Measure AI search visibility
If an assistant names your business today and leaves it out tomorrow, what changed? A repeatable set of buyer questions gives your team a way to describe the difference and decide what to investigate.
Key takeaway
Keep the question, engine, date, circumstances and cited page with every observation. A mention, a link and a qualified inquiry are different pieces of evidence.
Define the observation
You can measure what a defined set of questions returned, on a defined set of engines, at a defined time, under defined conditions. You cannot measure a single stable position the way a classic ranking report implies, because AI answers are generated, can change between sessions, and may differ by wording, location and signed-in state.
That limit shapes the whole program. The honest unit of measurement is an observation: one prompt, one engine, one session, one date, one captured answer with its visible sources. Everything else, including summaries and trends, is built from observations and should say how many were collected.
- Observed: whether the brand, a competitor or a third party was named in a captured answer.
- Observed: which URLs or publishers the engine showed as sources, where the engine displays them.
- Judged by a reviewer: whether the answer was accurate, current and appropriately cautious for the topic.
Build a question panel
A prompt panel is a fixed list of questions that stands in for real buyer research. Build it from sources you can defend: intake call notes, questions the intake team hears repeatedly, search-query data you already have permission to use, and the sales objections your team can describe. Do not fill it with prompts that merely flatter the brand.
Write each prompt the way a person would type or speak it, then keep the wording frozen. Changing wording between rounds makes comparisons meaningless. If you want to test phrasing variation, add labeled variants as new rows instead of editing old ones.
- Prompt text, frozen and versioned.
- Funnel stage: learn, compare, choose.
- Buyer role the prompt represents.
Capture buyer context
Geography matters when the service is local or licensed by jurisdiction. Law, healthcare and private ground transportation answers often depend on state, county or city. Test location in two ways: prompts that name the place in the text, and prompts asked without it. Record how the engine handled location, including whether it asked, assumed or ignored it.
Avoid placing real client matters, patient details or private account information in any prompt. Panels should describe generic situations.
Repeat and verify
Choose engines based on where your audience plausibly searches, and state the list in the report. For each engine, record the product mode used, whether the session was signed in, the interface language, the apparent location setting and the date. If an engine offers different modes, treat each mode as its own source of observations.
Run each prompt more than once. Repeat runs in separate sessions and, where practical, on separate days. The point is not to average away disagreement but to show it. If an answer names a provider in one run and omits it in the next, the honest finding is that presence is inconsistent for that prompt.
Every captured answer then gets a human read. The reviewer checks whether the answer is factually correct, whether it reflects current information, whether any claim about the brand is wrong or unsupported, and whether regulated topics are handled with appropriate care. Legal and healthcare answers need a subject-matter reviewer, not only a marketer.
- Store the full answer text and a screenshot or equivalent record, with timestamp.
- Record the cited or linked sources exactly as displayed.
- Mark each observation as accurate, partly accurate, inaccurate or unverifiable.
Trace each source
A mention means the brand or a competitor was named in the answer text. A citation means a page was shown as a source. They are different and should be logged in separate columns. An answer can mention a provider with no link, or link a page without naming the organization behind it.
For each cited URL, record the owner type: the brand's own site, a competitor, a directory, a news or trade publication, a professional body, a government or regulator page, a forum or a review platform. This is where source quality becomes useful. If answers on a compliance question mostly lean on regulator pages, the opportunity may be to make your own explanatory pages precise and clearly sourced, not to chase mentions elsewhere.
Resist converting these records into a single invented score. If a composite is wanted, define its inputs in writing, show the raw counts beside it and describe the question set and sample size beside it so a reader can judge what the score covers.
Turn gaps into actions
A coverage gap is a panel prompt where no page on your site is a reasonable answer. Map each prompt to the page that should address it. Where none exists, the action is a content or structure decision. Where one exists but is not indexable, thin, outdated or poorly structured, the action is technical or editorial. Google's SEO starter guide and structured data policies are the right reference points: structured data should match visible page content and follow Google's policies.
Use only signals you can identify. Analytics may record a referred session when the referrer is available. Google includes traffic from its AI features in the Web search reporting in Search Console. That aggregate does not identify which AI answer produced a given visit. Ask the receiving team to review fit for inquiries it can legitimately evaluate.
Reports should lead with decisions: which pages to create or revise, which facts to correct, which prompts to retire or add, and which findings need legal, clinical or compliance review before any change. Follow with the evidence table, then a limitations section that states sample size, engines, dates, location settings and what was not tested.
Worked example · sample records
Turn an observation into an owned action
What should the team do with a mention, citation or factual error?
These sample records change the action queue, not a visibility score. No engine results or business outcomes are reported.
Record the observation
- Observed sample
- The business name appears; no supporting URL is displayed.
- What it establishes
- A mention in this one recorded answer.
- What remains unknown
- Which source informed it and whether the same answer repeats.
Do not count a mention as a citation or an inquiry.
Save the answer and conditions; repeat the question under the same method.
Examples explain a review method. They do not report customer work, live measurements or a submitted inquiry.
Illustrative example
Private ground transportation
A hypothetical regional chauffeur and airport transfer company wants to know how AI answers describe booking, vetting and pricing for corporate travel. The company has no panel and only anecdotal reports from staff.
- Draft 30 to 40 prompts from intake notes, covering corporate travel managers, executive assistants and individual travelers, with and without the city named.
- Freeze wording, tag stage, role and location, and map each prompt to a target page or note that none exists.
- Run each prompt on the chosen engines in at least two separate sessions, saving full answers and displayed sources.
- Have an operations lead check each answer for accuracy on licensing, service area and booking process, marking unverifiable items.
- Classify cited sources by owner type and log mentions and citations separately.
- Produce a gap list, a correction list and a limitations note, and recommend next-round changes for human approval.
Prompt panel register: one row per observation with prompt ID, wording version, stage, buyer role, location setting, engine and mode, session and date, mentions, cited URLs with owner type, accuracy rating, reviewer, mapped page, gap status and recommended action.
Measurement program checklist
Find the gap in the answer.
Can a buyer tell what your service includes? Compare the question with the page and identify the specific detail that needs an answer.
Illustrative workflow: question, source, clearer answer.
Questions and answers
How many prompts should a panel include?
Enough to cover each buyer role, stage and location you care about, but few enough that every answer can be read by a person. HAI would propose sizing the panel to the review capacity available, then expanding in labeled rounds. There is no universal correct number.
Why run the same prompt more than once?
AI answers can differ between sessions, days and settings. Repeat runs show how consistent presence is for a prompt. A single capture is one observation, not a result, and the report should say how many runs each finding rests on.
Is a mention the same as a citation?
No. A mention is the brand named in the answer text. A citation is a page displayed as a source. Log them separately because they point to different actions: reputation and description versus page content and source selection.
Can analytics tell us how much traffic comes from AI answers?
Only partly. Some visits carry a recognizable referrer and some do not. Use analytics and Search Console as directional signals alongside the panel and inquiry notes, and avoid presenting them as a complete count.
Sources
- Google Search Central — AI features and your website
Eligibility for supporting links in Google AI Overviews and AI Mode, and how their performance is reported in Search Console.
