The answer is not on your website
Somebody is asking a machine about your company this week, and getting an answer you will never see. It is not built from your homepage, and it is usually older than you think.
August 5, 2026
Sometime this week, a stranger is going to ask an assistant about your company. Maybe they will type your name. More likely they will type the problem they have and ask who solves it. Either way they get one answer, in about four seconds, and it will sound certain.
That answer is not your website. It is assembled out of whatever a model can find about you, reconcile into a single organization, and decide to trust. Your homepage is one input. It is not the output, and it is frequently not the most recent thing in the pile.
You will never see the answer. Nobody forwards it to you. There is no analytics event for the moment a machine described your company to someone who was deciding whether to call you.
I found this out the embarrassing way, on my own site. In August I asked an assistant to describe my studio. It said we had spent five years on mission-driven work and that the team came out of political campaigns. Both lines came from a page I had deleted that morning, after two weeks rewriting every page on the site.
This stopped being a niche question
About half of US adults now use AI chatbots, up from roughly a third two years ago. Looking things up is the most common reason they open one.
I am not going to tell you that buyers have replaced their research with a chatbot. I have no evidence for that, and it would be a convenient thing for me to claim. What I will say is that what does this company actually do is exactly the shape of question people now type into one. So is who else works on this. So is are these people any good.
Those get asked casually. They get answered confidently. And whoever was left out of the answer never finds out they were in the running.
It is retrieval, not ranking
A search engine ranks documents. You compete for a position, and eighth place still sends you people. An assistant does something else. It gathers what it can and writes one answer. There is no eighth place. You are in the answer or you are not.
That sounds like a small distinction. It changes what is worth doing. Being the eighth most retrievable firm in your category earns you nothing at all.
Search gives you a rank. An assistant gives you a sentence, and either your name is in it or it is not.
There is a second difference, and it is the one that bites. A search engine can serve one good page from an otherwise incoherent site. An assistant has to work out what you refers to before it can say anything about you, and it does that by reconciling every mention of your name it can find. When those mentions disagree, it settles on whichever version is most consistent. That is rarely the version you would have picked.
I watched it happen to my own name in the same test. The results blended my studio with a foundation, a homeschool curriculum, and a web company in another state. Nobody had done anything wrong. There was simply nothing telling the model which of us was which.
The smaller you are, the more your own site matters
A model builds its picture of a subject from many sources, and it trusts sources that agree with each other. A household name has a Wikipedia entry, press coverage, filings, conference listings, and a decade of other people writing about it. Everyone else has their own website, a couple of directories, and maybe one local news story from 2019.
So the less famous you are, the more weight your own site carries, and the more damage a vague one does. That inverts what most people assume. The usual instinct is that a small organization's website matters less because fewer people see it. In retrieval terms it matters more, because there is nothing else to check it against.
40% of chatbot users go to one to look something up, which is the most common reported use of them.
Pew Research Center, June 2026It gets worse in Spanish. There is far less structured Spanish-language material for a model to draw on about a US organization, even one whose customers mostly read in Spanish. I have not yet seen anyone close that gap with a translation widget. Ask the same question in both languages and you will often get two different companies back.
What actually moves the answer
None of this is exotic. It is ordinary clarity, applied in the places a machine reads instead of only the places a person reads.
- Say what you do in plain words, in one obvious place. Not a mission statement and not a positioning line. The sentence someone on your team would say at a dinner party when a stranger asks where they work.
- Be one organization rather than nine mentions. One spelling of the name, one canonical address for the site, and links out to the profiles that already exist for you. Almost nobody has done this, and it is the highest-value hour on the list.
- Publish the specifics you would want quoted back at you: what you sell, to whom, where, since when, at what scale. Vagueness does not get summarized into something flattering. It gets skipped.
- Get named by sources that already carry weight. Partner sites, directories, coverage, customer pages. A link is not only traffic. It is somebody else asserting who you are, which is what settles the ambiguity above.
- Ask the question yourself, then ask it again next month, in every language your customers use. Write down what came back and the date. Answers drift, and you cannot manage what you have never measured.
That last one is the step everyone skips and the cheapest one on the list. It took me ten minutes to find out the internet was still describing my studio the way it had two years ago, and I only bothered because I had just rebuilt the site and wanted to admire the work.
What I am not claiming
Nobody can tell you your AI ranking, because there is no such number. The measurement here is genuinely immature. What you can do is write down what a handful of models say about you today and watch whether it improves, which is closer to reading a thermometer than reading a dashboard.
A word on the conventions being invented around this, since they are being sold hard. There is a proposed standard called llms.txt, a plain file describing your site for language models. A lot of people added one last year. I was one of them. This spring a study of 137,000 sites found that almost none of those files were ever fetched by anything, and Google has said it does not use the format and has no plans to. Believe that before you pay anyone for it.
The same caution applies to any product promising you an AI visibility score. There is no ranking to score. What there is, is a record of what several models said about you on a given date, which is useful precisely because it is modest.
This also replaces nothing. Search still matters. Referrals still matter more than either. And no amount of structured data will make a model recommend a company nobody has heard of and nobody has written about. It compounds slowly, the way reputation always has.
The reason I keep pulling on it is narrow. For most organizations the website is the one piece of the public record they fully control. That used to mean controlling what people read about them. It now means shaping what gets repeated about them in rooms they are not in.
