Readable without JavaScript

Is the text in the page itself, or does it only appear after JavaScript runs?

Readability · 45 of 100 in Readability · yours to do · content

What we look for

We fetch your home page and the sample of pages we check, your newest posts among them, the way most bots do: we read the HTML the server sends and don’t run any JavaScript. Then we ask whether the page’s words are there.

A page passes when its own text is in the HTML, however short. A link post of two sentences is the whole post. A page is charged only when there is a sign that words are missing:

  1. An empty shell. The page shows under about 100 characters of text, nothing to follow or look at, and loads script: a blank placeholder (<div id="root">, #app, #__next) that scripts fill in later. This scores 0.
  2. A notice. A <noscript> or the page’s own text says it needs JavaScript, on a page with little text.
  3. Script carries the words. The page has little text of its own, and its scripts hold sentences the HTML doesn’t show: a framework’s payload, or a blob of copy waiting to be rendered. Only sentences of the page’s own count: not a drawing’s path data, not a framework’s error messages, and not the page’s own title or meta description.
  4. Your feed promises more. The feed’s item for the page holds several times the text the page does.

A page that shows one of the last three signs scores 0.5 when it still has 200 characters of text, and 0 with less. The report names each charged page and the sign we saw.

The score is the average across pages. Pages served as Markdown or plain text are readable and count the same way; pages served as anything else that isn’t a web page (a download, for example) score 0. How much a page says is not this criterion’s question.

We measure the main content: what’s in your <main> or <article>, or, without one, the page outside its header, navigation, and footer. A <nav> inside <main> that holds the page’s cards, each a link with a sentence under its heading, is content, not a menu. A footer full of legal text doesn’t count as content, and neither does text that appears, word for word, on every page we read (a site-wide menu or an alert bar nobody marked as one). Hidden text doesn’t count either: what sits under hidden or aria-hidden="true", or in a dialog or a modal that isn’t open. Your <h1> and what directly follows it count wherever they sit, and so does a sidebar that holds the page’s heading or text no other page has.

Words in pictures count as no words. A page with fewer than 200 characters of text and three or more large images has its words in the pictures, and scores at most 0.5, whatever else it does; the report says so. An image counts as large if its width or height is 200 pixels or more; with no size given, any image counts unless it’s an SVG or its file name suggests something small (an icon, logo, thumbnail, avatar, sprite, badge, emoji, or favicon).

We also say how much of your home page’s HTML is text, and when it’s under ten percent, what the bulk is: inline CSS, inline script, inline SVG, the page’s own data that a framework sends along with it, tracking scripts, or the markup itself. That line is information only: it never moves the score.

Why it matters

Most AI crawlers and answer bots don’t run JavaScript. They read the HTML and move on. If your site is built so the words only appear once scripts run in a browser, those bots see a blank page, and an assistant has nothing to say about you. That is changing at the edges: in late 2026, OpenAI’s search crawler began running scripts on some sites, well after its first visit, and it has said nothing about it. The fetches made while someone is waiting for an answer still read the HTML alone, and so does everyone else’s. Words in the HTML are read by all of them, today.

A headline set in an image looks great to a person and says nothing to a machine; a page that’s a series of graphics with your pitch inside them is, to an assistant, a page with nothing on it. And a page whose HTML is mostly scripts and styles buries its words where a bot that reads only so much may never reach them.

This is the single heaviest criterion in Readability because nothing else here matters if there’s no text to read.

How to pass

  1. Check what a bot sees:
    curl -s https://example.com/ | sed -e 's/<[^>]*>//g' | tr -s ' \n' | head -c 1000
    If your headline and description aren’t in that output, bots can’t see them.
  2. Render pages on the server, or generate them ahead of time. Most frameworks can: Next.js (server rendering or static export), Nuxt, SvelteKit, Astro, Remix, and Gatsby all do it.
  3. If a full change isn’t possible soon, make sure at least the important pages — home, about, product, pricing — are prerendered.
  4. If your feed carries a whole post, put the whole post on its page too.
  5. Put your headline and description in the page as real text, styled with CSS, not in an image; if a graphic has to carry words, repeat them as text on the page.
  6. Move large inline styles and scripts into files, and drop tracking you no longer use, so the words come early in the HTML.

Evidence

  • Vercel with MERJ · December 2024 · strong

    In December 2024, OpenAI’s, Anthropic’s, and Perplexity’s crawlers didn’t run JavaScript; only Gemini, through Googlebot, and Applebot rendered pages.

    One month of server logs across Vercel’s network in 2024: 569 million GPTBot fetches and 370 million from Claude. Vercel sells server rendering.

  • Forge · September 2026 · moderate

    ChatGPT-User read 37,045 pages, Claude-User 10,028, ClaudeBot 915, and PerplexityBot 3,193, and none of them fetched a script. OAI-SearchBot fetched ten in 1,738 page requests.

    Thirty days of one news site’s server logs, to September 24, 2026, counted by the bot’s name and not checked against published addresses. Forge makes a bot-log plugin for WordPress.

  • Hisashi Space · September 2026 · contrary

    From September 25, 2026, OAI-SearchBot and GPTBot sent requests only a rendered Next.js page sends: 84% of OAI-SearchBot’s and 97% of GPTBot’s on September 28. OpenAI announced nothing.

    One developer’s server logs for four Next.js sites and a static one, every request checked against OpenAI’s published address ranges.

  • Genezio · October 2026 · contrary

    The call came, from an address in the ranges OpenAI publishes for its crawlers: OAI-SearchBot ran the script. The page gives no counts, and says it can’t tell what share of pages are rendered.

    A page whose script calls an address on the same server that nothing else links to. Genezio sells an AI-visibility platform.

  • Bogdan Ripa · October 2026 · contrary

    On the 26 pages where both were seen, the scripts and images were asked for a median of 32 minutes after the HTML.

    Two weeks of one static site’s logs, October 8 to 22, 2026: 560 requests from addresses on OpenAI’s published lists.

Last changed .

The whole rubric