Somewhere this week, a business owner asked an AI assistant to find them an accountant. The assistant read a handful of websites, picked out the firms it could understand, and came back with two or three names. Nobody clicked through ten search results. If your website is hard for an assistant to read, you were never in the running, and you would have no way to know.
The quiet shift in how customers find you
For twenty years the playbook was search: show up on Google, earn the click, and make your case on your own pages. That still matters. What is new is the layer in between. Tools like ChatGPT, Claude, and Google's AI answers now read websites on a person's behalf and summarize what they find. The person asking never sees your photos or your fonts. They see whatever the assistant understood, and assistants understand plain, well-organized sites far better than clever ones.
What we found on our own site
In late August we ran our own website through that lens, and it was humbling. The site looked fine in a browser, which is exactly why we had never caught the problems underneath. Our bare domain, the address without the www in front, showed an error to anyone who typed it. Pages that did not exist told crawlers they did, which muddies how tools like Google index a site. And our sitemap, the file that tells crawlers what pages exist, listed nine pages when the site had twenty. Every service page we most want people to find was missing from it.
Each fix took minutes once we knew where to look. We also added a file called llms.txt, a plain-text summary of who we are and what we do, written for the machines that read the site on your behalf. And we made sure the booking page works even when the embedded scheduler does not load, so there is always a plain link and a phone number in reach.
The experiment: a menu of actions for AI assistants
The newest piece is a web standard called WebMCP, backed by Google and Microsoft. It lets a website hand an AI assistant a short menu of things it can do there, the way a well-run front desk hands a caller the three options that matter. Our site now offers three: describe the firm, explain a service, and walk through booking a consultation.
Today this only works in test versions of Chrome, so nearly every visitor will never notice it is there. We built it anyway, for the same reason we test accounting software on our own books before recommending it to a client. The standard even changed while we were building, when Chrome moved the feature from one part of the browser to another. Early technology does that, and it is a far better lesson to learn on your own website than on a client's.
What this means for your business
WebMCP itself can wait until it ships in the browsers your customers use. The basics are worth doing now, and most small business websites we review are missing at least one of them: real text on the page instead of everything drawn by code, accurate contact details and hours, a sitemap that lists every page, and error pages that report themselves as errors. These fixes are quick, and they improve how you show up on Google today, well before AI assistants become the common way customers find you.
This is the homework we do inside our App & AI Advisory work. We try a tool on our own business first, and clients get a straight answer about whether it belongs in theirs. If you would like to know what an AI assistant sees when it reads your website, reach out and we will take a look.