consequence is that platform-reported ROAS cannot be summed, compared across platforms, or used as a basis for total budget allocation.

3. The identifiers the whole system runs on are disappearing

Rule-based and multi-touch attribution both depend on joining an ad click to a later conversion via an identifier. ATT removed most of that on iOS apps in 2021. Browser countermeasures have been stripping click IDs and shortening cookie lifetimes ever since. Server-side tagging and the Conversions API recover some signal, but they recover recording, not causality: a perfectly recorded click still does not tell you whether the ad caused the purchase. Every workaround in this category buys time against a trend that has moved in one direction for five years.

4. A growing share of revenue has no click trail at all

Dark social, word of mouth, podcasts, retail media halo, and now AI agents that buy through APIs without ever loading your storefront. None of these leave the trail attribution was designed to follow. They land in Direct or Unassigned, which quietly inflates those buckets and makes every channel that created the demand look weaker than it is. The bucket grows every year, and it grows fastest for the brands doing the most upper-funnel work.

5. Correlation and causation are genuinely hard to separate

Retargeting is the clearest case. You show ads to people who already visited your site; they convert at a high rate; the dashboard reports a spectacular ROAS. Most of those people were going to convert regardless. The same confound appears in branded search, in lookalike audiences, and in any channel whose targeting is good at finding people close to purchase. The better the targeting, the more existing demand it intercepts and relabels as new, which means the confound gets worse precisely as the platform gets better.

6. The methods that do measure causality all have real costs

Geo holdout experiments are the strongest evidence available, and they require withholding spend, weeks of waiting, and enough scale to power the test. Marketing mix modelling handles the no-click channels but needs years of history, a lot of it, and gives you channel-level rather than campaign-level answers. Causal inference on observational data is faster and cheaper but is bounded by its assumptions rather than by randomisation, which is why any honest implementation grades how far to trust each number instead of presenting one confident figure. There is no method here that is simultaneously cheap, fast, and as rigorous as an experiment. Anyone claiming otherwise is selling.

7. The organisational problem is harder than the technical one

Even with a correct number, someone has to act on it. Channel owners are compensated on the metrics the new measurement contradicts. Agencies report on the platform numbers that make them look effective. A causal read that says branded search is not incremental threatens a line item that somebody defended last quarter. In practice this is where most attribution projects die: not at implementation, but at the meeting where the finding contradicts an incentive. Any measurement change that does not come with a plan for that conversation will produce a nicer dashboard and identical spending.

What each method actually solves

Where we fit

Causality Engine is the sixth row: causal inference over the GA4 history you already have, returning what each channel caused next to last-click, with a data-health score on every channel, in 1–2 minutes for €99. It does not fix problem 7, and it is not a substitute for a properly powered experiment when the stakes justify one. It is the fast, cheap read that tells you which channels are worth the expensive test. The method: causal attribution, incrementality testing, and marketing mix modeling.