Claims and sources checker

It looks for the same claim shapes our own content gate looks for before a page of ours publishes, then reports what sits beside each one. It never opens a link and never decides whether a claim is true.

Give it a page and it splits the visible text into blocks, then into sentences, and flags every sentence carrying a percentage, a multiplier, a money figure, an “N in M” ratio, or a phrase that borrows authority without naming anyone.

For each one it reports what shares that block: a link off-site, a link back to your own pages, a cite element, a footnote marker, wording like “according to”, or a date. If nothing does, it says so plainly.

What this checks, and what it refuses to claim

The page is fetched once, server-side, as Attensira. Everything inside script, style, noscript, svg, template, head, iframe and canvas is discarded. What survives is split into blocks — one paragraph, list item, heading, table cell, blockquote or caption each — and then into sentences.

Each sentence is matched against five shapes: a percentage, a multiplier, a money amount, an “N in M” ratio, and phrases that assert somebody found something without naming who. Then the block around the flagged sentence is read for sourcing: links off-site, links back to the same site, in-page footnote references, a cite or time element, a bracketed footnote marker, attribution wording, and dates.

A link in the block is not verification, and this tool never treats it as one. We do not open the link. We cannot tell you whether the page on the other end says what your sentence says, whether it is still live, or whether it is itself sourced. “Linked source in the same block” means exactly that and nothing more — the reading is yours to do.

The reverse matters just as much: a sentence this tool did not flag is not a sentence it approved. It is a pattern matcher, not a reader. “The market leader”, “the fastest platform available” and “everyone in the industry agrees” carry no number and no hedge phrase, so they pass through untouched — and they are claims that need a source just as badly.

There is no score. Turning “eleven claims, four of them unsourced” into a number out of a hundred would need a weighting we would have to invent, and the invented part would be the part you acted on.

Why we built this for ourselves first

The claim patterns are our own content gate’s, turned outward. Every page Attensira publishes passes that gate before it ships: a number in prose has to have a claim entry behind it carrying a source URL and the date it was retrieved, or the sentence gets rewritten qualitatively. We wrote it after finding unsourced numbers in our own drafts — the kind that get repeated, cited and quietly become facts.

The sourcing half cannot be the same on your page, and we will not pretend it is. Our gate reads a declared source and a retrieval date we hold alongside the text; on a page we only fetched, all we can see is what shares the block with the sentence. That is a weaker test, and it is the strongest one available from the outside.

The rules that check produces are published rather than described: how Attensira reports the numbers it publishes states them, including the ones that make a page refuse to render rather than show a figure it cannot stand behind.

It matters more now than it did for search. An assistant asked “how big is this market” will happily lift a figure from a page that carries no source, attribute it to your brand, and repeat it to the next person who asks. The number becomes yours the moment it is quoted, whether or not you made it up and whether or not you can defend it.

How to fix what this finds

Nothing in the block sources it. Either link the primary source in the same paragraph — not in a references section three screens down — or rewrite the sentence qualitatively. “Most buyers now start in an assistant” needs no source; “63% of buyers do” does.
Only links to this same site. Fine when the link leads to your own published research with its method and sample on the page. Not fine when it leads to a product page, which is what most internal links on a claim turn out to be.
Names a source in words, links nothing. “According to a recent report” is unfollowable. Name the publisher, the year and the report, and link it.
A date, and nothing else. A date answers when, not who from. It is worth keeping — a 2019 figure presented as current is its own problem — but it is not the source.
An appeal to authority with no authority named. “Studies show” and “experts agree” are the two most quotable sentences you can write and the two least defensible. Name the study or drop the appeal.
Claims in the nav, header or footer. These are flagged too, and they are usually site-wide furniture. A price or a statistic repeated on every page is a claim repeated on every page.

Questions people ask about this check

Does this tell me whether the claims on my page are true?

No, and nothing that fetches a page once can. It tells you which sentences assert something checkable, and whether a reader standing at that sentence has anywhere to go to check it. Truth is downstream of that, and it needs a person who reads the source.

Why flag a sentence that has a link right next to it?

Because a link is presence, not proof. The verdict distinguishes between a link that leaves your site, a link back into your own pages, wording that names a source without linking it, and nothing at all. A claim with an off-site link is reported as exactly that: sourced in the block, unread by us.

Why does it skip whole sentences I know are claims?

It matches five shapes and nothing else. A qualitative claim, a comparison with no number, a promise about outcomes, or a statistic written out in words (“two-thirds of buyers”) will not match. Treat the list as a floor: these are the claims a pattern can find, not all the claims on the page.

The page looks empty to the tool but full in my browser. Why?

We do not execute JavaScript, and neither does a crawler that fetches your raw HTML. If your body is assembled in the browser, the text a model sees is the text we saw: very little. That is a finding about how the page is served, and the page token inspector shows the same gap from the retrieval side.

Do you store the pages I check?

Not against you. There is no account, no email field and no history you could be looked up in, and nothing here observes what any assistant fetches — we read the page ourselves, once, because you asked. The findings are held for about an hour in a shared cache keyed by the URL, so an identical check in that window is answered from the cache instead of fetching the page again.