Track whether AI search sends you customers by combining website referral data, a checked enquiry or booking event, and a simple question asking customers how they found you. Match these signals to unique enquiry records and completed sales. Report recorded AI referrals separately from customer-reported discovery, because neither gives a complete picture alone.
Seeing your business in an AI answer proves only that it appeared in that answer. A website visit is stronger evidence of interest, but it is still not a customer. The useful chain is appearance, visit, enquiry, suitable prospect and sale, with clear definitions at each step.
Define a customer before opening the traffic report
Choose the outcome that matters for this business. For a tour operator, it might be a confirmed booking. For a letting agency, it might be a signed management agreement. For a property maintenance firm, it might be an accepted job that is later completed and paid.
Keep the stages separate. An enquiry is a request for information. A qualified enquiry fits your services and basic requirements. A customer has reached the agreed buying stage. You may also need a later status for completed work, cancellation or refund.
Decide how value will be reported. Booking value, money collected and profit are different numbers. A deposit is not the full value of a stay. A signed agreement's possible future income is not revenue already earned.
Write one short measurement rule. An illustrative tour operator might use: “Count each enquiry once. Record whether it is suitable, whether it becomes a confirmed booking and the booking value. Update cancellations separately. Keep the date of first enquiry so later bookings remain attached to the correct group.”
This avoids changing the definition when the numbers look disappointing. It also lets the team compare AI search with other discovery routes using the same outcome.
Find recorded AI sources in Google Analytics
In Google Analytics 4, use the Traffic acquisition report and inspect Session source / medium. The session prefix refers to a visit's acquisition information. User acquisition uses first-user information, which answers a different question. Do not compare the two as though they were identical counts.
Choose a complete period, such as the previous calendar month. Search the source values for AI services that actually appear in your report. Keep the original source and medium in an export or worksheet, then add your own reporting group beside them.
An illustrative worksheet might contain chatgpt.com / referral or perplexity.ai / referral. These are examples of values to investigate, not a promise that every visit from those services will carry those exact labels. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to the links it shows, so those visits usually record chatgpt.com as the source, but the medium beside it can vary, and a link copied out of a chat and pasted elsewhere loses the trail. If the report shows no ChatGPT visits at all, ask your website maintainer whether robots.txt blocks OAI-SearchBot, the crawler that keeps a site eligible for ChatGPT search answers, before concluding that nobody finds you there. Use what your report records, and confirm unfamiliar sources before grouping them.
Do not restrict the investigation to one medium without checking your data. Also avoid matching every source containing the letters “ai”. That could collect unrelated websites. Start with an explicit list of verified source values and review new ones monthly.
- Record the date range and the source dimension used.
- Keep the unedited source and medium values.
- Add sessions and your selected enquiry or booking key event.
- Inspect the pages receiving those visits.
- Separate internal tests before interpreting the small totals.
Google Analytics describes direct traffic as traffic without a clear referral source. Some discovery will therefore remain unidentified. A rise in direct visits is not evidence that AI recommendations caused it. Leave unknown sources unknown unless you have another supporting signal.
For an illustrative maintenance firm, five of 12 apparent AI visits occur while the owner is testing prompts and opening the business's links. Note those test visits and improve the measurement process. Do not present all 12 as prospective customers simply because the source name looks promising.
Check that an enquiry event represents a real enquiry
A key event in Google Analytics is an interaction you mark as important. For this task, use an event that represents successful enquiry submission or a confirmed purchase, rather than a page view or button click that may never lead to contact.
Google's key-event setup tutorial describes creating a separate event from a confirmation-page view, using a name such as generate_lead, then marking it as a key event. This can suit a form with a dedicated success page. The setup must match how your own form actually works.
For that pattern, the current documented route is Admin, Events, Create event, then Create without code. The source event is page_view, matched to the confirmation URL. Have the person responsible for analytics check the condition and verify the resulting event in the Realtime report's key events card. Google says the new configuration can take from a few minutes to a few hours to apply, so an empty card straight after saving is not yet a fault.
Do not apply that recipe to a form that displays its success message without changing pages. Ask your website maintainer to measure the actual successful submission. Use the conversion tracking setup tutorial for the broader checks, even if you are measuring unpaid traffic.
Run three tests: an invalid submission should not count, a successful submission should count, and a refresh should not create a second business lead in your customer records. Analytics event counts and unique leads can differ, so retain the business record as the source for deduplication.
An illustrative estate agency discovers ten recorded enquiry events but only seven enquiry records. Two came from repeated visits to the confirmation page and one was a staff test. The correct report is seven recorded enquiries after reconciliation, with a task to improve event measurement. It is not ten new prospects.
If the contact journey finishes by phone, measure what you can honestly establish. A click on a phone link does not prove a conversation or booking occurred. The staff member handling the enquiry should record the outcome in the normal enquiry log.
Add one discovery question to the enquiry record
Ask “How did you first hear about us?” Keep it optional and brief. Suitable choices include search engine, AI assistant, recommendation, social media, returning customer and another route, with space for a short explanation. Ask without steering the person towards the answer you hope to see.
When someone chooses AI assistant, a useful follow-up is “Which one, if you remember?” You do not need the full conversation. Record their answer as customer-reported discovery, alongside the website source if one exists.
An illustrative guest house enquiry arrives by phone. The guest says, “I asked an AI assistant about places with an early breakfast, then searched for your name.” Record AI discovery as reported by the guest. Do not invent an AI referral session or replace a separately observed search-engine source.
A filled-in, privacy-conscious record might contain an internal enquiry reference, first enquiry date, requested service, recorded website source, reported discovery, qualification status, booking reference and value. Keep customer contact details in the business's approved customer system, not in the AI analysis sheet.
Illustrative reporting row
Enquiry reference: E104
First enquiry date: 6 October
Service: private walking tour
Recorded source: search engine
Reported discovery: AI assistant; service not remembered
Qualified: yes
Outcome: confirmed booking
Booking reference: B028
Booking value: $300
Evidence label: customer-reported AI discovery onlyThat row is more useful than forcing one “true” source. It preserves both the measured visit and what the customer remembers. A person can discover you in one place and arrive through another.
Connect source evidence to records before calculating sales
Do not assume that a website traffic report automatically places its source beside each enquiry in your customer system. That connection needs to exist and be tested. Ask whoever maintains the form or booking process what source information is currently saved, how it is collected and whether it survives the whole journey.
If the system supports it, have the maintainer save permitted source evidence alongside the internal enquiry reference. The requirement might include the entry page, the referring service when available, campaign values and the collection date. Keep the method documented; raw referring-page information is not automatically identical to Google Analytics attribution.
Use a labelled test to follow the handover. In an illustrative estate agency, a test visit reaches the appraisal page from a known source, submits the form and creates enquiry T001. The reviewer checks that T001 retains the expected source evidence, then changes its status to a test booking and confirms that the evidence remains attached.
A second test moves through two pages before submission. A third arrives without source information. The first two should preserve whatever your approved collection design intends to retain; the third should remain unknown. Do not make the form guess a source merely to avoid an empty field.
If this connection is unavailable, report aggregate AI-source visits and events separately from customer-reported bookings. You can still learn from both. Avoid matching an enquiry to a visit solely because their timestamps are close, and do not claim an observed lead-to-sale chain that your records cannot support.
Only calculate a ratio using analytics sessions and customer-record leads when their source rules, dates and coverage are compatible. If consent choices or collection differences make that uncertain, show the two counts separately with a note. A missing percentage is better than a precise-looking comparison of different populations.
A tour operator reconciles visits, leads and bookings
The following numbers are illustrative. A tour operator reviews one month's new enquiries after allowing 60 days for bookings to develop. Its website report contains 80 sessions from its verified AI-source group. Eight unique enquiries are linked through the tested enquiry process to that group.
Five of those eight enquiries fit the service, and three become confirmed bookings worth $300 each. The recorded group therefore has eight enquiries, five qualified enquiries and three bookings with a combined value of $900.
The enquiry log also contains six people who report discovering the operator through AI. Four are already among the eight recorded enquiries. Two are additional enquiries without a recorded AI source. One of those two becomes a $300 booking.
| Evidence group | Unique enquiries | Confirmed bookings | Booking value |
|---|---|---|---|
| Recorded AI-source group | 8 | 3 | $900 |
| Additional customer-reported AI discovery | 2 | 1 | $300 |
| Combined, after removing overlap | 10 | 4 | $1,200 |
The enquiry calculation is 8 plus 6 minus 4 overlapping records, giving 10. Do not report 14 enquiries. The booking total is four distinct bookings, not the sum of every source label attached to them.
Eight unique enquiries from 80 recorded sessions gives an illustrative lead-per-session ratio of 10%, provided the enquiry links and period are aligned. It is not necessarily the same as the analytics session key-event rate. A person can have several sessions, and an event can fire more than once.
The three bookings from eight recorded enquiries give a 37.5% enquiry-to-booking rate for that group at the review date. With only eight enquiries, a single booking changes the rate substantially. Report the counts beside the percentage and avoid ranking channels on tiny differences.
The $1,200 is booking value associated with the combined evidence, not proof of incremental revenue caused by AI search. Some customers might have found the operator another way. Deduct cancellations when they occur, and use contribution after delivery costs if you want to assess whether extra marketing work is worthwhile.
Budget an illustrative two hours for the first report and reconciliation, plus 30 minutes a month thereafter. At an assumed internal time value of $30 an hour, that is $60 initially and $15 a month. Analytics repairs, booking-system changes and supplier fees need a separate quote.
Handle booking systems and slower sales cycles separately
A campsite may send visitors to a booking system on a different domain. Google's cross-domain measurement guidance explains how a suitable setup can preserve measurement across that journey. It requires compatible tagging and configuration; do not assume your booking provider supports the arrangement you need.
In an illustrative failure, the campsite's booking report credits the booking-system domain with most sales. The owner concludes that AI visits never book. First ask the supplier to test whether the original source survives the journey. A broken handover can make the channel look weaker than it is.
Where a booking provider cannot support the required tracking, use an honest partial measure: recorded visits to the booking route, plus confirmed bookings with customer-reported discovery. Do not treat an outbound booking click as a completed booking or fill the missing link with an estimated conversion.
A letting agency has a different problem. An owner enquires in October but signs a management agreement in December. A report comparing October visits only with October agreements misses that outcome. Keep the first-enquiry group together and update its status at a stated later date.
For example, report “October enquiries, outcomes checked after 60 days”, and use the same window for comparison groups. Keep deals still in progress visible. Do not declare them lost because they have not closed by the first reporting date.
Ask your data-protection adviser when the tracking design involves new identifiers, customer-record matching or unclear consent requirements. Use the minimum data needed and keep access restricted to the people maintaining the report. A simple source question may be sufficient for a very small business.
Use Google's AI visibility report for the question it answers
Google announced separate Search Generative AI performance reports in Search Console on 3 June 2026, and a note on the announcement says they reached all websites by 31 August. They cover AI Overviews and AI Mode in Search, plus generative AI features in Discover. The launch post lists impressions as the metric, with breakdowns by page, country, device and date. It does not list clicks, and it says this visibility remains included in the overall Performance report as well.
This makes the old blanket advice that Search Console has no separate AI visibility reporting outdated. Check the official report announcement and the metrics available in your property. Do not assume an impressions view identifies individual enquiries, customers or their full discovery journey.
Keep those impressions in a visibility section of your report. Do not add them to website sessions or treat all traffic recorded as Google organic as AI traffic. The measures answer different questions and can overlap.
For an illustrative maintenance firm, 240 AI-feature impressions and 35 total search visits cannot be combined into “275 AI prospects”. The impressions describe visibility, while the visits are a different measure covering the recorded search traffic. Keep both labels intact and use enquiry evidence for customer claims.
Improving the sources that AI services can use is a separate task. Follow the AI recommendation tutorial when the measurement points to inaccurate or incomplete public information.
Let AI draft the monthly interpretation, then challenge it
Once the records are reconciled, AI can help turn the aggregate numbers into a short management note. Give it counts, definitions and known limits. Keep names, contact details and individual conversation histories out of the prompt.
Write a 120-word management note from these illustrative figures.
Recorded AI-source sessions: 80.
Unique enquiries linked to that source group: 8.
Qualified enquiries in that group: 5.
Confirmed bookings in that group: 3, total value $900.
Customer-reported AI discovery: 6 enquiries; 4 overlap with the 8.
The 2 additional enquiries produced 1 further booking worth $300.
Outcomes checked 60 days after the enquiry month.
Do not claim causation or estimate missing traffic.
Show the deduplicated total and explain one measurement limitation.An illustrative weak response says, “AI generated $1,200 and converted 50% of leads.” Both claims need correction. The evidence associates four bookings with ten distinct enquiries, which is 40% at that review date. It does not prove AI generated all the value.
A better note is: “Ten distinct enquiries had recorded or customer-reported AI discovery evidence after removing four overlaps. Four became confirmed bookings worth $1,200. Three bookings were in the recorded source group. The figures are small and do not establish how many customers would otherwise have found the business.”
Check the maths yourself. Then decide on a modest action: improve a frequently visited page, repair a measurement gap or continue collecting data. Use the marketing measurement tutorial to compare that decision with the rest of your marketing activity.
Keep the original report and the assumptions beside the note. Your next monthly review should explain whether the evidence improved, whether enquiries were suitable and whether the business earned useful work. An attractive chart of AI mentions cannot answer those questions on its own.
Questions about measuring a small AI search channel
Do I need a paid AI visibility tracker?
Start with your existing website reports and enquiry records. A paid tracker may help if regularly checking many questions is becoming a substantial job, but assess its coverage, method and export options before buying. It still needs to be compared with your actual enquiries and bookings; a mention count does not establish customer value.
Should I ask customers to share their AI conversation?
Usually a short answer about where they found you is enough. Do not require a screenshot or full conversation, which could contain unrelated personal information. If a customer volunteers a specific incorrect recommendation, record only the relevant business claim and handle any supplied material through your normal privacy process.
Can I combine different services in one customer-value figure?
You can show a business-wide total, but keep service-level figures underneath it. A single maintenance visit and an ongoing management agreement have different costs, timings and values. State the period and value definition, and avoid comparing one channel's confirmed revenue with another channel's speculative future contract value.
Further reads
- How to Audit Your Website's SEO With AI in One Afternoon — Check pages receiving search attention before changing your content.
- AI KPIs for Small Businesses: 12 Metrics Worth Tracking — Choose a wider set of useful business measures.
- How to Choose an SEO Agency That Understands AI Search — Assess a supplier's measurement promises and reporting method.
- How to Write Google Business Profile Posts With AI Every Week — Keep profile-post visits identifiable alongside other search activity.
- How Independent Hotels Win More Direct Bookings With AI — Five places direct bookings leak to online travel agents, what AI can fix at each, and the commission sum that tells you what it's worth.
- How to Get Your Products Recommended in ChatGPT Shopping — How ChatGPT picks products in 2026, how your data reaches it from Shopify, Etsy or your own site, and the fixes that make a recommendation more likely.
- How Wedding Businesses Get Found When Couples Ask AI for Suppliers — How AI assistants build supplier shortlists for couples, a test to see what they say about you, and the five fixes that get a wedding business named.
- How to Write Website Articles With AI That Still Rank on Google — A step-by-step method for AI-assisted website articles that rank: evidence packs, outline and draft prompts, the edit pass, and what to check in Search Console.
- How to Do Keyword Research With AI for a Small Business Website — Brainstorm and group search phrases with AI, check them against Google's free data, and build a keyword map for each page, with a bookshop worked through.
- What AI Can and Can't Do for Your Local SEO — Which local SEO jobs AI does well, which it can't touch, and an afternoon's worked example for a dry cleaner, mapped against Google's own ranking factors.
- How AI Overviews Are Changing Local Search for Small Businesses — How Google's AI answers now sit around the map pack, what they read about your business, and a 12-search visibility check you can run in an hour.
- What Is Generative Engine Optimisation and Does Your Business Need It? — Where the GEO idea came from, how AI answer engines choose sources, a decision table by business type, and five checks that cover most of it.
- What to Do When ChatGPT Gets Facts About Your Business Wrong — How to find and fix the listings, pages and directories behind a wrong ChatGPT answer about your hours, prices or services, and check it stays fixed.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Analytics Traffic acquisition report, traffic-source scopes, direct traffic, key-event setup and cross-domain measurement guidance; Google Search Central announcement of Search Generative AI performance reports (3 June 2026, rollout note 31 August 2026); OpenAI help page, Publishers and Developers FAQ. Checked 28 September 2026.