Skip to content
Compare · the same URL, three ways

Litmus is the predatathat matters.

Any tool can hand an agent the text of a page. The question that decides whether the answer is any good — should this be trusted, cited or dropped — is the one nobody was answering.

The same URL, three wayslitmus
Clean Markdownyesyes
Named authorshipnosometimes
Freshness signalnosometimes
Paywall flaggednosometimes
Same input, same outputyesno
Cost per callplantokens
extractmodellitmus
Side by side

The same URL, three ways.

"Reading it yourself" means handing the raw page to a model and asking what it thinks. It is the honest comparison, not a straw man — and it loses on repeatability, not intelligence.

For the same URLPlain extractionA model reads it Litmus
Clean Markdown Yes Yes Yes
JavaScript renderedSometimesDepends Yes
Source type + authoritySometimes Yes
Named author or anonymousSometimes Yes
Freshness + staleness warningSometimes Yes
Paywall flagged, not truncatedSometimes Yes
Error page / captcha detectedSometimes Yes
Syndicated copy detectedRarely Yes
Citations + reference densitySometimes Yes
Thin content + boilerplate ratioRarely Yes
Same input, same output YesNo Yes
Auditable in your own logsNo Yes
Cost per callPlan-basedTokens$0.002

$0.002 flat · charged only after a successful render

Structural blind spots

What a scraper cannot know, by design.

Not a criticism of extraction tools — they were built to do a different job. But the gap is real, and it is where agents fail.

It cannot see the author

A byline is markup, not text. Strip the page to Markdown and "By Jane Chen, Health Correspondent" either disappears or arrives as a stray line nobody parses.

It cannot see the date

Publish dates live in metadata, JSON-LD and headers far more often than in the body. The text of a 2019 page and a 2026 page look the same.

It cannot tell a gate from a page

A paywall teaser, a soft 404 and a captcha all produce text and a 200. Extraction has no opinion about whether that text is the article.

Determinism

Why not just ask a model?

Because a judge that changes its mind is not a judge you can log. A model can absolutely form an opinion about a page. The problem is what happens on the second run.

Same page · two runs
model run 1probably reliable
model run 2hedge, cite with caution
litmuscite · identical

Model in the path

  • Different verdict on the same page tomorrow
  • Unbounded cost per call, paid in tokens
  • No audit trail — "the model said so"
  • Confident when wrong, with no field to check

Deterministic in the path

  • Same verdict on the same page, always
  • Fixed cost — $0.002, metered per call
  • Every signal inspectable in the payload
  • Fails visibly — null and "unknown", never a guess
Litmus is not a smarter reader. It is a repeatable one — which is what you need before an agent acts on the answer.
Where we are honest

What Litmus does not do.

A trust product that overstates itself is worthless, so here are the limits, stated plainly.

We flag paywalls, we don't break them

If a page is gated, you get paywall.detected: true and a note — not the article. Bypassing a gate is not the product.

Authority is a heuristic

"Authority" is a score from observable signals about the domain and page, not a guarantee of correctness. A high score means "worth reading", not "true".

Unknown is an answer

When a signal cannot be determined it comes back null or "unknown". We would rather hand you a gap than a confident fabrication.

One test costs less than one wrong citation.

No account, no key, no plan. Send a URL and read the verdict.

10 free parses a month · no account · no API key