An AI visibility audit should show how your business appears in relevant answers and leave you with a supported next action. Begin with buyer questions whose origin you can explain. Record the product, date, market and collection method, then retain each answer and its visible sources.
Check what the linked pages actually support. Record mentions, recommendations, citations and factual accuracy separately. Put the observation, your interpretation and the proposed fix in different fields, with an owner and a way to check the work.
Repeat the same documented questions after changes. A score alone cannot explain an absence, estimate lost leads or prove that a page edit caused a better answer.
Decide what the audit is for
A consultancy may want to know whether it is represented accurately for a particular implementation project. An equipment supplier may want to inspect recommendations for a product application.
Write down the buyer, decision, market and eligible provider type. Decide which AI product you will observe. If you collect through an API, call it an API observation; do not silently present it as what someone saw in the consumer app.
Also decide what this audit cannot answer. A researcher entering questions does not establish how often buyers ask them or how many leads a company loses by being absent.
Choose and label the questions
Use enquiries, sales calls and customer interviews where you have permission. Add constructed questions where coverage is missing, but label their origin.
Keep branded accuracy questions separate from unbranded recommendations. For a shortlist, define which providers genuinely qualify. A software platform and a consulting firm should not automatically count as comparable rivals.
Freeze the initial question wording and keep a version history. New questions can be explored separately. Otherwise a changing question set can make a score look better without any change in the business's representation.
If the original records are missing, keep that gap visible. A score cannot replace the answer or establish the cause of an absence.
Save the answer and conditions
For each attempt, retain the exact prompt, date, product, collection route, market wording, search state where observable and full response. Record missing outputs and errors. Repeated attempts are observations, not independent survey respondents.
If personal context is enabled, record it. If the setup is unknown, say so. A useful audit can still have limits, provided a reader can see them.
A mention, a positive recommendation and a source citation need separate fields. Add an accuracy note when the answer describes the business. A company can be named and still be misrepresented.
Inspect the sources against the claims
Open the links and identify what each supports. Does the page establish the service being recommended, or only discuss a broad category? Is the statement current? Is it a company claim or independent evidence?
Use simple support labels: supported, partly supported, contradicted or unable to verify. “Unable to verify” does not mean false. It tells you what evidence is still needed.
For specialist facts, use an appropriate reviewer. An equipment configuration or Salesforce credential should not be judged solely from a marketing summary.
A filled example from a saved report
The following uses the New York–based Salesforce consultancy report dated 5 September 2026. Its collection was labeled ChatGPT, India, with questions asked once. It is a report review rather than a reproduced answer-level study; original prompts and responses were unavailable.
| Audit field | Recorded finding |
|---|---|
| Task | Inspect the firm's representation in reported CRM category answers |
| Observation | The report displays absence in five category questions |
| Comparability issue | Several first results are software platforms rather than consulting firms |
| Evidence limit | Category labels are visible, but exact underlying prompt records are missing |
| First action | Check provider intent and map verified project capabilities to buyer questions |
| Owner | Marketing prepares the map; a consultancy specialist checks capability scope |
| Verification | Repeat documented provider questions and retain answers plus source support |
That is enough to justify a focused investigation. It is not enough to assert a missing page caused the absence, or that the firm lost a particular volume of leads.
Prioritize the fix by buyer consequence
Put the observation, interpretation and action in different columns. “Not named in this answer” is an observation. “The page does not explain supported migrations” requires a page review. “Write a migration page” is a recommendation.
Fix inaccurate facts and broken access where evidence supports them. Improve relevant content when the buyer cannot establish fit. Avoid commissioning a new page because a dashboard row is red.
For an equipment business, one action could be to clarify operating conditions on a product page. For a consultancy, it could be to attach a permissioned case to an existing service page. Neither should be described as completed until it is implemented and checked.
A missing business name is an observation. The reason for that absence needs separate evidence.
Build the worksheet your team will use
Each row needs a question ID, origin, market, product, date, answer record, mention, recommendation, citation, checked claim, support label, proposed action and owner. Add a link to the page version so a later review can identify what changed.
Keep the report readable. Summarize the important decision in plain language, then make the detailed rows inspectable. A buyer should not need to understand a proprietary score to see the evidence.
If you use a percentage, show the numerator and denominator. Divide the recommendation count by the number of completed eligible answers in the same set. That set-specific result does not estimate market-wide share.
Repeat the test and review the work
Retain the same initial questions and conditions for later checks. Show each round and explain exclusions. Keep the change log beside the results.
A better answer can justify further investigation. Qualified enquiries and won work require separate records. Do not rename answer movement as return on investment.
Can one run be useful?
Yes, as an exploratory diagnostic. It gives you examples to inspect. Label it clearly and do not describe it as stable performance across a market.
Do I need to monitor every AI platform?
Start with the product experiences relevant to your buyers and the evidence you can review properly. Expanding the list without reviewing the answers produces more data, not necessarily a better decision.
Request an audit if you want help collecting and reviewing the evidence. TCC's services explain how we handle agreed implementation after diagnosis.