What an AI assistant actually reads when it lands on your site

The fetch is boring, and that is the whole point of it.

When an AI assistant needs something live about your business, it sends a request to your address and gets back one file. Text and tags. It is the same file your browser gets. The difference is what happens next.

Your browser spends the next second or 2 turning that file into a page. Fonts arrive. Photos load. Scripts run. The review widget fills itself in from another company’s server. The testimonial slider starts sliding. By the time you look at it, the page on your screen contains a lot of things that were never in the file.

Work on the assumption that the AI assistant read the file and stopped. Build the page so it does not matter which of those two it did.

Open your own homepage the honest way

Right click on your homepage and choose View Page Source. Not Inspect. Inspect shows you the page after every script has finished running, which is the flattering version. View Page Source shows you what came down the wire.

It will look like a mess. You are not reading it, you are searching it. Ctrl+F is the whole technique.

The first 20 lines carry more than the hero image does

Near the top sits your title tag, your meta description and your canonical address. The title is the single strongest one-line statement your page makes about itself, and on plenty of trade sites it reads Home | Smith Electrical. That tells a machine your business is called Smith Electrical and nothing else. No trade, no suburb, no hint of what you would be recommended for.

Then the body, in the order it was written

• The H1. One per page. It is the heading a machine treats as the subject of the document.
• The hero. Often a photo with words baked into it. If “24 hour emergency electrician” exists on your site only as pixels inside a JPEG, it is not in the file. The alt text is the one place that image gets to speak.
• Your nav labels. Real text, and a map of what you do. Menus that read Home, About, Services, Contact describe a website. Menus that name the 6 things you actually get paid for describe a business.
• The logo strip. Supplier and accreditation logos are images. With no alt text they are blank rectangles, and the accreditation you pay for every year says nothing.
• Your body copy. This part is genuinely readable, so make sure it says something.
• Accordions and FAQ toggles. Usually present in the file even while collapsed on screen. Search for one of the answers and check.
• Sliders and embedded review widgets. Frequently injected by a script that runs in the browser. Search for a phrase from a testimonial. If it is not there, that testimonial is not part of your page as far as the fetch is concerned.
• Your phone number. Search for it. It should appear as text, not only inside a button graphic.

The block at the very bottom nobody scrolls to

Under everything else sits your structured data, if you have any. It states your name, phone, hours, service area and services in a fixed format that a machine reads without guessing. It is the least ambiguous thing on the page.

Where the page starts arguing with itself

This is the failure worth 10 minutes of anyone’s morning, because it is free to fix. The visible page says one thing and the structured data underneath says another. A made-up example:

What the page shows	What the data underneath says
Mon to Fri, 7am to 4pm	Open 24 hours, 7 days
The mobile on the contact page	The landline from 2 vans ago
“Servicing the northern suburbs”	A list of 40 suburbs, half of them an hour away
6 questions in the visible FAQ	11 questions, 5 of which appear nowhere on the page

Give a machine 2 versions of your opening hours and you’ve no way of knowing which one it will repeat, if either.

Structured data (a hidden block that describes your business to machines) that contradicts the visible page makes your own site the source of the disagreement. Whoever built the site probably did not do this deliberately. A plugin filled the block in with defaults 3 years ago and nobody has read it since.

Things it does not read at all

Not a judgement on any of these. Several are worth having for the human. They just contribute nothing to being named.

• Colour, layout, spacing, how modern it looks
• Your hero video, and the drone shot of the ute
• Animations and anything that appears on scroll
• Words that exist only as pixels inside an image, including the photo of your licence certificate
• A star rating drawn by a third-party widget
• A PDF capability statement you never linked to from anywhere

The part nobody can tell you, us included

How much weight any one of these elements carries is not published by anybody. Not by OpenAI, not by Anthropic, not by Google, not by Perplexity. If someone hands you a percentage breakdown of what matters, they made the percentages up.

Nor does anyone outside those companies know how often a page gets fetched again, whether a given crawler runs JavaScript on the way through, or whether structured data is read directly or merely lines up with what the visible text already said. We write it because it is unambiguous and cheap, not because anyone has proven a mechanism.

What is actually known is 1 sentence long: an AI assistant can only repeat what it can read. Everything above follows from that. If you want the same treatment applied to the file that lets the crawler in at all, that is the robots.txt question.