Rebuilding acquisition attribution around observable evidence

The existing measurement could not safely support channel decisions

The prior acquisition-attribution system had accumulated multiple independent first-match classifiers, hard-coded campaign rules, partner and path heuristics, and incomplete tracking. Category, agent, medium, source, and campaign should have been one resolved story. Instead they were derived through partially independent logic.

Around March 2026, on a trailing 365-day view, about 42% of enrollments were unresolved or unknown. Channel mix and CAC could be reported, but not trusted for funding, cutting, or retiring acquisition channels.

Resolve observable evidence first, then apply business semantics

The rebuild changed the order of operations. The previous system classified business-facing dimensions from raw fields through separate first-match trees. The production model now establishes a canonical first-enrollment grain, normalizes raw acquisition evidence, evaluates an ordered evidence hierarchy, and only then maps the resolved evidence into channel and business-facing dimensions.

Lifecycle metadata is joined afterward. It can describe what happened after acquisition. It does not overwrite how the enrollment was attributed.

Four rules governed the model.

Facts before business semantics. Observable acquisition evidence is resolved first. Channel, category, and other business-facing labels are mappings of that evidence, not a parallel classification exercise.

Coverage without pretending certainty. Secondary signals can recover attribution when explicit tracking is missing. The model retains attribution provenance (the resolution method) instead of emitting only a channel label.

No-web-session is different from unknown session. A structurally unavailable web signal is not the same thing as a session that exists and still cannot be resolved. Collapsing both into “unknown” hid why coverage failed.

Lifecycle metadata enriches acquisition attribution instead of overriding it. Downstream engagement can explain later value. It should not rewrite the acquisition evidence.

Review the evidence under each classification, then watch the mix

A lower unknown rate was not treated as proof that the new labels were correct. Each channel decision was checked against raw session-level data, including what evidence was present, which rule in the hierarchy fired, and whether that method matched the classified channel.

After deployment, monitoring was set up to catch large swings in channel mix. Unexpected movement in reported mix is a reason to investigate the measurement, not an automatic conclusion that the underlying channels changed.

Incomplete measurement could have retired the wrong channel

The prior system left a large unknown bucket and also misstated channels that were already in the evidence.

For one partner population in Q4 2025, direct mail’s measured share of enrollments was about 4% under the previous measurement and about 38% after the corrected attribution. Direct mail had been scheduled for retirement. The corrected measurement changed the evidence being used to evaluate the channel, and direct mail stayed in use.

This does not say direct mail caused 38% of enrollments, and it is not an incrementality estimate. The prior measurement was incomplete enough that it could have supported the wrong channel decision.

Coverage recovered, and a retirement decision was reversed

On the trailing 365-day view, unknown attribution moved from about 42% to 8%. Channel and CAC reporting became usable measurement instead of a leftover bucket.

The direct-mail result is the clearest decision consequence. A channel slated for retirement stayed in use once the evidence used to judge it was no longer missing most of the attributed volume.

The unknown-rate drop is the headline. The supporting evidence is recovered coverage, retained provenance, session-level review, and mix-shift detection.

Attribution is not incrementality

This work answers a measurement question. Given the evidence that exists, how should an enrollment’s acquisition source be resolved, and how should that resolution be labeled for the business?

It does not answer the causal question. It does not say whether a channel produced enrollments that would not have happened otherwise, or where the next dollar should go. Moving direct mail from 4% to 38% of measured enrollments in one partner population means the previous system was incomplete. It does not mean direct mail caused 38% of those enrollments.

Incrementality, experimentation, and allocation under uncertainty are a different problem. That work is in development as independent research.