How a machine reads your website

Before a person sees your website, something else has read it. A search engine has decided whether to list it and what to say about it. More and more often an assistant has read it too, to answer somebody’s question in its own words.

Those readers are not given what your visitors see. No colours, no layout, not the photograph you paid for. They are given text with labels on it. If you are about to commission a website it is worth ten minutes to see a page the way they do. I think most of what gets called SEO begins there.

Two readers

I made up a business to show it: Cadena, a bicycle workshop in Palermo, Buenos Aires. It does not exist. Its page is an ordinary one: a good picture, a headline, three services and a friendly line of welcome. I gave it four faults on purpose, and a visitor would not notice any of them.

Figure 1 has the page on one side and, on the other, everything a machine is given by it.

Look at what became of the headline. To a visitor it is the first thing on the page: Bicycle repair in Palermo, in large white letters. But the letters are part of the picture, and the picture has no description. So the machine’s list has no main heading, and a picture it has been told nothing about.

The plainest sentence on the page, the one that says what this business is, is the one the machine is never given.

What the machine is given

That list is not a figure of speech. A page is sent as HTML: text, with marks around it that say what each piece is. This is a title. This is a heading. This is a paragraph. This is a picture, and here is a sentence describing it.

Some of these readers can make something of a picture on their own. None of them promises to, and Google’s advice is plain: make sure important content is available as text, and describe pictures in words.

Three of the lines are never shown on the page at all. The title appears on the browser’s tab. The description is written for search engines. The facts as data, which developers call structured data, are a small block that says things outright: this is a business, this is its name, this is where it is, these are its hours. A visitor reads none of the three. A machine is sent all of them.

Four things to mend

Figure 2 has a switch for each fault, and a sketch of a search result made from whatever the machine has at that moment.

With everything off, the result says: Home. We love bikes and we love this neighbourhood. I would not click on that, and I do not think somebody with a flat tyre would either.

Turn on the title and the result has a headline that says what the place is and where. Turn on the description and the line under it stops being whatever happened to come first on the page.

Two cautions, both from Google’s own documentation. It takes a result’s headline from several places, the title among them, and may write its own if yours does not match the page. And it builds the line underneath mainly from the page’s own words, using your description when it judges that to be the better account. So neither is an order. They are your best offer, and without them the machine makes do with what it finds.

The third switch changes nothing a visitor can see. The headline is the same words in the same place, now set as text over the picture and not painted into it, and the picture has a sentence saying what it shows.

The fourth is the facts as data. Google describes structured data as explicit clues about the meaning of a page. It asks that the data say nothing a visitor cannot also see, and it does not promise to use it. I add it anyway. Stating a fact costs less than hoping it will be worked out.

Before the scripts run

There is a fifth fault that no switch on that page can mend, because it lies in how the page is made.

Some websites send their words in the HTML. Others send a nearly empty page and a script, and the script puts the words in once it is running in a browser. To a visitor the two look the same. Figure 3 is the difference.

A reader that runs scripts gets everything either way. Google is one: it says it handles such pages in three phases, crawling, rendering and indexing, and that a page waits in a queue for the middle one. A reader that does not run scripts gets a title and an empty box.

Which readers are those? Google’s own guide puts it flatly: not all bots can run JavaScript. In December 2024 Vercel and MERJ published what they had seen in their own traffic: the crawlers of OpenAI, Anthropic, Perplexity and others fetched pages without running their JavaScript, while Google’s and Apple’s ran it. That finding is theirs and nearly two years old as I write. It may have changed. I would not build a business on the hope that it has.

The ones that answer

Assistants that answer a question outright are the newer reader, and there is a great deal of confident advice about how to please them. Very little of it is documented. This much is.

Google says there are no additional requirements for a page to appear in its AI features, and no special optimisation: the page has to be indexed and eligible to be shown with a snippet. In the same place it repeats the basics, among them that important content be available as text.

OpenAI documents a crawler for search, OAI-SearchBot, and says a site that opts out of it will not be shown in ChatGPT’s search answers. So one thing worth checking is that your site is not turning it away by accident.

Past that, I do not know how these systems choose what to quote, and I would be wary of anybody who says they do. My working assumption is that they begin where a search engine begins: with the text they are given. The list in Figure 1 is the raw material for an answer as much as for a result.

This page, read the same way

None of this is special to an invented bicycle workshop. Figure 4 reads the page you are on, in your browser, the way the other figures read the sample.

The words you are reading are in that count because they are in the HTML this page is sent as. I fetched it with a tool that runs no scripts, and every paragraph came back. Only the figures need a browser.

What to ask whoever builds yours

You do not need to check any of this yourself. You need to ask, before the work starts, and expect a straight answer.

• Does every page get its own title and description, and can I change them without calling you?
• Is every headline real text, and does every picture that carries meaning have a description?
• Are the words in the HTML the server sends, or put there by a script? Show me a page with JavaScript switched off.
• Is the business stated as data: its name, what it is, where it is, when it is open?
• Does anything on the site tell search engines or AI crawlers to keep out, and is that on purpose?
• Can I see what a machine is given for the home page before we launch?

None of it is a redesign. On the sample page three of the four changes cannot be seen at all, and the fourth only on the browser’s tab. A page can look exactly as good the day it becomes readable as it did the day before.