Defining success

What it represents

Success is the label. Every record in training is either a conversion or it is not, and you decide which. That means naming the object that records a conversion, the criteria that have to be true on it, and the date field that says when it happened.

The other half is quieter and matters just as much: the non-success cutoff. For a record that never converted, where does the observation stop? Without a boundary, a record that sat in the CRM for two years contributes two years of silence, and the model learns that silence is a signal.

How it’s calculated

All of it lives in one block, headed Success definition, on the Set up your score card. Four controls, top to bottom, and the same block is where you go back to change any of it later.

Set it up

1. Answer "What defines success?"

   The object that records a conversion. The picker offers your pinned defaults as chips plus a search over every object discovered on the connection, with record counts beside them, so you can see whether the object you had in mind actually holds anything.

   On most Salesforce orgs this is Opportunity. Pick the object that exists because the outcome happened rather than one the outcome gets noted on afterwards. Beneath it, the join field is how that object gets back to the record being scored; it fills itself in and is there to edit when your org joins them unusually.

2. Build the success criteria

   Criteria on that object's fields, combined with AND or OR, and groupable when the definition needs both. The builder stays disabled until a connection and a conversion object are set, because the field list comes from the object you just picked.

   Start with the narrowest definition your team would agree on out loud, and no narrower. Every criterion you add moves records out of the conversion class, and the conversion count is the real constraint on what the model can learn. When the existence of a related record is the whole definition, that is a row of its own rather than a field comparison.

3. Pick the success date field

   A date field on the conversion object: when the conversion happened. Choose the date the outcome settled rather than one stamped during cleanup afterwards, because this is the cutoff for every record that converted.

   Setting it does a second job. It is what makes engagement features available to every model on the score, so a score with no success date is a score whose activity counts cannot be dated and therefore cannot be trusted.

4. Choose the non-success cutoff

   A radio with two options. Auto-detect, marked recommended, measures the 90th percentile of your observed conversion times and applies that span from each non-converting record's creation date. Custom date field points at a real date you already keep: disqualified, archived, removed from the pipeline.

   Take auto-detect first, then look at the window it lands on. It is reading your history back to you, so if your real cycle runs far longer than the number it picked, that gap is worth understanding before you accept it.

Activity after a record’s cutoff is excluded from its features either way.

What it means for you

Two practical consequences. First, the success definition is part of what the model is · change it and you are training a different model, so retrain after you edit it. Second, if your real sales cycle is much longer than the auto-detected window, the number it picked is worth a look before you accept it.

Check your understanding

12,000 leads, 900 of which converted. Ninety percent of those conversions happened within 45 days of the lead being created, so auto-detect sets a 45-day window. A lead created 200 days ago that never converted is labeled a non-conversion, and only its first 45 days of activity become features. The nurture emails it received in month five do not count, because they cannot have caused an outcome that had already settled.

Read that back against the four controls. The conversion object and its criteria are what sorted 900 records from 11,100. The success date field is what dated each of those 900. And auto-detect is the whole reason the other 11,100 have an ending at all, rather than trailing on until today.