How measurement works

AI engines do not give the same answer twice. The same question can name different brands, in a different order, citing different pages, from one day to the next. Rixe therefore never relies on a single answer: it asks your questions repeatedly, records every answer in full, and reports how often something happened together with how sure that figure is. This page explains the cycle from question to score, so you know what every number on the screen stands for.

#The daily cycle

Measurement is daily on every plan. Each day goes through the same stages:

1. Planning (02:00 UTC). Rixe works out which questions to ask on which engines that day, for every workspace whose measurement is active. Questions are only planned for engines your plan includes in your market (see Engines and markets).
2. Asking. Each question is sent to each engine as a separate run. Rixe stores the whole answer, including the pages the engine cited as its sources.
3. Reading. Every answer is read to find which brands it names (yours and your competitors'), the position each brand appears in, whether it recommends them, the tone it uses about them, and which pages it cited. Brand matching follows your market's language rules — for example, Turkish suffixes attached to a brand name are recognised as the same brand.
4. Scoring. The day's readings are rolled up into rates per brand, engine and topic, each with its sample size and confidence interval. Scores are updated after the answers have been read, shortly after the nightly run.

Your very first measurement is an exception: it is queued as soon as setup is complete, so you do not wait for the night. See Quickstart.

#Runs, answers and samples

A run is one question asked once on one engine. Every run that completes produces one answer. A rate such as visibility is always answers that named you ÷ answers collected over a period — by default the last 30 days, and you can switch to 7 or 90.

That count of answers is the sample. It is shown next to the numbers it supports — for example "1,240 answers · 48 prompts · 3 engines · last 30 days" under the title of AEO Insights › Visibility, and the number of Runs in each table row.

When an engine answers without showing an AI answer at all — Google, for instance, does not show an AI Overview for every search — that still counts as a run in which no brand was named. It is a measurement result, not a failure.

#Why one answer is never a score

Rixe will not turn a small sample into a confident-looking number:

• At least 4 runs. Below four runs, a rate is shown as Not enough data (some screens say insufficient sample) rather than as a percentage.
• Every rate has a confidence interval. Next to each rate Rixe shows a Wilson confidence interval — the range the true rate is likely to fall in (95% confidence). It is wide on a small sample and narrows as answers accumulate. Two numbers whose intervals overlap have not really separated, however different they look.
• A change has to be real and large. Rixe only reports that something changed when both periods have at least 4 runs, the difference is statistically significant (p < 0.05) and it is at least 10 percentage points. Either condition alone produces alerts nobody can act on.

The maths is explained in Visibility score.

#When an engine fails

Engines time out, rate-limit, or go down. Rixe never hides this:

• A run that could not be completed is recorded as failed, with the reason (the engine did not answer in time, rate-limited us, returned an error, and so on).
• Failed runs are left out of the rates instead of being counted as "not mentioned". Your score is computed from the answers that actually arrived.
• The screen tells you: a degraded engine is named in a banner on AEO Insights › Visibility ("These engines are degraded, so their data is incomplete"), and each prompt's own page lists how many runs completed and how many failed in the period, and why.

#Shared answers

If two workspaces ask exactly the same question, in the same country, on the same engine within 24 hours, Rixe may reuse the answer it already collected instead of asking again. The answer is marked pooled on the prompt's page. This keeps measurement affordable without changing what is measured: it is still a real answer from that engine to that question in that market.

#What can slow or stop measurement

#Reading the evidence

Every number in Rixe can be traced back to the answers behind it. Open any question in Prompts › Your Prompts to see each engine's answer in full (Open answer), the brands it named, and the sources it cited. When a figure on a dashboard looks wrong, that is where to check it.

#Related pages

• Concepts — definitions of run, sample, confidence interval and other terms.
• Visibility score — how the 0–100 score is built.
• Engines and markets — which engines are measured and why.
• Plans and limits — what your plan includes.