How Do You Measure Whether AI Search Brought You Anything? From AI Referral to Enquiry

You measure whether AI search brought you anything by counting enquiries rather than clicks, and by accepting that the honest answer is a count rather than a proof: how many enquiries you logged last month, how many of those named an assistant when you asked, how many you never got round to asking, and what came of them. That is a real answer with a date on it, and you can have the first version of it this afternoon. Getting there means counting five stages, and only the last three of them are yours to count reliably: a visit you can recognise, a visit where the person did more than land and leave, an enquiry, a qualified enquiry, a sale. The first two depend on whether the platform that sent the visitor put anything in the link that identifies it, and the answer to that is not the same from one platform to the next. This article defines each of the five stages, says who controls it, and works through the enquiry routes a UAE business actually uses: the form, the phone and WhatsApp. It also sets out what cannot be traced back to a source at all. The template published with it is where you write it down.

This is general guidance on measuring your own marketing, not legal, tax, accounting or data-protection advice, and nothing here tells you what any UAE rule requires. The method below asks you to add a field to a form and to keep a record of what customers tell you. Settle what you may collect, store and record about a customer with somebody qualified before you change anything.

The short version: log every enquiry by hand, from every route, and ask each one how they found you. That log is the backbone. Nothing else connects a person to a route. Read your analytics for arrivals as well. One assistant documents a parameter on its links; another, on the two answers we counted, added nothing to the links we read. Attribution means deciding which earlier contact gets the credit for an enquiry. It has limits, so give them their own paragraph when you report rather than hoping nobody asks. None of this makes your business shown, named or recommended anywhere, and nobody can promise that it will. What it does is let you say, with a date on it, what actually arrived and what came of it.

AI Visibility is owned by Lunasol.

Does checking what AI says about you show whether AI search brought anything in?

Checking what AI says about you does not show whether AI search brought anything in. Asking the assistants what they say is a different job from measuring what those answers were worth. Our article on how to check what AI says about your company covers that first job properly. The prompt-side check tells you whether you are named, on what evidence, and against whom. It cannot tell you whether anybody acted on it.

The ladder our article on GEO, AEO and SEO sets out runs mention, citation, recommendation, visit, enquiry. The prompt-side check measures the first three, and it measures them by asking. This one measures the last two, which arrive at your own door wearing whatever the journey left on them. That is the whole problem.

What counts as a result from AI search? Five stages, defined once

Five stages are worth counting: a recognisable visit, an engaged visit, an enquiry, a qualified enquiry and a sale. Write the definitions down before you count anything, because the arguments later are always about definitions.

Three of the five need a decision from you before you count anything. An engaged visit: pick what counts, a second page, a certain time on the page, a scroll to the pricing block, a video played. Pick one, write it down, and do not change it mid-year. A qualified enquiry: the right service, a plausible budget, a real contact. That is your bar. A sale: paid, or the job confirmed. Again your own line.

The other two need no decision of that kind. Not every visit is recognisable, and that gap is decided by the platform rather than by you, which is what the next two sections are about. An enquiry counts whatever route it came in by, a form, a phone call, a WhatsApp message, an email or somebody walking in, which is why the template logs every one of them and not only the ones you think came from an assistant.

The five stages matter in that order because each one narrows, and because the further down you go the more the number belongs to you rather than to a platform. A platform decides what a visit looks like. You decide what a qualified enquiry is, and nobody can take that measurement away from you.

How do you recognise a visit from an AI assistant? What we counted in two answers

You recognise a visit from an AI assistant only when the arrival carries something that names where it came from. That is either a tag added to the end of the link, or the referrer: the note a browser normally passes to your site saying which page the visitor came from. Whether either is there at all differs from one platform to the next. In September 2026 we put two questions about one professional category to Google's AI Mode, one run each, and counted the links in the rendered answer. Numbers only: we read the address of every link in the page and recorded how many left the platform and how many carried anything a business could recognise. We did not click any of them, and the limits below say why that matters. One tool, one day, one connection, one sector. It is an illustration of link shape, not a measurement of anybody's traffic, and we name no business and no domain.

The first answer had twenty-five links in it, and four of them left the platform. The other twenty-one went back into the platform: further searches, a map card, profiles it holds itself. Of the four that left, three carried no parameter at all. One carried a single parameter.

The one parameter was not the platform's. It was a utm_source value, the standard tag for naming where a link was published. The business had evidently set it itself, on the website link inside its own business profile, pointing at that profile as the source. We are not reproducing the string, and we are not describing what else was in it, because either would narrow the business. The point is who put it there. The platform added nothing.

The second answer had forty-three links, twenty-three of them leaving the platform, across six distinct hosts. The same handful of sites was linked over and over, so a site that was one of the six carried a few of those links rather than all of them. Of those twenty-three outbound links, every single one carried nothing after the web address itself. No parameter of any kind. Had your site been one of the six, none of the links pointing at it would have carried a campaign source for your analytics to read.

Across the two answers, twenty-seven links left the platform and exactly one carried a tracking parameter, put there by the business itself.

That is the practical problem. On those twenty-seven links, as published in the answer, there was nothing in the address to say which surface the link came from. We did not click them, so we cannot tell you what a browser would finally send, including any referrer, and a platform is free to rewrite a link at click time. What we can say is that a business reading its own analytics can recognise only what actually arrives, and on this evidence there may be nothing there to recognise.

The contrast is documented rather than observed. OpenAI states in its publisher documentation that "ChatGPT automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs, enabling clear tracking and analysis of inbound traffic from ChatGPT search results", and that "Publishers who allow OAI-SearchBot to access their content can track referral traffic from ChatGPT using analytics platforms such as Google Analytics" (OpenAI Help Center). One platform documents a parameter. Another, on the evidence of two answers we counted, adds none. Perplexity's crawler documentation says its bot "is designed to surface and link websites in search results on Perplexity" and that when a user asks a question it "might visit a web page to help provide an accurate answer and include a link to the page in its response" (Perplexity), but names no referrer or parameter behaviour either way, so we make no claim about it.

One platform is not being generous and the other mean. What you can recognise differs by platform, is not fixed, and is not yours to control. Build your measurement so that it survives the difference.

What does your analytics actually show? Source, referral, direct and the leftover bucket

Your analytics shows four readings of where a visit came from, and each is worth something different: a session carrying a documented parameter, a referral session, Direct, and Unassigned. All four are built out of the same six words. Skim them, and come back if one of the readings below stops making sense.

• Session. One visit to your site, counted by your analytics from the moment somebody arrives.
• Source. The name of the place they came from, such as google or chatgpt.com.
• Medium. The kind of route it was, such as organic or referral. Where your analytics could not tell, it records the medium as none, which is a gap rather than a route.
• Channel group. The set of rules your analytics uses to sort every session into a bucket like Referral, Direct or Organic Search. The default channel group is the one Google supplies ready made.
• Referrer. The note a browser normally passes to your site saying which page the visitor came from. That page is allowed to switch it off.
• utm parameter. A tag somebody adds to the end of a link, like utm_source=chatgpt.com, so that your analytics can read where that link was published.

None of them needs memorising. What matters is that every one of them can arrive missing, and that your analytics cannot tell a missing one from a genuine one.

Sessions carrying a documented parameter. Where a platform adds something like utm_source, your analytics will read it. Google's documentation lists utm_source, utm_medium, utm_campaign, utm_content, utm_term and utm_id as the parameters it recognises, and states that "A session is processed as a custom campaign when custom campaign parameters are embedded in the document location field" (Google Analytics Help). This is the cleanest reading you will get, and it exists only where the platform chose to give it to you.

Referral sessions. Where the platform added nothing but the browser still passed on where the visitor came from, your analytics can record it as a referral. Google's default channel group puts a session in Referral when the "Medium is one of ("referral", "app", or "link")" (Google Analytics Help). This is the reading most people assume is complete. It is not, for the reason in the next paragraph.

Direct, which is where a lost signal ends up. Google's rule for Direct is that the "Source exactly matches "(direct)" AND Medium is one of ("(not set)", "(none)")" (same page). That is a statement about what analytics knew, not about what the person did. The web's own specification for referrers says the header "can be suppressed for links with the noreferrer link type", and that under the no-referrer policy "no referrer information is to be sent... The header will be omitted entirely" (W3C). When that happens there is nothing for your analytics to read, and the arrival lands in Direct alongside everybody who genuinely typed your address in. A rise in Direct is a question, never an answer.

Unassigned, the leftover bucket. Google describes it plainly, as "the value Analytics uses when there are no other channel rules that match the event data" (same page). Treat a rise in it exactly as you treat a rise in Direct.

And one reading that is not a click figure. Google states that sites appearing in AI features "are included in the overall search traffic in Search Console. In particular, they're reported on in the Performance report, within the "Web" search type" (Google Search Central). That search type does not separate AI features out, so there is no click number in it you can attribute to AI Overviews or AI Mode on their own. Impressions in those features are reported separately, in Search Console's generative AI performance report, which Google describes as showing "how your site performs in generative AI features on Google Search" and says "Not all properties have access to the report, as we're rolling out over time" (Search Console Help). Our article on Google AI search for UAE businesses covers what that report shows. This part of the product changes, so check the current documentation rather than treating any of it as settled.

How do you track enquiries from a form, a phone call and WhatsApp?

UAE businesses commonly take enquiries in four ways, by form, by phone, by WhatsApp and in person, and your analytics instruments only the form for you.

The form. Mark the submission as a key event. Google defines a key event as "an event that measures an action that's particularly important to the success of your business", and says that "Any event you collect can become a key event. To measure a key event, create or identify an event that measures the action and then mark the event as a key event" (Google Analytics Help). Add one field to the form: how did you come across us. Free text, not a dropdown, because a dropdown tells people what to say. Adding a field and keeping the answers is a decision about what you record about a customer. This article cannot make that decision for you, so take it to somebody qualified before you change the form. One more thing before you count on any of it. Marking an event as a key event is a switch in your analytics settings and takes a minute. Having an event that fires when your form is submitted is a different matter, and some website and form builders send one while others do not. Look in your events list first. If there is nothing there for a form submission, that is a job for whoever looks after your site rather than something to solve today, and the extra field and the log work either way.

The phone. Two different things, measured in two different places. A tel: link is the clickable phone number on your own website, the one that opens the phone app when somebody taps it. Counting taps on it is something you set up yourself, as a key event, and nothing does it for you. This is the step that usually needs somebody technical or a tag manager, so put it on a list for whoever looks after your site rather than losing today's hour to it. A call from your Google Business Profile is counted there instead: Google defines that metric as "The number of times a customer clicked on the call button on your Business Profile", alongside website clicks, "The number of times people clicked on the website link on your Business Profile", and directions, "How many people asked for directions to your business" (Google Business Profile Help). Note the shape of those definitions: they count a click on a control, not a customer, and Google states that "Performance data is available only for verified Business Profiles".

WhatsApp. A route plenty of businesses here rely on is also one of the least measured. A click on a chat link on your own site can be recorded as a key event in exactly the way a form submission can, and that is the part you control. It needs the same technical hand as the tel: link, so ask for both in one message. What happens inside the conversation is not visible to your analytics at all, so the log is the only place it gets recorded. You can start that today, without any of the click tracking: ask the question in the chat and write the answer down.

The walk-in, the forwarded message, the call to a mobile. None of these are measurable by any tool. They are measurable by asking, and writing down the answer.

What should you ask every enquiry? The one question that does more work than the tracking

Ask every enquiry how they came across you, and write down what they said, in their words.

The least technical thing in this article is also the most useful. It is the only signal that connects a person to a route, and it survives a suppressed referrer, a forwarded link, a screenshot, a conversation in a car. It also catches the sequence people actually follow, which is rarely one step.

Three rules for it. Ask it on every route, not just the form, because the routes that skip your website are exactly the ones your analytics cannot see. Record it as what they said, not as what you concluded, so that "my wife found you, I think she asked one of those AI things" goes in as that and not as a tidy category. And record when you did not ask. A log with the misses in it is honest; a log without them overstates every rate you calculate from it.

People misremember. They compress four steps into one and they name the last thing they touched. That is not a reason to skip the question. It is a reason to write the answer down as testimony rather than as data.

What cannot be attributed to AI search, and why?

Some things cannot be attributed to a source at all, and the whole of this measurement is only as good as your honesty about them: the reason somebody bought, an arrival that carried no signal, a platform that changed what it sends, a first contact older than the attribution window, and a figure the product withholds.

Two numbers moving together is not one causing the other. Somebody can ask an assistant on Sunday, read reviews on Monday, ask a friend on Wednesday and walk in on Saturday. Your log records the last recognisable thing. It does not record the reason, and no tool you can buy records the reason either. So record the sequence when somebody gives it to you, in their own words, in the column for what they said. Four steps written down as four steps is the nearest thing to a reason you will get, and you can report it, as testimony rather than as a measurement.

An arrival that carried nothing cannot be traced. Not partially, not approximately. If the referrer was omitted and no parameter was added, there is no information, and the arrival lands in Direct beside everybody else who lost their signal. So do not go looking for it, and never estimate your way into a share of Direct. Do the one thing that still reaches those people: ask at the enquiry. An arrival you cannot trace and a customer who tells you they asked an assistant are often the same person, and only the second version of them reaches your log.

A platform can change what it sends, and nothing obliges it to tell you. So record what you saw each month rather than assuming the method held still. A drop in a recognisable source can be a change at the platform rather than a change in your business.

Attribution only looks back so far. Google's own documentation on attribution settings says that "The key event lookback window determines how far back in time a touchpoint is eligible for attribution credit", with acquisition key events defaulting to thirty days and other key events to ninety (Google Analytics Help). A first contact older than the window is credited to nobody. The same page notes that "The first click, linear, time decay, and position-based attribution models are no longer available as of November 2023", which is worth knowing if somebody is proposing a model to you by name.

Some figures are withheld on purpose. Google says data thresholds "are applied to prevent anyone viewing a report or exploration from inferring the identity or sensitive information of individual users based on demographics, interests, or other signals present in the data", and that they "are system defined. You can't adjust them" (Google Analytics Help). Search Console says the same in its own way: "To protect user privacy, the Performance report doesn't show all data", and "Tables in the performance reports omit rare queries to protect user privacy" (Search Console Help). A blank is not a zero, and a small business hits these more often than a large one. Report a blank as a blank and name which reading it was. Then move that month's argument down to the enquiry log, where nothing is withheld because the figures are your own.

And nobody promises the outcome itself. OpenAI's help pages state that "Placement is not guaranteed" and that "Search results and citations can be incomplete, outdated, or incorrect" (OpenAI Help Center). Our article on whether you can guarantee a ChatGPT recommendation works through what that means for anything anybody sells you. Measurement does not change it.

The free measurement template published with this article

The measurement template comes with this article as a free download, as 17-ai-referral-measurement-template.xlsx. It opens in Excel, Numbers or Google Sheets. In Google Sheets, open it through File, Import and pick Replace spreadsheet, which brings the dropdowns through with it. Read the Start here tab, then go straight to the Enquiry log. That tab needs only your inbox. The Signals tab is the one that needs an hour with your analytics.

Measurement template developed by Lunasol Lunasol, which owns AI Visibility, built this template for the article and released it free, to use and to pass on. Six tabs: how to use it, the signals worth checking and what each is worth, an enquiry log, a monthly log that counts itself, a tab that tells you what to say when somebody asks what the figures prove, and the documentation behind every description in it. Fill the yellow cells only. The grey row at the top of the Enquiry log and of the Monthly log is an example, so leave it alone and start below it. The Signals tab has no example row, and its first row is a real signal to check.

The first ten minutes with it. Open the Enquiry log and start on row 6, the first row under the grey example. Reconstruct last month's enquiries from your inbox and your phone, one row each: the date in column A, how it arrived in column C, what they said when you asked in column D, whether there was any assistant signal in column E, whether it was qualified in column F, and what happened in column G. Columns C, E, F and G have dropdowns, and so does column H if you want to band enquiries by size. Pick from the list rather than typing. Column D is free text on purpose: write down what they actually said. Enter the date the way your spreadsheet writes dates, then check that column B has filled itself in as a year and a month, like 2026-08. If it says the date was not read as a date, reformat column A before you go on, because the monthly counts match on that column. Where you never asked how somebody found you, choose Not asked. The sheet counts that separately, and a log that pretends is worth nothing.

Then open the Monthly log, go to row 6, and type the month in column A exactly as the Enquiry log writes it, 2026-08 rather than August 2026. The count columns fill themselves from your log, and the last column on the row is the one to read next. It says either that the month is too thin to read as a rate, or that it found no enquiries for the month you typed. The second one usually means you have spelled the month differently here from the way the Enquiry log spells it. The two visit numbers beside the month are yours to type in from your analytics later, and the per 100 column stays blank until you do. That is ten minutes, and a count and a denominator you did not have this morning.

The Monthly log is built to be read several months at a time, and three months of rows is where a trend starts to show. Quarterly is also the rhythm of Lunasol's own published process, whose last step is named track and described as "Numbers, every quarter" (Lunasol).

What do you report to whoever asked, and how do you say it?

Report the count, the period, the reading it came from, and the thing you did not ask.

Put your own figures into a sentence of this shape. The ones below are invented, to show the shape: "Fourteen of the fifty-one enquiries we logged last month named an assistant when we asked, and we did not ask nine of them". That survives a sceptical reading. "AI brought us fourteen enquiries" does not, and the person you say it to will work that out eventually. All three figures come off one row of the Monthly log, including the count of the ones you did not ask.

Three habits that make the report credible rather than defensive. Give the denominator every time, because a number without one is a mood. Show the same five stages every month, in the same order, even in the months where the bottom three are zero. And put the limits in the report itself, in their own short paragraph, rather than waiting to be asked: a report that names its own weaknesses is read as careful, and a report that hides them is read as marketing the moment anybody finds one.

One more, on patience. A rate calculated on a handful of enquiries moves on a single enquiry: one qualified enquiry out of one that named an assistant reads as a hundred per cent, and it is still one enquiry. The last column of the Monthly log marks any month thin enough for that to happen and tells you to give the counts instead. Read it before you quote a percentage off that tab. Three months of a small number is a trend worth discussing. One month of it is noise with a decimal point on it.

Where to start this month

Start with the log, not with the analytics. The enquiries you are not writing down are gone the moment you forget them, so reconstruct the last month from your inbox, your phone and your calendar, one row each, and ask the question of every new one from today. Reading the analytics reports takes about an hour and will still be there next week. The click tracking on your phone and chat links may need somebody technical, which is a reason to ask for it now rather than a reason to wait. Start the log today regardless: it is the half that works without anybody's help.

If you would rather have the measurement set up with you, Lunasol offers a free AI visibility check as a starting point. Its published process and its contact routes, WhatsApp and email, are at Lunasol: there is no form, and the stated reply time is within the hour.

Common questions

Nine questions about measuring whether AI search brought you anything.

How do I know if AI search sent me traffic? Look for the parameters a platform documents, read your referral sources, and then ask every enquiry how they found you. On the two Google AI Mode answers we counted, reading the address of every link rather than clicking it, twenty-seven links left the platform and only one carried any parameter, which the business had put there itself. The technical route alone will miss arrivals. OpenAI documents that ChatGPT adds utm_source=chatgpt.com to referral URLs, so that one is readable in your analytics.

Can I count clicks from AI Overviews as a result? Not on their own. Google states that sites appearing in AI features are included in the overall search traffic in Search Console, reported in the Performance report within the Web search type. That search type does not separate them out, so no click figure there belongs to AI features alone. Impressions in those features are reported separately, and not every property has that report yet. Check the current documentation before relying on either.

How do I track ChatGPT traffic in Google Analytics? OpenAI states that "ChatGPT automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs, enabling clear tracking and analysis of inbound traffic from ChatGPT search results". Where that parameter arrives, your analytics reads it as a campaign source, so there is nothing to set up. Open your traffic acquisition report, look at session source, and chatgpt.com will be in the list if anything came that way. Your analytics reads only the visits that carry the parameter, and a visit that lost its signal lands in Direct instead.

Why did my direct traffic go up? It may mean nothing about your customers. Direct is where analytics puts an arrival when it had no source and no medium, which includes every arrival whose referrer was suppressed. Treat it as a question to investigate, not as a result.

What should I count as a conversion? Whatever you can define and will not change. A form submission is easy to count and easy to overvalue. A qualified enquiry is harder to count and worth far more, and it is the one stage in this article that is entirely yours to define.

How do I know where a WhatsApp enquiry came from? By asking in the chat and writing the answer down. A click on a chat link on your own site can be recorded as a key event, but nothing inside the conversation is visible to your analytics, so the log is the only place a WhatsApp enquiry gets a source at all.

Can I prove that AI search made me a sale? No. Nothing in this method establishes that one thing caused another, and no tool you can buy establishes it either. What you can show is a count with a date on it, the route the customer named when you asked, and the number of enquiries you did not ask.

How long before the numbers mean anything? Longer than one month. Read the five stages as a trend across months, and treat any month with fewer than ten enquiries as a number you do not yet have.

Do I need new tools to measure AI search? Nothing you have to pay for. A spreadsheet, the analytics and search reporting you already have, and the discipline to ask one question of every enquiry. Everything in the template runs on figures those products already give you. Counting taps on your phone and chat links is the one part that may need a tag manager and somebody technical. The log works without it.

What this measurement cannot do

A measurement plan cannot prove that AI search caused a sale. It cannot recover a visit that arrived with no signal on it. It cannot separate any clicks that AI features sent you from the rest of Google search. It cannot tell you what a platform will send next quarter.

What it can do is stop two failures that cost more than they look. The first is claiming a result you cannot support, which works until the first person checks. The second is the opposite, and more common: doing the work, receiving enquiries, and keeping no record of what those enquiries said about where they came from, so that when the budget conversation comes round the honest answer is that nobody knows.

Method, sources and notes

This article and the template published with it are general guidance on measuring your own marketing, current at September 2026. They are not legal, tax, accounting or data-protection advice, and nothing in them describes what any UAE authority requires or permits you to collect, store, ask for or record about a customer. Check your own obligations with somebody qualified before changing what you collect. Nothing here makes a business shown, named, cited or recommended in any AI product, and nobody can promise that it will.

Every description of an analytics, search or business-profile figure in this article and in the template repeats how that product's own published documentation describes it, as fetched on 21 September 2026. That documentation changes without notice, so check the current version before relying on it.

The live check was one tool, one day, one connection, one professional category, two questions and one run each. We counted the links as they were published in the rendered answer, by reading the address of every link in the page. We did not click any of them, so a platform that rewrites a link at click time would not show up in our counts, and the figures describe the links as published rather than the final address a browser would reach. It is an illustration of link shape, not a study, and nothing in it is a finding about any business. We name none of them, and we name no domain or host. The one tracking parameter we found is described by who set it, never reproduced, and its contents are not described, because either would narrow the business. We did not attempt a second assistant: this environment produced no usable answer from it across three earlier articles in this series, and that decision was taken before these runs rather than after them. The logged runs are available on request. Assistant answers vary by account, location and time.

Product names are used here for identification only. Neither this publication nor its owner is affiliated with, endorsed by or sponsored by any of them, nothing here is an endorsement by us of any of them, and none of them has reviewed, approved or contributed to this article. ChatGPT is a product of OpenAI, and the OpenAI Help Center is a publication of OpenAI. Google Search, Google Search Central, Google Analytics, Google Search Console, Google Business Profile, AI Mode and AI Overviews are products or publications of Google. Perplexity is a product of Perplexity AI. WhatsApp is a product of Meta. Excel is a product of Microsoft, Numbers of Apple, and Google Sheets of Google. The World Wide Web Consortium is an independent standards body. All are trademarks of their respective owners.

AI Visibility is owned by Lunasol.