Echo Lab

How AI answers really work.

Echo Lab is where we publish what we measure about how ChatGPT, Perplexity, Gemini, Claude, Grok and Google's AI search, read, pick and cite their sources. Our own data, the method next to every number, and what it doesn't prove.

Studies

What we're studying

The first studies, in the order we'll publish them. Each one answers a question teams ask us every week.

  1. 01

    How stable is an AI answer?

    Ask the same question several times on the same day: how often do the brands, their order and the sources change? And how many runs does a visibility number need before you can trust it?

    Next
  2. 02

    Retrieved is not cited

    Of the pages an assistant fetches while answering, which end up cited, and what do they have in common: page type, structure, length, freshness?

    Next
  3. 03

    Five assistants, five sets of sources

    How much do the assistants agree on sources for the same question, and which kinds of sites does each one prefer?

    Next
  4. 04

    What the AI searches before it answers

    Which web searches does an assistant run behind a question, how many, and how far are they from the words people typed?

    Planned
  5. 05

    From crawl to citation

    When an AI crawler reads a new or changed page, how long until it shows up in answers, if it does at all?

    Planned
  6. 06

    Does fixing the page change the answer?

    Pages that were changed, against similar questions whose pages weren't: what moved, and how sure can we be?

    Planned

Rules

How we publish

Our own data

Only what Echo measured. Every figure says when it was measured, how many answers it rests on and which assistants.

Method and limits

Every study says how it was done and what it doesn't show. Attribution is called attribution; cause only when an experiment shows it.

No brand names from customers

Customer data only ever as anonymised totals. Never a workspace you could recognise.

Data you can check

The totals behind every chart as a CSV, a date and a version on every study, and updates when we have more weeks of data.

Next

Also on the way

How Echo measures

Every metric's definition, how each assistant is asked, what counts as a citation, and what Echo can't see.

AI citation index

A monthly table of the most cited sites and kinds of sites, per category and per assistant.

Engine changelog

Short, dated notes when an assistant changes how it searches or cites.

AI crawler list

The AI crawlers Echo recognises, as a public file, kept current.

Join the Lab

Do this work with us

We're hiring researchers, not developers: people who design a measurement that holds up and know when a result is noise.

  • Measurement scientistMake Echo's numbers ones a statistician would sign: how AI answers are sampled, how sure each number is, and experiments that show what a change moved.See the position
  • AI search researcherFind out how ChatGPT, Perplexity, Gemini, Claude and Google's AI search, read and pick their sources, and turn it into what Echo measures and recommends.See the position

A question we should study?