AI visibility can be measured. The harder questions are what the measurement covers, what it leaves out, and what you should do with the result.
When I first wrote about this, my concern was the gap between knowing that AI systems surface your content and having enough detail to improve it. That remains the practical problem. A dashboard can show activity without explaining whether the right people are discovering you, whether the answer is accurate, or whether anyone becomes a customer.
A useful measurement process connects four distinct observations: mentions, citations, referrals and enquiries. It keeps the evidence behind each one.
| Measure | What it records | What it does not establish |
|---|---|---|
| Mention | An answer names your business or you personally. | That it links to your website, describes you correctly or recommends you. |
| Citation | An answer identifies a source, such as a page on your website. | That someone clicked it or that the answer endorses your business. |
| Referral | A visit arrives with evidence connecting it to an AI service. | Every earlier AI interaction that influenced the visitor. |
| Enquiry | Someone submits a genuine enquiry, books a conversation or contacts you. | That AI caused it, unless you have supporting attribution evidence — and even attribution has limits. |
A system can mention your business while citing someone else. It can cite your guide without mentioning your brand in the answer. A reader can see either, remember your name and visit later. That is why one “AI visibility” total cannot tell the whole story.
Checking whether an assistant knows your name is useful for testing identity. It is a different test from discovering whether the assistant introduces you to someone who has never heard of you. Create a small, fixed set of questions covering:
These are examples of test design, not claims that the questions have measurable search demand. Build your own set from customer conversations, enquiries and search data. Keep questions that name you separate from those that do not — otherwise, repeatedly asking about your own brand can make discovery appear stronger than it is.
Choose the services and markets relevant to your audience. Start a fresh conversation for each test, use consistent settings and avoid supplying your website or coaching the system towards you in discovery tests.
Record the exact question, date, service, visible model or mode, whether web search was enabled, language and relevant location settings. Record personalisation or memory settings where you can control them. Naming a city in a question is not proof that the test ran from that city.
For each response, retain:
Open the cited links. Check whether they support the claim attributed to them. A plausible-looking citation is not enough. Repeat on a declared schedule — weekly checks can provide a manageable starting point; additional repeats help reveal variability. Fresh conversations reduce carry-over but do not eliminate personalisation or randomness.
You can calculate a mention rate within your test set: responses mentioning the correct entity divided by successfully assessed responses. Do the same for responses citing your domain. Report failed runs separately and count each response once for these rates.
For example, 6 mentions in 20 assessed responses is 30% of that sample. It is not 30% of all AI searches or market share. Keep results separated by service and question type.
Manual checks show what happened in your tests. Platform reports describe activity within their own reporting scope. Use both, without treating them as interchangeable.
Open your verified property and review the Generative AI performance report for Search. Google documents impressions for URLs appearing in features such as AI Overviews and AI Mode, with page, country, device and date breakdowns. Export the available data for your chosen period.
Review the ordinary Search results report alongside it. Keep name-based searches separate from subject and service searches, and inspect which articles attract impressions and clicks. Do not label all Google traffic “AI traffic”: Google says the AI-feature data is also included in its overall performance reporting, so adding the two reports together would double-count overlapping activity.
For your verified website, review AI Performance. Examine citations, cited pages and grounding queries. Microsoft describes grounding queries as phrases used to retrieve supporting content; the reported sample is not a complete transcript of users’ questions. If Citation Share is available, read its definition carefully: it describes your portion of citations for a grounding query, not traffic share or a quality score. Coverage applies to Microsoft’s supported experiences, not every AI system.
OpenAI documents that ChatGPT search referral links include utm_source=chatgpt.com. That provides a useful signal for identifying arriving traffic. In GA4, open Traffic acquisition, select Session source / medium, and inspect the sources actually recorded. Look for identifiable AI-service sources, including chatgpt.com, without assuming every service uses the same source or medium. Check that redirects preserve tracking parameters and that analytics records a test visit correctly. This measures identifiable visits, not everyone who saw an answer — and a crawler request in your server logs is not a human referral or proof of a citation.
Decide what a meaningful outcome is before looking at the numbers. For a consultancy, that might be a successfully submitted contact form or a completed appointment booking. Record those actions and mark the appropriate events as key events in GA4. Test the setup so a failed submission or repeated button click does not inflate the result.
Keep an email-link click or booking-button click separate from a received email or completed booking. Then check your actual enquiries for relevance and quality. An optional “How did you hear about me?” question can add context — retain the person’s answer alongside analytics rather than replacing one with the other. Someone may first encounter you in an AI answer, then search Google for your name before contacting you. Describe that as a reported discovery path; avoid presenting the last recorded click as the whole journey.
My entity case study documents the work of distinguishing my identity from other people with the same name. To judge its practical value, I also need to report the website’s actual search activity. The Search Console export dated 29 September 2026 covers 30 August–26 September 2026:
| Observation | Recorded result |
|---|---|
| Site-wide Google Web search clicks | 30 |
| Site-wide impressions | 463 |
| “doug lord” | 15 clicks; 179 impressions; average position 2.23 |
| “douglas lord” | 8 clicks; 78 impressions; average position 4.06 |
| Entity case-study page | 3 clicks; 28 impressions; average position 6.11 |
The two disclosed name queries account for 23 of the 30 clicks. That supports a modest conclusion: people searching for my name are finding the site. It does not establish broad discovery for AI visibility services. Nor does this export identify which clicks, if any, originated in AI answers. The remaining seven clicks should not automatically be classified as non-branded: the query table does not disclose every query.
The site’s reported authority score has also moved from 31 to 48. That is a separate instrument reading. It is not a count of visitors, citations or customers, and the supplied data does not establish which changes caused that movement. This is an observational case study, not a controlled experiment. Content, links, indexing, search systems and demand can change together; a favourable result after an edit does not isolate that edit as the cause.
The next useful test is whether relevant subject queries reach more pages, whether repeat AI checks identify and cite the correct material, and whether identifiable visits lead to genuine enquiries. Each needs its own evidence.
| Finding | Useful next step |
|---|---|
| Answers confuse you with another entity | Correct visible identity information and structured relationships; check the sources behind the confusion. |
| Relevant questions surface other sources | Assess whether your page answers the question clearly and offers evidence worth citing. |
| A page is cited but receives few visits | Check the answer context and referral measurement before changing the page. The answer may satisfy the user without a click. |
| Relevant visitors arrive but do not enquire | Review page relevance, contact routes and measurement; small samples may not support a conclusion yet. |
| A score improves but traffic stays flat | Inspect the score’s inputs and keep monitoring actual discovery and business outcomes. |
Record the changes you make and their dates. Compare consistent, preferably non-overlapping periods, and note campaigns or other events that could affect the result. At low volumes, report counts alongside percentages and allow more time. A score can help organise investigation when its method and evidence are clear, but it cannot replace retained answers, platform data or customer outcomes.
The conviction behind the original article remains: measurement matters when it helps you decide what to fix. Keep the question, the observation and the conclusion connected — and make the conclusion no stronger than the evidence.
Platform documentation checked 29 September 2026. Interfaces and reporting scope can change.
Douglas Lord is the creator of the Periodic Table of Digital Authority™ and founder of Digital Dominator. His work examines digital authority, AI visibility and how businesses are understood across search and AI systems.