# How a measurement is made

The sampling, scoring and limits behind a reading.

Source: https://praised.co/docs/methodology
Updated: 2026-10-14

Praised collects answers for tracked questions on selected engines. The measurement retains the engine and model information behind an answer, rather than inferring an AI answer from a search ranking.

## Collect and judge the answers

Scoring extracts mentions, position, sentiment, recommendations, cited domains and factual claims. Human labels can be used to calibrate the judge.

## Estimate rather than overcount

Answers to the same question are correlated. The statistical method groups answers by question when estimating headline uncertainty, rather than treating every answer as an independent buyer question.

The method keeps grounded answers and answers from model memory separate. Location information also needs a stated mechanism: a query parameter and a location hint do not measure the same thing.

## Respect the measurement boundary

A result measures the selected engine configuration and sampled questions. A consumer interface can use different routing or settings. The reading should not be treated as a census of every answer every buyer receives.

Modelled authoring previews are labelled separately and do not count as measured visibility. A sample or simulation is not a live measurement.

A changed answer surface or collector is a dated marker on the chart, and the line continues. Copilot measured Copilot Search inside Bing until 29 September 2026, and the Copilot app from that date.