How to measure AI search visibility without overstating the results
Build a repeatable AI visibility log with fixed buyer questions, citation checks, accurate denominators, and a separate view of qualified inquiries.
Choose a buyer decision before choosing a metric
Start with questions a potential customer would ask while selecting a service, comparing alternatives, or checking an implementation constraint. A small consulting firm might choose twenty questions covering those decisions. Avoid filling the set with broad definitions just because they are easy to test. Being cited for a basic definition can be useful, but it answers a different business question from appearing in a vendor comparison.
Maintain a fixed baseline set and a separate discovery list. Put newly suggested questions in the discovery list until the next planned revision. Otherwise, changing the sample can make visibility appear to improve even when nothing changed. Record the intended country and language: US and Canadian results should remain separate when the service requirements differ.
Keep mentions, citations, and inquiries separate
A mention means the response names your brand. A citation means it links to a specific source. Accuracy describes whether the response gets your offer, limitations, and evidence right. An inquiry belongs in your analytics or CRM. These are related observations, but a mention does not establish that someone visited, and a visit does not establish a qualified opportunity.
Google includes traffic from its AI search features within overall Search Console reporting. A change in total search clicks therefore cannot be labeled an AI-only result. Bing’s AI Performance reporting covers citations across its supported experiences. Record the scope of each platform report alongside its numbers rather than presenting it as a census of every AI assistant.
Make the observation repeatable
Use the same question wording, language, search settings, and checking schedule. Start each question in a new conversation without mentioning the brand you hope to find. Save the full response, source URLs, date, system name, and the model label if the interface provides one. If the model version is unavailable, record that limitation rather than guessing.
Repeat important questions several times under comparable conditions. Preserve every response, including failures and contradictory answers. A brand appearing once in three attempts should not become a screenshot implying consistent visibility. If a citation disappears, check whether the interface changed before concluding that the source was dropped. Separate recurring patterns from observations that need another check.
Use a denominator the reader can inspect
Consider an illustrative sample of twenty fixed questions, each tested once. Six responses mention the brand; four link to its website. One of those four responses misstates the billing model. Mention incidence in this sample is 30%, and citation incidence is 20%. Neither figure measures market share, purchase intent, or how often real customers encounter the brand.
The following week produces eight mentions. Before calling that progress, inspect which questions changed and whether the responses remained accurate. A gain in introductory questions might matter less than losing a citation in a buying decision. Preserve the raw counts with the percentages so a stakeholder can see how small the sample actually is.
| Observation | Illustrative result | Next check |
|---|---|---|
| Brand mentioned | 6 / 20 = 30% | Which buyer questions? |
| Website cited | 4 / 20 = 20% | Which source URLs? |
| Inaccurate linked answer | 1 of 4 linked responses | Which claim needs correction? |
| Qualified inquiries | Tracked separately | Agreed CRM qualification criteria |
Turn evidence into a specific content change
When an answer misstates your pricing, inspect the public pricing page and any outdated documents that describe the offer. When a general overview is cited instead of your comparison guide, check whether the guide explains concrete selection criteria and limitations. Assign one page owner and one change to each recurring issue. Retest the original question after the update.
The site’s geo-check tool can help assess whether a supplied page is clear enough to cite. It does not observe every live AI response or verify that your brand already appears. Bring the actual inaccurate excerpt and the correct source into the review. That gives an editor a solvable task instead of an unexplained visibility score.
Report business signals without inventing attribution
Track assistant referral visits where source information is available, then assess inquiries using your normal CRM criteria. Some journeys lose referral information; others return through branded search later. A self-reported discovery question can add context, but it should remain a separate signal rather than being used to allocate all revenue retroactively.
Common mistakes include changing questions between checks, combining different systems into one unexplained percentage, and treating every mention as a recommendation. Close the report with a decision: repair an ambiguous claim, strengthen a useful source page, or collect another observation. When there is insufficient evidence to estimate business impact, leave the impact unpriced.
Put it into practice
Frequently asked questions
How often should a small team measure AI visibility?
A weekly check is a reasonable starting schedule, not a universal requirement. Match the frequency to your publishing pace and question set. Daily checks of a tiny sample may mainly reveal variation between responses.
Can this measure our share of voice across all AI systems?
No. A manual log covers the questions, settings, systems, and repetitions you tested. Platform dashboards also have defined coverage. State those boundaries next to any percentage.