What Tracking Loss Actually Costs You

“We only see about 60% of conversions now” is said in a tone that suggests the other 40% is the problem.

It usually is not. A program losing 40% of its conversions at random is in much better shape than one losing 10% in a pattern that tracks who converts. The first has lost precision and can buy it back with volume. The second has lost validity, and there is no volume of additional data that repairs it, because every new observation is missing in the same biased way.

The number everyone quotes is the size of the loss. The number that decides whether the measurement survives is its shape.

Two kinds of missing, and only one is dangerous

Loss unrelated to the outcome. A sampling rate, a logging service dropping one request in fifty, a region excluded for an unrelated reason. The observed data is a smaller version of the same population. Estimates stay centered where they should be, intervals get wider, and more traffic fixes it.

Loss that correlates with the outcome. Consent refusal, ad blocking, cross-device journeys, browsers that expire identifiers. These are not random draws. They select on privacy behavior, on device ownership, on age, on platform, and every one of those correlates with purchasing. The observed data is a different population, and it is different in the direction that matters.

The three losses are different problems wearing one word

Consent refusal removes people who declined. Refusal rates differ by region, by browser, and by prompt design, and in practice by age. What is left is a population that agreed to be observed, which is not the population you sell to.

Identity resolution failure merges or splits people. Its effect is subtler: a journey broken into two half-journeys makes early touches look like they never led anywhere and makes later touches look like they arrived from nowhere, which systematically credits the bottom of the funnel.

Cross-device loss is identity failure with a specific shape. People who research on a phone and buy on a laptop are the ones who disappear, and they are disproportionately the considered, higher-value purchases.

Only the first is usually measured, because only the first arrives as a number in a dashboard.

What each one costs, in the estimate rather than in the dashboard

For a channel comparison, the damage is proportional to how differently the loss falls across channels. If two channels lose the same share of conversions, their ratio survives even at heavy loss. If one channel skews toward mobile Safari users and another toward desktop, the loss rates differ, and the comparison is distorted by that difference.

That is why the most common real-world symptom of tracking loss is not that totals fall. It is that channel mix appears to shift, gradually, in a direction nobody can explain from anything they did.

What helps

Measure the loss rate as a series, by segment. It is not a constant and it moves when a browser ships an update. A trend in observation rate that is read as a trend in performance is the most expensive version of this mistake.

Model it where the loss is measurable and ignorable within groups. If the rate varies by device but is flat within device, reweighting by device recovers the total. That is a real technique with a real assumption attached, and it does nothing at all when the loss depends on something unobserved.

Move the question to a grain that does not need the individual. This is the one that works. A geo holdout, a mix model, or a market-level readout never observed anyone in the first place, so it cannot lose them. Attribution is not incrementality argues this from the counterfactual side; tracking loss is a second, entirely separate reason the aggregate designs keep earning their place.

Why more data fixes one and not the other Technical

With loss that is independent of the outcome, the observed sample is a random subsample. The estimator remains unbiased and its variance scales with the number of observations, so doubling traffic restores the precision that a 50% loss removed. Nothing about the conclusion changes; it just takes longer to reach.

With loss that depends on the outcome, the observed sample is drawn from a different distribution. The estimator converges, as sample size grows, to the wrong number. That is the whole difference: one error goes to zero with volume and the other goes to a constant. Reporting a tighter interval around a biased estimate is a more confident statement of the same wrong thing.

The intermediate case is loss that depends on the outcome only through variables you observe. Reweighting by those variables recovers the estimate, which is the same inverse-probability logic used elsewhere on this site for off-policy evaluation. It rests on there being no unobserved driver of observation, which is untestable and is what privacy-motivated loss tends to violate, since the reason somebody blocks tracking is not in your data.

Context

Best suited for

Programs whose reporting depends on observing individuals, and any team watching a channel mix drift without a change in spend to explain it.

Data requirements

Observation or consent rate as a time series, split by device, browser, and region
A tracking-independent count of the outcome, from finance or billing, to compare totals against
Loss rates by channel rather than in aggregate, since the comparison is what breaks first

What changes by context

B2C ecommerce: Short cycles limit cross-device loss, so consent is the dominant term. Because refusal rates differ sharply by platform, the channels that skew toward one platform are the ones whose measured performance moves when a browser changes policy.
B2C subscription: The outcome that matters lands months later, so the identifier has to survive far longer than the session. Identity decay is the dominant loss here, and it grows with the horizon that makes the measurement worth doing.
Product-led: Authenticated usage means the outcome side is well observed while the acquisition side is not, so the loss is concentrated entirely in the join between them and shows up as unattributed signups rather than as missing conversions.

When it breaks down

Fully authenticated funnels where nearly every outcome is tied to a known account, in which case the loss is small and the effort belongs elsewhere.
Programs already measured at aggregate level, where individual observation was never the input and the question does not arise.
Cases where the loss depends on something unobserved, where reweighting is not available and the answer is a design that does not need individuals.

What this does not establish

That a high observation rate makes an estimate trustworthy. A well-tracked comparison of two channels can still lack any counterfactual, which is a separate and larger problem.
That reweighting repairs consent loss. It repairs loss that is ignorable given observed variables, and privacy-motivated refusal is the case least likely to satisfy that.
That the loss rate you measure is the loss rate you have. It is a lower bound, established on the dimensions still visible to you.
Any particular figure for how much loss is tolerable. That depends entirely on whether it falls evenly across the comparison being made, which is a property of a specific program rather than a general threshold.