Segment composition change analysis
See which audience composition changed and split it into movement between units and change inside them.
End of analysisNext steps and reference material below. Not part of the analysis result.
Next step
This week
Full journey
Why did a segment's share of total signups change?
Compare two periods' segment shares and split the shift into movement vs internal change.
Method and FAQOpen only when you need the methodology
See calculation conditions and limits
When to use it
Use it when overall CPA, signups, or conversions moved and per-campaign numbers alone do not explain it. It separates whether the added users look different from before, whether a campaign started bringing different people, or whether budget simply moved between campaigns.
What it computes
Per axis it computes member shares for both periods, the total distribution shift, and each member's percentage-point change with real head counts. With an analysis unit declared it splits the change into mix, rate, and interaction, and the three terms add up exactly to the total change. When the population grew it also profiles the added slice.
How to use the result
If movement dominates, look at budget allocation first; if internal change dominates, look at targeting, creative, and landing. Axes with small samples or without exclusivity return a reason instead of a number. To test an intervention, move on to incrementality analysis or an experiment.
Frequently asked questions
My gender and age columns are named differently in every file. Is that a problem?
No. Column names do not need to match a standard schema; inside the tool you declare which column is the segment axis and which one carries head counts.
Can I see the composition of the users who were added?
If the population grew, the tool profiles the net-new slice by subtracting the two periods. That is arithmetic between periods, not a causal profile of incremental users.
Does it work for axes where one person can hold several tags?
You can read per-value adoption change, but the overall distribution shift is not computed. Applying a distribution distance where shares do not add to 100 percent would make the number false.