ChatGPT and Claude still can't read JavaScript pages

Gabe asked each assistant whether it could read three pages and to give him each page's first and last paragraphs. On the first page all the main content was drawn by JavaScript. On the second the header and footer were plain HTML and the main content was drawn by JavaScript. The third was a parallax site built with WebGL and animation. ChatGPT, Claude and Perplexity failed all three, returning navigation, a site shell, cookie notices or an error message instead of the text. Gemini succeeded on all three. Copilot succeeded on the first two and saw only an error message on the third.

Gabe puts the difference down to Gemini and Copilot drawing on Google's and Bing's search systems, which can see content drawn by JavaScript most of the time. It repeats what he found in August 2025, when a site that depended on JavaScript was often not shown as a main source in ChatGPT or Perplexity and turned up only in ChatGPT's "More" list or Perplexity's "Reviewed" list, where Perplexity showed the page's title and no snippet. Before that, Vercel reported in December 2024 that none of the major AI crawlers it studied rendered JavaScript, OpenAI's and Anthropic's among them; they fetched JavaScript files without running them.

Until that changes, treat it as the working rule: if the words on a page only exist after a script runs, assume ChatGPT and Claude do not see them.

We took the 50 brands from our Google UK brand search reading of 6 October, across nine sectors from supermarkets to travel, and loaded each homepage twice in an automated browser: once normally and once with JavaScript switched off, which is close to what a crawler that does not run scripts receives. Seven sites answered with a bot check or an access-denied page and one sent us its US site, which left 42 UK homepages we could read.

All 42 carried real text without JavaScript; the smallest still had 166 words. Twenty-eight kept at least 80% of the words the full page showed, and twelve kept between half and 80%. Two kept less than half, and one of those kept 7%: its full homepage showed 5,463 words and the version without scripts 381. Word counts include menus and cookie notices, so a lower share does not prove the brand's main copy is missing, only that some of what a visitor reads arrives by script.

These are homepages of large companies with large web teams. We did not test product, pricing or comparison pages, which are the ones assistants quote when a buyer asks what you cost or how you compare.

The pages to worry about are the ones Gabe's test describes: a site built as a single-page app that draws everything in the browser, and a page whose frame is plain HTML but whose main content is fetched by a script after it loads, such as a pricing table or a product comparison. Google and Bing render JavaScript, so a page like this can rank well in Google and still look blank to ChatGPT. An SEO tool that renders pages before checking them will not show the problem either.

Bot walls are a separate risk. Seven of the 50 sites showed our automated browser a security check or an access-denied page instead of the homepage. Whether a given AI fetcher gets through depends on how each site's security service is set to treat it, so check yours rather than assume.

Open the page you most want quoted, such as your pricing or main product page, and turn JavaScript off in your browser's settings, or fetch it with curl. If the prices, product names and claims are still there as text, a crawler that does not run scripts can read them too. If you see a loading spinner or an empty frame, ask your developers to deliver that content in the page's HTML rather than drawing it with a script; Gabe's advice is not to rely on JavaScript for content you want AI search to see. You can also repeat his test: give ChatGPT or Claude the address, ask whether it can read the page, and ask for its first and last paragraphs.

Then make sure the crawlers are allowed in at all. The free AI crawler checker reads your robots.txt against the main AI crawlers and checks for an llms.txt file. Our ChatGPT visibility tracker page covers how ChatGPT sources what it says about a business.

On 21 October 2026, between 06:09 and 06:11 UK time, we loaded https://www. plus each brand's main domain from our 6 October reading in headless Chromium, with a desktop Chrome user agent and a British English locale, from a US internet connection. Each homepage was loaded once with JavaScript on, waiting for the page to load and then six seconds more, and once with JavaScript off. For each load we counted the words of visible text in the page body, leaving out scripts and styles, and compared the two counts. A site counted as blocked when either load returned a bot check or an access-denied page. Each homepage was read once, so a page that changes by visit, region or time could give a different count on another day.