A useful AI visibility audit leaves the team with a decision. If the final deck contains forty screenshots and the sentence “we need more authority,” the audit has mostly documented its own workload.
Here is a more practical way to run it.
The audit begins with a decision
Write down why you are looking now. Perhaps a product is entering a new market, Sales keeps answering the same implementation question, or monitoring has exposed a competitor around an important prompt.
Then name the person who can use the finding. An SEO lead may own the observation, while a product marketer or subject-matter expert owns the facts. Without that handoff, even a sharp diagnosis can sit untouched.
At this stage, define the brand, product, audience, market and decision. Do not force a URL into the brief yet. The relevant content becomes clearer after you see what the model answers and which sources help it do so.
Capture an observation you can repeat
Choose a small prompt set around the decision. Include enough variety to see the problem from more than one angle: a diagnostic question, a comparison, an implementation concern or a risk. The separate guide to choosing buyer prompts goes deeper into that work.
For each run, save the exact prompt, language, market, model, date, full answer, named brands and available sources. It is mundane record-keeping, and it prevents a great deal of confident nonsense later.
Suppose one person runs “best contract software for a mid-sized procurement team” in English and another tests “contract tools” in Polish two weeks later. Those are two observations, not a before-and-after result. Without context, they can easily become a misleading chart.
Read the sources before opening the CMS
A mention tells you who appeared. The source tells you what helped construct the answer.
Read each relevant source with the buyer’s question in mind. Did it define a term the model needed? Did it give comparison criteria, name a limitation, show a process or support a claim? A documentation page and an industry article can both appear for the same prompt while doing completely different jobs.
This is also the point to discard distractions. A well-ranked directory may be useful to the model but irrelevant as a content pattern for your product page. A competitor may make a claim you cannot truthfully make. An audit that recommends copying either one has missed the assignment.
The page appears after the gap
Once the source difference is clear, find the content that should carry the answer. It might be a product page, documentation, pricing, an integration page or an article. Sometimes the right material does not exist yet.
Read the relevant passage as a skeptical buyer would. Where is the direct answer? Which noun is vague? Which claim lacks a boundary? Can the paragraph be understood when it is lifted out of the page?
A useful diagnosis is specific enough to make someone slightly uncomfortable:
The security section says “enterprise-grade protection,” but the prompt asks about data residency and the page never names a region, policy or supporting document.
“Content quality is weak” is easier to present. It is nearly impossible to edit.
A good result fits into an assignable task
The final recommendation should contain the prompt, source insight, destination passage, proposed change and fact owner. It should also state what the editor must not invent.
Keep the possible actions broad enough for judgment. The answer may be to clarify a sentence, add evidence, create supporting material, move information from another page or leave the content untouched. Forcing every finding into a new article is how an audit becomes a publishing quota.
Before closing the task, write one acceptance sentence: “A buyer reading this section can now find X, its conditions and the evidence behind it.” That sentence is often more valuable than another score.
Retesting without theatre
After publication, preserve the changed wording and repeat the observation under comparable conditions. Review the page quality first. Then inspect the model answer and sources.
If the answer changes, record it. If it does not, record that too. A single retest cannot isolate causality, but it can tell you whether the page now answers the question more clearly and whether the external observation moved in the same period.
That is enough for the next decision. An audit does not need a victory lap every time.
Frequently asked questions
How many prompts should the first audit include?
There is no universal number. Start with a small set that represents a real decision and has a clear owner. A few prompts you can examine properly are more useful than a long list nobody will act on.
Should results from different models be combined?
Compare them, but keep the raw answers and sources separate. An aggregate score can hide the exact difference that explains the next content decision.
How often should an AI visibility audit be repeated?
Repeat it after a meaningful content change or on a cadence that matches the importance of the topic. Keep the prompt, model, date and scope so the comparison remains interpretable.