Skip to content
FactCheck by GeriAI

How it works inside.

This is the public version of the method: it shows the steps and the principles. Internal thresholds and settings are not published, so that nobody can craft a story to score high.

The trust score measures how much a headline is corroborated in the digital press over a period: how many independent sources report the same thing and whether they support or rebut it. It does not say whether a story is true. Next to the score, every result shows all its evidence: which pieces it found, from which outlets, on which dates, and why each one supports, rebuts or is neutral.

The steps

  1. The story

    A link, or the headline and the standfirst.

    Detour If it cannot be read, the headline and standfirst are requested.

  2. Can it be analysed?

    Only a verifiable claim is measured.

    Detour An opinion, a question, someone else’s fact-check or a forecast gets no score: the result explains why.

  3. The keys AI

    The people, organisations and places it names.

  4. Search the digital press

    Across the public sources TrawlingWeb processes, over the chosen period.

  5. Keep what reports the same AI

    By similarity of meaning, not of words.

  6. Count independent sources

    Copies, the same agency or the same outlet count as a single source.

  7. Look at time

    When it appeared and how far it spread.

  8. Supports, rebuts or neutral? AI

    Piece by piece, against the headline.

  9. Trust score

    Combines corroboration, independence, source quality and time.

  10. Limits

    With little corroboration, or with more rebuttals than support, the score cannot be high.

  11. Result

    The trust score and all the evidence.

Diagram of the eleven steps, from the story to the result, with the two detours where no score is given.

The principles

Corroboration, not truth.
A much-repeated headline is not true because of it, and a true exclusive may have little corroboration at first. The score measures the second, not the first.
Independence.
A hundred outlets reproducing the same wire story are one source, not a hundred.
Supporting is not the same as mentioning.
A piece that covers the topic but rebuts the claim counts against it.
Limits.
Without enough corroboration, or with more rebuttals than support, the score cannot be high.
Transparency.
Every result carries its evidence, and whoever reads it can check it.
AI that classifies, not AI that writes.
The steps marked «AI» use models that recognise names, compare meanings or classify support and rebuttal. None of them writes the result or supplies sources. The rest is rules and calculation.

What is not published

There is more detail than you see here: the thresholds, the settings of each reading and the calibration tests. They are held back so the score cannot be gamed. If you research disinformation and need to know them, write to us. info@trawlingweb.com

Try a headline