Intent-mix drift
Intent-mix drift is what happens when the mix of what users want shifts because the population coming in has changed. The cause can be on the team's side (new acquisition channels, campaign shifts, partnerships) or external (market maturation, shifting user expectations), but the felt experience inside the company is the same: signups want different things now.
Definition
A mechanism of correlation drift in which the incoming user base's intent distribution shifts, breaking the correlation a behavioral proxy held with the commercial outcome under the prior distribution.
Two sources of the shift
The shift comes from one of two directions:
- Provider-side acquisition decisions: channel changes, campaign shifts, partnership launches — anything that changes who arrives at signup. The team controls these and usually knows when they happen, though the consequences for proxy correlation are rarely the thing being planned for.
- External context shifts: market maturation, technology baseline changes, evolving user norms. The team doesn't control these and often doesn't notice them until something downstream has already moved.
From the proxy's perspective, the failure state is the same regardless of cause: signups want different things now.
The asymmetry with offering drift
Offering drift attacks proxies internal to the offering's surface — when a UI element gets reorganized or a feature gets retired, proxies anchored to those elements lose their meaning along with them. Proxies causally tied to the user's transition rather than to the offering's interface are less susceptible to that kind of drift.
Intent-mix drift attacks at a different layer. The intent-aligned proxies that survive offering drift are anchored to a specific intent's transition; when the intent distribution itself shifts, those proxies shift along with it. Scale-testing under the prior distribution validates the proxy for that distribution. It says nothing about whichever distribution the user base actually has a year later.
Legibility and the early canary
The framework offers two defenses, both at the detection layer.
The first is legibility. Model-level reasoning, applied to a portfolio of milestones, lets the team identify which intent's referent has moved when a correlation begins to shift. The portfolio doesn't prevent drift at the per-milestone level, but it makes drift diagnosable rather than mysterious.
The second is the early canary. Intent detection reads ambient signals plus sorting-hat answers on each user, surfacing intent-mismatch in real-time routing performance before first paid conversion resolves the question through declining numbers. A team running a deployed sorting hat sees the intent-mix fit shift while the commercial outcome is still catching up.
Example
An invoicing tool's value milestone for "sent first invoice" was scale-tested on a signup cohort dominated by working freelancers, where it correlated reliably with first paid conversion. The team adds paid acquisition on a small-business-tools roundup and starts pulling in people comparing accounting platforms. The new signups still cross the milestone at near-normal rates — anyone exploring sends a test invoice — but their first-paid-conversion rate is much lower because they're not in the same pursuit. The milestone keeps firing, the count keeps moving, and aggregate first-paid-conversion drifts down without any individual intervention having broken. The proxy held its correlation under the freelancer-dominated distribution; the new distribution changed the aggregate correlation.