Someone is looking for a business like yours right now. Increasingly, they aren't starting with Google. They're asking ChatGPT, and it comes back with a handful of names.

Something decided those names. And whatever it was, it has already visited your website.

I've seen this firsthand from the sites I manage. When I looked at where the visits were coming from, a large share weren't people at all. They were AI crawlers, reading the site to work out whether it was worth using as a source. They arrive, they read, they leave. Later, when someone asks a question in that industry, the sites they found useful are the ones that get named.

This is happening at a scale most people haven't registered. Earlier this year, automated traffic overtook human traffic on Cloudflare's network for the first time, and AI crawlers account for a large share of it.11Cloudflare Radar data shared by CEO Matthew Prince, 3 June 2026. Summarised in Digital Applied's AI crawler statistics reference.22Cloudflare put AI crawlers at 20.3% of verified bot traffic in May 2026, with AI search bots adding a further 6.5%. Via Nobori.

The crawling happens whether you're ready or not. All you control is what they find when they arrive. That decides whether your business gets named when someone asks a question you should be the answer to.

Most of the businesses I look at have never thought about this once.

Why an automation consultant keeps ending up in website work

I get hired to build AI systems. Automations, integrations, internal tools. I came into this from web development, and for a while I treated that as the part of my career I was growing out of.

I had it backwards. Those skills turned out to be the other half of the job, and the two together do more for a business than either does on its own.

Most small businesses I work with don't get their leads from their website. Work comes through referrals, reputation, people they already know. But the referral still looks you up. Someone recommends you over coffee, and the next thing that happens is your name gets typed into a search bar. Your website doesn't bring you the lead. It decides whether you keep it.

What's changing is who makes the recommendation in the first place. When someone asks ChatGPT for a good firm in your field, it doesn't invent an answer. It assembles one from what the web already says about you, hands over a short list of names, and then the person goes and checks whether those names hold up.

That is word of mouth with a machine in the middle, and it draws on two different things.

The first is whether anyone vouches for you. Reviews, directories, industry mentions, the ordinary trace a working business leaves behind.33Acromatico's 2026 AI recommendation study ran 100 live category searches through Gemini with search grounding. The engine named an average of 12.53 businesses per query and cited 569 distinct domains, concentrated on directories and review sites.

The second is what you actually do, and that almost always comes from you. Reputation establishes that you're credible. Your website is where a machine learns what you're credible at, and that's what decides whether you match the question being asked.

A model needs both, and it needs them to agree. Strong reputation with a vague or unreadable site makes you a business nobody can place. A precise site with no external trace makes you a business nobody can confirm.

Which makes your website more important than it used to be.

Think about what that visitor already knows before they arrive. They asked a question, they got a short list, and your name was on it. They aren't browsing. They're comparing. And the other names on that list are one tab away.

So the site has two jobs it didn't used to have. It has to be readable and specific enough that a machine can tell what you're for, and good enough for the person who arrives that they stop comparing and make contact. Almost every site I look at was built for neither.

Whether a machine can tell what you're for

Start with who your website was written for.

If your work comes through referrals, your homepage was almost certainly written for people who already know who you are. It doesn't need to explain the business, because whoever sent them did that over coffee. So it can afford to talk about values, approach, the way you like to work.

A machine gets no introduction. It reads you cold, trying to work out whether you're the answer to something specific that somebody just asked.

And the questions are specific. Nobody asks for a good accountant. They ask for an accountant who works with tradespeople. For you to be the answer, the words describing that situation have to exist somewhere on your site.

This is where otherwise decent sites lose. The usual versions:

  • The service list is generic. Tailored solutions and bespoke advice describe nothing a question can attach to.
  • Nowhere states who the client is. Not the industry, the size, the location, or the stage they're at.
  • The real work sits in a PDF, an image, or only comes up on a call.
  • Services carry in-house names without ever saying, in plain words, what they are.
  • The homepage explains how you work before it says what you do.

Here's a test you can run yourself. Take your last new client. Before they knew you existed, they had a problem they would have described in their own words. Is there a sentence on your site that would have matched it?

This is most of what a rebuild actually involves. Working out what the business is for, then making the pages say it plainly enough that a person and a machine reach the same conclusion.

Whether it can read you at all

All of that assumes a machine can reach your site in the first place. A surprising number can't, and nothing on your end would tell you.

This part doesn't earn you a recommendation. It only keeps you in the running. Two things decide it, and both usually go wrong by accident.

The file that decides who gets in

Every website has a small text file called robots.txt. It's a set of instructions for automated visitors, telling them which parts of the site they're allowed to look at. It was invented for search engines, it gets set up once when the site is built, and almost nobody opens it again.

The AI companies don't send one visitor. They send several, and each does a different job. One collects material to train the model. Another builds the list of sources the model draws on when it answers a live question. They arrive under different names, and your file treats them as separate visitors, which means you can allow one and block the other without realising you've done it.44OpenAI's are GPTBot (training), OAI-SearchBot (the live answer index) and ChatGPT-User (fetching a page on request). Anthropic runs the same three-way split as ClaudeBot, Claude-SearchBot and Claude-User. See Anagram's AI crawlability checklist.

Turn away the wrong one and you disappear from the answers.

This usually isn't something anyone chose. Hosting platforms and security services add bot rules of their own during setup, and the owner is rarely told which ones.55A 2026 analysis found Anthropic's crawler implicitly blocked across a meaningful share of common hosting providers, without the site owners' knowledge. Configuring robots.txt for AI crawlers.

Pages that arrive empty

On some sites, the words aren't in the page when it's sent. The page arrives almost empty and the visitor's browser fills it in a fraction of a second later. Nobody notices, because by the time you're looking at it, everything is there.

Machines tend to read what arrived rather than what appeared.66Anagram, 2026 crawlability checklist.

So a site can look completely current to you and to your clients, and still be a blank page to the systems deciding which names to hand over. It's a common way to build a website, and you can't tell by looking at one.

One to watch out for. If you start asking around about any of this, sooner or later someone will try to sell you a file called llms.txt. It's meant to be a summary of your business written for AI to read. I'll save you the money: there's no good evidence anything is reading it, and sites that have one get named about as often as sites that don't.77Gary Illyes confirmed in July 2025 that Google does not support llms.txt; John Mueller compared it to the keywords meta tag. Crawl data from Limy's analysis of 500M+ bot events. A separate three-month study by Otterly found the file accounted for 0.1% of AI crawler visits.88MaxAEO citation tracking found a 0.2 percentage point gap between adopters and non-adopters, inside noise. It takes ten minutes and costs nothing, so I still set one up in case that changes. But if it's the centrepiece of somebody's proposal, ask them what they're basing that on.

When they arrive

Say all of that works. Your name comes up, and someone clicks through with a specific question and two other tabs open.

What most sites offer at that moment is a contact form. It asks the visitor to write out their question, hand over their details, and wait, on the chance that the reply is useful. Against two competitors one click away, that's a lot to ask.

What I've had the strongest results with is a chatbot, though not the kind you're picturing. It's built on the business itself: the services, the products, the documents, the questions that come up over and over. It answers the question there and then, and it books the meeting inside the same conversation.

I built one for a superannuation and financial advice firm in Australia. Meetings were booked through it in the first week.

Before that, an interested visitor went onto a list. Someone had to notice them, work out whether they were worth a call, and get in touch. In practice most never got that far.

Over its first 30 days, 83 Australian visitors landed on the site. Fourteen of them opened the chatbot. Five booked a meeting and three more left their details. For context, website chat typically gets 2-8% of visitors to engage; this was closer to 17%.99Chatbot KPI benchmarks, 2026.

It's a small sample and I'd hold the percentages loosely. What isn't in doubt is the five meetings, none of which had a way of happening before.

Nothing about the firm changed. The same people were arriving with the same questions. They just stopped needing somebody to notice them.

The half that takes longer

Reputation is the other input, and it's the one nobody can install for you. It comes from doing good work and having clients say so somewhere other people can see it.

What can be built is the asking.

Most businesses I work with have plenty of satisfied clients and very few reviews. That gap isn't reluctance. Roughly seven in ten customers will leave a review when a business actually asks them to. The problem is that asking depends on somebody remembering, at the end of a job, while they're already moving on to the next one.

This is exactly the sort of job automation is good at, and it's the kind of thing I get hired for. Attach the request to the event instead. A finished engagement, a paid invoice, a completed job sends the message, and it goes out within the hour rather than whenever someone gets to it. Timing does most of the work here, and a single automatic reminder a few days later does much of the rest.1010ReviewTrackers data indicates requests sent within an hour of service completion reach roughly a 32% open rate against 12% at 24 hours, and Reputation.com puts manual request response rates at 5-8% versus 18-35% for automated triggers. Summarised in US Tech Automations' review automation guide. On follow-up reminders doubling completion, see Plaudit.

This half compounds slowly, but it feeds everything above. Reviews make it more likely you get named at all. Your website decides what you get named for.

Where I'd start

None of this announces itself. Nothing breaks, no error appears, and no report flags it. Someone asks a question you should be the answer to, a few names come back, and yours isn't one of them. You never find out.

So before fixing anything, find out whether you have a problem. It takes two questions and about a minute.

Open ChatGPT and ask it for a firm like yours. Your city, your field, the thing you actually do. See whether you come up, and pay attention to who does.

Then ask it about you by name, and read the answer as though you were a prospective client.

If you come up and the description is right, you're in better shape than most businesses and this is maintenance.

If you don't, the reason is one of the things in this article. Working out which one, and what it will actually take to put right, is the part that needs somebody to look properly. That's the work I do.

Either way, it's worth knowing. Your name is already coming up in these conversations or it isn't, and that has been true for a while now. The only question is whether you've seen what they're finding.