What ChatGPT Actually Sees When It Looks at Your Website (and How to Check)

You paid for a beautiful website. It looks incredible on your phone. And when you ask ChatGPT to recommend someone who does what you do, it has never heard of you.

The usual explanation is that you need more content, or more backlinks, or to post more. Sometimes that is true. But there is a more basic possibility to rule out first: the assistant may be receiving a nearly empty page.

Most AI crawlers do not run JavaScript

Your browser does a lot of work on your behalf. It downloads the HTML, then runs the JavaScript, which fetches your content and paints it onto the screen. What you see is the rendered page, not the raw one your server actually sent.

A crawler does not have to do that second step, and most of the ones that feed AI assistants do not. They take the raw HTML your server sends and read that. If your words only appear after JavaScript runs, the crawler gets the wrapper and none of the substance.

Google is the exception people cite: it does render JavaScript, eventually, on its own schedule and its own budget. That exception is doing a lot of work in most SEO advice. It does not extend to the assistants your customers are now asking.

What that looks like on real sites

Measured on 2026-08-20 with the command further down this page, counting only the words a non-rendering crawler receives from the homepage. Run it yourself and you should land within a few words of these, since live sites change under you:

• An award-winning design studio: 753 words, structured data yes, llms.txt yes
• A well-known design awards site: 751 words, structured data yes, llms.txt no
• A small business on a hosted builder: 999 words, structured data NONE, llms.txt no
• A hand-built client site, short page: 701 words, structured data yes, llms.txt yes
• This site: 6,049 words, structured data yes, llms.txt yes

The interesting row is the third one. That is a real small business with a genuinely nice site, and it ships no structured data at all: nothing telling a machine what the business is, where it is, or what it sells. The owner did nothing wrong — nobody added it, and the platform does not add it for you.

Word count is not the metric

Look at the fourth row. Seven hundred words, and it is completely fine. It is a short, focused landing page, and every word that exists on it is in the raw HTML. Nothing is hidden from anything.

That is the actual test — a comparison, not a number: does the content that exists on your page appear in what your server sends? A short page that is fully readable beats a long page that arrives empty, every time.

If someone quotes you a word count target without looking at your site, they are selling a package, not a diagnosis.

Check your own site in one command

You do not have to take my word for any of this. Open Terminal on a Mac, paste this, and swap in your own domain. It fetches your page exactly as a non-rendering crawler would, strips the markup, and counts what is left:

curl -sL https://YOURDOMAIN.com \
  | perl -0777 -pe 's/<script.*?<\/script>//gsi; s/<style.*?<\/style>//gsi; s/<[^>]+>/ /g; s/\s+/ /g' \
  | wc -w

Now compare that number to how much text you can see on the page. If the page is full of writing and the number comes back small, your content is arriving after JavaScript, and the assistants are not waiting for it.

Two more checks worth thirty seconds. Visit yourdomain.com/llms.txt and see whether anything is there. Then view your page source and search it for ld+json: that is the structured data that tells a machine what your business actually is. Finding neither is common — and fixable without touching your design.

What to do about it

If your content is in the HTML already, you do not have a rendering problem, and rebuilding would be a waste of your money. What you probably need is the machine-readable layer on top: structured data, an llms.txt, explicit crawler permissions, clean headings. That is an addition, not a rebuild.

If your content genuinely is not in the HTML, that is a build decision, and it is worth understanding what it costs you before you decide whether to change it. Plenty of businesses reasonably choose a platform for other reasons. The point is to choose it knowing.

The honest caveat

None of this makes an assistant recommend you. Being readable makes you eligible to be recommended, which is a different and smaller claim. Anyone promising that a file will get you cited by ChatGPT is overselling it. The numbers above are a snapshot from one day, and they will drift as these platforms change.

What I can say plainly: a page a machine cannot read will never be cited by one, and most people have never checked which kind of page they have. The command above takes ten seconds and settles it.