Is AI Search Bringing You Customers? What You Can Measure, and What You Can’t

Your analytics can't fully tell you whether AI search is bringing you customers, and for most businesses no tool can honestly prove which customers it brought in. What you can measure is three things: whether AI answers name or cite you when buyers ask your questions, a floor of visits you can identify as coming from AI, and what happens on the pages AI answers cite.

That is less satisfying than a revenue number. It has the advantage of being true. Below is why your analytics miss most of it, a week of our own numbers (including the ones that don't flatter us), and what you can put in front of a board without crossing your fingers.

Disclosure: I build Alice at TranX, which checks AI answers for businesses. Weigh the advice with that in mind. The numbers are from our own site.

Four terms, in plain words

• Cited: the AI answer links to your website as a source.
• Mentioned: the answer names your business, with no link. A different result from cited, so count it separately.
• Known AI visits: visits your analytics can tell came from an AI assistant. A floor, not total demand.
• Outcomes on AI-cited pages: signups, purchases or leads on the pages AI answers cite. A page segment, not attribution.

Why your analytics can't answer the question

Picture a buyer asking an AI assistant who to hire for the thing you sell. The answer names you. What happens next depends on how they act on it, and only one route reaches your analytics labelled as AI.

The first path surprises most people. Google's Analytics channel definitions put clicks from AI Overviews and AI Mode in Organic Search, and the newer AI Assistant channel explicitly excludes them. So clicks from Google's own AI answers are reported as ordinary search. The third path is the phone problem. A browser normally passes along a referrer, a note saying which site sent the visitor, but many links tapped inside assistant apps arrive without one, and unless the link carries a tracking tag, they land in Direct. The fourth and fifth are the quiet ones. The answer did its job, and the credit went to a branded search, or to nobody.

Asking GA4 whether AI search brought you customers is like asking your doorbell how people heard about your shop. It knows who rang. It has no idea who read the sign last week and came back today.

Then there is scale. Even when the labels work, a small business often sees a handful of AI visits a week, and a handful of visits cannot show that anything caused anything. If you want the GA4 side done properly, our guide to measuring AI search traffic covers the three signals to combine. Keep the label on the result: at least this many.

A week of AI attention on our own site

Here is what this looks like on tranx.io. We pulled the raw request logs from the network that delivers our website (our CDN) for October 7 to 13, 2026. Unlike GA4, these logs record every request, including the AI crawlers and assistants that never run analytics code. We counted only requests that returned a page (no images, no scripts, no bots probing for password files), and counted OpenAI and Perplexity crawlers only when they came from the IP ranges those companies publish.

Three things stand out. First, AI crawlers read us all week: OpenAI's search crawler alone fetched our pages 161 times. Reading is not citing. It only means we are eligible.

Second, OpenAI says its ChatGPT-User agent visits a page when someone asks ChatGPT a question, and is not used to crawl the web automatically. That agent opened our pages 116 times, 99 of them the homepage. We can't tell how many were people asking about us and how many were automated tasks, and some may be our own tests.

Third, in the same week, 5 visits arrived in a browser with a chatgpt.com referrer. Whatever the mix behind them, ChatGPT's page fetches outnumbered labelled visits by more than 20 to 1. GA4 saw none of those 116. It could have seen the 5, minus anyone who declined analytics cookies. Judged by GA4 alone, AI search barely exists for us. It does exist. It mostly happens where analytics can't look.

We also ran our free check on ourselves

On October 14, 2026 we put tranx.io through our own free AI visibility check, which asks two questions live on Google AI Overviews, ChatGPT and Claude: one about us by name, and one a buyer would ask.

Asked about TranX by name, ChatGPT and Claude both cited tranx.io. Google AI Overviews did not. Asked “How do I get my website to appear in AI-generated answers?”, which is roughly the question our product exists to answer, none of the three cited us. They cited six other sites, including TechRadar, Quora and Wikipedia. The cobbler's children do have shoes, but only if you ask for them by name.

That gap, known by name but absent when a buyer asks, is exactly what you can measure. It is also exactly what GA4 will never show you, because a buyer who never saw your name never clicks.

What you can measure honestly

No single number answers “is AI search bringing us customers?”. The three measurements from the top of this post, plus two signals that come from people rather than software, get you close enough to make decisions. Each one shows one thing and not the next.

The first signal is the only one that measures the AI answer itself, so it carries the most weight. This is the method we use, and you can run it by hand:

1. Write down 5 to 10 questions a buyer asks before they know your name, plus one that uses your name. For an invented example, “best accounting firm for a small manufacturer in Ohio” beats “Smith and Co reviews”.
2. Ask each one on each engine, more than once. AI answers wobble. Profound compared about 80,000 prompts per platform a month apart in 2025 and found that 40 to 60 percent of cited domains had changed. One reading is weather. A week of readings starts to be climate.
3. Record one of four results per answer: cited (links you), mentioned (names you, no link), not cited (write down who was cited instead), or no sources. An answer with no sources is unknown, not a miss.
4. Judge each question over a window, for example “cited on 5 of 7 days”, never on one check.

This is what Alice automates. Paid plans check Google AI Overviews and ChatGPT every day and Claude every week; the free plan checks five questions weekly. Each question is judged over the last seven days, and you can read what each engine actually answered. Doing it by hand works too. It is just the kind of chore that quietly stops in week three.

How to tell if a change you made is working

Say you rewrite a page so it answers a buyer question directly. How do you know it worked? Not from revenue, for all the reasons above. From the answer.

• Keep a change log. The date, the page, and the question it was meant for. Without it, a month from now you will be guessing which edit did what.
• Keep a few similar questions you did not touch. If everything rises together, the engine changed, not your page.
• Give it time. An engine has to fetch the page again before its answers can change. Google says AI Overviews have no special requirements beyond a page being indexed and eligible to show with a snippet, so your change reaches them only after Google has crawled the new version. Judge in weeks, not days.
• Watch for the fingerprint. The strongest evidence a small business can get is the engine starting to cite the exact page you changed, for the question you targeted. It is not proof, because engines change for their own reasons. But a citation rate that moves from 2 to 3 days out of 7 is noise. Your edited URL turning up in the sources is a signature.

Alice keeps asking the same questions on paid plans (daily on Google AI Overviews and ChatGPT), so you can see in the trend when a page you changed starts getting cited, and mark the change on your charts. For how much evidence each kind of decision needs, see is AI visibility a vanity metric?

What to tell your board

Boards ask for ROI because it is the number they are used to getting. For AI search, an honest ROI figure usually isn't available, and a made-up one is worse than none: the first time someone asks how it was calculated, the whole update loses credibility. Give them evidence instead.

If you need a paragraph to paste, here is one. Swap in your own numbers; these are invented:

We track whether AI answers name or cite us for the ten questions buyers ask before they know us. This month we were cited for four of them, up from one. Visits we can identify as coming from AI assistants were at least 40; the real number is higher, because Google counts AI Overview clicks as organic search and many app clicks arrive without a source. We do not attribute revenue to AI search. We judge it by citations, by what new customers tell us, and by searches for our name.

Why we stopped promising the revenue number

For a while, Alice's own homepage said it linked AI visibility to revenue. In October 2026 we took that promise off the homepage and product pages. The data is still interesting. The claim just didn't survive our own numbers. Google's AI clicks arrive as organic search. App clicks lose their referrer. Most of the influence happens without a click. And a business with a handful of AI visits a week cannot show cause, however clever the dashboard. Our week above is the example: five labelled visits can't carry a revenue claim, and for most small businesses a week of theirs can't either.

What Alice says now is narrower and checkable: which buyer questions AI answers with you or with a competitor, what to change next, visits from AI assistants as a minimum, and what happens on the pages AI cites. If a tool tells you exactly how much revenue AI search made you, ask it which of the five paths in Fig. 1 it can see. Then ask about the other four.

To see where you stand today, run the free AI visibility check. It asks two questions live on Google AI Overviews, ChatGPT and Claude, and checks whether AI crawlers can read your site. No signup. It won't tell you your AI revenue either. It will tell you whether AI knows your name, and whether it thinks of you when a buyer asks.

Questions people ask