Refused, never guessed.
A question too large to count exactly is refused rather than sampled.
MEASURE · ANALYTICS · HOW WE COUNT
The rules behind every number in Praxis Analytics, as the product holds them in October 2026. One query engine applies them to the Overview, Reports, the copilot's answers, the CSV downloads and every alert, and funnels run on an engine built to the same rules, so the same question gets the same number wherever you ask it.
| Where you read a number | The rules on this page that make it |
|---|---|
| "Overview" tiles and "Reports" | People and visits · No sampling · Consent |
| "Sources" and the "Campaigns" tab | Attribution · Channel groups |
| Revenue questions and campaign revenue | Revenue |
| "Goals", and any answer about a goal | Goals |
| "Funnels", "where did they go?" and "why?" | Funnels · What drives drop-off |
| "Do visitors come back?" | Retention |
| "Set an alert" | Alerts |
| Explore | People and visits · Consent |
| The copilot's answers, the CSV downloads | One engine |
Behind "why?" on a funnel step, every signal is tested in the same way.
| Part | Method | Values |
|---|---|---|
| The table | A two-by-two of did or did not, against converted or dropped | for each signal |
| The size | A risk ratio, with a Katz log interval; 0.5 is added to every cell when one is zero | 95% |
| The test | The N-1 chi-squared test, or Fisher's exact test when the cells are small | |
| Many signals | A Benjamini-Hochberg correction across every signal | q 0.1 |
| Floors | Enough people behind a signal before it is judged: 25 people, or 2% of the step's people when that is fewer, and 5 outcomes on each side | 25, 2% and 5 |
| Refusal | No judgment when one outcome outnumbers the other | 10 to 1 |
| The label | A solid finding reads "Correlated, not proven"; an early one is "a hint, not a conclusion" |
A question too large to count exactly is refused rather than sampled.
Visitors who never consented are never modelled back in, and every answer says who is missing.
A goal or measure that changes is versioned, and an answer built on it says so.
A driver of drop-off is tested, corrected, and labelled for what it is.
No query can reach another site's data.
The copilot asks the same deterministic engine as the screens.
It counts only visitors who allowed analytics, it honours Global Privacy Control and Do Not Track, it never counts your previews, and it never estimates the visitors it could not see.
Yes, when you edit what the goal matches. The goal takes a new version, its history is restated, and every answer that uses it says so.
Their raw events are deleted, what remains of them is anonymised, and they leave every retention cohort.
Days are UTC days, and the screens say so.
First touch, written once for each visitor, and last touch, the latest tagged arrival, side by side.
No. The copilot asks in words; the same deterministic engine as the screens does the counting.
Praxis never converts money, so it will not add amounts in different currencies together.