Correlation drift
Correlation drift is what happens when a behavior that used to reliably predict whether users would pay, renew, or stick around stops predicting it, even though the behavior is still measurable and the team can still influence it. The cause sits outside the behavior itself: the mix of users coming in has shifted, or the product has changed enough that doing the behavior now means something different than it used to.
Definition
An established behavioral proxy's correlation with the commercial outcome dissolves as conditions shift.
Drift vs proxy failure
Both describe a proxy that stops predicting the commercial outcome, but the underlying problem differs. Proxy failure means the proxy was never a usable lever — the original correlation reflected which users did the behavior organically rather than a causal property of the behavior driving the outcome. Correlation drift means the proxy was usable; conditions changed.
The diagnostic response follows from this. Proxy failure indicts the candidate behavior itself, and the team should abandon it and look for a different one. Drift indicts the current applicability of a proxy that was sound, and the team's work is to identify what shifted and decide whether the proxy can be re-anchored, re-validated, or replaced.
Two mechanisms
Two distinct mechanisms produce drift:
- Intent-mix drift: The incoming user base's intent distribution shifts. This can come from provider-side acquisition decisions (channel changes, campaign shifts, partnerships) or external context shifts (market maturation, technology baseline, user norms). The proxy was correlated with the outcome under one intent distribution; the new distribution changes the aggregate correlation. From the proxy's perspective, the failure state is the same regardless of cause: signups want different things now.
- Offering-surface drift: The offering itself evolves — features added, removed, or reorganized. The proxy's meaning shifts with the offering's surface; a proxy tied to a UI element may track different user pursuits before and after a UI change.
Defenses against each mechanism
The two mechanisms threaten correlation differently, and the framework's apparatus answers them differently.
Against offering-surface drift, external-impact events are generally more durable than feature-usage events. They're anchored to the user's pursuit outside the offering's surface rather than to the surface itself, so they don't track different pursuits when the surface changes.
Against intent-mix drift, no proxy is inherently robust. Any intent-anchored proxy shifts with the intent distribution, no matter how well it survived earlier scale-testing. What the framework offers here is legibility — model-level reasoning can identify which intent's referent moved — and intent detection as an early canary: intent-mismatch surfaces in real-time routing performance before the commercial outcome resolves.
Example
An invoicing SaaS has scale-tested "sent an invoice" as a value milestone, historically a strong correlate of first paid conversion because the signup base is mostly freelancers for whom sending an invoice represents committed use of the product. A partnership shifts the acquisition mix toward agency owners, who use invoicing as one piece of broader workflows. The proxy keeps firing — agencies do send invoices — but the conversion correlation softens: agencies who send one and bounce now look the same at the proxy as agencies who adopt the platform deeply. The intent distribution the proxy was correlated against has shifted underneath it.