When your agent scrapes a page, it gets text back — but text alone doesn't say whether that text is worth trusting. Was this written yesterday or five years ago? Is this the original source, or a copy of a copy sitting three syndication hops away from the truth? Is this the whole article, or the first three paragraphs before a paywall cuts it off?
Litmus answers those questions on every single parse, automatically. It reads the page the way a skeptical researcher would, then hands your agent a short, plain-language brief instead of making your agent guess and burn up tokens.
Litmus is heuristic, not model-based. Most "trust" layers on competing tools work by calling another LLM to judge the content, which just stacks a second hallucination risk on top of the first. Litmus can't hallucinate its own metrics — every signal is computed with deterministic rules at the edge. Your agent gets a second opinion on whether to trust the data, instantly, with no extra LLM call and no extra hallucination risk. Scrapers extract. Litmus interprets.
Every domain scored and classified — government, academic, news, forum, social. Your agent ranks sources before it reads them.
High, medium, or low — scored from source authority, freshness, outbound link quality, and content signals. Decide whether to cite, verify, or skip.
Scores the quality of every domain the page links to — strong external citations are a trust indicator that can't be faked.
Checks the Wayback Machine for a snapshot of the page — if the live page disappears, your agent knows a cached copy exists.
Publish dates extracted from meta tags, schema, and headers. No date found? Your agent gets an explicit staleness warning.
Flags when a page has been revised or corrected after publication — your agent knows whether it's reading the original version or an updated one.
Flags when content matches wire copy or near-identical text already seen on other domains — so your agent does not mistake five copies of one story for five independent sources.
Flags when a page was cut short — paywall, subscription gate, or partial render. Your agent knows when "clean Markdown" is not the whole story.
The page's own summary, structured and ready to cite. Extracted from the page itself — no LLM paraphrasing.
Prices, percentages, stats — extracted upfront into a structured array. No scanning 3,000 words for the numbers.
News, tutorial, product page, forum, academic paper, and more — classified upfront so your agent can reason about the page the way it is meant to be read.
Estimated minutes to read the full page, computed upfront — your agent knows the depth of the content before it commits.
Measures how much of the page's language is hedged or uncertain. High hedge density warns your agent that claims may be speculative even from high-authority sources.
Flags thin content — pages that are mostly filler wrapped around a small amount of actual information. Even fresh, authoritative pages can be content-poor.
Measures how much of the page is nav, footer, and cookie-banner noise versus real content — your agent knows the extraction was clean.
Correction detection ships inside the freshness object — a correction is a page modified after publication, so the signal is scored against its own dates.