DEV Community

Cover image for I Kept Asking ChatGPT to Read My Website Until Its AEO Got Better
shrey vijayvargiya
shrey vijayvargiya

Posted on

I Kept Asking ChatGPT to Read My Website Until Its AEO Got Better

#seo #ai #webdev #programming

I found a pretty simple way to test whether a website is actually ready for AI search.

I stopped looking at my website like an SEO.

Instead, I started looking at it like ChatGPT, Claude and Perplexity.

The process was basically:

Can you fetch my website?

Then:

Fetch the latest articles from https://ihatereading.in

Then:

Fetch more details about https://ihatereading.in

Then:

Get all links from https://ihatereading.in

And every time the answer was incomplete, wrong, empty or failed completely, I went back to the code and fixed something.

I kept doing this until the answers started getting much better.

That sounds almost too simple.

But there is something important behind it.

AEO has a very practical test

We talk a lot about AEO — Answer Engine Optimization.

We discuss:

  • structured content
  • entity understanding
  • llms.txt
  • schema
  • internal linking
  • crawlability
  • semantic HTML
  • content structure
  • AI search visibility

All of that matters.

But there is another test I don't see people talking about enough:

Can an AI actually read your website?

Not theoretically.

Actually.

Give the URL to ChatGPT and ask it to retrieve something.

If it can't, you've found a problem.

I used is-agentic to test the website

I also used is-agentic, the npm package from Vercel, to test the site's agentic readiness.

I managed to bring the website to 96.

But the more interesting part wasn't the number.

It was what happened when I started manually testing the website through AI chatbots.

Because an automated score can tell you that something is wrong.

An AI chatbot can show you how the website is failing from the AI's perspective.

That's much more useful.

My AEO debugging loop

This became my basic workflow:

Website
   ↓
Ask ChatGPT / Claude / Perplexity
   ↓
See what it can actually retrieve
   ↓
Find what is missing
   ↓
Fix the website
   ↓
Ask again
   ↓
Repeat
Enter fullscreen mode Exit fullscreen mode

For example, I would ask:

Fetch the latest articles from https://ihatereading.in

If the AI couldn't find the articles, I didn't immediately assume the AI was bad.

I checked my website.

Maybe the article content wasn't available in the initial HTML.

Maybe the links weren't discoverable.

Maybe the routing wasn't exposing the content correctly.

Maybe the page returned an unexpected response.

Maybe the content structure made it difficult to extract.

So I fixed it.

Then I asked again.

Then I started asking harder questions

Once the basic request worked, I started increasing the difficulty.

For example:

Fetch more details about https://ihatereading.in

Then:

Get all links from https://ihatereading.in

Then I could ask things like:

What topics does iHateReading cover?

What are the latest articles?

Find articles about AI agents.

What developer tools are mentioned on the website?

Give me the URLs of relevant articles.

The point isn't that these are sophisticated prompts.

The point is that they expose the actual machine readability of your website.

Your homepage isn't enough

This is where I think AEO gets misunderstood.

A website can look perfect to a human.

The navigation works.

The animations work.

The articles look great.

The content loads when you open the page.

But an AI system may see something very different.

Imagine an AI trying to answer:

What is iHateReading?

It needs to discover the relevant information first.

Then imagine:

What are the latest articles on iHateReading?

Now it needs to find article URLs, understand dates and titles, and retrieve the content.

Then:

Find me an iHateReading article about AI agents.

Now it needs to connect the topic with the site's content.

These are different retrieval problems.

And each one can expose a different technical issue.

The scary errors are the boring ones

One thing that stood out to me was how important basic HTTP responses are.

If an AI crawler gets something like:

400
500
204
content not found
empty response
Enter fullscreen mode Exit fullscreen mode

there isn't much AEO magic you can do afterward.

The AI can't recommend content it can't retrieve.

This is why I think AEO starts much lower in the stack than people think.

Before worrying about whether an LLM understands your brand positioning, make sure it can actually get the content.

HTTP response
     ↓
HTML
     ↓
Links
     ↓
Content
     ↓
Structure
     ↓
Meaning
     ↓
Answer
Enter fullscreen mode Exit fullscreen mode

If you fail at the first few layers, the later ones don't matter.

AI search is still search

This is probably the biggest takeaway for me.

We sometimes treat AI search like something completely separate from SEO.

But there is a huge overlap.

If an AI needs to recommend your website, it needs to:

  1. Discover it.
  2. Fetch it.
  3. Read it.
  4. Understand it.
  5. Connect it with a user's question.
  6. Retrieve the right page.
  7. Use that information in an answer.

That's not completely different from search.

The interface changed.

The retrieval and understanding problems didn't disappear.

So I started using AI as an AEO debugger

This is the part I'd actually recommend trying.

Don't ask ChatGPT:

Is my website AEO optimized?

That question is too broad.

Ask it to do something with your website.

For example:

Fetch the latest articles from https://yourwebsite.com
Enter fullscreen mode Exit fullscreen mode

Then:

Get all discoverable links from https://yourwebsite.com
Enter fullscreen mode Exit fullscreen mode

Then:

Find the pages about [topic] on https://yourwebsite.com
Enter fullscreen mode Exit fullscreen mode

Then:

Summarize what this website does.
Enter fullscreen mode Exit fullscreen mode

Then:

Find the most relevant article on this website for [question].
Enter fullscreen mode Exit fullscreen mode

Then:

Give me the URL and title of that article.
Enter fullscreen mode Exit fullscreen mode

Now you're testing multiple parts of the system.

And when something fails, don't just try another prompt.

Fix the website.

Then ask again.

This is basically AEO unit testing

That's the mental model I like most.

Instead of treating AEO as a checklist you complete once, treat it like testing an API.

You have an input:

"Find articles about AI agents"
Enter fullscreen mode Exit fullscreen mode

You expect:

{
  "title": "...",
  "url": "...",
  "content": "..."
}
Enter fullscreen mode Exit fullscreen mode

But the AI gives you:

I couldn't find relevant articles.
Enter fullscreen mode Exit fullscreen mode

That's a failed test.

So you investigate.

Maybe the links aren't discoverable.

Maybe your HTML is mostly an application shell.

Maybe your article content isn't server-rendered.

Maybe your URLs aren't consistent.

Maybe your metadata is missing.

Maybe your internal linking is weak.

Maybe the response itself is broken.

You fix it.

Run the test again.

That's a much more practical way to think about AEO.

Don't optimize for a chatbot's answer

There is one important distinction, though.

I'm not trying to make ChatGPT say nice things about my website.

That's not the goal.

The goal is to make the underlying website easy for machines to retrieve and understand.

If I ask:

Fetch all links from my website.

and it returns them correctly, that's useful.

If I ask:

Find my latest article about AI agents.

and it finds the correct page, that's useful.

If I ask:

What does iHateReading publish?

and it can accurately explain the site, that's useful.

The answer itself is the test.

Try this on your own website

You don't need some massive AEO tool to start.

Open ChatGPT, Claude or Perplexity.

Give it your URL.

Then keep asking questions.

Start simple:

Fetch my homepage.
Enter fullscreen mode Exit fullscreen mode

Then:

Fetch my latest articles.
Enter fullscreen mode Exit fullscreen mode

Then:

Get the links to those articles.
Enter fullscreen mode Exit fullscreen mode

Then:

Find articles about [your topic].
Enter fullscreen mode Exit fullscreen mode

Then:

Explain what this website is about.
Enter fullscreen mode Exit fullscreen mode

Then:

Which page on this website best answers [your customer's question]?
Enter fullscreen mode Exit fullscreen mode

Every failure is a debugging clue.

And every successful answer is another indication that your website is becoming easier for an answer engine to use.

The simplest AEO test might be a conversation

This is probably my favorite part of the whole experiment.

Instead of asking:

"Is my website optimized for AI?"

just ask the AI to use your website.

If it can't use it, fix the website.

Ask again.

If it still can't, fix something else.

Keep going.

That's what I did with iHateReading.

I used is-agentic to get a measurable signal, but then I used actual AI conversations to find the problems that mattered.

Your AEO score is interesting.

Whether an AI can actually find, retrieve and understand your content is more interesting.

And the easiest way to find out is to simply ask.

That would be enough for today, for more details visit our website iHateReading
Cheers
Shrey

Top comments (0)