Most Business Websites Are Invisible to AI Assistants

A growing share of the people deciding whether to contact your business never see your website. They ask an assistant, and the assistant reads your site on their behalf and summarises it.

Which means your site now has a second reader with different capabilities from the first, and most sites are built entirely for the first one.

Ranking and being cited are different problems

Search ranking answers: which pages are most relevant. Citation answers: which source do I quote when composing this answer.

Those select for different things, and the research is fairly consistent about what the second one rewards.

A study running over 250,000 trials across six large language models tested which content changes actually move citation rates. The interventions that produced the biggest gains were adding statistics, citing sources, and including quotations. Separate analysis found content carrying original data and verifiable claims sees materially higher visibility in generated answers.

The pattern is fact density. An assistant composing an answer needs something specific to attribute. A page of well-written general prose gives it nothing to point at, so it gets absorbed into the model's general knowledge and no link appears. A page with a number, a source, and a date gives it a reason to name you.

There is also an entity effect: models cite sources they can verify as real organisations, which correlates with consistent presence across other places on the web rather than anything on your own domain.

The failure that happens before any of that

All of which is irrelevant if the crawler cannot read the page.

We crawl customer websites for a living, and the most common outcome is not a bad page. It is an empty one. A client-rendered application serves a nearly empty document plus a bundle, and a fetcher that does not execute JavaScript sees a shell. Not a warning, not a partial read. Nothing.

The site looks perfect in a browser. It is blank to anything that does not run a full engine, and not every assistant does.

Four other reliable failures:

Text baked into images. Pricing tables, specification sheets, and comparison charts are frequently pictures. Every fact in them is unreadable.

Content behind interaction. Accordions and tabs where the content is not in the DOM until clicked. A crawler does not click.

PDFs carrying the real answers. Specifications and policies parked in a two-column PDF that extracts as interleaved gibberish.

Answers scattered across a journey. Information a human assembles from four pages by navigating. A retriever gets one chunk at a time and cannot assemble anything.

What to actually check

Fetch your own most important page without JavaScript:

curl -s https://yoursite.com/pricing | wc -c

Read what comes back. If your prices are not in it, they do not exist to a meaningful set of readers.

Do the same for the pages that answer your five most common sales questions. That is a fifteen minute audit and it is worth more than most content strategy.

llms.txt, and what it does not do

The convention is a markdown file at your root pointing to your most useful content in a clean form. We publish one, at /llms.txt, alongside a fuller /llms-full.txt.

Be clear about its status. It is a proposed convention, not a standard, and adoption by the major assistants is partial and changing. Publishing it is cheap and low-risk, and it is not a mechanism that makes an unreadable site readable.

The honest ordering is: make the HTML readable first, then add llms.txt as a convenience. A site whose pricing lives in a JavaScript bundle is not fixed by a text file describing where the pricing would be.

The writing changes too

Assistants quote passages, not pages. A sentence that begins "This means that..." cannot be lifted, because out of its paragraph it refers to nothing. A sentence that names its own subject can be.

Which produces a small, unglamorous discipline: write self-contained sentences. Answer the question in the first two or three lines of the page, before context and before narrative. Give the specific number rather than the reassuring adjective. Put a date on anything that will age.

None of that is new advice. Technical writers have said it for decades. What is new is that a second reader now enforces it, and that reader has no patience, no scroll, and no ability to infer what you obviously meant.

The uncomfortable version

Most of this is not an AI problem. Text in images, content that requires clicking, answers spread across four pages, prose that never states a fact plainly: all of it was already bad for screen readers, bad for people on slow connections, and bad for anyone in a hurry.

Assistants did not introduce a new requirement. They removed the tolerance for failing an old one, and made the failure quantifiable for the first time.