Your past winners, re-analyzed by Probant’s engine.
Import your test history from AB Tasty, VWO, Kameleoon, or another tool. Probant re-analyzes each test from its visitors and conversions, and shows which of your winners hold up.
Three steps, and nothing is saved before the last one.
In your workspace, go to Library → Import. Importing is limited to editors and owners, and recorded in the audit log.
1. Export
- AB Tasty: the results export, whose columns you map on screen
- VWO: a test’s “Summary” report, as CSV
- Kameleoon: an experiment’s results returned by its API, as JSON
- Optimizely: the CSV export of an experiment’s results page
- Convert: the report’s “Tabular CSV, Aggregated” export
- GrowthBook: the CSV export of the results
- Any other tool: Probant’s generic CSV, one row per variant (template below)
2. Upload and review
Semicolon or comma separator, French or English decimals, UTF-8 or Latin-1, up to 768 KB and 5,000 rows. The preview shows each test with its recalculated verdict, and each rejected row with the reason (more conversions than visitors, missing original, end date in the future…).
3. Confirm, or undo
The import is saved in one go: the whole file, or nothing. It can be undone afterward, which deletes every test it created. A management API does the same from a server (POST /api/v1/imports, as CSV or JSON).
Generic CSV template
The first line of the file, then one row per variant. Semicolon or comma separator.
test;zone;variante;role;debut;fin;visiteurs;conversions;revenu;revenu_ecart_type;devise;part_prevue;annonce;gain_annonceProbant starts from the raw data, not the conclusions.
Imported
For each variant: visitors (or sessions, or visits, named as in the export), conversions, revenue and, if the export includes it, its spread per visitor. Dates, the goal, and the planned traffic split when known. And what the tool reported, winner or not and the gain shown, displayed alongside under the tool’s name.
Not imported
The original tool’s “chance to win”, probability to be best, p-values, and “significant” labels. Revenue without a known spread gets no range: Probant flags it instead of computing one. Neither the visitors themselves nor the events are imported: an imported test cannot be relaunched and does not count toward your plan.
The same test, read differently.
Probant re-reads each test at a two-sided 5% threshold, split across variants when there are several, with a single look at the end. The original plan is unknown: if the test was checked daily and stopped as soon as it seemed to win, its winners are even weaker than the re-analysis shows.
The difference and its range
Computed from visitors and conversions, with the same engine as the tests you launch in Probant.
The winner’s curse
On average, a winner overstates its effect. The corrected gain brings it closer to what your tests actually deliver. The winner’s curse
The false positive risk
For each winner, the share of winners obtained under these conditions that would in fact have no effect, based on your tests’ success rate.
The traffic split
Checked when the planned traffic split is known: a sample ratio mismatch distorts the measurement, whatever verdict was reported.
In the demo: 7 announced winners, 3 hold up after recalculation.
Maison Lune, the demo’s fictional store, imported 12 tests run with AB Tasty. Every figure in its migration summary is computed by the engine from its visitors and conversions.
- Announced winners that hold up
- 3 of 7 · 3 winners in total according to Probant, announced or not.
- Average announced gain
- +9.5% · Average of announced winners (7), as the tool showed it.
- Gain corrected for the winner’s curse
- +4.6% · The same tests, pulled toward what the workspace’s tests bring in on average.
- Average false-positive risk
- 12% · Among winners that hold up, assuming each test was sized for 80% power, the usual convention. With less power, it would be riskier.
Imported history is useful from day one.
Imported tests join the learnings library with a badge for their original tool. They count toward the decision quality score, where anything that cannot be verified (sealed plan, planned sample size, decision made) is marked as such rather than as passed. They feed the prior used for the winner’s curse correction, and the one used by the Bayesian reading, which needs 10 stopped tests.
Conclusive Or not yet…
Create your workspace, or explore the demo first with data from a fictional store.
- Open demo, no sign-up
- Free to start, no overage charges
- Database and compute in Amsterdam, with Scaleway