A website can look perfect to customers and still be half-invisible to ChatGPT. People see your site through a browser that runs scripts and loads images. Many AI fetchers just download the raw code and read what is there. Here is what they can and can’t see, a two-minute test for your own site, and how to write an opening they will actually use.

Can ChatGPT read every website?

No. When Vercel and MERJ studied AI crawler traffic across Vercel’s network (December 2024), they found that “none of the major AI crawlers currently render JavaScript”. That included OpenAI’s crawlers, Anthropic’s ClaudeBot, Meta’s crawler, ByteDance’s Bytespider and PerplexityBot. The exceptions were Google’s Gemini, which uses Googlebot’s infrastructure and renders pages fully, and Applebot, which renders through a browser-based crawler.

Does ChatGPT run JavaScript?

No. ChatGPT’s crawlers download JavaScript files but don’t execute them, so anything your site only shows after a script runs is invisible to them. JavaScript is the code that makes pages interactive, and on some sites it builds the content itself. Typical casualties are price lists or treatment menus pulled in from a booking system, reviews widgets, tabs and accordions whose text only loads when clicked, and sites built as single-page apps without server rendering.

Vercel’s advice is to prioritise server-side rendering for critical content. In plain English, the server should send the finished text as part of the page. Chat widgets and other extras can still load with JavaScript.

How can I tell if ChatGPT can see my content?

Use the view-source test. It takes two minutes and needs no special tools.

  1. Open your page on a computerStart with your homepage, then your most important service page.
  2. View the page sourceRight-click and choose View Page Source, or press Ctrl+U (Cmd+Option+U in Chrome on a Mac). Don’t use Inspect: it shows the page after scripts have run, which is exactly what non-JavaScript fetchers miss.
  3. Search for your key factsPress Ctrl+F (Cmd+F on a Mac) and look for a sentence from the middle of the page, your phone number, a price and a service name. If they’re in the source, a fetcher that doesn’t run JavaScript can read them.
  4. Ask ChatGPT directlyPaste your page address into ChatGPT, ask it to fetch and summarise the page, and compare its summary with what’s really there.

Or let our free Can ChatGPT read my website? tool check a page for you.

Is there a page size limit for ChatGPT?

Yes, at about four megabytes. RESONEO, a French search consultancy that reverse-engineered ChatGPT’s retrieval in July and August 2026, found that “past four megabytes, a page is not read at all”, adding: “No partial read, no truncation: the page is rejected outright.” The study also notes that “in practice almost no page reaches that size”, so this is rare. It is still a good reason to keep pages lean, which makes them faster for people too.

What does ChatGPT read first on a page?

For quick ‘instant’ answers, very little: the page title and a snippet cut off just past 200 characters. RESONEO found that snippet is “taken from the start of the page’s rendered body, not from your metadata”, and that “your meta description is ignored”. Of 463 pages with a main heading (an H1), 83.6% had it inside the snippet. In practice, that means your H1 and the first sentence or so beneath it.

Getting past that stage matters. In RESONEO’s data, pages ChatGPT actually opened and read were cited 74% of the time (326 of 440), while pages that only appeared as search results were cited 7% of the time (4,303 of 57,853). When a model may see only a title and 200 characters, those characters need to carry your answer.

How do I write an opening ChatGPT can use?

Answer the obvious question straight away: who you are, what you do, where, and one concrete detail. Kevin Indig’s analysis of ChatGPT citations (February 2026) found that 44.2% came from the first 30% of a page’s text. Here is a made-up example for a fictional business, Harper & Co Plumbing in Guildford.

  • Before: “Welcome to Harper & Co! We’re passionate about delivering a first-class service to all our valued customers, and we take pride in quality workmanship.”
  • After: “Harper & Co Plumbing is a Gas Safe registered plumbing and heating firm in Guildford, Surrey. We repair leaks, service boilers and fit bathrooms across Guildford, Godalming and Woking, with same-day call-outs.”

The second version names the business, the service, the town, the area covered and a concrete detail in about 200 characters. It reads as a direct answer to “who’s a good plumber in Guildford?”, which is exactly what an assistant is trying to work out. The first gives a machine nothing it can quote.

Carry the habit down the page. Ahrefs’ study of 1.4 million ChatGPT prompts (April 2026) found that cited pages had titles closer in meaning to the question (a similarity score of 0.60, against 0.48 for pages not cited), and results with readable, descriptive web addresses were cited 89.8% of the time, against 81.1% for those without. Use customers’ questions as headings, answer each in the first sentence underneath, and prefer addresses like /boiler-repair-guildford to /page?id=214.

Could my website be blocking AI without me knowing?

Yes, and the cause is often a security service in front of your site rather than the site itself. Cloudflare, one of the largest, began blocking AI crawlers by default for new domains on 1 July 2025 and has changed its defaults since. It now asks new customers at sign-up how to treat AI bots, under settings called ‘AI Crawler and Agent access policies’ (Search, Agent and Training).

The bigger catch: Cloudflare’s Training block also applies to crawlers that do search and training together. When it announced this (July 2026, effective 15 September), it named Googlebot, Applebot and Bingbot, and said its older “Block AI bots” option is covered too. A switch meant to stop AI training could shut out Google, Apple and Bing too. If your site uses Cloudflare, check its AI crawler settings. And ‘checking your browser’ pages are a dead end for any fetcher that can’t run JavaScript.

Your robots.txt file can also turn bots away. Our guide to which AI bots to let in has a template, and our free Is my website blocking AI? tool checks for you.

Do broken links stop ChatGPT finding my pages?

They waste its visits. In the Vercel and MERJ study, 34.82% of ChatGPT’s fetches hit pages returning a 404 (‘not found’) error, and Claude’s figure was similar at 34.16%. Every fetch spent on a dead page is one not spent on a live one. Keep your sitemap (the file listing your pages for search engines) current, redirect old addresses to their replacements and fix broken internal links.

What should I fix first?

Fix whatever stops AI reading your pages before you polish what it reads.

  1. Run the view-source testOn your homepage and top three service pages. Anything missing from the source needs to be in the HTML your server sends.
  2. Rewrite your title and first 200 charactersWho, what, where and one concrete detail, in plain words.
  3. Check bot accessLook at your robots.txt file and any Cloudflare or firewall bot settings.
  4. Clear out dead linksFix 404s, update your sitemap and redirect old addresses.
  5. Keep pages leanWell under 4MB, and quick to load.

Frequently asked questions

Does Google’s AI have the same problem with JavaScript?

Less so. According to the Vercel and MERJ study, Gemini uses Googlebot’s infrastructure, which renders JavaScript, and Apple’s Applebot renders pages too. In that December 2024 study, crawlers from OpenAI, Anthropic and Perplexity didn’t, so content that only appears after scripts run is still a risk. See which AI uses which search engine.

Is my meta description still worth writing?

Yes, because Google can still show it in search results. But RESONEO found ChatGPT’s instant snippets ignore it and take text from the top of the page instead, so don’t rely on the meta description to carry your key message.

I’ve updated my page. Why does ChatGPT still describe the old version?

It may be reading a saved copy. RESONEO found ChatGPT stores pages it has fetched as Markdown (a stripped-down text format) and doesn’t reopen them for about 30 minutes, and it saw some old copies served weeks or even months later. Also check the old wording isn’t still live elsewhere, such as on a duplicate page.

Will a website builder pass the test?

It depends on the builder and how each page was put together. Extras such as booking widgets, reviews feeds and some menus can be loaded by script even when the main text isn’t. The view-source test is the only way to be sure.

Sources