Notes from the build/

What AI assistants actually read on your website

When someone asks an assistant about your business, the answer is assembled from whatever is legible about you, which is often not your website. Here is what these systems can actually use, and what they skip.

By

A growing share of the questions that used to go into a search box now go to an assistant, and the answer comes back as a paragraph with no links to argue with. If your business is described in that paragraph, you were chosen by something you cannot log into. If it is not, you were skipped without a rejection.

This is worth understanding mechanically rather than mystically, because the mechanics are duller and more actionable than the discussion around them suggests.

What can these systems actually use?

Text that exists in the page's source, in plain sentences, that states a fact.

That sounds obvious until you look at how most business websites are built. A great deal of what a company believes it says about itself is carried in images, in a slider that assembles itself after the page loads, in a video, or in a PDF nobody linked to. None of that is reliably legible.

The pages that get used are the boring ones. A services page that says what the service is in a sentence. An about page that states where the business operates and since when. A pricing page with actual numbers on it. A frequently-asked-questions page written in the words a customer would use rather than the words the industry uses.

Why do specifics beat adjectives here?

Because an assistant summarising your business needs something to summarise, and adjectives compress to nothing.

"We deliver bespoke solutions tailored to your needs" carries no facts. There is nothing in it to extract, nothing to compare against a competitor, and nothing a system could use to decide that you are the right answer to a specific question. It is not that the sentence is disliked. It is that it is empty, and empty text is skipped in favour of text that says something.

"We install and maintain solar systems for households in the Western Cape, typically between 5kW and 12kW, with an average install time of two days" is a worse sentence by the standards of a brochure and a far better one by the standards of being found.

What is legible, and what is not
  • Plain sentences stating facts, in the page source · readable

    Where you work, what you charge, how long it takes, who it is for.

  • Questions and answers in a customer's own words · readable

    The closest thing to how the question will actually be asked.

  • Claims made only inside an image or a video · invisible

    Design assets. Treat anything only visible there as unsaid.

  • Copy that assembles itself after the page loads · invisible

    A slider or a tab panel built entirely in the browser may never be read.

  • Adjectives without a fact attached · invisible

    Bespoke, innovative, world-class. Nothing to extract, so nothing gets extracted.

Does structured data help?

It helps, and it is worth doing, and it is not the main thing.

Marking up your organisation, your services and your frequently asked questions in the machine-readable format search engines already understand makes the facts easier to lift and harder to garble. It is a couple of hours of work on most sites and it removes ambiguity.

But structured data describes what the page already says. It cannot manufacture facts that are not there. Businesses that add markup to a site with nothing specific on it have made a vague page slightly easier to parse.

What about the places you do not control?

They matter more than the discussion usually admits.

Directory entries, review platforms, industry listings, an old profile with your previous address on it: these are all sources, and several of them will be more crawlable than your own site. If they disagree with you, there is no rule that says your version wins.

So the maintenance job has two halves. Say the facts clearly on your own pages, then go and correct the third-party records that state them differently. The second half is unglamorous, and it is usually where the actual errors live.

Can you measure any of this?

Partially, and I would be careful of anyone who tells you otherwise.

You can see whether traffic is arriving from assistant products, because it usually carries a referrer. You can watch your branded search volume, which tends to move when people encounter you somewhere and then look you up. You can check, by hand, what a handful of assistants currently say about you, and repeat that check monthly.

What you cannot do is get a ranking report. There is no position one. There is a paragraph that changes between users and between days, and a business plan that assumes otherwise is built on sand. The related problem of measuring what your ads are really doing is in the three receipts.

What should you actually do this month?

Write the twenty questions your customers ask before they buy, and answer each one in plain sentences on a page you own. Put real numbers in the answers. State where you work and what you charge, or the range if it varies.

That is most of it. It is also, not coincidentally, the same work that makes a website convert, which is the useful part: unlike most things that chase an algorithm, this one pays even if nothing about the algorithms turns out the way anyone expects.

If you want it done rather than described, it is part of the search work I do.

Get the next essay in your inbox.

Notes and guides on ads, measurement and building with AI. One email a week. No noise.

Read next

The builds behind this kind of work, done for your business. Custom AI, tools and apps

← Back to writing