AI visibility
How AI assistants decide which brands to mention
Understand AI brand mentions, citations and recommendations, then build a practical workflow for stronger evidence, reliable monitoring and accountable improvements.
OppRise Editorial · Reviewed · 8 min read
The short answer
AI assistants may mention brands using learned information, retrieved sources and conversation context, depending on the provider and mode. There is no shared public selection formula. Improve the relevance, accessibility and verifiability of your public information, then measure mentions, citations and recommendations separately. None of these improvements guarantees inclusion.
What to take away
- A brand mention, source citation and buying recommendation are different outcomes.
- Prioritise accurate, accessible evidence that answers specific buyer constraints.
- Measure a repeatable question panel, not an assumed share of all AI answers.
- Connect observations to owned corrections and measure net team effort.
1. Separate a mention from a citation or recommendation
AI assistants do not share one public brand-selection formula. An answer may draw on learned information, retrieved web pages, connected tools or conversation context. Which sources are available, how they are selected and whether citations appear depend on the provider and mode. Treat this as a diagnostic model, not a disclosed ranking recipe.
A mention simply names your brand. A citation links to a source supporting an answer. A recommendation presents a brand as suitable for the request. These can occur independently: an assistant might cite your educational guide while recommending no supplier, or name your business without linking anywhere.
Hypothetical example: an acoustic-panel supplier is cited for its installation guidance but omitted from a product shortlist. That is evidence of informational visibility, not proof that assistants consider its products a good purchase.
- Record mentions, linked citations and recommendations separately.
- Check whether the cited page actually supports the nearby claim.
- Classify prominence and sentiment rather than counting every appearance as a win.
2. Start with the buyer's constraints, not your brand name
The question determines what a useful answer needs to resolve. Location, budget, compatibility, availability and exclusions can change which brands are relevant. Brand-led prompts test recognition; unbranded prompts test discovery. Mixing them makes a visibility score difficult to interpret.
Build question groups from real sales enquiries and support conversations. Preserve meaningful constraints rather than manufacturing dozens of near-identical prompts. A broad category query and a detailed purchasing query deserve separate reporting.
Hypothetical example: a Perth payroll consultancy tracks both ‘When should a business outsource payroll?’ and ‘Payroll support for a Perth hospitality business operating weekend shifts’. The second question exposes whether its public pages explain service coverage and support hours. Adding those facts is more useful than repeating ‘leading payroll consultancy’.
- Separate discovery, evaluation and branded verification questions.
- Include disqualifying conditions, such as unsupported locations or integrations.
- Assign each question a buyer need and a relevant destination page.
3. Verify source access before rewriting everything
A useful page cannot contribute through a retrieval route that cannot access it. Check response errors, accidental indexing restrictions, login barriers, internal links and whether essential information is available as readable text. Different providers use different systems, so eligibility in one search product does not establish access everywhere.
Google's AI-features guidance says normal SEO practices apply: there is no special AI markup requirement, and inclusion is not guaranteed. For supporting links in Google's AI features, pages must meet its search eligibility requirements. This guidance concerns Google, not every assistant.
Hypothetical example: a machinery supplier publishes model compatibility only inside a customer portal. A public compatibility guide could help buyers verify suitability. Keep confidential service records private; visibility does not justify exposing restricted information.
- Inspect the exact useful URL, not only the homepage.
- Check access controls before requesting broader crawler access.
- Fix broken or inaccessible evidence before commissioning more content.
4. Make the brand and its claims unambiguous
Give readers a consistent explanation of who you are, what you offer and where limitations apply. Keep trading names, product names, locations and service scope aligned across pages you control. Consistency reduces ambiguity for readers and source checking; it is not a guaranteed assistant-ranking signal.
Hypothetical example: a software business calls the same service ‘LedgerFlow’, ‘Ledger Flow Cloud’ and ‘LF Pro’ across its site. A product overview can explain the current name and genuine legacy names, then link to current documentation. Do not invent aliases to capture unrelated searches.
Maintain a claim register with the claim, supporting evidence, responsible owner and review date. Distinguish your own product specifications from independently verified statements. Google recommends helpful, reliable, people-first content, which supports this evidence-first editorial approach rather than writing primarily to manipulate systems.
- Qualify ‘supports’ with the relevant version, plan or configuration.
- Remove expired offers and superseded service claims.
- Use credentials only when genuine, current and relevant.
5. Publish passages that remain accurate when extracted
A useful passage answers a specific question and carries its important qualifications with it. Put the answer, scope and limitation together. Clear headings and compact tables can improve comprehension, but no formatting pattern forces an assistant to quote a page.
Hypothetical example: a commercial dishwasher supplier replaces ‘rapid installation available’ with a verified explanation of eligible postcodes, required plumbing checks and when scheduling is confirmed. A concise answer helps only if those conditions remain attached. Separating the caveats several screens below invites misunderstanding.
Publish original evidence where it genuinely helps: a documented testing method, an accurately scoped service process or a product compatibility table. Explain how the information was established. Avoid unsupported statistics, invented research and interchangeable question pages.
- Can the answer stand alone without overstating the offer?
- Are units, dates, versions and exceptions next to the claim?
- Does the page help a buyer decide, including deciding against you?
6. Audit external context without manufacturing endorsement
Your website is not the only possible source of brand information. When an observed answer cites a directory, publication or review page, inspect that source directly. Record what it says and whether it is current. Do not assume every external mention improves visibility, or that a backlink proves influence over an answer.
Hypothetical example: an assistant describes a venue as offering private dining because an old listing says so. The venue has discontinued that service. Correct the owned page and request a factual correction from the listing owner, then retest the question. The update may not propagate immediately.
Prioritise corrections that could mislead customers over speculative mention-building. Genuine editorial coverage may provide useful independent context, but buying flattering lists or fabricating reviews undermines the evidence you want people to trust.
- Keep the observed answer, source URL and disputed statement together.
- Distinguish a correction request from a request for favourable coverage.
- Never report unexamined external links as verified findings.
7. Measure a defined sample, not an imagined market share
Use a stable question panel and document provider, visible model or mode, date, location settings and conversation state. Start fresh conversations where possible. Repeat observations because a single answer is not a dependable baseline. Record failed runs separately rather than treating them as brand absences.
Calculate mention rate as eligible sampled answers naming the brand divided by eligible sampled answers. Calculate linked-citation rate separately. Neither represents your share of all AI conversations. Keep providers and intent groups visible instead of hiding differences inside one average.
Hypothetical example: a brand appears in 12 of 40 valid observations. Report ‘30% of this panel’, with the sample window and method, not ‘30% AI market share’. Also inspect whether those appearances contain incorrect pricing or unsuitable recommendations.
- Save the full answer and destination links for review.
- Tag unsupported claims and misidentifications.
- Keep an unchanged core panel; report new questions separately.
8. Connect visibility to outcomes without inventing causation
Keep observed answer visibility, website visits and qualified enquiries as separate measurement layers. A mention may produce no click, and an enquiry may have several influences. Referral data alone cannot recover every assistant-assisted journey.
Search Console can help diagnose search demand and page performance, but its average position is an aggregated search metric, not a universal AI recommendation rank. Google's AI-feature reporting is included within its search reporting as described in its documentation, rather than proving which mention generated an enquiry.
Hypothetical example: citation frequency and enquiries both rise after a service-page revision. Report the association while checking seasonality, paid activity and other changes. A period comparison cannot establish that the revision caused either result. Track a defined page cohort and change log to make subsequent investigation more useful.
- Agree what counts as a qualified enquiry before reviewing results.
- Separate factual correction success from commercial performance.
- Report missing attribution explicitly rather than assigning uncertain revenue.
9. Turn the evidence into owned work
Use one evidence packet per issue: question, observed answer, source, proposed correction, owner and acceptance criteria. Batch source checks and reuse the template. Measure preparation, review and rework minutes per completed issue before and after adopting the workflow. Faster drafting is not a saving if verification takes longer.
RiseVisibility is described as bringing search performance, AI answer mentions and site health into visibility actions. Those actions can become assigned RiseProjects tasks with evidence attached.
Provider access, supported modes, permissions and reporting availability affect coverage. Do not assume exhaustive monitoring or guaranteed inclusion. Brand Vault stores internal brand context; it does not submit that context to public assistants. Explore RiseVisibility, RiseProjects and /the-solution for the proposed workflow, with people retaining responsibility for changes.
- Prioritise misleading claims and inaccessible high-value evidence.
- Approve factual changes before publication.
- Measure net time saved alongside correction quality, without promising an efficiency gain.
Frequently asked questions
Does structured data guarantee an AI citation?
No. Google says its AI features require no special markup. Useful structured data does not create a guarantee of inclusion, and other providers have their own dependencies.
Why can an assistant mention us without linking?
The answer may use learned information or context rather than a cited retrieval result. A mention alone does not reveal its source or prove that your website influenced it.
How often should we check answers?
Choose a cadence your team can repeat consistently. Preserve a stable core panel and retest important corrections, recording dates and provider settings rather than interpreting isolated answers.
Can we pay to guarantee organic recommendations?
Do not treat any such promise as reliable. Advertising placements, where offered, are distinct from organic recommendations. Ask for the exact placement, disclosure and measurement terms.
Put this into practice with