AI visibility action plan

Use this worksheet with AI visibility: where businesses should start. Turn a customer question into a check, an assigned action and evidence that the work is complete.

Guide sources checked: 20 October 2026. Set your own assessment and review dates below. A completed action does not guarantee inclusion in an AI answer.

Set the scope

Field	Your entry
Business, offer and website
Customer group and market
Decision you want to help
Useful outcome and its definition
Customer questions and where they came from
Services or AI products and interfaces in scope
Important page and profile URLs
Content/data to make available, restrict or keep private
Intended permissions for search, user-requested agents and training
Expected return for each use, and when to reconsider access
Assessment date and responsible person

Record each finding

Copy this block for each customer question or issue. Use specific page URLs and service names so another person can repeat the check.

Field	Your entry
Customer question or business need
Check and expected result
Actual finding and evidence link
Result: pass / needs attention / unknown / not applicable
Action or decision required
Priority, commercial reason and estimated effort/cost
Named owner and any dependency
Target completion date
Evidence required to close the action
Actual completion date and evidence
Next check date and trigger for an earlier review

Pass means the evidence meets the defined expectation for the named check. Needs attention means an observed issue needs action. Unknown means evidence is insufficient; assign a check rather than guessing. Not applicable needs a short explanation. These results describe individual checks, not an overall visibility score.

Prioritise a material access problem or inaccurate offer information, then improvements to the evidence and reporting. Keep optional experiments separately identifiable.

Worked example: hypothetical Sydney office cleaner

All findings below are invented to demonstrate the method. No business, website or platform account was tested.

Customer need	Illustrative finding	Next action and owner	Completion evidence	Next check
Find an after-hours cleaner in Parramatta	Needs attention: the service page omits after-hours availability; operations confirms it is offered	Marketing lead adds accurate hours and coverage, with operations approval; high priority because it affects service fit	Approved page text is visible on the relevant URL and matches controlled business profiles	Immediately after publication; then at the next monthly review or when hours change
Obtain public service information through a selected search service	Unknown: no access evidence has been supplied	Website owner checks the service page against the agreed policy, provider guidance and applicable security rules	Named URL/service, method/date, response and useful content, available index evidence, identity-verification scope and gaps; retest any fix	After any fix; after a security-policy change
Compare the provider with other suitable suppliers	Needs attention: the page claims extensive experience but gives no relevant example	Content owner obtains permission for a representative office-cleaning example and checks every claim	An approved example explains scope, constraints and support for any result claimed	Before publication; quarterly or when evidence changes
Know whether identifiable visits produce useful enquiries	Unknown: form submissions have not been checked against enquiry quality	Analyst and sales lead define a qualified site-assessment enquiry and check tracking	Documented test and baseline report, with attribution limits and unavailable data stated	After setup; review outcomes after a suitable sales-cycle interval

Replace each role with a named person and each relative review point with a calendar date before assigning the work. Change an unknown result only when evidence supports the new finding. The business may close an action by accepting uncertainty; retain the unknown result and record the decision, owner and review date.

Follow one change through to a decision

This continues the invented cleaner example. Dates are illustrative planning dates; none of this work or testing has occurred.

Step	Illustrative record
Question and baseline	On 20 October, five of eight substantive answers in a fixed branded accuracy check say the cleaner does not cover Parramatta after hours. Two give the correct coverage; one does not address it. Operations confirms the service; the website omits the detail. All eight runs are saved, with no failures.
Evidence check	Two of the five incorrect answers cite an old directory listing that repeats the error. The other three have no visible supporting source. Do not assume they used the directory. Save each claim, source, supporting passage and missing condition; distinguish factual accuracy from citation support.
Action and effort	Marketing corrects the service page and controlled profiles on 22 October, with operations approval. Allow half a working day for the work and QA. A request to correct the directory is a separate action, subject to approval.
Implementation evidence	The dated page and profile records show accurate coverage. The website owner verifies that the intended service can obtain the information, recording the method and limits. This closes the page-correction task.
Later observation	At an illustrative 19 November review, the same questions, interface, settings and repeats produce eight substantive answers: one repeats the error, six give the correct coverage and one does not address it. No failures. The remaining error cites the old directory. This is an observed change, not proof that the edit caused it.
Decision	Keep the factual correction, follow up the directory issue and continue the accuracy check. Defer optional rewriting expenditure. Observe qualified site-assessment enquiries separately over the agreed sales-cycle window; this sample establishes no sales effect.

The dates demonstrate an assigned review, not a promised AI response time. A factual correction is worth making because the information is wrong. Optional experiments need a separate investment decision.

Count the answer sample

Agree the rules first. Count a relevant business mention once per substantive answer. For an answer-level citation measure, count answers with at least one relevant link to the business’s site; record third-party links separately. A recommendation must present the business as an option for the specified need. These categories may overlap. Verify important claims against approved facts and cited evidence; mark inaccessible evidence unknown.

For example, ten unbranded questions repeated three times produce 30 attempts: three technical failures, two unanswered requests and 25 substantive answers. Eight mention the business, four link to its site and two recommend it. Report the counts and denominators: 8/25 answered runs mention the business; 2 unanswered requests and 3 failures are reported separately. A valid answer with no mention stays in the denominator.

These invented counts explain the calculation, not a required sample size or market-share estimate. Keep branded checks separate. Preserve question mix, name-matching rules, repeat weighting and exclusions across comparisons; repeated prompts are not independent customers.

Keep the baseline beside the actions

Record	What to save
Platform reporting	Product, report, metric definition, actual values, period, time zone, market, filters, aggregation level, export and coverage limits
Answer sample	Question-set version, exact prompts and their source, product/interface/model if known, context and personalisation, date, search mode and observed use if exposed, repeats, raw answers, citations, failures and counting rules
Website and sales outcomes	Baseline values and period, source classification, outcome definition, tracking coverage, attribution rules and observation window
Change record	URLs, what changed, when, expected mechanism and concurrent activity that could affect results

Use unavailable for missing data. Enter zero only when the observation actually supports zero within known coverage. Keep impressions, citations, sampled mentions, visits and qualified enquiries as separate measures.

Check exports against the original report. Google’s generative AI report converts unavailable markers to zeros on download; preserve their unavailable status before comparing periods. Google’s export rules.

For an optional experiment, set the primary measure, observation window and comparison before making the change. Record what result would justify continuing or stopping. If no credible comparison is possible, describe the result as an observation.

At review, decide whether to keep the change, investigate a gap or run a defined experiment. Record the observation and its limits before describing an improvement as caused by the work.