How to Run an AI Visibility Audit for a New Client
A scoped workflow for checking prompts, verifying evidence, and presenting findings without promising AI recommendations.
A first AI visibility audit should answer a narrow business question: when potential customers ask a set of relevant questions, what does the client’s brand appear alongside, which sources support the answers, and what should the team investigate next? It is a snapshot, not a forecast or a promise that a platform will recommend the client. Sell a documented baseline and a prioritized work plan.
Before the working session: get the inputs
Ask the client to complete a short intake before the working session. Request the target market and language, priority products or services, competitors, high-value customer questions, any regulated claims, and a person who can verify product facts. Ask for a site URL and permission to review public pages. If Search Console access is available, request read-only access; it can add conventional search-performance context, but it does not measure every AI product.
Set scope in writing: named platforms, market, prompt count, collection date, and deliverables. A pilot might cover a defined set of prompts across selected AI experiences, with a repeatable prompt set and a concise findings deck. Choose scope according to your tooling and review capacity; don’t imply that a small sample represents all users or all answers. Allow time beyond the working session for intake, platform access or setup, and final quality checks.
A four-hour working session
- 00:00–00:25 — Confirm the brief. Translate the client’s business goals into prompt groups such as category discovery, comparison, and problem or use-case questions. Remove prompts that are branded-only if the objective is non-brand discovery.
- 00:25–00:50 — Freeze the test plan. Record exact prompt wording, platform, collection date, and any market, language, or other settings available in that product. Note whether web search is available and enabled. Settings and controls differ by product, so record what you can see and reproduce; don’t imply that every platform offers the same controls. Save the plan so a later audit can be compared more fairly.
- 00:50–02:00 — Run the prompts and capture evidence. For each answer, record whether the brand is named, the order of any brand mentions in that specific answer, the surrounding context, competitors named, visible citations or links, and material factual errors. Save the answer or a dated screenshot, not just a score.
- 02:00–02:45 — Verify important claims. Open cited pages and check whether they support the answer. Note whether sources are the client’s own pages, third-party coverage, directories, or other sources. Separate ‘brand absent’ from ‘brand present but poorly described’ and from ‘answer cites a page that does not substantiate the claim.’
- 02:45–03:25 — Add site and search context. Review relevant landing pages for accessibility to search engines, clear product facts, and useful supporting content. If access is granted, use Search Console’s Performance report for conventional Google Search performance data. Google’s documentation says traffic from its AI features is included in the overall Search Console Performance report; this is not a separate measure of visibility across AI products or a complete account of AI responses. See Google’s Search Console Performance report documentation for access and report details: https://support.google.com/webmasters/answer/7576553?hl=en.
- 03:25–04:00 — Prioritize and prepare the readout. As one possible format, list a few actions, each with an evidence note, owner, effort estimate, and a way to check progress. Share the draft with the client fact-checker before presenting.
Make the output usable, not theatrical
Keep the readout to a concise summary plus an evidence appendix. The summary should state the scope, date, sample size, what appeared in this sample, what did not, and possible next actions. The appendix should contain prompt wording and answer evidence. If using a mention rate, define its denominator plainly—for example, brand-mentioned answers divided by prompts tested—and label it as a result for this exact sample, not a market-wide share.
Use careful language: ‘The brand was named in 5 of the 12 responses collected’ is auditable when those counts match the documented sample. ‘AI recommends you in 42% of cases’ overstates what a prompt sample establishes. Answers can vary, and products may show different sources or citations. OpenAI’s ChatGPT Search help explains that search responses may include links to sources and that search results can vary; citations should be inspected rather than treated as proof that every claim is correct: https://help.openai.com/en/articles/9237897-chatgpt-search-faq.
Turn findings into a client conversation
Lead with one verified example, then explain the business implication and the proposed next test. If the brand is absent, avoid jumping straight to ‘publish more content.’ First ask whether the prompt reflects a real buying question, whether competitors are being surfaced through sources the client can influence, and whether the client’s own pages clearly substantiate the relevant facts. If a claim is wrong, route it to the appropriate product, legal, or communications owner before recommending a public response.
Close by distinguishing controllable work from platform outcomes. The agency can improve and clarify source material, strengthen relevant pages, and track whether a defined sample changes. It cannot guarantee that an AI system will cite, mention, or recommend a brand.
Scope the work and make the process repeatable
Time-box manual work and price the audit around its defined scope, not an open-ended promise to ‘check AI.’ Track actual analyst time for the pilot, including intake, evidence review, quality checks, and presentation. If delivery exceeds the planned hours, reduce prompt volume, automate capture where permitted, or agree on the price and scope of additional analysis before proceeding. Keep human review for interpretation, source verification, and client-specific claims; templating should remove repetitive formatting, not judgment.
As the service grows, standardize the intake form, prompt taxonomy, capture fields, naming convention, quality checklist, and report layout. Store prompt sets with version dates, and log any change rather than silently editing the baseline. Use a shared rubric for mention context and source type, with a second reviewer checking a sample of records. This helps the agency compare its own delivery quality and effort without pretending every client’s market or prompt set is interchangeable.
For a follow-up engagement, define the work in advance: rerun the same prompts where possible, report what changed in the observed sample, and investigate an evidence gap. That makes the audit a repeatable service with a clear scope—not a claim about what every AI system will do.
Sources
Running this for clients?
The Agency plan is priced per fresh scan, not per location — the client list is unlimited, so a sleeping client costs nothing.
See plans