What this check tests
The audit never runs JavaScript. It reads the HTML the server returns for each crawled page (the homepage plus a sample of same-site pages) and asks whether there is real content in it. For each page:
- It looks for framework markers anywhere in the raw HTML:
__NEXT_DATA__,window.__nuxt__,ng-app,ng-version,data-reactroot,data-v-app,__svelteandember-application. These are plain text matches, so they can also match in a code sample. - It decides whether the page is a JS-only shell. All three of these must be true: the HTML has an empty mount element, meaning a
divwith id__next,root,appor__nuxt(or adata-reactrootdiv) with nothing inside; there are fewer than 5<a>elements with anhref; and the visible text, at about 5 bytes per word, is under 20% of the size of the<script>tags. A shell page never counts. - For other pages it counts the words of visible body text. Script, style, noscript and template content are not visible text. Words inside
<noscript>are counted separately and added. The page counts as good when the total reaches 50 words, or 100 words if a framework marker was found in step 1.
The score is good pages divided by all crawled pages, rounded.
- Pass: 70 or more.
- Partial: above 0 but under 70.
- Fail: 0.
If markers were found, the result lists the frameworks. A marker is not penalised on its own: a server-rendered Next.js, Nuxt, Vue or Angular site with enough text passes, but it needs the higher 100-word bar.
The check counts words, not usefulness, so a page of cookie-banner text can reach 50. A thin page that is not a shell can fail on the count alone. It does not compare the raw page with a rendered one.
Why it matters to agents
Whether a tool sees your content depends on whether it runs scripts. Google's documentation says that Google Search runs JavaScript with an evergreen version of Chromium, but that rendering happens in a queue after crawling, and that server-side or pre-rendering is still a great idea because it makes your site faster for users and crawlers, and not all bots can run JavaScript.
That last point is the general one. Any agent or tool that fetches the raw HTML and does not render it sees only what this check sees. For a client-rendered site that can be an empty page: no text, few links, nothing to cite. This page does not name any AI vendor's fetcher, because the sources cited here do not state which ones render JavaScript. The safe assumption is that some do not. None of the other Technical checks asks whether your content exists in the first response. Weight is 1.
How to fix it
- Look at what the server returns. Run
curl -s https://example.com/ | head -c 5000and read the output. If the main text and links are not there, a non-rendering tool cannot see them. - Render on the server or at build time: server-side rendering, static generation or prerendering, depending on your framework. Start with the pages the audit samples: the homepage and the pages your sitemap and homepage link to.
- Keep real
<a href>links in the HTML. Google notes that it can only discover links that are<a>elements with anhref, and that SPAs should use the History API instead of URL fragments for routing. - Put the main text of each page in the initial HTML, not behind a fetch call after load.
- A
<noscript>block is counted by this check, and MDN describes it as the content shown when scripting is off. Use it for a real fallback, not as a substitute for rendering. - If you cannot render server-side, offer another path to the text, such as a markdown alternate or an llms.txt that links to static pages.
Page structure inside the rendered HTML is covered by Semantic HTML elements and Heading hierarchy.