The practical answer: AI visibility becomes useful when you can explain what was tested, what the assistant actually showed and what the buyer could do next. Build a small evidence log before you build a dashboard.
A useful test needs a stable question
Olivia Moore's LinkedIn shopping-agent example offers a useful starting point: experience discovery as a customer and observe where options disappear. Her reported retailer results are a case observation, not a measurement standard.
The marketing temptation is to turn one answer into a score: “We are third in AI.” But third in which surface, for which prompt, in which region, on which date? A reproducible test begins with those details.
Separate your business hypothesis from the platform. For example: “Buyers looking for a small-team lifecycle specialist cannot find a clear description of our scope.” The experiment should tell you whether that problem appears and where the missing evidence belongs.
Build a ten-prompt panel from customer language
Take questions from sales calls, customer support and search data you are authorized to use. Remove private information. Group ten prompts across problem discovery, option comparison, buying constraints and next-step questions. Include at least one scenario where your offer is genuinely a poor fit.
Write the buyer, job, budget and constraint as naturally as a customer would. “Find a CRM onboarding consultant for a five-person B2B team” is a better test than repeatedly asking “Is our agency the best?”
Keep branded accuracy checks in a separate group. Asking about your brand directly measures whether the assistant can describe it; it does not establish that the assistant would discover it unaided. Preserve the initial wording when comparing later runs.
Use a repeatable protocol
Choose two relevant AI surfaces and run each prompt three times per surface: 30 runs on each, 60 total. That is a suggested starting sample, not a statistical threshold. If the effort is too large, reduce the panel and label the smaller sample honestly.
Use a fresh session, available memory controls and a consistent language, region and delivery location. Record account state and the model or product version when shown. A clean session reduces some prior context; it does not create a neutral laboratory.
Save the first answer before probing it. Follow-ups can change the recommendation. Keep baseline results separate from any later debugging conversation.
Compare options for [buyer] who needs [job] under [constraints]. Link the pages supporting your comparison. Mark unknown prices, capabilities or terms as unknown. Do not assume a provider is suitable because it uses the right keywords. Do not submit a form or contact anyone.
Record outcomes with clear denominators
| Measure | Definition | What it tells you |
|---|---|---|
| Mention rate | Runs mentioning the brand / completed eligible runs | Presence in this prompt sample |
| Citation rate | Runs linking to your domain / completed eligible runs | Whether your pages support the answer |
| Accuracy | Correct verified facts / facts checked | Reliability of the description |
| Next-step completion | Journeys reaching the defined step / journeys attempted | Usability after recommendation |
| Failure count | Blocked or incomplete runs, with reasons | What the other metrics exclude |
Specify what counts as eligible before testing. Do not quietly discard unfavorable answers. Break out “no recommendation,” timeout and access failure. Report results by platform before making any combined summary.
For example, 6 mentions in 30 completed runs means 20% of that sample contained the brand. It does not mean 20% market share, a 20% chance that every buyer sees you, or proof of future traffic. This example is illustrative.
Treat explanations as hypotheses
Ask which visible source supports a claim, not what the exact secret ranking weights are. A tool's explanation may be useful, but it is not an independent audit of its internal process.
Label each finding as observed, verified from a source, or inferred. A visible error screen is observed. A price checked against the store is verified. “We lost because the competitor has more Reddit mentions” is an inference unless you have much stronger evidence.
When a source is stale, record the source URL and the conflicting first-party fact. Do not assume a model “trained on” the cited page merely because it linked to it.
Let the evidence choose the next content brief
If the assistant cannot explain who your service is for, improve the offer page. If it cannot compare alternatives, publish a factual comparison with fit, limitations and decision criteria. If a delivery promise is unclear, fix the operational information before creating another awareness article.
A useful brief names the buyer question, missing fact, approved evidence, page to improve, owner and review date. Add one explicit next action for the reader. Avoid manufacturing best-of lists in which your own service always wins.
Google's AI search guidance says ordinary SEO foundations remain relevant. It also includes AI-feature traffic within the Web search reporting in Search Console. Do not label all organic search visits as AI traffic or claim a clean AI-only attribution split from that report.
Run a two-week learning cycle
In the first week, collect the baseline and verify the most consequential errors. Choose one page or journey repair, write down the expected change and publish it. Leave other variables as stable as practical.
In the second week, repeat the same prompt panel and inspect the evidence. Track qualified enquiries, sales and identifiable referrals as separate business outcomes. A higher mention rate with no improvement in buyer understanding may have limited value.
Keep the decision simple: retain the repair, investigate another source of friction, or revise the hypothesis. This is a learning cycle; a before-and-after comparison alone does not establish causality.
Use the Marketing Clarity Session if you need help choosing one audience, offer and next experiment. For technical and journey diagnosis, see Waymaker Agency's audit guide.
Sources and editorial note
Sources checked October 20, 2026. The inspiration is a reader-supplied LinkedIn excerpt by Olivia Moore (profile); the original post permalink was not supplied. Her reported shopping results are attributed observations. The audit methods and hypothetical examples here are our editorial recommendations, not measured client results.
Keep exploring
How to read analytics reports: a visual walkthrough for beginners