Rebuilding Lark’s acquisition attribution

Incomplete tracking and conflicting channel labels

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 were derived through partially independent logic, which could produce conflicting labels for the same enrollment.

Around March 2026, on a trailing 365-day view, about 42% of enrollments were unresolved or unknown. That left channel-mix and CAC reports too incomplete to use for budget decisions.

How the attribution model works

The production model uses one record per first enrollment. It normalizes the acquisition data and applies an ordered set of attribution rules to resolve source, medium, and campaign. It then derives channel and other reporting categories from that result.

Lifecycle metadata is joined afterward to describe activity following acquisition, without changing the original attribution.

The model handles missing tracking and later activity as follows.

Source resolution. Channel and category labels are derived from the resolved acquisition source, so they no longer rely on separate classification rules.

Resolution method. When explicit tracking is missing, the model uses secondary signals and records which method produced the attribution. This makes the classification reviewable.

Missing sessions. The model distinguishes enrollments for which web tracking is unavailable from existing sessions whose source cannot be resolved. These require different follow-up checks.

Lifecycle data. Later engagement can be analyzed alongside the acquisition source without changing its classification.

Session checks and channel monitoring

I checked channel classifications against raw session-level data, reviewing the available evidence and the rule that produced each label. This was a separate check from measuring the reduction in unknown attribution.

After deployment, I set up monitoring for large swings in channel mix. Unexpected changes prompt a review of the attribution before being interpreted as a change in channel performance.

Why direct mail stayed in use

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 was going to be discontinued before my attribution work was complete. My findings reversed that decision, and the channel stayed in use. Annual spend on the direct-mail channel is $550K.

The increase reflects corrected attribution. It does not establish how many enrollments were caused by direct mail.

Reporting after the rebuild

On the trailing 365-day view, unknown attribution moved from about 42% to 8%, improving coverage in channel and CAC reports.

The reports retain each attribution’s resolution method, so a channel classification can be traced back to the evidence and rule that produced it.

What the attribution results establish

The model identifies acquisition sources from the available tracking and enrollment data. It does not estimate how many enrollments would have happened without a channel. Direct mail’s corrected 38% share for one partner population is an attribution result, not a measure of causal lift.

I am now designing randomized geographic experiments for Google and Meta paid advertising at Lark. That work has not shipped and does not supply a causal result for this case. Separate independent research on incrementality and allocation is also in development.