AI visibility is difficult to measure because an answer can mention your brand, link to a page, use information that resembles your content, or send a referral visit. Those are different signals. If you put them in one metric, you will not know what actually happened.
Use a small, repeatable monitoring system instead: run the same questions in the same AI tools, save the evidence visible in each answer, and compare it with referral data. The goal is not to infer hidden model behaviour. It is to maintain an auditable record of citations and mentions that a practitioner can check again later.
Define what counts before you start
Create separate fields for citation, brand mention, and referral visit. A citation is the strongest observation in this workflow: your domain or a URL from your domain appears as a source link in the answer interface. A brand mention is text naming your company or site without a visible source link. A referral visit is a session arriving at your site from an AI-related referrer in analytics.
Do not promote a brand mention to a citation just because the answer sounds similar to your page. Search AI notes that a site may be used as a reference without a user clicking a link, which is precisely why a visible-link record and traffic record should remain separate Search Ai.
Set these rules in your spreadsheet before testing:
- Cited: a visible source link resolves to your domain or a subdomain you control.
- Mentioned: your brand, product, author, or article is named, but no source link to your domain is visible.
- Not observed: neither a qualifying source link nor a qualifying mention appears in the saved answer.
- Unverified: the answer contains a claim similar to your content, but there is no clear source link or name tying it to you.
This last category prevents a common reporting error: treating apparent content reuse as proof of citation.
Build a prompt set from pages you want cited
Start with 10 to 15 questions. ZDNET recommends asking 10–15 questions where you expect your content to appear, then noting who is cited and how often ZDNET. Keep the set small enough that you can run it carefully and consistently.
Choose prompts from real pages, not only broad category terms. Each question should have one intended landing page. For example, turn a service page into a buyer question, a guide into a how-to question, and a location page into a local-intent question. Include questions where competitors are plausible sources; an all-brand prompt set will overstate your visibility.
- Use the language a customer would use, including product, problem, industry, and location terms where relevant.
- Map every prompt to one target URL on your site.
- Label each prompt by intent: definition, comparison, how-to, troubleshooting, local service, or recommendation.
- Keep the wording frozen for a reporting period. If you rewrite it, record it as a new version.
- Remove prompts that do not have a clear business or content owner.
Search Ai similarly recommends manually querying several AI tools with prompts related to your business and then checking for domain, brand, or article mentions Search Ai. The important operational detail is to preserve the exact prompt text. Otherwise, a change in the answer may result from a different question rather than a change in visibility.
Run the manual citation test across multiple engines
For each prompt, test the same set of tools available to your team. The supplied sources identify ChatGPT, Perplexity, Gemini, Claude, Copilot, and Google’s AI-generated search experiences as places to check, but availability and source presentation can differ by tool and interface Search Ai. Do not assume that a citation in one engine will appear in another.
Perplexity is especially practical for manual review because it automatically lists numbered source citations, according to ESEOspace ESEOspace. For interfaces where sources are not obvious, save the full response and any visible source panel. You may ask an engine what sources it used, as ESEOspace suggests for ChatGPT, Claude, and Gemini, but classify the result as cited only when the interface provides an identifiable link or domain ESEOspace.
For every test, save the following fields in a sheet:
- Date and time of the test.
- Engine and interface used.
- Exact prompt text and prompt version.
- Whether your site was cited, mentioned, not observed, or unverified.
- The cited URL, if visible, plus the page title where available.
- The answer excerpt containing your brand or source.
- All competitor domains cited in the same answer.
- A screenshot or exported copy of the response.
Do not rely on memory or a simple yes/no score. The cited URL matters because it tells you whether the engine surfaced the page you intended to monitor or a different asset, such as an old article, documentation page, or homepage.
Calculate a simple, auditable baseline
After completing one run, calculate three measures from your log. First, count the prompts with at least one visible citation to your domain. Second, count prompts with a brand mention but no visible citation. Third, count citations by target URL and by engine.
Report the raw counts alongside the denominator: for example, “cited on 4 of 12 tested prompts in Engine A.” Do not turn this into a claim about your overall market share or all possible AI queries. It only describes the defined prompt set on the date tested.
Then review competitor domains that recur. This is not proof that their pages are universally preferred. It is a practical content-research list: inspect the cited pages, compare their format and topical coverage with your mapped target page, and record gaps you can verify on your own site.
Use analytics to measure visits, not citations
If Google Analytics is installed, review referral traffic from AI-related domains. ZDNET specifically suggests filtering referrals for sources such as chatgpt.com, perplexity.ai, and gemini.google.com over several months to see whether that referral presence changes ZDNET.
Keep this report separate from the prompt log. A referral session confirms that analytics received a visit with that referral information; it does not identify the exact prompt, prove a particular cited answer, or show that every citation generated a click. Conversely, a visible citation with no referral session is still a citation in your manual record.
Use a monthly analytics export with:
- Referral domain.
- Sessions and landing pages as reported by your analytics setup.
- Date range.
- Any annotation for publishing changes or major page updates.
This makes a useful cross-check. If a page is repeatedly cited in your test and also appears as an AI-related referral landing page, record both observations. Do not merge them into an unsupported causal claim.
Repeat on a fixed schedule and preserve the evidence
Run the complete prompt set monthly at minimum, or more often only if you can keep the process consistent. Paid monitoring products can automate multi-platform tracking, history, reporting, and competitor review, according to Search Ai Search Ai. Before buying one, use the manual log for at least two rounds. You will then know which prompts, engines, URLs, and evidence fields a tool must support.
At each review, compare only like with like: the same prompt version, engine, and classification rules. Flag a new citation, lost citation, changed cited URL, and newly recurring competitor. Retain screenshots and answer copies, because a later interface view may not reproduce the earlier result.
The output is deliberately modest but useful: a dated record of what was visibly cited, what was merely mentioned, and what referral visits appeared. That is enough to identify pages worth reviewing without claiming access to an AI engine’s internal selection process.