Crafte Sessions analysis 2026 09 29 2026 10 29
sessions_analysis-2026-09-29-2026-10-29.csv · 2,246 rows · 15 columns · definition v3 saved 29 Oct 2026 14:08
Segments v13 built 29 Oct 2026 14:08 on feature set v6
A newer feature set (v8) exists. These segments were built on v6.
Segments found
9
best of 3 to 10
In a segment
99%
2,222 of 2,246 snapshots
Unknown
1.1%
0 history · 1 between · 23 novel
Rules match the model
95%
for named segments
What to do
From these segments' numbers. Setting buttons fill in the Build segments tab (then Build segments); merge buttons change the table below (then Save labels, no refit).- Worth doingS8, S9 are very small (1 snapshot)
Under 30 snapshots, a segment's rates swing with a couple of people, and here the fit is mostly splitting off a few odd snapshots. Merge each into the segment it's most like (S8 → S6, S9 → S6). Keep one only if it's a group you specifically care about.
- Worth doingordered: use S6, S5, S2 as the targets on Outcomes
S6, S5, S2 hold 45% of ordered in 21% of the snapshots (2.2× the overall rate). Leaving out S3 (10% of snapshots) would also leave out 5% of ordered: that's the trade-off of the reduction.
- Fine as isUnknowns: 1.1% of snapshots
A small unknown share means nearly everyone fits a segment, while the odd ones aren't forced in.
- Fine as is9 segments is a clear best fit
The score is clearly better at 9 than at the counts either side. After merging the small ones you'd have 7 that matter.
Outcome concentration
Share of each segment's snapshots whose entity eventually had the outcome. Lift = rate ÷ overall rate.S6 has the highest rate: 16% for ordered, 4.2× the overall 3.8%.
- S6 has the highest rate: 16% for ordered, 4.2× the overall 3.8%.
- Rate: share of a segment's snapshots whose entity eventually had the outcome. Everyone together: 3.8% for ordered.
- Lift = rate ÷ overall rate. S6 has 16% for ordered, so 4.2× the average; above 1 is better than average.
- Captured: the segment's share of all the outcomes (each Captured column adds to 100%). Compare it with Share: S1 is 55% of snapshots and holds 30% of ordered, so it dilutes it.
- Segments that capture much more than their share are the ones to take to Outcomes as targets.
| Segment | Snapshots | Share | ordered | ||
|---|---|---|---|---|---|
| Rate | Lift | Captured | |||
| S1 | 1,246 | 55.5% | 2.1% | 0.5× | 30% |
| S2 | 243 | 10.8% | 5.8% | 1.5× | 16% |
| S3 | 228 | 10.2% | 1.8% | 0.5× | 5% |
| S4 | 196 | 8.7% | 4.1% | 1.1× | 9% |
| S5 | 145 | 6.5% | 9.0% | 2.3× | 15% |
| S7 | 88 | 3.9% | 3.4% | 0.9× | 3% |
| S6 | 74 | 3.3% | 16.2% | 4.2× | 14% |
| S8 | 1 | 0.0% | 100.0% | 26.1× | 1% |
| S9 | 1 | 0.0% | 0.0% | 0.0× | 0% |
| Unknown: between segments | 1 | 0.0% | 0.0% | 0.0× | 0% |
| Unknown: fits no segment | 23 | 1.0% | 21.7% | 5.7× | 6% |
| Everyone | 2,246 | 100% | 3.8% | ||
How each segment differs
Average of each input, in standard deviations from everyone. Red higher, blue lower.landingPage=/product/79-chocolate-fruit-cake-500g-boxed separates the segments most: from -0.2 sd in S4 to +5.0 sd in S7.- Each column is a segment and each row an input. A cell is how far the segment's average sits from everyone's, in standard deviations: +1 is clearly higher than typical (red), −1 clearly lower (blue), and near 0 is typical. Below, "slightly" means under 0.6 and "much" means over 1.5.
landingPage=/product/79-chocolate-fruit-cake-500g-boxedseparates the segments most: from -0.2 sd in S4 to +5.0 sd in S7.- S1 (1,246 snapshots): slightly more often device is mobile (+0.6); slightly more often landingPage is / (+0.5).
- S2 (243 snapshots): much longer-standing (returning) (+1.7); much more visits so far (+1.7).
- S3 (228 snapshots): much more often device is tablet (+2.6); slightly more often landingPage is / (+0.4).
- S4 (196 snapshots): much more often device is desktop (+2.7); slightly more often landingPage is / (+0.3).
- S5 (145 snapshots): much more often landingPage is other (+2.5).
- S7 (88 snapshots): much more often landingPage is /product/79-chocolate-fruit-cake-500g-boxed (+5.0).
- Read a segment down its column to describe it; read across a row to see which segments differ on that input.
| S1 | S2 | S3 | S4 | S5 | S7 | S6 | S8 | S9 | Between | Novel | |
|---|---|---|---|---|---|---|---|---|---|---|---|
actions[add_search] | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | +47.4 | -0.0 | -0.0 | -0.0 |
actions[sub_start] | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | +47.4 | -0.0 | -0.0 |
pages | -0.1 | +0.2 | -0.2 | -0.1 | +0.2 | +0.4 | +0.7 | +6.3 | +4.3 | +1.3 | +2.6 |
newVisitor | +0.1 | +0.1 | +0.1 | +0.1 | +0.1 | +0.1 | -0.9 | +0.1 | +0.1 | +0.1 | -5.1 |
landingPage=/product/79-chocolate-fruit-cake-500g-boxed | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 | +5.0 | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 |
minutes | -0.1 | +0.1 | -0.1 | -0.1 | +0.2 | -0.0 | +0.7 | +2.4 | +2.3 | +4.9 | +1.5 |
source=Facebook | +0.2 | +0.2 | +0.2 | +0.1 | +0.2 | +0.2 | -4.8 | +0.2 | +0.2 | +0.2 | -3.3 |
source=Direct | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 | +4.7 | -0.2 | -0.2 | -0.2 | +1.4 |
device=desktop | -0.4 | -0.0 | -0.4 | +2.7 | -0.2 | -0.1 | +0.6 | -0.4 | -0.4 | -0.4 | +0.4 |
device=tablet | -0.4 | -0.1 | +2.6 | -0.4 | -0.0 | +0.1 | -0.3 | -0.4 | +2.6 | -0.4 | +0.8 |
rows | -0.3 | +1.7 | -0.3 | -0.3 | +0.3 | -0.0 | -0.1 | -0.3 | -0.3 | -0.3 | +2.5 |
landingPage=other | -0.2 | -0.2 | -0.2 | -0.2 | +2.5 | -0.2 | +1.2 | -0.2 | -0.2 | -0.2 | +0.7 |
landingPage=/ | +0.5 | -0.5 | +0.4 | +0.3 | -2.2 | -2.2 | -1.1 | +0.5 | +0.5 | -2.2 | -0.6 |
actions[add_recommended] | -0.1 | +0.3 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | +2.1 |
tenure_days | -0.3 | +1.7 | -0.3 | -0.3 | +0.3 | -0.0 | -0.1 | -0.3 | -0.3 | -0.3 | +2.0 |
landingPage=/product/474-port-and-prune-pudding-800g-log-in-cloth | -0.1 | -0.1 | -0.1 | -0.1 | +1.7 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 |
device=mobile | +0.6 | +0.1 | -1.7 | -1.7 | +0.2 | -0.0 | -0.1 | +0.6 | -1.7 | +0.6 | -0.9 |
landingPage=/product/75-christmas-cake-boxed | -0.1 | -0.1 | -0.1 | -0.1 | +1.5 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | +0.3 |
Movement between segments
Where each entity's next snapshot lands, as a share of the rowThere are 232 steps from one snapshot to the next. 31% stayed in the same segment, so people often change segment between visits.
- Rows are where an entity's snapshot was; columns are where its next snapshot landed. Each row adds to 100%. The diagonal is staying put; anything else is a move.
- There are 232 steps from one snapshot to the next. 31% stayed in the same segment, so people often change segment between visits.
- Most common moves: S1 → S2 (85); S3 → S2 (18); S4 → S2 (12).
| From ↓ to → | S1 | S2 | S3 | S4 | S5 | S7 | S6 | S8 | S9 | Between | Novel | moves |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| S1 | 85% | 11% | 4% | 100 | ||||||||
| S2 | 85% | 13% | 2% | 52 | ||||||||
| S3 | 82% | 9% | 9% | 22 | ||||||||
| S4 | 86% | 14% | 14 | |||||||||
| S5 | 27% | 60% | 7% | 7% | 15 | |||||||
| S7 | 56% | 44% | 9 | |||||||||
| S6 | 11% | 67% | 22% | 9 | ||||||||
| Unknown: between segments | 100% | 1 | ||||||||||
| Unknown: fits no segment | 10% | 90% | 10 |
Segments over time
Rebuild to see how the segments change over time.
How good are these segments?
Rebuild to measure separation and stability.
Choosing the number of segments
The lowest BIC is at 9 segments, so that's what was used. The next best is 10 (+29,157).
- Each bar is the fit score (BIC) for one number of segments. Lower is better: it rewards a close fit and penalises extra segments.
- The lowest BIC is at 9 segments, so that's what was used. The next best is 10 (+29,157).
- It's inside the range you tried, so the range was wide enough.
Lower BIC is a better balance of fit and simplicity. 9 segments scored best.
The unknowns
24 snapshots (1.1%) are unknown: 0 too little history, 1 between segments, 23 fit no segment.
- 24 snapshots (1.1%) are unknown: 0 too little history, 1 between segments, 23 fit no segment.
- Too little history (0, 0.0%): fewer rows so far than the minimum you set.
- Between segments (1): no segment was a clear best fit; often someone changing pattern.
- Fits no segment (23): the oddest snapshots. Worth a look: new behaviour (or fraud) tends to show up here first.
- Nearest segment says which segment each would have joined if forced.
| Kind | Snapshots | Nearest segment |
|---|---|---|
| Unknown: between segments Sits between two or more segments; often mid-transition. | 1 | S2 (1) |
| Unknown: fits no segment Doesn't resemble any segment. Worth a look: new patterns and fraud often start here. | 23 | S6 (16), S5 (4), S2 (2) |
Boundary rules
How to recognise each segmentTogether the rules agree with the model 95% of the time: good enough to use as a description.
- Read each line as IF these conditions THEN this segment, using the inputs in their own units (minutes, page counts).
- Covers is the share of the segment the rule finds; correct is how many of the snapshots it finds really are in that segment.
- Together the rules agree with the model 95% of the time: good enough to use as a description.
Each segment gets its own short rules for picking it out from everyone else, in each feature's own units. Taken together, simple rules reproduce 95% of the assignments. Rules show where the lines fall, not why: use the profile above for what defines a segment. Unknowns have no rules; they come from how well a snapshot fits, not from where it sits.
S1 1 rule covering 100% of the segment
- device = mobile and landingPage = / and rows ≤ 1.5 · covers 100% of the segment, 98% correct
S2 1 rule covering 70% of the segment
- rows > 1.5 and landingPage is not /product/79-chocolate-fruit-cake-500g-boxed and landingPage is not /product/474-port-and-prune-pudding-800g-log-in-cloth · covers 70% of the segment, 85% correct
S3 1 rule covering 100% of the segment
- device = tablet and tenure_days ≤ 0.272 and landingPage is not /product/79-chocolate-fruit-cake-500g-boxed · covers 100% of the segment, 93% correct
S4 1 rule covering 100% of the segment
- device = desktop and rows ≤ 1.5 and source is not Direct · covers 100% of the segment, 92% correct
S5 1 rule covering 100% of the segment
- landingPage is not / and landingPage is not /product/79-chocolate-fruit-cake-500g-boxed and source = Facebook · covers 100% of the segment, 58% correct
S6 3 rules covering 100% of the segment
- source is not Facebook and source is not Crafte and source is not Google · covers 85% of the segment, 100% correct
- source is not Facebook and source = Crafte · covers 10% of the segment, 70% correct
- source is not Facebook and source is not Crafte and source = Google · covers 5% of the segment, 80% correct
S7 1 rule covering 100% of the segment
- landingPage = /product/79-chocolate-fruit-cake-500g-boxed · covers 100% of the segment, 100% correct
S8 no clean rule: this segment overlaps others on these features
S9 no clean rule: this segment overlaps others on these features
Compare runs
Find the best segmentationThe pick is v20: its targets hold 83% of ordered in 25% of snapshots (3.3× the rate), with target stability 0.52.
- Each row is one segmentation run. Click the version to open it; the highlighted row is the one shown above.
- Separation and Stability describe the whole segmentation: are the segments distinct, and do they come back when refitted?
- The last four columns are about ordered: Targets are the segments at 1.5× the overall rate or more; Captured is the share of all ordered they hold (and the share of snapshots they are); Lift is how concentrated they are; Target stability is whether those particular segments come back when refitted.
- Read Captured and Lift together: a high lift on a sliver of the outcome is less useful than a slightly lower lift that holds most of it. pick marks the best mix: captured × lift, with full credit only when the targets' stability is 0.5 or more. ★ marks the best value in each column.
- Change one thing per run (the group handling, the number of segments, the shape or the feature set) so you can see what made the difference. Runs built before this table existed show '-'.
| Run | Features | Groups | Shape | Segments | Unknown | Separation | Stability | Targets | Captured | Lift | Target stability |
|---|---|---|---|---|---|---|---|---|---|---|---|
| v29 30 Oct 2026 12:50 |
v8 | device evened out | Full | 8 | 1.1% | 0.64 | 0.88 | S2, S5, S6 | 33% in 12% | 2.8× | 0.96 ★ |
| v28 30 Oct 2026 09:20 |
v7 | device evened out | Diagonal | 8 | 1.3% | 0.54 | 0.92 ★ | S2, S3, S5, S6 | 81% in 30% | 2.7× | 0.66 |
| v27 29 Oct 2026 17:17 |
v6 | device evened out | Diagonal | 8 | 1.3% | 0.54 | 0.92 ★ | S2, S3, S5, S6 | 81% in 30% | 2.7× | 0.66 |
| v26 29 Oct 2026 16:29 |
v6 | - | Diagonal | 5 | 1.0% | 0.45 | 0.59 | S3, S4 | 19% in 5% | 3.4× ★ | 0.21 |
| v25 29 Oct 2026 16:28 |
v6 | device evened out | Diagonal | 5 | 1.0% | 0.49 | 0.41 | S2, S3 | 83% in 31% ★ | 2.7× | 0.39 |
| v24 29 Oct 2026 16:27 |
v6 | device separately | Diagonal | 15 | 1.0% | 0.66 ★ | 0.54 | S5, S6, S7 | 31% in 10% | 3.1× | 0.62 |
| v23 29 Oct 2026 16:26 |
v6 | device ignored | Full | 5 | 0.7% ★ | 0.63 | 0.37 | S2, S3 | 41% in 18% | 2.2× | 0.42 |
| v22 29 Oct 2026 16:26 |
v6 | - | Full | 5 | 1.0% | 0.44 | 0.60 | S3, S4 | 19% in 6% | 3.4× | 0.21 |
| v21 29 Oct 2026 16:26 |
v6 | device separately | Full | 15 | 1.0% | 0.65 | 0.51 | S5, S6 | 27% in 9% | 3.0× | 0.57 |
| v20 pick 29 Oct 2026 16:24 |
v6 | device evened out | Full | 5 | 1.0% | 0.50 | 0.27 | S2, S4 | 83% in 25% ★ | 3.3× | 0.52 |
| v19 29 Oct 2026 16:22 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v18 29 Oct 2026 16:20 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v17 29 Oct 2026 16:20 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v16 29 Oct 2026 16:19 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v15 29 Oct 2026 14:10 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v14 29 Oct 2026 14:09 |
v6 | - | Diagonal | 4 | 1.0% | - | - | - | - | - | - |
| v13 29 Oct 2026 14:08 |
v6 | - | Diagonal | 9 | 1.1% | - | - | - | - | - | - |
| v12 29 Oct 2026 14:08 |
v6 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v11 29 Oct 2026 14:01 |
v5 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v10 29 Oct 2026 13:10 |
v4 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v9 29 Oct 2026 13:10 |
v4 | - | Diagonal | 7 | 89.8% | - | - | - | - | - | - |
| v8 29 Oct 2026 13:10 |
v4 | - | Diagonal | 10 | 1.0% | - | - | - | - | - | - |
| v7 29 Oct 2026 13:09 |
v4 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v6 29 Oct 2026 12:51 |
v3 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v5 29 Oct 2026 12:51 |
v3 | - | Diagonal | 10 | 1.0% | - | - | - | - | - | - |
Search results
49 setups · 30 Oct 2026 12:43 · judged on orderedBest: K-means · device evened out (8 segments). Its target segments (S2, S3, S4) hold 77% of ordered in 20% of the training snapshots, 3.8× the overall rate, and come back at 0.96 stability when refitted.
Every setup was fitted on the same 1,828 training snapshots (1,611 people); the 20% of people held out for testing weren't used, so Outcomes can still test the winner honestly.
- Every setup was fitted on the same 1,828 training snapshots (1,611 people); the 20% of people held out for testing weren't used, so Outcomes can still test the winner honestly.
- Score = share of ordered captured × lift, with full credit only when the target segments' stability is 0.5 or more. Setups with over 25% unknown are halved.
- Stability is for the whole segmentation, Target stability just for the target segments. A setup with a slightly lower score but much higher stability can be the safer choice.
- Build on full data makes the real segmentation with that setup on everyone. It then appears below as a normal run: name the segments, compare, take it to Outcomes, and export it.
| # | Setup | Segments | Unknown | Separation | Stability | Captured | Lift | Target stability | Score | |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | K-means · device evened out | 8 | 1.3% | 0.64 | 0.96 | 77% in 20% | 3.8× | 0.96 | 2.93 | |
| 2 | Bayesian Gaussian mixture · Diagonal · device evened out | 8 | 1.0% | 0.54 | 0.63 | 84% in 31% | 2.7× | 0.75 | 2.25 | |
| 3 | Bayesian Gaussian mixture · Full · source ignored | 8 | 1.0% | 0.60 | 0.69 | 78% in 27% | 2.9× | 0.54 | 2.23 | |
| 4 | K-means · device separately | 18 | 1.3% | 0.73 | 0.90 | 57% in 14% | 3.9× | 0.94 | 2.23 | |
| 5 | K-means · source evened out | 8 | 1.1% | 0.63 | 0.76 | 59% in 16% | 3.7× | 0.72 | 2.19 | |
| 6 | Gaussian mixture · Diagonal · device evened out | 8 | 1.0% | 0.55 | 0.63 | 81% in 31% | 2.6× | 0.72 | 2.14 | |
| 7 | Gaussian mixture · Full · device evened out | 8 | 1.0% | 0.55 | 0.61 | 81% in 31% | 2.6× | 0.78 | 2.13 | |
| 8 | Bayesian Gaussian mixture · Diagonal · device ignored | 8 | 1.0% | 0.67 | 0.81 | 81% in 31% | 2.6× | 0.77 | 2.11 | |
| 9 | Gaussian mixture · Full · device ignored | 8 | 1.0% | 0.66 | 0.77 | 81% in 33% | 2.5× | 0.72 | 2.00 | |
| 10 | Bayesian Gaussian mixture · Full · device evened out | 8 | 1.0% | 0.55 | 0.59 | 78% in 31% | 2.5× | 0.74 | 2.00 | |
| 11 | Gaussian mixture · Diagonal · device ignored | 8 | 1.0% | 0.67 | 0.84 | 78% in 31% | 2.5× | 0.75 | 1.99 | |
| 12 | Gaussian mixture · Full · device separately | 19 | 1.1% | 0.67 | 0.82 | 62% in 20% | 3.1× | 0.56 | 1.96 |