A customer is looking for a plumber who also comes out on Saturdays. They do not open ten websites. They ask their phone. The answer has three names, and one of them is yours, with two lines about what you do and where you work. They tap the number they see and call you.

They never opened your website.

In your statistics, this customer does not exist. They did not come from Google, they did not come from Facebook, they did not come from anywhere. The visit never happened. But the customer did come.

This article explains what a customer like that means for the numbers you look at every month.

The assistant came by, the person did not

When someone asks an artificial intelligence assistant something, the assistant often goes and reads pages at that very moment, on behalf of the person who asked. OpenAI, the company behind ChatGPT, says so in its documentation: its automated agent named ChatGPT-User carries out actions a user asks for, and does not roam the internet on its own collecting pages.

These readings leave a trace, outside the visit statistics. On a single day in October, across the websites we host, ChatGPT-User requested pages 278 times and its counterpart from Perplexity, another answer engine, 214. Behind these readings are people who asked something, and the assistant came to read so it could answer them. The people themselves often never came.

In other words, your website works even when nobody is looking at it. It is just that statistics such as Google Analytics, the Google tool that counts visits, count people who open pages, not assistants that read them on someone’s behalf.

Why the numbers fall when business does not

The change is not a hypothesis. In 2025 the Pew Research Center measured, in the US, that when Google shows an artificial intelligence summary above the results, people click a link almost half as often. We have gone through the numbers in the article on what the difference between SEO and GEO actually is. Since 8 October 2025, Google’s AI Mode, where search becomes a conversation and links move to the margins, has been available in Greece, and in Greek.

So if your visits fell this year, that alone does not mean fewer people are finding you. The same people, or even more, may be finding you, and no longer need to pass through the website to decide.

Search Console, the Google tool that shows how you appear in search, does not tell you how much of the drop is due to artificial intelligence answers. Impressions (how many times you appeared in the results) and clicks from these answers are included in the overall numbers, without being separated. In June 2026 Google announced a separate report for these answers. Since late August it has been available to all websites. But it shows only how many times you appeared, not how many clicked. In other words, you see how visible you are, not how many visits were lost.

And when they do come, you often do not know from where

There is also the opposite case: the customer clicked the link in the answer and came. You would expect Analytics to show it clearly. Often it does not.

Links from assistants often open without telling the website where the visitor is coming from. The visit is then counted as “direct”, as if someone had typed the address, in the same group as the people a friend sends over Viber, as we explain in the article “The post got three hundred likes. How many came to the site?”. ChatGPT adds a small marker with its name to its links, and that marker is often the only trace. On our own website, most visitors from ChatGPT arrived with that marker and nothing else.

And that trace is fragile. On 8 October 2026, around the release of a new version of ChatGPT, the marker stopped arriving, because of a bug that began to be fixed on 15 October. Those watching their statistics saw visitors from ChatGPT drop sharply and “direct” visits rise. The people were there. They had simply moved to a different column.

Which numbers no longer tell the whole story

I am not saying you should throw away your statistics. I am saying some numbers no longer mean what they used to.

Total visits. They still have value as a trend over time, not as a measure of your success. A drop needs explaining before it becomes a worry.

The comparison with last year. Last year the same customer might have opened three of your pages before phoning. This year the assistant may have “read” them on their behalf. You are no longer comparing the same thing.

The visitor who left after one page. If they came from an assistant, they already knew what they wanted. They may have seen your number and left to call you. That is not a failure of the page.

Which matter more than before

Whatever corresponds to business. Phone calls, messages from the form, appointments, orders. These numbers show the business that came in, whichever way the customer came. Provided they are measured correctly, and often they are not, as we explain in the article “Analytics says “conversion”. But did a customer arrive?”.

How many people search for you by name. Anyone who heard about you from an assistant and wants to check you out often types your name into Google. You can see this number in Search Console, and it usually rises when the assistants recommend you.

The customer’s own answer. Often only the customer knows that they found you through an assistant. No tool records it for you. Turning that information into a number you can trust, rather than an anecdote you remember, takes a method that is learned in practice.

What the assistants read on your website. There is a way to see this, even when the person does not come: which pages the assistants request and how often. It does not tell you how many customers you won. It tells you which of your pages are being used to answer questions, and which go unread.

One question for whoever does your measurement

If someone does your measurement, ask them whether the monthly report shows, next to visits, the phone calls and the messages. If it shows only visits, it measures the road and not the destination.

Whether the assistants recommend you is a different measurement, one that does not wait for the visit. We explain it in the article “How You Know Whether Answer Engines Recommend You”, and how you become a source for the assistants’ answers is in the guide to answer engines.

The customer who did not stop at the shop window

For years the website was the shop window, and statistics counted how many people stopped in front of it. Today some people do not even walk past the shop window. They send their assistant to look, hear what it saw, and come straight to you.

If your statistics are falling and you do not know whether to worry, we look at it with you. Telling what was really lost apart from what simply moved to a different column takes more than a glance at Analytics, and that is our job. We tell you in writing what we see and which numbers are worth looking at from now on. If you want measurement that runs on our own server in the European Union and shows separately the visitors from answer engines that can be identified, there is MS-Analytics, and I say this with the caveat it deserves: it is ours. For how the assistants find you, see the SEO & GEO service, or send us your website’s address through the form.