How to measure whether clients come from AI
Investing in visibility for language models runs into one question: is there any return. Measuring it is harder than measuring search traffic, but not impossible. Here is what can already be counted and what you will have to check manually.
Visits from ChatGPT, Perplexity and similar services appear in analytics as distinct sources — you can group them into a segment and count them like search. But most of the effect never reaches your reports: someone reads the answer, remembers the name and arrives later directly. So two manual practices are added to analytics: asking «how did you hear about us» at the point of enquiry, and checking your own questions in the models monthly.
In this article7 sections
Why ordinary reports understate it
The model shows an answer and names a company. Some people click the link and reach you — that visit is visible. But many simply remember the name and come later through a branded search or directly. In the report that will be a «direct visit», even though AI brought them.
The practical conclusion: the number of visits from AI is the lower bound of the effect, not all of it. Growth in branded queries helps you see the upper bound.
If direct visits and searches for your company name are growing with no advertising and no offline activity, the likely cause is that you are being mentioned more often — in models, in directories, in recommendations.
What is visible in analytics
Visits arrive from recognisable addresses, and they can be grouped into a single set of sources.
| Source | How it appears in the report | What it means |
|---|---|---|
| ChatGPT | chatgpt.com | A click from a link in an answer |
| Perplexity | perplexity.ai | A click from an answer with sources |
| Gemini | gemini.google.com | A click from a model's answer |
| Copilot | copilot.microsoft.com | A click from an answer in Bing |
| AI overviews in search | usually as a search visit | Not identified separately |
The last row matters: answers a search engine shows directly in the results usually fall into general search traffic and are not counted apart.
How to set up the tracking
- Create a segmentGroup the known model addresses into one set of sources in your analytics.
- Add goalsFor that segment count enquiries rather than visits: form, call, messenger.
- Look at behaviour separatelyDepth of reading and share of enquiries for this audience usually differ from search.
- Watch branded queriesGrowth in searches for your company name is an indirect sign you are being named more often.
- Record direct visitsTheir trend alongside branded queries shows the overall awareness effect.
Checking mentions by hand
Analytics shows visits but not whether you are named at all. That is checked manually, and it is worth doing monthly.
Draw up a list of 10–15 questions your client would ask: «where can I get X in city Y», «what does X cost», «who should I choose for X». Then for each question:
- Ask it in two or three services in a private window.
- Record whether you were named, in what position, and what was said about you.
- Note who is named instead of you and what the model cites.
- Check the answer for outdated or incorrect facts about you.
Model answers are unstable: the same question gives different companies on different days. Look at the share of mentions across several checks, not at a single answer.
A simple table for tracking
| What we record | How often | What for |
|---|---|---|
| Visits from models | Monthly | The lower bound of the effect |
| Enquiries from that segment | Monthly | The actual return |
| Searches for the company name | Monthly | Growth in awareness |
| Share of mentions in answers | Monthly | Visibility in models |
| Factual errors about you | Monthly | What to fix on the site |
Five rows filled in monthly give a sufficient picture. Paid mention-tracking services make sense at scale, but you do not have to start with them.
What to do with the results
- You are never named. Check that the site is open to model crawlers and that pages carry direct answers and prices as figures.
- Named, but with errors. The model picked up outdated data — update it on the site and in directories.
- Competitors are named. Look at what the model cites: usually catalogues and reviews you are absent from.
- Visits but no enquiries. The problem is not visibility but the page people land on.
What volume to expect
Today the share of visits from models for most local businesses is a few percent of total traffic. But it is a fast-growing share, and the audience quality is usually higher: the person arrives with intent already formed, because the model answered their question before the click.
The sensible position is not to expect the main flow here today, but to take the ground while the channel is still empty — and to learn how to measure it.
Frequent questions
Can clients from AI be counted precisely?
Not precisely. Only link clicks are visible; the rest of the effect shows through direct visits and branded searches. Count the three metrics together rather than one figure.
Is it worth asking clients how they found us?
Yes, it is the cheapest way to catch what analytics misses. «I asked ChatGPT» is an answer heard more and more often.
How often should we check mentions by hand?
Monthly, using the same list of questions. Less often and you miss changes; more often and you react to random variation in the answers.
Do we need paid tracking services?
Not at the start: a list of questions and a spreadsheet give the same picture. They pay off when mentions are numerous and regular automated monitoring is needed.
Why are the answers different every time?
Models do not produce deterministic output and take the query context into account. That is why you judge the share of mentions across a series of checks, not a single answer.
Should AI overviews in search count as a separate channel?
Technically they arrive as search traffic and are not separated out. An indirect sign of their effect is a fall in clicks while positions stay the same.
The materials answer general questions. We will look at your specific case — free and without obligation.