How to measure AI search visibility
An AI visibility audit asks two different questions: can systems access accurate information about your business, and what do selected AI products actually say when you test them? Keep those questions separate. A technically sound site may not be cited in a particular answer; a frequently mentioned business may still be described incorrectly.
The useful output is a record of evidence, prioritised corrections and a repeatable method. A single visibility score cannot show the full picture. The comparison with traditional SEO explains why prompts are not equivalent to a stable search position.
Define the sample before you run it
Choose the countries, languages and customer needs you want to understand. An English query from a buyer in Germany can differ from a Finnish query about the same service. Record those conditions rather than combining them into one Europe-wide percentage.
Create a fixed prompt set with three groups: questions about your brand, unbranded questions about a service and questions where location or industry changes the answer. Use neutral wording. “Which suppliers can manage a business WordPress website?” is more informative than asking why your company is the best.
Use fresh conversations. Record the product, visible model or mode, whether search was used and the test date. Repeat enough to see whether results fluctuate, and report the number of observations. Do not simulate a representative customer survey by merely changing a VPN location.
Twelve checks for a practical audit
Website access and structure
- Availability: do the important pages return usable content without errors?
- Crawl policy: are the intended search crawlers allowed, and does the firewall agree?
- Indexing signals: are noindex directives, canonicals and sitemap entries intentional?
- Content readability: can a visitor find the main answer without relying on images or inaccessible interactions?
- Structured data accuracy: does markup match the visible business information?
Treat llms.txt as a separate optional experiment if you use it. Its presence is not a pass mark for AI discovery, and its absence should not count as a technical failure.
Business identity and evidence
- Company facts: do the legal name, trading name, address and contacts agree?
- Offer accuracy: are service inclusions, exclusions and current prices clear?
- External consistency: do authentic business profiles match the website?
- Evidence quality: do material claims link to primary evidence, and are old documents marked or replaced appropriately?
A Wikipedia page is not a universal requirement. Review relevant external sources without encouraging promotional edits that do not meet a platform's own rules.
Observed answers and outcomes
- Mention rate: in how many recorded responses is the business named?
- Citation and accuracy: which responses link to your pages, and are their factual statements correct?
- Commercial follow-through: what identifiable visits, enquiries or sales follow, with the limits of attribution stated?
For example, if your business is mentioned in 4 of 20 recorded responses, report “4 of 20 responses in this test”. That is an illustrative calculation, not a customer result or an estimate of all users' experiences. Keep the prompt list attached so someone else can review what was tested.
Use a report that supports decisions
| Finding | Evidence to save | Next action |
|---|---|---|
| Wrong price in an answer | Answer, prompt, date and cited URL | Correct the old source and schedule a retest |
| Important page unavailable | URL, response and time | Ask the technical owner to repair it |
| Name confused with another business | Exact conflicting statements | Clarify identity and official profiles |
| Citation without qualified traffic | Cited URL and analytics context | Check whether the page meets buyer needs |
Give each correction an owner and a due date. Separate directly testable defects from hypotheses. “The page returns an error” is a defect; “more articles may increase mentions” is a hypothesis that needs a proportionate experiment.
Choose tools around the evidence
Search Console and other webmaster tools help with search discovery and indexing. A schema validator checks markup syntax; it does not certify the business facts. Browser and server checks help establish access. Manual AI tests can be sufficient for a small initial prompt set.
Crawler rules should follow the product's own documentation. For example, Perplexity documents its search crawler and user-initiated fetcher separately. Perplexity crawler documentation.
Paid monitoring can save collection time, but ask whether it records prompts, locations, modes and citations. A dashboard without an exportable method is difficult to verify.
Review at useful intervals
A quarterly review is a reasonable working rhythm, supplemented by checks after major changes. Retest corrected facts and retain previous results. Do not claim causation from two observations when the model or search product also changed.
Start with our AI visibility checklist. For the underlying site, WordPress maintenance provides a defined technical service. Confirm any separate AI measurement work in the proposal rather than assuming it is included in a hosting plan.