Definition

The two checks applied to a single intervention designed to amplify a candidate behavior — lift (the intervention increased the share of the signup cohort performing the behavior) and hold (the conversion rate among performers stayed within the conversion-rate metric's natural variation band).

What lift and hold measure

Both readings come from running multiplier analysis on a post-intervention cohort and comparing it against the pre-intervention baseline.

• Lift: the pre/post change in percent-of-signups. Did the intervention actually move more users to do the candidate behavior?
• Hold: the stability of the multiplier between pre- and post-intervention cohorts. Did the conversion rate among performers stay within the metric's natural variation band?

Lift is about whether the intervention worked as a behavior-amplifier. Hold is about whether the correlation that originally made the behavior interesting survived the amplification. The two are independent — moving the behavior count and preserving its predictive power are different operations, and an intervention can succeed at one while failing the other.

Accumulation across interventions

A single passing lift+hold test reinforces the candidate behavior's status but doesn't promote it. What accumulates a candidate into value milestone qualification is repeated lift+hold passing across multiple diverse interventions — different surfaces, framings, or sub-populations. The unit of diagnosis is accumulated evidence, not the single test; scale verification is the discipline that runs this accumulation.

Failure modes

• Single-test failure: either lift didn't move the behavior, or lift passed but hold didn't follow. Read as single-test failure — the charitable interpretation is intervention-quality variance: the intervention may have been incomplete, miscalibrated, or applied to the wrong population. The candidate behavior's status is unchanged; what needs reconsidering is the intervention design.
• Accumulated lift failures: across multiple diverse interventions, the share of users doing the behavior doesn't move. The proxy is unindicted as a signal; what's failing is the team's ability to design interventions that serve the underlying intent.
• Accumulated hold failures with lift passing: proxy failure. The behavior amplifies across diverse interventions but the conversion correlation doesn't follow. The original correlation reflected which users did the behavior organically, not a causal property of the behavior driving the outcome.

Hold's tolerance band

Hold's tolerance is a practitioner-judgment call calibrated against the conversion-rate metric's long-term volatility. Jumpier historical trends warrant wider bands; stable trends tighter. The band is set against the metric's own variance because the question hold answers is whether the post-intervention multiplier moved more than the metric ordinarily moves — anything inside that envelope reads as noise.

Example

A B2B SaaS team has identified "invited a teammate in first session" as a candidate behavior. Pre-intervention, 18% of signups perform it, with a 3.1x multiplier on first paid conversion within 30 days.

The team ships an onboarding prompt nudging the invite. Post-intervention reading:

• Percent-of-signups moves from 18% to 32%. Lift passes.
• The multiplier among performers comes in at 2.9x, within the band the team set against the metric's historical volatility. Hold passes.

Both checks passing reinforces the candidate behavior. The team logs the result and runs a different intervention next — a different surface, a different framing — to accumulate evidence across diverse interventions. If across five more attempts lift kept passing but the multiplier compressed toward 1x, the team would read accumulated hold failures as proxy failure: inviting a teammate was a signal of users already intending to expand collaboratively, not a lever that produced that intention.

See also

• Correlation drift