Our position: share of voice is a comparison inside a test, not a census of everything AI users saw.
What you should leave with
- Define an eligible recommendation opportunity.
- Keep citations and mentions separate.
- Freeze weights before collection.
- Show prompt-family distributions.
How is AI share of voice calculated?
Count your brand's qualifying recommendation outcomes and divide by all qualifying brand recommendation outcomes in the same controlled set. Define whether multiple brands per answer, order, and prompt value affect the calculation.
A simple unweighted formula is brand recommendations divided by total recommendations among tracked brands. A weighted version can reflect prompt value or platform importance, but weights must be justified and locked before results. Publish both when stakeholders need transparency.
The competitor universe matters. Include real direct alternatives and allow discovery of unexpected brands, but do not change the denominator silently. Report newly discovered competitors as a separate finding until the next baseline version.
- Defined qualifying recommendation
- Visible competitor universe
- Prompt and platform denominator
- Precommitted weighting policy
Evidence used in this section
What should count as share of voice?
Count direct recommendations and clearly defined shortlist inclusions; report comparisons, mentions, and citations as separate measures. Use sentiment and buyer-fit review so negative or irrelevant appearances do not become wins.
Answer engines often name several options with caveats. Decide whether list order or recommendation strength receives extra weight and document the rule. A brand listed as unsuitable for the target buyer should not earn the same credit as the top fitted choice.
Resolve products and parent brands consistently. If the buyer selects a product but the dashboard tracks the parent company, state the roll-up rule. Entity ambiguity is a measurement issue, not a rounding error.
- Direct answer and shortlist role
- Buyer-fit and sentiment
- Position or strength policy
- Product-parent entity policy
Evidence used in this section
How do you build a defensible share-of-voice baseline?
Map valuable buyer decisions, approve prompts and competitors, freeze classification and weighting, run controlled platform tests, review ambiguous results, and repeat high-impact outcomes. Store the raw evidence behind every count.
Balance breadth and depth. Prompt families should cover segments and constraints without multiplying cosmetic wording variants. Repeats should concentrate on decisions where one fluctuating answer would materially change the conclusion.
Keep an exploratory set outside the baseline. It can discover emerging language and competitors without contaminating trend continuity. Promote changes only at a documented version boundary.
- STEP 1
Model
Define buyers, decisions, competitors, qualifying roles, and entity rules.
- STEP 2
Freeze
Approve prompts, platforms, weights, scope, and repeat policy before results.
- STEP 3
Review
Validate recommendation roles, fit, sentiment, entities, and uncertain answers.
- STEP 4
Publish
Show totals, distributions, evidence, confidence, and method version together.
Evidence used in this section
How should AI share of voice be reported?
Show the overall result beside unweighted counts, prompt-family breakdowns, platform results, competitor distribution, repeat stability, and changes in the denominator. One percentage should never stand alone.
A five-point gain can come from one valuable new recommendation or several low-value variants. The distribution tells stakeholders whether the market position actually improved. Include gross wins and losses rather than only the net change.
Confidence should reflect prompt coverage, repeats, reviewer certainty, and stability. Small denominators need plain-language caution even when the arithmetic is exact to two decimal places.
| View | What it reveals | Required disclosure |
|---|---|---|
| Overall share | Relative presence in the tested market | Denominator and weights |
| Prompt family | Where the brand wins or disappears | Family size and value |
| Platform split | Environment-specific divergence | Comparable run context |
Evidence used in this section
What can share of voice not tell you?
It cannot reveal total answer impressions, prove why a model chose a brand, measure every personalized response, or establish revenue impact. It is an observed relative outcome within a disclosed test.
Do not compare percentages across vendors unless prompts, platforms, competitors, roles, weights, and run policy align. Two tools can use the same label for fundamentally different measurements.
Use share of voice to locate the gap, then inspect recommendation reasons and sources. Improving the percentage without understanding the evidence can reward low-value mentions instead of better buyer outcomes.
- No audience estimate
- No causal ranking explanation
- No universal category benchmark
- No direct revenue attribution
Method boundary: A share-of-voice result is only as representative as its prompt set. Disclose who and what the test does not cover.
Evidence used in this section
Questions that change the decision
Frequently asked questions
Should list position affect AI share of voice?
It can, if the policy is set before collection and the answer format makes order meaningful. Also publish simple recommendation counts for transparency.
Do citations count as share of voice?
Track citation share separately. A source link may support background information without placing the brand in the recommended set.
How many competitors should be tracked?
Track the direct set buyers realistically consider and log unexpected brands separately. Too narrow a set overstates share; an indiscriminate set dilutes relevance.
What is a good AI share of voice?
There is no universal threshold. Compare against direct competitors and your own stable baseline on valuable prompt families.
Primary sources and research
Platform documentation supports factual statements. Where we describe an audit method or prioritization rule, that is AnswerMentions' operating judgment and is labeled as such.
- [1]Aggarwal et al.: Generative Engine OptimizationThe KDD 2024 paper evaluates generative-engine visibility in a controlled benchmark; it is evidence that visibility can be studied, not a universal ranking recipe.
- [2]NIST: AI Risk Management FrameworkNIST frames AI risk work around governance, mapping, measurement, and management, a useful model for separating observations from decisions.
- [3]OpenAI Help: accuracy and citationsOpenAI warns that ChatGPT can produce incorrect facts and fabricated references, so consequential claims should be checked against reliable sources.
- [4]FTC: advertising and marketing basicsThe FTC states that advertising claims must be truthful, non-deceptive, and supported by evidence when appropriate.