How to measure AI search visibility
Measure a defined set of answers under a stated counting rule. Keep the original records, show collection gaps and compare the same scope over time.
By Carqo
Updated
An answer rate needs its counting rule, sample and missing evidence beside it.
Which questions belong in the sample?
Start with the customer decision and relevant offer. Choose questions you can explain and review. Keep branded questions separate from questions that do not name the business.
Save the wording with a question-set version. If the offer changes, preserve the old set and date the revision. Do not overwrite the scope behind an earlier report.
Which collection conditions matter?
Record the provider, experience, location and available request settings. A consumer-app answer and a reply from a software interface are different observations. Do not combine them without explaining the comparison.
Note whether the conversation was fresh or included earlier turns. If a condition is unavailable, keep it unknown. A plausible default cannot replace a setting you did not observe.
What evidence should each attempt retain?
Keep a capture identifier, the exact prompt, time, full reply and source links. Record a failed attempt with its failure status rather than inventing an empty answer.
Where an export is incomplete, state which context is missing. The original record should remain available to an authorised reviewer. Public examples should omit unrelated personal or account information.
What does the counting rule include?
State what qualifies as a mention or citation before counting. Specify whether repeated names count once per answer. Record explicit ranked positions as a separate measure when they matter.
Keep an own-site citation apart from a link to another publisher. A reply may name the business without linking to its website. Those observations should not collapse into one positive flag.
How should missing captures be reported?
Show planned attempts and readable answers separately. A reply without a mention is an observed absence. A failed request is not. The report should preserve both distinctions.
Inspect which questions or surfaces are missing before interpreting a rate. Uneven gaps can change the sample. If there is no readable evidence, leave the answer rate unavailable.
When is a comparison meaningful?
Use equivalent question sets, methods and reporting windows where possible. Keep the filters and counting rules with an export. A chart should not hide changes to its own scope.
If conditions changed, identify the comparable group or report the break. Describe the observed difference precisely rather than claiming that every customer now sees the business more often.
How should you record an improvement task?
Write down the page, disputed detail and current factual reference before editing. Record who publishes the correction and when. A prepared draft is different from a live change.
Inspect the public result after publication. Keep that check alongside later answer observations without replacing earlier evidence. A correct page and a changed answer are different completion conditions.
Which outcomes need separate measurement?
Use authorised referral and customer-action records for outcome reporting. Their definitions and reporting windows may differ from the answer sample. Read the records alongside each other without summing unlike events.
Google says visits from AI Overviews are included in overall Search Console traffic under the Web search type.
Google also documents a Generative AI performance report. Read its scope and definitions separately from the overall Web report.
Without a connected measurement source, the outcome is not measured. It is not zero. A before-and-after difference does not prove causation. Keep that limit beside any reported change.
A worked example
A business repeats the same customer questions after correcting an outdated service page.
The page check confirms the correction. A later saved reply uses the current description, while another capture fails.
Report the page result, the readable reply and the coverage gap separately. Leave the cause of the changed answer open.
Common questions
Should failed attempts count as answers without a mention?
No. Record them as collection gaps. The mention measure should follow its stated rule for readable answers, with coverage reported separately.
Can a later answer prove a page change worked?
It can show what that later reply said. It does not establish the cause. Keep the publication record and the new answer as separate evidence.
Before reporting a change
- Fix the question set and counting rule.
- Keep comparable collection conditions.
- Separate missing attempts from readable answers.
- Preserve the publication and observation dates.
- Report customer outcomes from their own measurement source.