Method
How the check works
Three steps: the record, the automatic marks, and a written verdict.
1. The record
The address is reduced to its domain. If that domain is one we have already investigated, or is part of a network we have, you go straight to the listing. The listing is the answer: it holds the evidence, the operator's reply if there is one, and every sister domain.
2. The automatic marks
If the domain is new to us we fetch the page as a normal browser would and look for the marks that pay-to-play copy tends to carry. Each mark has a weight, and the weights add up to a score out of 100. Among the marks:
- The page's own labels: "Commissioned Research", "Includes paid placements", "Vendors set the topic", "funded jointly by the vendors", "We influence ChatGPT".
- A ranked-list schema with many positions and scores, often paired with an FAQ schema, which is the shape AI answer boxes prefer.
- Three or more named roles on one page: written by, reviewed by, fact-checked by. On a template page these are decoration.
- "Best" or "Top 10" plus a year in the title, and many "out of 10" scores in the text.
- Phrasing typical of machine-written copy, a methodology that says AI agents compiled the ranking, a link to a partner or get-listed page.
- Assets loaded from a known network's servers.
- No company name and no contact route anywhere on the page.
The marks are public so that you can judge them. They are also blunt: a real newspaper's "best laptops" page can trip two or three of them. That is why there is a third step.
3. The written verdict
The page text and the marks go to Claude with one instruction: judge whether the page exists to be cited by machines or to rank its owner first, be fair to ordinary blogs and real review sites, and never invent a fact. It returns a verdict (likely, possible, unlikely, or not enough to judge), a confidence, a short summary and its reasons. We show all of it.
Limits
- A check looks at one page. A network reveals itself across many pages and domains, which is what the listings are for.
- Sites behind a bot wall cannot be read. The check says so, and you can report the site instead.
- A low score is not a clean bill of health, and a high score is not a listing. Listings are made by a person, from evidence.
- Checks are kept for seven days so that repeat requests for the same page do not run twice. They are not indexed and not published.
For developers
A read-only endpoint answers whether a domain is on record and returns its last check, if any:
GET https://llmspam.com/api/check?url=example.com
It returns JSON with listed (name, kind, listing URL, or null) and last_scan (score, verdict, confidence, summary, or null). Please keep to a few requests a minute.