Quick start
- On Matrices, choose New matrix, pick True segments and choose Continue.
- Give it a Name, pick the transactions to use (or Connect your CRM or upload a file), and choose Continue. Three or more years work best.
- Watch it build: we profile every person, try different numbers of segments, keep the one that fits your data best, place everyone and spot who's likely to move up.
- Read each segment's plain-language description and where it starts. Choose Rename segments to give them names your team uses.
- In Likely to move up, pick a segment, review the strongest five, then choose View records or Export N IDs to work the list in your CRM.
What True segments are
True segments (in Beta) group your donors or customers by how they actually behave — how recently, how often and how much — instead of by score bands. Machine learning finds the groups in your own file, puts them in order from the top, and describes each in plain words, such as “Gives monthly · typical year $260” or “About 2 purchases a year · typical year $1,450”. Everyone with nothing in 18 months forms a gray dormant base at the bottom (when your file goes back that far).
It then looks at everyone below the top segment and flags those who already look like the segment above, learning from your organization's own past upgrades when your file has enough history. That's the Likely to move up list: IDs to review with your team for upgrade asks, loyalty offers, closer relationships or reactivation.
- Segment 1 Last 3 months · about 2 a year · typical year $6,800 96 people · 46% of value · top segment
- Segment 2 3–4 a year · typical year $1,450 488 people · 26% of value · 37 likely to move up
- Segment 3 Monthly · typical year $260 1,210 people · 15% of value · 112 likely to move up
- Segment 4 About 1 a year · typical $120 3,410 people · 13% of value · 268 likely to move up
- Segment 5 Nothing in the last 18 months 4,666 people · 0% of value · 403 likely to move up
Note: Beta means the method may still change. Segments are learned from your file alone and learned again with every build. Treat “likely to move up” as a list to review with your team, not a prediction of any one person's next gift or purchase.
True segments or named segments?
| Named segments | True segments | |
|---|---|---|
| Segments come from | Scores 1–5 and a map of which cell is which segment (you can repaint it) | Your data: groups of people who behave alike |
| Number of segments | 11 standard | 3 to 7, chosen by the data |
| You can change | Names, cells, window and breakpoints | Names |
| Best answers | Who to thank, nurture, win back, and who moved since last year | Who behaves alike, and who already looks like the next segment up |
Many organizations build both from the same file. See named segments.
Creating True segments
The setup page asks only for a Name (up to 200 characters; it starts as “Who's likely to move up”) and the transactions to use: every finished import, newest first, or Connect your CRM or upload a file to bring them in on the next page. There are no scales, windows or segments to set. Choose Continue and the page opens and starts building. See Uploading files and Connecting your CRM.
The data it needs
True segments build from any file with counted transactions. They work best with three or more years and about 500 people. With less, we still build everything and warn you:
| Your file | What happens |
|---|---|
| Fewer than 500 people, or less than 3 years | A notice: These segments may be less reliable. |
| Less than 18 months before the latest transaction | No dormant base, because no one can have gone 18 months without a transaction. |
| Less than about two years, or too few past upgrades | “Likely to move up” rests on resemblance alone, and the list says so. See learning from past upgrades. |
| A small number of recent people | Fewer segments are tried — at least about 20 recent people per segment. |
| No counted transactions at all | The page says there's nothing to learn from. |
Which transactions count
The same as every matrix: your organization type's counted types, net of refunds — money received for nonprofits and government agencies (open pledges and planned gifts left out), purchases, subscriptions and tickets for businesses. See which transactions count. IDs that differ only in letter case (C-5 and c-5) are the same person.
How the segments are learned
1. A profile for each person
The date of the latest counted transaction in the file is the as-of date; “the last 12 months” end on it. For each person we work out their typical year (the median of their yearly totals over the last 3 years, counting only years they were active), what they gave or spent in the last 12 months, transactions per year, the recurring share (recurring gifts or subscriptions), and days since their last transaction — plus, for the upgrade model, years active, the trend against the year before and their typical single transaction.
2. The dormant base
Anyone whose last transaction is 18 or more months before the as-of date goes in the dormant base, drawn in gray at the bottom. They're left out of the grouping so the active segments describe how people behave now.
3. Grouping people who behave alike
Everyone else is grouped on five measures: typical year, last 12 months, transactions per year, recurring share and recency. Amounts and counts are put on a log scale so $100 to $1,000 is as big a step as $1,000 to $10,000, each measure is scaled so none swamps the rest, and recency counts half. The grouping is k-means clustering, using Microsoft's ML.NET.
4. Choosing how many segments
We try every number from 3 to 7 in all, counting the dormant base, and keep the one where the groups are tightest and most clearly apart (the highest silhouette score); on a tie, fewer segments win. The build shows “Trying 4 segments…” and then, for example, “5 segments fit your data best”.
5. Ordering and placing
Segments are ordered by their middle member's typical year, highest first. Each person active in the last 18 months is placed in the group whose middle they're nearest to; everyone else is in the dormant base.
Reading the segments
- Segment
- Its name (“Segment 1” until you rename it), its plain-language description, and View records.
- Donors or customers
- How many people are in it.
- Of giving or spend
- Its share of the last 12 months' total. The dormant base's share is 0%.
- Where it starts
- “From about $X a year” (where the bottom tenth of the segment is), “Monthly giving, any amount”, “Recurring giving, any amount”, or for the lowest active segment, any transaction in the last 18 months.
- Likely to move up
- How many already look like the segment above, with See the list. The top segment shows “Top segment”.
Descriptions describe each segment's middle member, so not every member matches every word. Amounts are rounded to read as typical, not exact.
Rename segments
Choose Rename segments once the build has finished. Each segment has a box labeled with its position, its description underneath, and room for a name up to 100 characters. Choose Save names. Names are only labels: saving doesn't rebuild anything. Clear a box to go back to “Segment N”.
Careful: Names are saved by position. If a later build finds different segments — after a CRM refresh, say — a name stays on its position even if those people moved. Check names against the descriptions after each rebuild.
Likely to move up
The card says how many people already look like the segment above theirs. A menu switches between All segments and each segment's own list; the table shows the five strongest with their ID, Now, Looks like, Why (one or two reasons from their own history, such as “Giving up 3 years running; $3,900 in the last 12 months”) and Signal (Strong, Moderate or Early). View records opens everyone on the list; Export N IDs downloads them.
Everyone below the top segment is considered, including the dormant base: someone there on the list has an old pattern that looks like today's active people — a reactivation candidate.
How close they are
For each person we measure how far they've come from the middle of their own segment toward the middle of the one above, counting only the differences that separate those two: 0 is typical of their own segment, 1 is right at the edge of the next.
Learning from your own past upgrades
When your file reaches back about two years before its latest transaction, we rebuild everyone's profile as it stood a year (and two years) ago, place them with today's segments, and note who was in a higher segment a year later. A logistic regression model learns which profiles tended to move up, and each person's chance is compared with how often others in the same segment moved up. It needs at least 50 past examples, including at least 10 who moved up and 10 who didn't; without that the list says it's based on resemblance alone.
Score and signal
The score starts from similarity and is raised or lowered by the model. Scores of 0.60 or more are flagged: Strong from 0.85, Moderate from 0.72, Early from 0.60.
At most 10% of a segment
Each segment's list keeps to its strongest 10% at most (rounded up), highest scores first, so the list stays something a team can work.
While it builds
The stages are reading and sorting transactions, building profiles, finding segments, placing everyone and spotting who's likely to move up. Segments appear once they're found, then fill in as people are placed; you can leave the page, and we email you when it's ready if you left. A CRM refresh or a new import learns the segments again.
If it doesn't finish
The page says Learning the segments didn't finish with a Try again button. If it keeps failing, upload the file again or email support@truerfm.com.
View records and export
View records on a segment or the likely-to-move-up list shows one row per person — ID, value in the last 12 months, total, transactions, first and last, and segment — expandable to their transactions, with your kept detail columns. Export CSV at the top downloads the segment table. See Viewing records.
History
With Plus, True segments get a History tab: monthly snapshots of each segment and who was in it. See Matrix history.
Questions
Why did the number of segments change after a refresh?
Segments are learned again from the new data, and the best fit may now be a different number. Check your names against the descriptions.
Why is someone who gave or spent a lot in a lower segment?
Segments look at the whole pattern — how recently and how often too — and at their typical year, not one big transaction.
Why is there no dormant base?
Your file doesn't reach back 18 months before its latest transaction.