Two new leads came through our own website in less than seven days because somebody in Atlanta asked a chatbot a question and the chatbot said our name. Not Google. Not an ad. There is not one line in the traffic report that says ChatGPT, because there is no column waiting for it.
That is the part most business owners have not run into yet, and it changes what you do next. If a customer found you through an AI assistant last month, you probably cannot prove it from your dashboard. If you cannot prove it, you cannot decide whether to do more of it or leave it alone. Two leads is a start, not a finding. It is a signal, though, and it is one you can check yourself in an afternoon without buying anything.
Where an AI referral goes to hide
The assistant answers inside the conversation. It reads pages from the open web, names two or three businesses, and then the person acts. They call. They type your name into a new tab. Or they tap a link.
When the tap comes through, the request usually arrives with no referrer at all. An assistant is not a browser, some of the handoffs drop the query string, and the systems doing the reporting are not the systems your analytics was built around. Google Analytics 4 drops all of it into Direct, which is the bucket it uses when it has no better idea. Direct has always been a pile: bookmarks, typed addresses, an email app, a QR code on the side of a van, a link in a group text. Now the fastest growing source of new business on the web is sitting in that pile under a label that means we do not know.
Most owners look at the number, shrug, and go back to the ads report, which has a tidy column for everything and a bill attached.
There is a second shape worth knowing about. Sometimes the visit does carry a referrer, from a search engine or from a page the assistant pointed at, and it looks completely ordinary. The assistant disappears in the middle of the trip instead of at the end of it. Your report credits the last click, the assistant gets nothing, and everybody goes back to arguing about attribution models that were built for a world where a person clicked one link at a time.
The log knows before the report does
Before an assistant can name you, it has to read you, and the reading shows up in your raw server log, the dated record of every request for your pages. The readers announce themselves in it. OpenAI's arrive as GPTBot and OAI-SearchBot. Perplexity's arrive as PerplexityBot. Anthropic's arrive as ClaudeBot, and every other assistant worth naming has its own. Your host keeps these files and will hand them over if you ask.
Read them once a month and you learn something no dashboard will tell you. Say the count falls from a steady few dozen requests a month to a trickle. Something changed in how the machines see you, and that is worth an afternoon now instead of a quiet quarter later.
Three different things hide in that log. Being read is not the same as being named. Being named is not the same as being clicked. Being clicked is not the same as being called. Only the last one pays the electric bill, and the first three are how you find out which way it is going before the phone tells you.
Ask the machine the question your customer asks
You can read the answer yourself, by hand, in about ten minutes. Write down the five questions your best customer would ask before hiring you. For a plumber in Decatur that looks like who fixes a slab leak in Decatur, what it usually costs, and can somebody come this week. Then ask those questions to the assistants people actually use, on a device you are not signed into, because being signed in changes what you get back. Personalization is doing a lot of quiet work in those replies.
Read the answers the way a customer would. Are you in them at all. Are you named first, third, or never. Which competitor keeps showing up beside you. Which page of yours the answer leaned on, if it says. Save every round with the date on it, because the same five questions in January are the only way to see movement in June. It is the same discipline as the short list of numbers we actually run a website by, and we wrote those down here: the five we watch. Ten minutes, once a month, no tool required.
Nobody outside your shop will check that for you, and no software we have tried does it honestly. It is the nearest thing this new front door has to a position report.
Ask the person who called
Then do the thing almost nobody does. Ask.
Put the question in the intake form, or ask it on the call and write the answer down the same day. When we ask, the answers land in one of three places: a friend, a search, or an assistant. That third answer is the datapoint. If a few of your next twenty leads say it, you are in the answers and the report will never admit it.
Put the answer in a field in your CRM the same way you record the job type, so you can count it at the end of a month instead of leaning on memory. Three lines in a spreadsheet on Friday is the whole system, and it settles arguments that analytics cannot settle.
A tracked phone number settles part of the argument, because a call is a call no matter what the browser did with the referrer. And when a crawler shows up in your logs, shows up again the following week, and then a call comes in from somebody who says they asked an assistant, you have a chain you can show a partner or a lender instead of a hunch.
What to fix if you are not in the answer
If you asked the five questions and you are nowhere in the replies, the fix is not a trick. Assistants repeat what they can read plainly, and they repeat it more when the same facts appear in the same words in more than one place.
So give the machine something it can hold. What you do. Where you do it, and the towns you actually serve. Your hours. Your phone number. A price range, if you are willing to say one out loud. The same business name, address and phone, spelled the same way on your site, on your Google Business Profile, and in the directories still carrying the number you had in 2019. Pages that answer one question each instead of one long page that answers nothing. Labels a machine can parse, which is what schema markup is for when it is done honestly instead of sprinkled on at the end. We went through the markup stack for service businesses in a separate note, and it is the least glamorous item on this list and the one carrying the most weight.
None of that is new. It is the same boring list that worked before, done with more discipline, because the thing reading it now has no patience for a page that makes it guess.
What to do with the numbers
Put the three checks on a calendar. The log once a month, the five questions once a month, the intake question forever. Then read your report knowing what Direct actually holds. If Direct climbs while everything else sits still, and a few of your newest leads say they asked an assistant, that is your answer and it is worth spending on. If Direct is flat and nobody says it, you have work to do on the site before you have a case to make.
We took a bet three years ago that answering would replace the ten blue links, and the phone is starting to agree with us. You do not have to take our word for any of it. Ask the machines about your own business this week and see what comes back. If you would rather have somebody run the checks and fix the parts that come up short, that is the work we do for Atlanta businesses. Bring us the five questions.