Skip to content

Trace a passage

Find which of your pages a piece of text came from, word for word, and which licences covered it.

On this page

You found a paragraph of yours somewhere it shouldn't be: an AI answer, a forum post, a content farm. Trace a passage tells you which of your pages it came from, word for word, which naulon licences were sold for those pages, and whether your page says what the text claims.

Open Trace a passage in the dashboard, paste the text, and press Trace. Found it on a web page? Switch to Link, paste the address, and we read the page's main text and trace that. Before your first trace, Try a passage fills in a few lines from one of your own pages so you can see what a match looks like.

What you get back

  • Your paste, with every copied run highlighted. Each source page has its own colour, and a bar above shows how much of the paste each page supplied.
  • For each page: its title, how many words of the paste came from it, and its licences: how many were sold, for how much, and when. Open the list to see each one with its amount. The terms are explained once below the list (usually AI input: read, quote and summarise for the person who paid, no republishing).
  • A Verify link per licence. It opens the public verifier, which checks the signed licence against your site's keys in the reader's own browser. Show all lists every licence on the page, fifty at a time.
  • Sold for this exact text on a licence whose recorded fingerprint is identical to the text you traced. A licence records a fingerprint of what it paid for, so this one covered these very words. When the fingerprints differ we say nothing, because the buyer may simply have taken the page as HTML.
  • Copy summary, which puts the source URLs, the copied excerpts and the licence links on your clipboard as plain text, ready for an email.

Does your page say it?

Under the trace, each sentence of what you pasted comes back beside the paragraphs of your page closest to it. It starts on its own with the page the trace credits with the most words. Pick another matched page, or Another page and paste an address of yours, when the text is about a page it does not quote. An AI that rewords usually leaves nothing to trace, so this is the common case for an answer about one of your articles.

Each sentence gets one mark:

  • A figure your page doesn't give. A number in the sentence (40,000, 1453) that none of the closest paragraphs states. This is the commonest way an answer overreaches its source, so it leads.
  • Not in your page. Fewer than half of the sentence's words appear anywhere in your page.
  • Partly in your words. Half or more, under three quarters.
  • Close to your words. Three quarters or more, and no missing figure.

Next to each mark is the count it rests on, such as 7 of 9 words in your page, and the paragraphs are shown with the shared words marked. Words like "according to" or "the article says" are left out of the count, because they describe your page rather than repeat it.

This compares words and figures, not meaning. A sentence close to your words can still turn them around, so read each one beside its paragraph before you act on the mark.

We read the page from your site's own server, the one you proved under Connection, as it is at that moment. Neither the page nor your text is kept. A site can check 30 pages an hour. The address must be a page on one of your sites, with its domain: a bare path such as /articles/x is refused, because you may run more than one site.

Keeping it as evidence

Save as evidence stores the trace with today's result: the text, what matched, and the licences as they stood. It's listed under Saved traces for everyone on your team, and it doesn't change if the page or its licences change later.

Turn on Share with a link to send it to someone without an account, such as a lawyer or the site that copied you. The shared page never shows who bought the licences. Turning sharing off breaks the link, and turning it on again makes a new one.

A page with no licence is called out. If an agent never paid to read it, any copy you found didn't come from a paid read.

What counts as a match

A match is a run of at least 8 identical words. Case, punctuation, quote marks and formatting are ignored, so a copy that was re-styled or turned into markdown still matches. In Japanese, Chinese and other languages written without spaces, each character counts as a word and a run needs 16.

Reworded text doesn't match. That's deliberate: every highlighted word is on your page in that order, so a match is never a guess.

A match shows that text was copied. It doesn't show that a model was trained on your work, and it says nothing about who did the copying.

What gets searched

Every page whose text we've captured for your catalogue, on every site your role can read. Fingerprints are built on their own once a page's text is captured, and rebuilt whenever the text changes. We store fingerprints, not your text a second time. If you turn off body search for a site, its fingerprints are deleted with its text.

Authors have the same page under Trace a passage in the author menu, with the same example, link reading and saved evidence. It keeps only the pages that paid them and lists only the licences that paid them. An author's saved traces are theirs alone.

Checking a quote from the other side

Anyone holding a licence link can check a quote against it. Open the link, and under a valid licence paste the passage an AI attributed to your page. The verifier shows which words are on the licensed page, and says so when the licence was sold for exactly that text. It never shows who bought the licence.

A link can carry the quote itself. Add #quote= and the passage, percent-encoded, and the page checks it as soon as it opens; Copy link with this quote under a result builds one for you. The quote sits after #, so it is never sent to a server with the page request.

The public trace

Anyone can ask which naulon pages a passage came from, without an account: through naulon's free research plugin in Claude or ChatGPT, or through the endpoint below. It is a smaller answer than yours, on purpose.

  • It searches only sites listed in the public directory. Turn off List in the public directory in your site's settings and your pages drop out of it.
  • For each page it found, it gives the title, the link, and whether the page holds most of the passage or part of it. It never shows your text, the copied words, or who bought a licence.
  • It names a page only when at least 25 words match, and it takes 40 words to 5,000 characters.
  • It shows how many licences a page has sold only when you turned on Credit this site on the homepage ticker.
curl -s https://gate.naulon.app/_naulon/public/trace/search \
  -H 'content-type: application/json' \
  -d '{"text": "<a paragraph of at least 40 words>"}'
{
  "searched": 12,
  "matches": [
    {
      "article": { "title": "The status code nobody used", "url": "https://example.com/blog/the-status-code-nobody-used/" },
      "coverage": "most"
    }
  ]
}

coverage is most when the page holds at least half the passage and part otherwise. A short passage answers 400 {"error": "too_short", "min": 40}, a long one 413 {"error": "too_long", "max": 5000}. Calls are limited per address and per day across everyone, and a refusal carries retry-after. The passage is matched and dropped: it is not stored.