Would ChatGPT recommend us? Not at first.

Earlier this year we built a way for other people’s AI assistants to book a discovery call with us. We did not build a chat window on our website for people to do the same. Nobody decided that; it is simply the order the work came in, and it tells its own story. We think the next client is more likely to ask an assistant to find them a software firm than to search and then work through our website. So we built the door the assistant would use first. Then we asked the obvious question: would an assistant actually send anyone through it?

Three jobs, not one

For an assistant to bring you a client it has to do three separate things. It has to find you. It has to be able to talk to you. And it has to decide to recommend you. We had been treating the second as the whole problem. It is the easy one.

Being found is still search

An assistant that is asked for a software firm near its user looks one up the way a person would, with a search tool, and reads what it finds. If your site does not say the thing, the assistant cannot know it. Everything the last fifteen years taught about being findable still applies; the reader has changed, not the rule.

How many people start this way is genuinely unsettled, and surveys in 2026 disagree by a wide margin depending on who is asked. Semrush’s study of American shoppers who had already tried AI found Google still the more common first stop, 33% to 26%, with only 4% relying mostly on AI. Other surveys put the AI share higher. Read all of them as direction. The direction is that a growing minority of your prospects now ask a machine first, and the machine reads your website more literally than any human visitor ever did.

Being talked to

Once an assistant has found you it needs a way in. Ours is an agent-facing door on our website: a published description of what an outside assistant may do, which is read about us and book a provisional discovery call. The same door serves a person using the assistant on our site and a machine calling from somewhere else, and either way the result is one enquiry record for a human to confirm.

There is an opinion doing the rounds that none of this is necessary, because modern assistants are clever enough to poke at any website until they work out how to use it. The opinion is half right and misses the point. A patient agent can learn a well-described system, for reading at least. But we want to be found, and we want other people’s agents talking to ours, so why make it harder? The door is also where the manners live. Ours lets an agent book, and only book: the call is created as provisional and a person confirms or declines it, nothing an agent does carries more permission than an anonymous visitor would have, and there is a daily spending cap after which the assistant politely points to the contact form instead.

In July we let an AI agent we had not built, and told nothing about how the door worked, try to book a call for a client and then try to misbehave. It booked the call. Every attempt to read other people’s enquiries, confirm its own booking or smuggle instructions into a form field was refused. It also found a real defect: the booking form expected a date and time in computer format, and an agent, like a person, says “Thursday afternoon”. We changed the form. That is what the test is for.

Being recommended

This is the job we had not thought about, and it turned out to be the interesting one. We did not ask ChatGPT whether it could book a call with us. We described a project, gave it our website, and asked whether it would recommend us. At first it would not.

The first gap was geography. Our site did not say that we have three offices in the UK, or that we work with clients across the country and abroad. To a person skimming, a British software firm is obviously available to a British client. To an assistant weighing options, a firm that does not say where it works is a firm that might not work where you are, and it went to the next name on the list. We added the facts and asked again. There were more rounds after that, and each one turned up something we knew perfectly well and had never thought to write down, because no human reader had ever needed to be told.

That is the lesson we would hand to any business owner. The facts an assistant needs before it will put you forward are not the facts a person needs, and you cannot guess them. You have to ask. Give the assistant a realistic brief and your website, and ask what it would need to know to recommend you with confidence. Then put the answers where it will read them. It is the cheapest piece of market research you will do this year, and the machine will tell you honestly where you fall short, which is more than most prospects do.

How we do it

Every system we build can be driven by an outside assistant under exactly the permissions a person would have, and we test with more than one maker’s assistant because a client will arrive with whichever one they use. For our own business we now treat “would an assistant recommend us” as a question to re-ask whenever the site changes, the same way we re-run a test suite. The chat window for people is on the list. It will reuse the door we built for the machines, which is how it should have been built the first time.

The question to ask on Monday

Open whichever assistant you use, describe a job your firm does well, and ask it whether it would recommend you. If it says no, ask it why. Everything it tells you is a sentence your website is missing.