SEO and marketing

How to measure AI search visibility without mistaking citations for trust

Track sampled AI mentions and citations alongside answer accuracy, observed visits and qualified enquiries. A practical measurement plan for small businesses.

By Clairevue · · 5 min read

A search-answer panel links to a source page under a magnifying glass, with a separate enquiry envelope.
AI-generated illustration of checking search citations separately from enquiries; not actual search or analytics results.

Suppose a small consultancy’s AI visibility score rises after it publishes a new guide. More tracked answers mention the company, but enquiries haven’t changed. Should the business pay for another month of content work?

Before deciding, find out which questions produced those mentions, whether the answers describe the service correctly, and what happened when people reached the website.

Marketers often call efforts to improve visibility in AI answers generative engine optimization, or GEO. For this consultancy, the goal is qualified enquiries about work it can deliver.

Define the sample before reading the score

OpenSEO’s AI visibility page describes tracking brand mentions and cited pages by prompt, engine and market. Its own FAQ says the result doesn’t measure every AI conversation or traffic to your site.

Choose questions a plausible buyer would ask before knowing your company’s name. A local consultancy might track questions about reducing invoice-processing work or choosing an implementation partner in its area. Keep explicitly branded questions separate; asking what your company does tests a different problem from asking which supplier to hire.

For an illustrative check, if five of 20 collected answers mention the business, the sampled-answer mention rate is 25%. That describes this set of answers, not a quarter of the market. Adding easier, brand-specific questions next month would change what the rate measures.

Keep the same questions for comparisons, along with the market and language settings. Record when each check ran, the product or model used, and whether web search was active. Keep refusals and answers without search visible; report failed checks separately rather than quietly changing the denominator.

OpenSEO’s Prompt Explorer describes comparing API model responses and their available citations. An API answer is a test under that configuration; don’t assume it reproduces what a customer sees in a consumer app with a different context or settings.

Record mentions and website citations in separate fields. An answer can recommend a company while linking to a directory, or cite its technical guide without recommending the company’s services.

Check what the citation actually supports

In his October 7 account of StateGlobe, Metehan Yeşilyurt reports filling a site with invented statistics that subsequently attracted links and AI-referred visits.

It’s a self-reported experiment by a researcher working at Peec AI, a visibility vendor; we haven’t audited its traffic or independently checked the reported backlinks. For business content, publish only figures whose underlying evidence you can verify.

The Tow Center’s March 2025 study asked AI search tools to identify news articles from excerpts. Researchers found fabricated URLs and links to copied or syndicated versions instead of original sources. That was a historical news-identification task, with each excerpt queried once; it doesn’t establish a current error rate for business recommendations.

For your own sample, open the cited page and check the claim attached to it. If an answer says your consultancy serves a region you don’t cover, count that as a description error even if it includes a link to your homepage. A citation to an old pricing page deserves a different response from a correct recommendation that sends a buyer to a current service page.

Save the answer and the destination URL with the date. Review the underlying evidence before turning a mention into a testimonial or repeating a statistic in your next article.

Keep visits and qualified enquiries separate

A tracked answer appearing on a dashboard doesn’t show that anyone clicked. Use website analytics to inspect observed visits and landing pages, then check whether those visitors made relevant enquiries.

Google’s current Analytics channel definitions include an AI Assistant channel for sources such as ChatGPT and Gemini. Google AI Overviews and AI Mode sit within Organic Search instead. Check source details as well as the headline channel; don’t label all organic traffic as AI traffic or assume every AI-related visit falls into one bucket.

Google Search Central also says AI Overviews and AI Mode traffic is included in Search Console’s Web performance data. That aggregate isn’t a separate count of AI-driven visitors.

For the hypothetical consultancy, decide what a qualified enquiry means before evaluating the content work: a business with a relevant process problem that the team can realistically help solve. A newsletter signup and a request for an implementation proposal are different outcomes.

Keep enquiry details in the appropriate customer records, with consent and access controls. Don’t put email addresses or private customer problems into analytics events or tracking URLs. An optional “How did you hear about us?” question can capture an AI-assisted discovery story, but report that self-reported assistance separately from an observed website source. Neither proves the new guide caused a sale.

Run a bounded test before expanding the spend

Check access first. Google says pages need indexing and snippet eligibility to appear as supporting links in its AI features, with no additional special schema requirement. Verify that public service pages contain readable, accurate information and that crawler restrictions match your intended policy; keep private pages protected.

Bot requests are useful for diagnosing access, but they aren’t website visitors or enquiries. OpenAI distinguishes its search crawler from its training crawler and from user-triggered page visits. A request in a server log doesn’t establish that the answer cited your page.

Set a time and spending limit for one content change, such as correcting a service page that obscures what the consultancy does and where it works. Save the initial answer sample, repeat the checks on a stated schedule, and review description accuracy alongside attributable visits and qualified enquiries.

A before-and-after increase can coincide with platform changes or competitors’ activity, so keep the result observational. Before approving more content work, ask which relevant enquiries it appears to have helped generate and whether those opportunities justify the time spent.