One Answer, Four Brands: What Should AI Share of Voice Count?

An AI answer recommends your company, names three competitors, and repeats your company name in a comparison table. Did you win one mention, two mentions, a quarter of the answer, or the entire answer?

Each result can follow a different counting rule. The useful question is which rule the report applies, whether it applies that rule consistently, and whether the number answers the business question you actually have.

For Generative Engine Optimization, or GEO, a share-of-voice figure needs a defined population, a brand universe, and a counting unit. A percentage without those three things cannot be reproduced from an answer archive.

This article develops a small, fictional example. The companies and answers below are invented to explain the arithmetic; they are not observations of any platform or customer.

Start with the decision the metric should support

A marketing director may want to know whether buyers encounter the brand at all. A competitor analyst may want to know how often it appears relative to a fixed group. A sales team may care whether the answer recommends the company for the requested application.

Those are three different questions. Answer-level presence, competitive share, and qualified recommendation should therefore remain separate columns, even when they appear on one dashboard.

Peec AI's dashboard documentation distinguishes visibility from share of tracked-brand mentions, position, sentiment, and use as a source. That is one vendor's published terminology. It does not establish a universal implementation for every system using the same labels. In particular, ask how repeated occurrences, product aliases, and parent companies affect the counter.

A four-answer dataset

Suppose the review population contains four complete, valid answers to agreed buyer questions. The tracked companies are Aster, Beacon, and Cedar. Drift is visible in one answer but has not been added to the tracked group.

The fourth answer belongs in answer-level visibility. The absence of a tracked brand is a substantive observation, not a reason to delete the answer. The source-card title in R3 needs a different field from a narrative mention.

An incomplete response or a login failure belongs in a collection-coverage report. It should not silently become another complete answer with no brand.

Three reproducible calculations

Answer-level visibility

Count each complete answer once for each tracked company present in its narrative. Aster appears in R1 and R2, so its visibility is 2/4, or 50%.

Beacon is also 2/4. Cedar is 1/4. These percentages sum to 125%. That is not necessarily an error: several companies can appear in the same answer. They are overlapping presence rates, not slices of a single exclusive pie.

Share of brand-answer incidences

Define one incidence as one tracked company appearing in one answer, regardless of repetition within that answer. The incidence counts are Aster 2, Beacon 2, and Cedar 1, for a total of five.

Aster's incidence share is therefore 2/5, or 40%. Its presence rate remains 50%. Neither number changes the underlying answer archive.

Share of name occurrences

Now count every recognized narrative occurrence. If Aster appears twice in R1 and once in R2, its count becomes three. Beacon has two and Cedar one. Aster's occurrence share is 3/6, or 50%.

That calculation rewards repeated name text within an answer. It may be useful for a particular analysis, but it should not be presented as the same measure as incidence share. A comparison table, a heading, or a repeated summary can alter it without introducing a new recommendation.

These are explicit definitions for this example. They are not claims about a particular vendor's internal calculation.

A competitor-setting change can move the score

Add Drift to the tracked universe without collecting a single new answer. The incidence denominator rises from five to six. Aster's incidence share falls from 40% to approximately 33.3%, although Aster still appears in exactly two of four answers.

This is why competitor settings belong in report metadata. A chart that changes after a competitor is added may be showing a population-definition change, a historic recomputation, or both. It does not automatically show that the AI has become less likely to mention your company.

Keep the earlier calculation and the updated calculation with their respective universe versions. If a client asks for a trend, either hold the universe constant over the comparison or explicitly label the break. Do not splice the two definitions into an uninterrupted performance story.

Recommendations require a second review

Consider an answer that says Aster sells filtration components but is unsuitable for the requested fluid. That may be a correct narrative mention. It is not a favorable supplier recommendation for that application.

For a qualified-recommendation count, retain the relevant passage, identify the entity, and verify the individual suitability reason. A source link alone does not complete that review. A correctly spelled name with an invented capability should not receive qualified credit.

The guide to classifying AI answers explains that distinction. The counting question here comes afterward: once a record has been classified, what precisely enters each metric?

Put the definition beside the chart

A reproducible report should identify the exact question version, collection surface, period, valid-answer count, exclusions, tracked-brand universe, entity-matching rules, and counting rule. If a result combines markets or models, show the component counts as well as the aggregate.

Retain answer-to-brand rows separately from answer-to-citation rows. Joining both one-to-many tables before counting can multiply records. Two brands and three citations in one answer may create six joined rows; that does not mean six answers or six independent endorsements.

For a software acceptance test, give the provider these four fictional answers and an explicit brand universe. Ask it to explain the numerator and denominator, export the supporting rows, and show what happens when Drift is added. The objective is a calculation you can inspect, not a particular percentage.

Semvi by Xindar can be considered within that procurement process. Its product team confirms complete-answer and citation-URL exports; the calculation rules and an actual account export still need acceptance. A product description does not prove that the exact counting convention proposed here is implemented.

A practical acceptance record

Write one short definition before collecting the pilot: “A brand-answer incidence is a verified narrative mention of one tracked entity in one valid answer; repeated occurrences count once.” Then specify whether negative mentions qualify as presence, how product names roll up to a company, and how ambiguous names are held for review.

Have a second person reconstruct the four-answer example from the export. Require the same 50% presence rate and 40% incidence share under the original universe, then the 33.3% share after adding Drift. Preserve the original universe and outputs. This is a small calculation test, not a statistical validation of market visibility.

Questions buyers should ask

Should visibility percentages add to 100%?

Only if the measure assigns each observation exclusively to one category. Brand presence usually allows multiple companies in an answer, so overlapping rates need not sum to 100%.

Should a company receive credit for every linked page?

A citation count can record every qualifying link. Answer-level brand presence should still count the answer once under the stated rule. Keep citation frequency and brand presence separate.

Can I compare two vendors' share-of-voice scores?

First align their questions, dates, surfaces, valid-answer population, competitor universe, entity mapping, and formula. If the definitions differ, compare the underlying records or report the scores with their different definitions.

Does more share of voice prove more sales?

No. The metric describes the sampled answer population. Visits, inquiries, qualified opportunities, and orders require their own evidence and appropriate attribution boundaries.

*Method and source review: November 9, 2026. Written by Daoyu Guan. All numerical examples are fictional, independently reproducible arithmetic; no vendor account test or customer performance is claimed.*