Measure AI-search visibility with four separate records: brand mentions, citations to your website, attributable visits and qualified enquiries. Use a repeatable question sample for answer observations, platform reporting where available, and analytics for visits. Do not combine these into a single unexplained “AI ranking”.
This guide provides an operational scorecard for a small business team. It is a reporting method we recommend, not a claim about Esperto’s own visibility or client results. For implementation, pair it with our ChatGPT checklist and Google AI Overviews guide.
Define the outcome before choosing a tracking tool
A team selling website development may care about buyers discovering a service. A publisher may care about readers finding an article. A software company may care about trial sign-ups. Choose the business outcome first so a rising mention count does not distract from a broken enquiry form.
- Mention: the answer names the business or product.
- Owned citation: the answer links to a URL on your website.
- Referral session: analytics attributes a visit to an AI source.
- Qualified enquiry: a submitted request matches your agreed business criteria.
Keep mentions linked to third-party pages separate from citations to your own domain. Also distinguish a general list of options from an explicit recommendation. They answer different reporting questions.
Build a small, repeatable prompt panel
Our suggested starting point is 15 questions: five discovery questions, five comparison questions and five implementation questions. This is a manageable research sample, not a statistical representation of all demand. Use language from actual sales conversations and support queries.
For a website agency, a discovery question might ask what kind of provider can connect a website to a CRM. A comparison question might ask when a website builder is sufficient. An implementation question might ask how to prevent duplicate payment confirmations. Add relevant market details naturally; do not fill every prompt with your own brand name.
Save the exact wording. Keep a separate branded group if you want to check whether the assistant identifies your company accurately. If you change a question, give it a new identifier so the comparison does not silently change.
Use this observation record
For every run, record the following fields in your tracking sheet:
- Observation date, prompt identifier and exact question.
- Platform, mode or model when visible, language and relevant country.
- Whether a usable answer was returned. Mark failures as failures, not as zero visibility.
- Brand mentioned: yes or no. Note whether it was a recommendation or just a reference.
- Owned URL cited: yes or no, with the exact URL.
- Other cited sources and a saved answer or screenshot.
- Any important limitation: different account context, a follow-up conversation or a changed product mode.
Use a fresh conversation for comparable discovery tests. Keep repeat runs separate rather than choosing only the response that mentions your brand. Record the conditions you can control and acknowledge those you cannot.
A worked example: calculate rates without overstating them
Hypothetical example: you run 15 prompts twice on one platform. Two runs fail, leaving 28 valid observations. Your brand appears in seven answers; five answers cite your own site.
- Observed mention rate: 7 ÷ 28 × 100 = 25%.
- Observed owned-citation rate: 5 ÷ 28 × 100 = 17.9%, rounded.
- Failed runs: 2, reported separately.
These figures describe that panel under those test conditions. They are not the percentage of all users who see your brand. If an answer cites three of your pages, count one cited answer in this rate and record three URLs in a separate URL total. State your denominator every time.
Add Google’s own visibility reporting
Search Console’s Generative AI performance report provides impression information for Google AI Overviews and AI Mode, with dimensions including pages, countries, dates and devices. The report can be absent when impressions are insufficient. Chart and table totals can differ because of aggregation. Use its documented definitions rather than treating these impressions as a citation rate from your prompt panel.
For a USA and UK campaign, review those countries separately. Use the same date ranges when comparing periods and note the reporting timezone. Keep ordinary search clicks alongside these observations; do not label all Google organic traffic as AI-generated traffic.
Connect visits to enquiries without inventing attribution
Review the source, medium and landing page recorded by your analytics system. Check whether recognised AI referrals reach relevant pages and whether those sessions complete a useful action. Attribution can be incomplete: apps, privacy choices, redirects and later return visits may prevent a clean source label.
Do not reclassify all direct traffic as ChatGPT traffic. If a customer says they found you through an AI assistant, keep that as self-reported attribution. It is useful evidence, but a different field from an analytics referral.
For forms, count a conversion after the server confirms a successful submission, not when someone clicks the button. Check the saved record and the admin notification. For privacy, do not send names, email addresses, phone numbers or the enquiry message as analytics event parameters.
A useful conversion event can identify the form type and landing-page path. Your team can then review whether the enquiry is relevant using the appropriate private business system. Our website development service includes planning the enquiry journey; our SEO service can help connect landing-page work with discovery goals.
Review the scorecard and choose one next action
If mentions rise but owned citations do not, inspect which sources are being cited and whether your own pages provide useful evidence. If citations rise but visits stay flat, consider whether the answer already meets the reader’s need; do not assume tracking is broken. If visits arrive but enquiries do not, inspect relevance, the offer, mobile usability and form delivery before publishing more articles.
If there is no change after a small number of observations, keep the sample consistent and review the page’s usefulness. One answer changing is not proof that your last edit caused it. Keep a dated implementation log and avoid making many unrelated changes at once when you want to learn what helped.
Frequently asked questions
Can one visibility score compare every AI platform?
Only if its scope and calculation are made clear. Different prompt panels, products and definitions can produce different scores. Preserve platform-level results alongside any summary.
How often should we check?
A weekly observation run and a monthly business review is a reasonable starting workflow for a small team. Adjust it to the volume of evidence and available resources, rather than checking one prompt every hour.
Are citations the same as leads?
No. Visibility, visits and enquiries are separate stages. Report them separately and improve the stage that is actually weak.
Make AI visibility a useful business report
Start with the AI-answers strategy guide, select a small group of important pages, and agree what a qualified enquiry means. Discuss a measurement and website improvement project with Esperto.
