Getting found by AI assistants, not only by Google.

A growing share of people never see your search result. They see a sentence an assistant wrote about you, or they do not see you at all.

For twenty years the job was to rank. Somebody typed a question, got ten blue links, and clicked one. Everything the industry built, keywords, backlinks, meta descriptions, existed to win that click.

The behaviour has shifted. People ask an assistant, read the answer it composes, and click through only when they want to verify something or buy something. The page still matters, but it now competes to be the source of a sentence rather than the destination of a click.

That has a name now, generative engine optimisation, and most of what is written about it is guesswork dressed up as method. What follows is the part that is actually mechanical, and therefore checkable.

The part nobody mentions: most AI crawlers cannot run your JavaScript

This is the single biggest technical difference between search engines and assistants, and it decides whether the rest of the work is even worth doing.

Googlebot renders pages. It downloads your JavaScript, executes it, waits for the framework to build the DOM, and indexes what a browser would have shown. A React app that ships an empty div and fills it on the client can rank perfectly well.

The assistants' crawlers do not do this. They fetch the HTML and read it as it arrived.

Read that against how most agency sites are built and the problem is obvious. A single-page application that renders on the client is, to ChatGPT and Claude and Perplexity, a blank document with a script tag in it. Not badly ranked. Absent.

There are exceptions worth knowing. Applebot renders JavaScript, and Gemini inherits Google's rendering infrastructure. But building for the two that render while three of the most used assistants read raw HTML is a strange bet to place.

The check takes ten seconds and needs no tools beyond the ones already on your machine.

# What an assistant's crawler actually receives
curl -s https://yoursite.com | grep -c "the headline you expect"

# 0 means your content is not in the HTML.
# It exists only after JavaScript runs, and it will not run.

Write in units that survive being lifted out

An assistant is not summarising your page. It is looking for a passage it can carry into an answer about something else, and passages that only make sense in sequence do not travel.

Compare two ways of saying the same thing. The first is normal editorial prose and is close to useless here:

As we saw above, this depends heavily on the factors already discussed, and in our experience the answer is usually somewhere in the middle.

The second says what it means with no dependencies: a custom marketing site in Latin America usually costs between 750 and 25,000 dollars, and the four things that move the price are whether the design is original, who writes the copy, whether there is functionality beyond display, and who answers when it breaks.

Only one of those can be quoted. The test is blunt and works well: cut the sentence out, show it to someone who has not read the page, and see whether it still means anything.

Some practical shape for that, all of it cheap to do:

• Put the direct answer immediately under the question it answers, not three paragraphs later after the throat clearing.
• Repeat the subject instead of leaning on it, this and that and the above are invisible to anything reading one paragraph in isolation.
• Give numbers with their units and their year, since a figure with no date decays into a liability.
• Keep one idea per paragraph so an extract never arrives half formed.

This is also, and not by coincidence, better writing for humans. The pages that assistants quote well tend to be the ones people finish reading.

Say where the number came from

Assistants are built to be careful about attribution, and that has a consequence most people miss. Given two pages making the same claim, the one that names its source is the safer thing to repeat, so it gets repeated.

This is the cheapest advantage available on the whole list, and almost nobody takes it. Most marketing pages are full of numbers with no origin at all. 73% of users expect this. Studies show that. Which studies, from when, measuring whom.

Our own rule on this site is that a figure ships with a link to where it came from, or it does not ship. It cost nothing to adopt and it removed a whole category of copy we would have had to defend later.

Structure the page like a document, not like a canvas

Design tools encourage thinking in boxes on a plane. Everything reading your page programmatically thinks in a document tree, and the gap between the two is where meaning gets lost.

A heading that is a heading, rather than a div that happens to be large and bold, tells a parser that everything under it belongs to it. A list marked up as a list is a set of parallel items. A table is a relationship between rows and columns. None of this is visible to a visitor, and all of it is the difference between a page that can be understood and a page that has to be guessed at.

Structured data is the same idea made explicit. An Article block that states the headline, the date, the author and the language, and an FAQPage block that pairs each question with its answer, hand over the reading rather than leaving it to inference.

One caution, since it is a real risk rather than a theoretical one. Structured data that disagrees with the visible page is worse than none: it is a signal that the page cannot be trusted. Generate it from the same source as the content, never by hand alongside it.

What has not changed

It is tempting to treat this as a new discipline that replaces the old one. It is not. Assistants are still fed largely by search indexes, so a page that no crawler can reach is invisible in both worlds.

Speed still matters, for the ordinary reason that people leave. Being genuinely worth citing still matters most of all, and no amount of markup manufactures that. A page that says something specific and true, in a form that can be lifted, from a site that loads, is the whole of it.

The uncomfortable part is that this rewards sites with something to say. There is no technical arrangement that makes a page of adjectives worth quoting. That is a content problem wearing an engineering costume, and it is the one most companies would rather not look at.