Fireshift / Method / AI use

How we use AI

By the Fireshift team · Updated 2026-11-05

Short answer

AI agents do a large part of the work at Fireshift: they research, draft and edit pages and emails, write and test code, and run the studies. They don't make up the findings. The scan is fixed rules in code, and every number we publish comes from our own scan output or a source we name, and is checked against that file or source before it goes live.

Where AI helps

  • Research. Reading documentation, forum threads and vendor pages, and summarising them with links. Facts we use get a source and the date we read it.
  • Writing. Drafting and editing guides, help articles, this page, and emails to customers and prospects.
  • Building. Writing the scanner code, the site and the tests that check them.
  • Running studies. Starting the scans, counting the results and writing them up.

Where it doesn't

  • Scoring. The Leak Score and every finding come from fixed rules in code (the rules). The same container gets the same result every time. No model reads your site and gives an opinion.
  • Numbers. No number is written by a model from memory. Each comes from a scan output file or a cited source.
  • People. We don't use AI to invent authors, reviews, quotes or customer stories.

How a number gets checked

  1. Every study writes an aggregate file with the scan date, the sample, the counts and the limits.
  2. A page quotes the number from that file, with the date and the sample size next to it ("of 206 sites on 5 Nov 2026").
  3. The number is checked against the file before the page goes live. Outside facts carry their link and the date we read them.
  4. The site build refuses a page with a broken internal link, a missing title or description, or phrasing on our banned list.
  5. The scanner has regression tests built from real false positives, and the browser scan is tested line by line against the server one.

What can still go wrong

Checks catch a wrong copy of a number. They don't catch a rule that measures the wrong thing. That happened: our first duplicates figure used a loose method, and the scanner once read a video event as a trial ending. Not every page is read by a person before it goes live, so when we find an error, we fix it and log it in public: corrections. If you spot one, email hello@fireshift.ai.