Pages cite their sources
Do the site’s articles link to other sites as sources?
What we look for
On each sampled article (the same pages the author and dates criteria judge), we count links in the post’s own content that go to other sites. A post’s own content is the body of its first <article> or h-entry (its e-content or articleBody when it marks one; else the article without its header, footer, byline, tags, and share controls); a page with neither is read from its <main>, less the same parts. Nothing after the post counts: not a second entry under “Related”, not a form the site puts under every article. Links and text that repeat on every page we read are your site’s furniture, whatever element they sit in (a second copy of the menu in a <div>), and don’t count either. An author’s box is a byline: its LinkedIn or X link names the author and cites nothing. Your own sites are not sources: a site your home page’s navigation links, or one your home page links anywhere when the post’s link to it carries a campaign tag. One domain is one site, however many of its hosts you link, and a host’s short form is that host (youtu.be is youtube.com). Links in the navigation, header, footer, sidebar, comments, widgets, or a form don’t count, and neither do links to your own profiles (your sameAs and rel="me" pages, a profile whose handle is your domain, the repository or podcast page your home page’s navigation or footer links), a byline’s link to its author, a footer or a menu marked only by its class, comment links, sponsored or campaign-tagged links (a sponsor, an affiliate, a tribute store), share buttons (posting to X, Facebook, LinkedIn, Reddit, Bluesky, Threads, and the like), sign-in links, or ads. Your own www and bare domain count as your site, not another one, and so does any other host under your domain: a link to your own subdomain is not a source. But a post whose text links only your own subdomains isn’t failed for it; it is left out of the count. A post that links nowhere at all is judged.
Two or more other sites linked from an article scores it 1; one scores 0.6; none scores 0.2. (Sites are counted, not links, so three links to one site are one site.) The score is the average across articles. If no sampled page is an article, we skip the criterion. This criterion reads pages we already fetched, so it costs no extra requests.
We only judge real articles, and only when at least two are left to judge once the short ones and the ones that link only your own sites are set aside; otherwise we skip, and when you do publish posts and we read too few of them, we say we couldn’t read enough, never that you have too little writing. A post that holds only a paywall’s notice isn’t judged. Neither is a post whose own text is under 300 words: a note, a link post, or a release note has nothing to cite. A post’s length is its own text, its title and its body; the byline, the tags, and the share buttons don’t count toward it. If every post we read is that short, we skip the criterion and say so. Never a post, whatever it carries: your about page, a press release, reference docs (/docs/, /help/, /guides/, /api/), a login, signup, cart, thank-you, or test page, a landing page your navigation lists as a section, a page of five or more dated entries, or the home page. Any kind of site that publishes posts is asked, a company with a blog or a changelog as much as a magazine; a page is a post only by its own content: it shows a date or a byline of its own (in the article itself, or in the structured data that describes the page as an article), or a feed that your site or the page itself links has an entry for it, or it has a place among your posts: an <article> at an address where writing lives (/blog/a-post, /2026/11/a-post), or a page one of your indexes of posts labels as an article. Such a page is a post even with no date and no byline, which is how we can tell you to put a date on it. A label alone, anywhere else, makes no post: not an <article> your template wraps around every page or around an intro paragraph, not og:type, not an author card every page carries in its footer, and not a date we judged a build’s stamp (such a page leaves the count; it is never an undated article). A page under your about section, or one your navigation lists as a product, tool, or landing page, is never a post, and neither is a page that only carries its own dateModified or a legal page’s “Last updated”: a modified date never dates a post. The pages we sample include your two newest posts whenever your sitemap lists any; translations, logins, test pages, and legal pages are sampled only when nothing else is left. Reference and docs pages, press releases, tools, store pages, and logins are never judged.
Why it matters
Links to sources show your work. Writing that cites where its facts come from is easier for a reader — or a machine — to check and to trust, and it places your page in the web of things already known to be reliable.
AI assistants are built to cite their sources, and they lean toward pages that do the same. Models cite sites that cite.
How to pass
- When an article states a fact, a number, or someone else’s idea, link to where it came from — in the text, where it’s used:
<p>Clinics lose about a fifth of appointments to no-shows, according to <a href="https://example.org/study">a 2025 study</a>.</p> - Link to the original source rather than to someone else’s summary of it.
- Prefer a couple of good sources over a long list; two on a page is enough to pass.
Evidence
-
Aggarwal et al., “GEO: Generative Engine Optimization” (KDD 2024) · August 2024 · moderate
Adding quotations, statistics, and sources raised a page’s share of the answer by about 30 to 40%, once the page had already been retrieved.
10,000 queries in a simulated answer engine.
-
Martinez, a survey of 45 studies · July 2026 · contrary
Replications are weak. The evidence is strong that a retrieved page’s wording can sway an answer, and thin that these tactics help a page get found at all.
A review of the published work on optimizing for answer engines.
Last changed .