AI Search Share of Voice, Done Right
Key takeaways
- AI search share of voice is meaningless unless the prompt set, engines, locales and scoring event are defined.
- A brand mention, recommendation, citation and linked citation are different events and should be scored separately.
- Keep the prompt panel stable for trend reporting, and maintain a separate discovery panel for new demand.
- Record the answer evidence, not only a percentage, because generated results can change between runs.
- Connect visibility to qualified visits, leads and revenue without pretending every mention caused a conversion.
What is AI search share of voice?
AI search share of voice measures how often a brand appears across a controlled set of AI-generated answers. A basic unweighted formula is:
brand-visible prompts / eligible prompts x 100
The numerator must name the event. If brand-visible means any plain-text mention, that is mention share of voice. If it means a clickable source link, that is linked-citation share of voice. Combining both into one score hides a material difference.
The denominator matters just as much. Ten prompts written by the brand team do not represent a market. A useful panel reflects actual customer questions across discovery, comparison, validation and purchase stages.
Our judgement is that one universal AI visibility score is too blunt for executive decisions. Use a small family of clearly named rates instead.
Which events should you score?
Score mentions, recommendations, citations and linked citations as separate binary events at prompt level. This preserves what the engine actually did.
A practical event model is:
- Mention: the answer names the brand.
- Recommendation: the answer presents the brand as an option for the stated need.
- Citation: the answer attributes a claim to a brand-controlled page.
- Linked citation: the answer provides a usable link to that page.
- Accurate attribution: the cited claim is supported by the linked page and assigned to the correct source.
Do not award partial credit because a competitor was cited near an uncited brand mention. Save the answer, link target, engine, model or mode where visible, date, time, account state and location setting.
SAGEO's analysis of AI citations without brand attribution explains why citation presence and brand recognition must not be treated as synonyms.
How do you build the prompt set?
Build prompts from customer language and group them by intent before collecting results. Sources can include search queries, site-search logs, sales calls, support tickets, community questions and approved keyword research.
Each prompt should have:
- a stable ID;
- the exact wording;
- market and language;
- intent stage;
- topic cluster;
- business priority;
- eligibility rules;
- an owner and review date.
Include non-brand prompts that allow genuine competitive discovery. Avoid loading the brand name into every question. Include realistic local qualifiers where the business serves specific markets.
Keep a locked trend panel for comparable reporting. Add emerging questions to a separate discovery panel. If the core panel changes, publish the old and new denominator so the apparent trend cannot be mistaken for performance.
Which engines and modes belong in the test?
Include only engines and answer modes that matter to the audience, and report each one separately before calculating any blend. Search-integrated summaries, conversational assistants and research modes can retrieve and cite differently.
Google says the same foundational SEO practices apply to its AI features and that no special AI file or schema is required. Its official AI features guidance also makes clear that normal indexing and preview controls continue to apply.
Bing's 2026 AI Performance preview provides site owners with citation activity and cited-page information. That first-party evidence is valuable, but it is not a substitute for a controlled cross-engine prompt panel.
Do not merge engines before checking coverage. A blended score can rise because the weighting changed, even when every individual engine stayed flat.
Should prompts be weighted?
Weight prompts only when the business rationale is documented and the unweighted result remains visible. Weighting can reflect revenue potential, audience size or strategic priority, but it can also turn a measurement into a preference disguised as data.
Use the unweighted rate as the audit baseline. Then show the weighted rate with the weight for every prompt group. Cap extreme weights and review them on a fixed schedule.
For example, a procurement comparison may deserve more commercial weight than a broad definition. That does not make the definition invisible. It means the dashboard answers two questions: how often are we present, and how often are we present where the business cares most?
How often should you run the panel?
Run the panel often enough to identify a sustained change, but not so often that normal answer variation becomes a crisis. Weekly or monthly cadence can be appropriate depending on query volume, engine volatility and decision speed.
Repeat samples help estimate instability. If one prompt produces three different brand sets in three runs, a single screenshot is weak trend evidence. Report prompt-level persistence alongside the headline rate.
Preserve raw outputs and calculate the metric from stored evidence. Automation can reduce collection work, but human review is still needed for ambiguous recommendations, broken links and inaccurate attribution.
How do you calculate a defensible score?
Calculate each event rate from eligible prompts and expose the numerator, denominator and missing observations. Never silently treat a failed collection as a zero.
A minimum reporting table contains:
| Metric | Numerator | Denominator |
|---|---|---|
| Mention share | Prompts naming the brand | Eligible prompts observed |
| Recommendation share | Prompts recommending the brand | Eligible prompts observed |
| Citation share | Prompts citing a brand page | Eligible prompts observed |
| Linked-citation share | Prompts linking a brand page | Eligible prompts observed |
| Accurate-attribution rate | Supported brand citations | Brand citations reviewed |
Report engine, locale, intent and topic cuts only when sample sizes remain visible. A 100 percent result from one prompt should never be presented like a stable market-level finding.
How should competitors be compared?
Compare competitors against the same answer and scoring rules, not against separate prompt panels. Record all eligible brands found in each answer and preserve their order only if position is a defined metric.
Entity resolution is essential. Brand names, parent groups, product lines and domain variants can otherwise split one competitor into several rows. Maintain an alias table and review mergers or rebrands.
Do not assume the most mentioned company receives the most clicks or leads. Share of voice describes answer presence. It does not measure audience size, sentiment, commercial fit or conversion.
How does share of voice connect to business outcomes?
Connect AI visibility to referral, assisted and self-reported outcomes as separate evidence layers. A cited page may generate trackable sessions. A mention without a link may influence later brand search. Neither path is fully observable.
Use SAGEO's guide to tracking ChatGPT traffic and conversions to preserve source, landing page and CRM outcomes. Keep crawler requests outside human referral counts.
The AI search dashboard framework separates technical eligibility, answer presence and business impact. That separation prevents a visibility gain from being reported as revenue before the evidence exists.
What should an executive dashboard show?
An executive dashboard should show trend, evidence quality and business consequence without hiding methodology changes. Include:
- mention, recommendation and linked-citation share;
- prompt count and successful observation rate;
- engine and market coverage;
- persistence across repeat samples;
- cited landing pages and attribution accuracy;
- qualified referral sessions and accepted leads;
- prompt-set or scoring-rule changes.
Show selected answer evidence beside the chart. A metric that cannot be traced to the underlying output should not drive a content or investment decision.
Want a measurement baseline? SAGEO can design the prompt taxonomy, evidence store and reporting rules before optimisation begins.
What are the limitations of AI share of voice?
AI search share of voice is a sampled observation, not a census of every answer shown to every user. Results can vary by time, model, location, personalisation, account state and product changes. Prompt panels also reflect researcher choices.
Report the metric as directional evidence within a declared method. Do not convert it into estimated market share, reach or revenue without separate validated data.