This is an illustrative worked framework, not a client result. It shows how a biotech company could audit AI visibility without turning a handful of screenshots into a performance claim.

The example situation

Assume a clinical-stage biotech company has a corporate site, one lead program, several pipeline pages, scientific publications, and limited non-brand search visibility. The marketing team believes AI interfaces misunderstand the relationship between the company, target, indication, and trial. Nobody has measured the problem consistently.

The audit question is not simply, does the brand appear? It is: which questions produce an answer, which sources support that answer, is the representation accurate, and which owned page should have been useful?

Step 1: Build the prompt inventory

Create prompt groups for the company, lead asset, target, indication, mechanism, clinical stage, competitors, publications, leadership, and investor questions. Separate factual prompts from comparison and recommendation prompts. Record the exact wording because small changes can produce different retrieval.

The first pass might include 50 prompts across five groups. That number is an example, not a universal benchmark. The inventory should reflect the company's real audiences and evidence.

Step 2: Define the evidence record

For every test, record the interface, date, account or location condition when relevant, answer summary, citations, brand inclusion, factual concerns, and expected owned page. Save the evidence capture with a stable identifier.

A simple score can help triage, but keep the underlying observations. One possible scale is:

  • 0: No relevant answer or no brand inclusion where the brand has a legitimate role.
  • 1: Brand appears, but the context is incomplete or unsupported.
  • 2: Brand appears with generally accurate context but weak or indirect sourcing.
  • 3: Accurate representation with a relevant, accessible supporting source.

The score is an internal comparison tool. It is not an industry standard.

Step 3: Trace cited and repeated sources

Group citations by domain, content type, author, freshness, and question group. A publication may dominate scientific prompts while a trial registry dominates development-stage questions. A competitor may win broad explanation prompts because its educational page is clearer than the company's corporate page.

Do not assume every third-party citation should be replaced. Some sources are appropriate precisely because they are independent or primary. The goal is to find questions where the company has a legitimate first-party source and has failed to make it useful.

Step 4: Check entity accuracy

Review company names, aliases, program names, targets, indications, development stages, leadership, and relationships. Compare the answer with current approved corporate sources. Flag uncertainty and route clinical or regulatory concerns to qualified reviewers.

Then inspect the site. Does the relevant page state the relationship directly? Does the title describe the page? Is the named author or publisher visible? Is the source linked? Does structured data mirror what the page says?

Step 5: Map gaps to page owners

Every remediation item needs an intended URL. A missing corporate definition may belong on the company page. A mechanism explanation may belong in a scientific resource. A leadership fact belongs on the leadership page. Creating one catch-all AI page usually produces another ownership problem.

Use the structured data guide for entity markup and the AI search optimization framework for content and measurement.

Step 6: Prioritize by accuracy, legitimacy, and effort

Correct material factual problems first. Next, fix missing or inaccessible primary pages. Then improve weak content, sources, internal links, and structured data. Low-value cosmetic changes come later.

Google's generative AI search guidance keeps the technical foundation inside ordinary SEO. OpenAI separately publishes publisher and crawler controls. Test live access instead of assuming one configuration covers every system.

Step 7: Retest the same groups

Retest after the changed pages have been crawled and enough time has passed to observe a new state. Compare prompt-level inclusion, citation changes, accuracy, and owned-source use. Preserve unchanged and negative results.

A useful audit does not end with a flattering score. It leaves a repeatable inventory, dated evidence, assigned page owners, and a clear record of what changed next.