AI Share of Voice: How AIsee Frames Competitive Visibility

Define AI share of voice before calculating it

AI share of voice (AI SOV) is a relative visibility measure. It compares one brand's qualifying mentions with the total qualifying mentions earned by all brands in a specified competitive universe.

The basic formula is:

AI SOV = brand qualifying mentions / all tracked brands' qualifying mentions × 100

Suppose 50 generated responses contain 80 qualifying mentions across five tracked vendors. ExampleCo receives 20 of those mentions. Its unweighted AI share of voice is 20 / 80 × 100 = 25%. This is an illustrative formula, not AIsee product or customer data.

Another valid design uses eligible responses as the denominator: responses mentioning brand / all eligible responses × 100. That is better named brand mention rate, not competitive share of voice. Both metrics are useful, but mixing their denominators makes comparisons misleading.

AI SOV describes observable presence in the sample; it does not by itself prove awareness, preference, accuracy, traffic, or revenue impact. Pair it with citation quality, answer accuracy, sentiment, referral sessions, and conversion measures from the broader GEO KPI framework.

The denominator determines what the metric means

Document six boundaries before reporting AI SOV:

1. Prompt universe: the exact questions included and their intent categories.
2. Competitive set: the brands eligible for the denominator, including rules for newly discovered competitors.
3. Surfaces: the engines, modes, or products tested.
4. Market: language, geography, audience, and product segment.
5. Time window: test dates and aggregation period.
6. Qualification rule: what counts as a mention and how aliases, parent companies, product names, and duplicate mentions are handled.

A practical rule is to count each eligible brand at most once per response. This avoids giving extra weight to repetitive prose. Map common aliases to a canonical entity, but do not merge distinct products unless the research question calls for a parent-brand view.

Decide whether untracked brands join the denominator. A closed-set SOV compares only a predeclared competitor list and supports stable trend reporting. An open-set SOV counts every qualifying brand found in the responses and is better for market discovery. Report them separately; adding a new competitor can lower historical-looking share even when the focal brand’s mentions do not change.

Exclude prompts for which no tracked solution is relevant, or retain them in a separate “no qualifying brand” rate. Never force a mention into the denominator when the correct answer is category advice without a vendor recommendation.

Use weighting only when the business logic is explicit

Unweighted AI SOV treats every prompt and every qualifying mention equally. That is transparent and should remain the default trend line. A weighted version can reflect strategic differences, but the weights must be defined before reviewing the result.

One prompt-weighted formula is:

Weighted AI SOV = Σ(brand mention × prompt weight) / Σ(all brand mentions × prompt weight) × 100

Example: a purchase-comparison prompt has weight 3, a category-education prompt has weight 2, and a broad discovery prompt has weight 1. ExampleCo earns two purchase mentions, four education mentions, and five discovery mentions. Its weighted mention points are (2×3) + (4×2) + (5×1) = 19. If all brands earn 60 weighted points, ExampleCo's weighted SOV is 19 / 60 = 31.7%. This is an illustrative calculation, not AIsee product or customer data.

Common weighting dimensions include journey stage, qualified demand, market priority, product segment, and answer prominence. Avoid weights based on the outcome you hope to show. Do not multiply several weak estimates until the metric becomes impossible to explain.

Keep a companion unweighted SOV and show results by segment. If weighted share rises while unweighted share falls, the brand may be gaining in high-priority questions while losing broad discovery. That is a useful strategic signal, not a reason to hide one series.

Build a stable and representative prompt set

AI SOV is only as representative as its prompts. Start with observed audience language: Search Console queries, customer interviews, support cases, sales-call questions, category research, and on-site search. Remove prompts that differ only cosmetically, then stratify the set.

• Problem prompts: describe the need without naming a solution category.
• Category prompts: ask what a solution type is or how it works.
• Comparison prompts: evaluate approaches, features, or vendors.
• Validation prompts: ask about evidence, limitations, integrations, security, or fit.
• Purchase prompts: request recommendations for a defined scenario.

Separate branded prompts from unbranded prompts. “What does AIsee do?” tests entity accuracy and owned-source control; it should not inflate discovery SOV. Also separate prompts whose wording names a competitor, because the mention is partly induced.

Give each prompt a stable ID, version the set, and freeze the core cohort for trend reporting. Add a smaller exploratory cohort to discover emerging questions and brands. Promote a new prompt into the core set only through a documented review, then annotate the series.

Generated answers vary. Repeat each prompt under consistent conditions and report the number of observations. Document the surface, date, login state, language, location, and conversational context. The goal is not to simulate every user; it is to create a reproducible measurement frame.

Use an auditable tracking protocol

Start with the AIsee report to identify which competitive, provider, content, or entity questions deserve a controlled sample. Then store the raw answer and cited URLs before assigning labels. A compact record includes:

Use deterministic alias rules where possible, then manually review ambiguous names. Calibrate reviewers on the same sample and preserve "not sure" rather than forcing a classification.

Calculate SOV by engine, prompt stage, market, and product segment before rolling up a total. Pair mention records with the AI citation tracking dataset so the team can see which owned and third-party sources accompany the visibility.

Keep the AIsee Competitor Landscape value beside the formal SOV series, not inside its numerator or denominator. The diagnostic can prioritize an investigation; the controlled prompt panel supplies the evidence for a competitive mention rate.

Interpret movement before taking action

When AI SOV changes, decompose the movement:

1. Sampling: Did the prompt set, engine mix, market, run count, or qualification rule change?
2. Competitive mix: Did a new brand enter the open set, or did one competitor gain across a narrow cluster?
3. Answer quality: Are new mentions accurate, neutral, and supported, or are they associated with incorrect claims?
4. Source pattern: Which URLs and domains are cited when the brand appears or disappears?
5. Content coverage: Does the brand have an authoritative page that directly answers the losing prompt cluster?

Choose the intervention that matches the cause. Repair inconsistent entity facts when descriptions are wrong. Strengthen evidence and direct passages when third-party pages are cited for facts the owned site should establish. Create a new page only when a material audience question lacks a useful canonical answer; do not manufacture near-duplicate pages for every prompt variation. Improve internal links and crawlability when the source exists but is isolated.

Set targets as directional hypotheses tied to a fixed protocol, not as universal benchmarks. For example: “Increase accurate closed-set SOV for the 30 unbranded evaluation prompts over three monthly samples while maintaining a material-error rate below the documented baseline.”

Finally, keep SEO outcomes beside AI SOV. Google’s official guidance says standard SEO practices remain relevant to its AI search features, and eligible pages still require search indexation. Use the GEO vs SEO vs AEO comparison to connect competitive answer visibility with rankings, clicks, and qualified demand.

Frequently asked questions