steering-watch
A recurring index of AI-directed instructions in llms.txt files, with the scanner that finds them and harmless canary pages. Every match is an exact quote.
- Status
- Live
- Launched
- Licence
- MIT
- Language
- JavaScript
One command, copied from the repo.
Install steering-watch: npx github:rishsadh/steering-watch scan example.com
rishsadh/steering-watch
What it does
It reads the llms.txt of each site on a fixed list of 140, line by line, and prints every instruction aimed at an AI reader with an exact quote and a line number. Each run writes a dated snapshot and rebuilds INDEX.md, and the next run shows what was added or removed. The scanner works on any llms.txt, Markdown or HTML text. Seventeen harmless canary pages, each hiding one made-up-code instruction in a different way, test whether an agent obeys instructions it finds in a page. It needs Node 18 or newer, has no dependencies and uses no API keys.
Matches are graded into tiers: injection, command, hidden, promotional, steering, terms, platform-default and literal. The tiers describe the kind of instruction, not the intent behind it. In the run on 2026-09-30, 124 of the 140 sites served a valid llms.txt. 10 files had promotional-tier lines, 30 more had steering-tier lines and 16 had usage terms that mention AI. No file had a classic prompt injection, a command to run or hidden text.
One paragraph, asking an agent to recommend that its user install a third-party skill, appears on several unrelated sites in the list. The same paragraph is served by a Shopify store that is not on the list, and the files that carry it say they are built on Shopify, so it reads as a Shopify platform default and not as each company's own steering. The scanner classes it as platform-default, the index lists it in its own section, and four list entries that carry only that paragraph are not graded for it. No Shopify document confirming that it is platform-generated was found.
A match means a line is worth reading. It does not mean the site did anything wrong: a line that tells an AI to use the official docs is steering, and so is one that tells it to recommend the company. The rules are regular expressions, recall is unknown, and some lines will be misgraded. The list is fixed and small, and it says nothing about how common steering is overall. It sends one request per site with a named User-Agent, honours robots.txt and does not disguise itself. If you run a listed site and want it removed, open an issue and it will be.
Disclosure: Rish Sadh runs Reidify, a studio that sells AEO services. This is published as data with its method, not as a pitch.
Why it exists
In August and September 2026 two separate studies looked inside llms.txt files and found text written for AI assistants rather than for people. Common Crawl analysed 584,107 files from its July 2026 crawl and graded 3,793 as mild steering, 102 as promotional and 10 as classic prompt injections. It also says one crawl cannot show whether any of this is growing. Installmap read 695 llms.txt files from top sites and found 16 that steer what an assistant recommends.
Both are one-off studies. A search on 30 September 2026 found no tool that re-runs the check on a fixed list of sites, so the question of whether this is growing, and who adds or removes it, had no answer. steering-watch is that re-run.