AI visibility
Choosing the questions to track for AI visibility
Build a defensible AI visibility question set using buyer evidence, intent coverage, repeatable checks and clear rules for turning findings into work.
OppRise Editorial · Reviewed · 8 min read
The short answer
Track questions grounded in real buyer decisions, covering unbranded discovery, evaluation and branded accuracy. Prioritise commercially relevant, actionable topics; preserve exact wording and observation conditions; and report mentions, citations and accuracy separately. Keep a stable core so comparisons remain interpretable.
What to take away
- Customer evidence is a stronger starting point than prompts designed to elicit praise.
- Separate branded verification from unbranded discovery.
- Record provider conditions, failed checks and question changes.
- Treat visibility movement as observation, not proof of marketing impact.
- Measure review time and quality together before claiming efficiency gains.
1. Start with the buying decision, not a flattering prompt
AI visibility tracking is a sampling exercise: you select questions and observe how particular answer services respond under recorded conditions. It is not a census of what every buyer sees. A useful question set exposes gaps in discovery, evaluation and factual understanding, rather than repeatedly inviting an assistant to praise your brand.
Start by naming the decision you want to support. In a hypothetical payroll software business, that might be making the shortlist for Australian employers with casual staff. “Which payroll tools handle changing weekly rosters?” tests discovery. “Why is our software excellent?” does not.
Google says normal SEO practices apply to its AI features, with no special markup requirement or guaranteed inclusion. That guidance concerns Google, not every answer provider. Choose questions to investigate customer needs, not to chase a supposed universal AI optimisation formula.
- Write the buyer, decision and relevant product or service beside each candidate question.
- Identify what an unfavourable answer would prompt you to investigate.
- Reject questions whose only purpose is producing an impressive screenshot.
2. Build candidates from evidence of real buyer language
Collect candidate questions from sales enquiries, support conversations, on-site search and Search Console. Record the source and date, removing personal information. Separate verbatim customer wording from questions your team has inferred.
Hypothetically, a commercial cleaning firm might hear “Can you clean after our medical centre closes?” in enquiries. That supports testing questions about after-hours availability and clinical premises, rather than only “best commercial cleaner”. Several similar enquiries strengthen the rationale for the topic, but do not establish its market-wide popularity.
Search Console supplies search evidence, not AI prompt volumes. Its performance report uses aggregation rules and average position, so a query's reported position should not be treated as a fixed ranking or evidence of AI recommendation. Use queries to discover vocabulary and needs, then validate them with customer-facing colleagues.
- Tag candidates as customer-sourced, search-sourced or team hypotheses.
- Keep strategically important hypotheses, but label their weaker evidence.
- Do not manufacture AI search-volume estimates from ordinary search impressions.
3. Cover distinct intent and brand-awareness groups
Separate unbranded discovery, evaluation and branded verification. Each answers a different business question. Discovery asks whether you enter consideration; evaluation tests suitability; branded verification checks whether an answer represents you accurately.
For a hypothetical accommodation operator, discovery could be “Where can a family stay near Hobart's waterfront?” Evaluation could be “Which waterfront stays have connecting rooms?” Branded verification could ask whether the named property offers parking. Treat these as separate reporting groups: a brand already supplied in the prompt has not been independently discovered.
Include practical barriers such as service areas, accessibility information, compatibility and cancellation terms where relevant. Avoid a panel dominated by broad “best” questions. Those can reveal shortlists, but tell you little about the constraints that determine whether a recommendation is useful.
- Label intent, brand awareness and the buyer constraint.
- Keep branded accuracy results separate from unbranded mention rates.
- Check whether important offerings are missing because the easiest questions received all the attention.
4. Prioritise questions within a realistic checking budget
Rank candidates using buyer evidence, commercial relevance and actionability. A simple low, medium or high judgement for each is enough if reviewers write down their reasoning. Avoid multiplying subjective scores into a number that looks more precise than the underlying evidence.
Hypothetically, an industrial supplier could prioritise “Which pump suits abrasive slurry?” over “What is a pump?” because the first concerns a served application and a meaningful selection decision. However, exclude applications the supplier cannot genuinely support, even if the associated question looks commercially attractive.
Set the panel size from review capacity. Twenty questions across three providers with two repeats means 120 scheduled observations per cycle, before investigating errors. This is illustrative arithmetic, not a recommended minimum. Broader coverage trades off against the time available to inspect answers carefully.
- Estimate collection, classification and quality-check time before adding questions.
- Reserve capacity for high-consequence factual errors.
- Park low-actionability topics with a reason and reconsideration date.
5. Use natural wording and controlled variants
Write prompts a buyer could plausibly use. Include a location, budget or technical constraint only when it changes suitability. Overloaded prompts can make the test unnaturally specific and effectively reveal the brand you hope to see.
For a hypothetical Adelaide solar installer, “Which installers service the Adelaide Hills?” tests a different requirement from “Which installers offer battery retrofits?” Track both if both services matter. Replacing “installers” with “companies” repeatedly is usually less informative than testing a distinct buyer constraint.
Assign a question-family ID to related variants. Keep one canonical wording fixed and use a smaller exploratory group for alternatives. Otherwise, a topic with six near-duplicates can dominate the headline result while a commercially important topic with one prompt barely counts.
- Record exact wording, language and any location supplied.
- Do not insert your own selling points into an ostensibly neutral discovery prompt.
- Version changed questions rather than silently overwriting them.
6. Define what counts as a comparable observation
Record the provider, interface, date, available model identifier, browsing setting and whether the session starts fresh. Consumer interfaces and API responses should not be assumed equivalent. Availability, account settings, provider changes and supported data access constrain what can be observed.
Hypothetically, a restaurant question asked in a fresh session is not directly comparable with one asked after a conversation about vegan dining. Keep conversational tests in their own group, with the preceding context saved.
Retain the answer and visible source links where permitted. Distinguish a successful answer containing no brand mention from an unavailable or failed check. Repeating selected questions can reveal variation, but there is no universal repeat count that makes a small panel representative of all users.
- Use a fixed collection window and document exceptions.
- Report completed checks alongside scheduled checks.
- Flag provider or interface changes before interpreting movement.
7. Score presence, citation and accuracy separately
A mention is not necessarily a recommendation, and a citation is not necessarily support for the claim beside it. Define separate fields for brand presence, recommendation status, owned-site citation and factual accuracy. Keep unknown accuracy separate from verified accuracy.
In a hypothetical equipment-hire answer, your business could appear but be assigned an incorrect delivery area. That is a presence success and an accuracy problem. Conversely, your safety guide could be cited without the business being recommended as a supplier.
Define mention rate as successful eligible answers mentioning the brand divided by successful eligible answers in the stated panel. Show the denominator and provider breakdown. A rise from four mentions in twenty answers to six in twenty is an observed change, not proof that your latest page update caused it.
- Agree classification examples before reviewers begin.
- Double-check ambiguous recommendations and disputed claims.
- Avoid a blended visibility score that hides accuracy failures or changing panel composition.
8. Connect each finding to an evidence-backed action
Investigate before rewriting. An absent mention could lead you to check whether the relevant service is clearly documented, whether the page is accessible, and whether the business actually meets the question's requirements. Absence alone does not diagnose a technical fault.
For a hypothetical accountant, an answer gap around not-for-profit reporting might justify reviewing the firm's service page and supporting expertise. Google's people-first content guidance supports useful, reliable material. It does not justify publishing thin pages for every wording variant.
Its described RiseVisibility workflow connects search evidence, AI answer mentions and site health to visibility actions that can become RiseProjects tasks. Provider coverage and connected capabilities depend on supported access and provider availability. Do not assume identical observations across services or guaranteed inclusion.
- Attach the question, answer evidence, proposed investigation and owner.
- Link actions to buyer usefulness, not merely obtaining a mention.
- Review the planned workflow through RiseVisibility and /the-solution before committing.
9. Maintain a stable panel and measure review efficiency
Keep a stable core for comparison and a labelled exploratory group for new services or emerging customer questions. Retire obsolete questions with their history intact. When the panel changes, report the continuing-question cohort separately rather than presenting the new total as like-for-like.
Measure team efficiency independently of visibility. Track collection, classification, quality assurance and handoff minutes per completed question-provider check. Also track rework and the proportion of findings with enough evidence for an owner to act.
Hypothetically, reducing a comparable review cycle from 180 to 135 minutes saves 45 minutes. Verify that coverage and checking quality remained intact before claiming an efficiency improvement. Templates and evidence-linked tasks may reduce administration, but savings are not guaranteed, and additional provider checks can increase total workload.
- Review relevance when offerings or customer priorities change.
- Keep question ownership and retirement reasons visible.
- Report operational savings separately from mentions, enquiries and revenue.
Frequently asked questions
How many questions should we track?
Start with the distinct buyer decisions your team can review reliably. Calculate the workload across questions, providers and repeats before expanding. There is no universal ideal panel size.
Should we track exactly the same wording each time?
Keep canonical wording stable for comparisons. Test alternative phrasing in a separate exploratory group, preserving question-family IDs so near-duplicates do not distort reporting.
Can Search Console tell us which AI questions buyers ask?
Its performance report provides search query evidence, not a cross-provider inventory of AI prompts. Use it to identify relevant language, then validate candidate questions with customer evidence.
Does being mentioned mean our content changes worked?
Not by itself. Answers, providers and observation conditions can change. Record content updates and compare consistent panels, but do not infer causation from a before-and-after mention increase.
Put this into practice with