What actually predicts conversion
Across trial-based products the strongest signals are behavioural and early: whether the user reached the point where the product does its main job, how many distinct sessions they had in the first days, and whether more than one person from the organisation logged in.
Firmographic data - company size, industry, job title - usually adds less than expected once behaviour is included. It is useful before the trial starts, when behaviour does not exist yet, and fades quickly afterwards.
The single most valuable piece of work is often defining activation properly: the specific action that means someone has experienced the value. That definition improves onboarding whether or not a model gets built.
The self-fulfilling loop
The obvious use of a conversion score is to route high scorers to sales and ignore the rest. Within a quarter the model looks excellent - low-scored trials convert at almost nothing - because nobody contacted them.
The model is now measuring your sales process rather than the customer. Worse, it is learning from data it generated, so the bias compounds with each retraining.
- Always keep a random holdout that receives standard treatment regardless of score.
- Record what treatment each trial received, so future training can account for it.
- Watch whether the score distribution is drifting because of your own targeting.
Likely to convert is not the same as worth contacting
A trial almost certain to convert on its own needs no phone call. A trial that will never convert cannot be rescued. The value is in the middle, where an intervention changes the outcome.
This is the uplift question rather than the propensity question, and it needs experimental data - some trials contacted, some not, at random - to answer properly. Running that experiment for a few weeks before building the model produces something far more useful than a pure propensity score.
| Group | Converts alone | Converts if contacted | Action |
|---|---|---|---|
| Sure thing | Yes | Yes | Do not spend sales time |
| Persuadable | No | Yes | Contact - this is the value |
| Lost cause | No | No | Leave to self-serve |
| Do not disturb | Yes | No | Avoid - contact hurts |
Acting on the score without annoying people
Scores can drive more than a call list. Lower-touch responses are often better value and carry less risk of the self-fulfilling loop.
- Trigger an in-product prompt when a trial stalls before activation.
- Send a targeted help article matching the feature they got stuck on.
- Offer a short onboarding call only where the model suggests an intervention would change the outcome.
- Feed the stall points back to product, which is often where the real fix is.
If a trial score only ever produces a call list, most of its value is going unused.