AI visibility: build a manual prompt panel you can repeat

Start with a question you can actually answer

The useful question is narrow: “Does this assistant mention our business for these selected questions, in these recorded conditions?” That question can be answered by a manual ledger. “What share of all AI recommendations do we own?” needs a population, a sampling design and much more evidence.

Choose the decision before writing prompts. Perhaps you want to identify factual errors about your opening arrangements or see which source pages an assistant cites. Those are different jobs from measuring competitive recommendation share. Your recording fields should follow the decision you intend to make.

For this guide, Peak Plumbing is a fictional Sheffield business used to illustrate the method. No assistant was asked about it, and no output below is a measured result. The name helps show where branded and unbranded questions belong without pretending to publish a customer case study.

Build four kinds of questions

Use a balanced starting structure: branded questions, comparisons, generic needs and person-specific situations. Three prompts in each group make a twelve-question editorial panel. Twelve is a manageable starting budget, not a statistically representative sample or a required industry standard.

Branded prompts test what the assistant says when the business is already named. Generic prompts test a broader information need without planting the brand. Keep those groups separate in the report. A business named in your question has an advantage that an unprompted recommendation does not.

Comparison questions can also steer the answer. Record exactly which alternatives the question names and why you chose them. Do not later describe a deliberately restricted comparison as the assistant independently choosing among every business in the market.

Here is the proposed panel. These are illustrative prompts, not observed customer queries or measured search demand:

• Branded: What information would you need to verify Peak Plumbing's service area?
• Branded: What should I check before contacting Peak Plumbing?
• Branded: Which public sources would you use to confirm Peak Plumbing's services?
• Comparison: How should I compare Peak Plumbing with another local plumber?
• Comparison: What evidence matters when comparing two plumbing service pages?
• Comparison: How can I compare enquiry options without relying on unsupported reviews?
• Generic: What details should I give a plumber about a boiler-pressure problem?
• Generic: How can I check whether a plumber covers my area?
• Generic: What should a clear plumbing contact page explain?
• Persona: I am a landlord arranging a repair; what should I confirm first?
• Persona: I am a tenant reporting a plumbing issue; what information helps?
• Persona: I manage several properties; how should I organize a repair enquiry?

Some prompts ask for an evaluation method rather than a list of businesses. That is deliberate: use prompts that match the decision you actually need. If recommendation monitoring is your goal, design a separate set of non-leading recommendation questions and record their selection rationale.

Freeze the wording before collection

Give every question a stable identifier. Save its exact text, category and intended geography. Keep a version date for the panel. If a question changes, create a new version instead of quietly replacing the wording in an old trend line.

Write an observation rule before the first answer. Choose the interfaces, whether web search is enabled, how many attempts you will make and how you handle unavailable responses. Do not discard inconvenient answers or keep trying until you receive a favourable one.

Use a fresh conversation for each question when your aim is to reduce carryover from earlier context. Record whether you were signed in and any relevant settings you can observe. A fresh conversation does not guarantee an unpersonalized or identical environment; it simply makes your procedure more explicit.

If multiple people collect observations, give them the same written instructions. Otherwise one collector may ask follow-ups while another records only the first response. That difference can look like a change in visibility when it is a change in method.

Run the panel by hand and preserve evidence

Open the chosen assistant interface and enter one frozen question. Save the answer text or an appropriate screenshot, the date and the visible source destinations. Avoid collecting personal information that is irrelevant to the exercise. Keep any account identifiers out of a public ledger.

Record the actual interface. A ChatGPT web observation and an API response are different collection surfaces. Do not label an API-generated answer as consumer-interface evidence, and do not merge the two into a single trend without explaining the difference.

A useful ledger has one row per attempt, including failures:

Panel version: prepared 21 October 2026
Prompt ID: branded-01
Host and interface: Not observed
Run time and timezone: Not observed
Search setting: Not observed
Business mentioned: Not observed
Citation destination: Not observed
Evidence location: No run collected
Collector note: illustrative ledger structure only

The repeated “Not observed” values matter. They prevent a reader from mistaking an empty template for a result. When you later run a real panel, replace them with recorded observations and retain the original evidence so another person can check the coding.

Code mentions, citations and claims separately

Define a mention as the business being named in the answer. Define a citation as an identifiable reference or source link. Define a recommendation using a written rule that another reviewer could apply. None of these automatically establishes that a user clicked or became a customer.

Record where a citation actually goes. A link might lead to your site, a directory, an unrelated business or a source that does not support the sentence beside it. Inspect the destination before treating the citation as helpful evidence.

For factual accuracy, compare each business claim with a verified source. Mark supported, contradicted or unresolved rather than asking an assistant whether its own answer was correct. An unsupported claim about a fictional example remains illustrative; a claim about a real business needs real verification.

Keep disagreement visible. If two reviewers interpret a recommendation differently, refine the rule and preserve the original judgement. A neat percentage is less useful than a ledger whose categories can actually be repeated.

Report change without overstating certainty

Show the count of qualifying observations and the number of eligible attempts, grouped by prompt type and host. Do not mix branded and generic questions into a headline that implies unprompted discovery. State how failed or unavailable attempts were handled.

A hand-picked panel does not form a random sample of every customer question. Repeated answers may also depend on one another or on conditions you cannot observe. For those reasons, a conventional confidence interval can imply more precision than the design supports.

If an analyst calculates an interval, require them to state the sampling unit, dependence assumptions and target population. Otherwise present the raw observations and describe the uncertainty in words. Do not manufacture a confidence range merely because a reporting template requests one.

When a new observation differs, inspect its evidence before celebrating. The source page may have changed, the interface may have retrieved different material or your settings may differ. Those explanations need investigation; the ledger cannot identify causation on its own.

Use the evidence to improve a page

The most actionable result may be a factual gap rather than a visibility score. If an answer cannot establish your service area, inspect whether a reliable public page states it clearly. If a cited source is outdated, determine which current source you control and whether it explains the relevant facts.

Use Claude or ChatGPT to organize the evidence after collection:

Using only this observation ledger, group unsupported business
claims and missing-source questions. Quote the relevant evidence
row for each suggestion. Separate factual corrections from
visibility hypotheses. Do not infer rankings, demand or revenue.

Then ask the business owner to verify the correction. Improving public information is work you can inspect. Whether that change affects later assistant answers requires another observation period, and a favourable answer still does not prove commercial impact.

Keep the before-and-after page versions and repeat the frozen panel on the schedule you chose. Report concurrent changes and missing observations. If you change the questions to reflect a new business need, start a new series rather than preserving a misleading trend line.

Know when to use a different approach

A manual panel is useful for learning the questions, coding rules and evidence you care about. It becomes cumbersome when you need broad coverage, many markets, frequent collection or formal comparative research. Evaluate an established specialist service against those requirements rather than pretending this template replaces it.

AdvisorSEO can be discussed as part of a review workflow over your own records, but automated answer tracking is planned. Static AI-readiness checks examine page properties; they do not ask engines questions. Keep those two activities in different columns so a readiness change never becomes a fabricated citation result.