Share of voice is a familiar idea in advertising and SEO: what portion of the conversation in your market belongs to you. In AI answers, the concept is just as useful, but the counting rules need more care. Answers are generated, they vary, and they mention brands in very different ways.
This article sets out a method you can apply with a spreadsheet, and explains the decisions that will make your numbers either reliable or misleading.
The basic definition
For a given question set, a given list of brands, a given set of assistants and a given period:
Share of voice = your brand mentions / total mentions of all tracked brands
If assistants mention your brand and your four competitors 200 times in total across the period, and 40 of those mentions are yours, your share of voice is 20 percent.
That formula is simple. The method around it is where the work lies.
Step 1: Fix the scope
Share of voice only means something relative to a scope. Write down:
- The question set. The exact questions, in the exact wording.
- The tracked brands. You plus your direct competitors.
- The assistants. For example ChatGPT, Gemini and Perplexity.
- The market. Country and language.
- The period. A month, two weeks, a week.
Change any of these and you have a new measurement. A share of voice of 20 percent on 30 category questions in the UK is not comparable to 20 percent on 80 questions in France.
Step 2: Define what counts as a mention
This is the step most teams skip, and it causes most of the noise. Agree on rules before you count.
Brand variants. Decide which names map to each brand: “Acme”, “Acme CRM”, “acme.com”. Keep a list of aliases per brand.
Products vs company. If a competitor has several product names, decide whether each counts toward the parent brand.
Citations vs mentions. An answer can name a brand in its text, or only link to its website as a source. Track these separately. A brand that is cited as a source but never named in the text is in a weaker position than one that is recommended by name.
Negative mentions. “Avoid X, it has poor support” is still a mention. Count it, but record its sentiment, so it does not inflate a positive reading.
One mention per answer or every occurrence? Counting every occurrence rewards brands that the assistant discusses at length. Counting once per answer measures breadth. Both are valid. Choose one and stick to it. For most teams, once per answer is easier to interpret.
Step 3: Choose a weighting
Raw counts treat a first-place recommendation and a passing mention the same. A weighted version reflects prominence.
| Approach | How it works | When to use it |
|---|---|---|
| Unweighted | Each mention counts as 1 | Starting out, small question sets |
| Position-weighted | Higher weight for earlier positions in a list (for example 1st = 3, 2nd = 2, others = 1) | Recommendation-heavy questions |
| Intent-weighted | Higher weight for questions closer to purchase | When some questions matter more commercially |
| Sentiment-adjusted | Negative mentions count as zero or are reported separately | Markets where reputation is contested |
The weights in the table are an example, not a standard. Whatever you choose, document it and keep it stable over time. Changing the weighting mid-year breaks your trend line.
Step 4: Sample properly
AI answers vary between runs. The same question asked twice can return different brands or a different order. That means a share of voice figure is an estimate from a sample, not an exact count.
Practical ways to reduce noise:
- Use more questions rather than fewer. Variation on individual answers evens out across a larger set.
- Measure on a schedule. Monthly, twice-monthly or weekly runs let you see whether a movement persists.
- Keep conditions identical. Same wording, same market, same language, each time.
- Report the sample size. “20 percent across 150 answers” is more honest than “20 percent”.
Be wary of reacting to one period. A competitor jumping by a few points in a single run may simply be variation.
Step 5: Break it down
A single overall share of voice hides the useful detail. Break it down by:
- Assistant. You may be strong on one assistant and nearly absent on another. That often points to different sources being used.
- Topic group. Group questions by product line, use case or problem. Low share on one group is a concrete content or positioning gap.
- Intent. Category discovery (“best X”), comparison (“X vs Y”), and validation (“is X good?”) behave differently.
- Market. Each country or language has its own competitive set and its own sources.
A worked example
Imagine a bookkeeping software brand tracking 40 questions in the UK across three assistants, with four competitors. Across the period:
- Overall share of voice: moderate, second of five brands.
- By assistant: strong on one, weak on the other two.
- By topic: strong on “invoicing”, very weak on “VAT returns”.
- Cited sources on “VAT returns” questions: mostly comparison articles and a public forum where the brand is not discussed.
The action is not “improve AI visibility” in general. It is specific: publish a clear VAT returns page, and work on being present in the comparison articles that assistants cite for those questions.
Step 6: Pair share of voice with sources
Share of voice tells you where you stand. Cited sources tell you why. For each topic group, list the domains that assistants cite most often, and check whether your brand is present and accurately described on them.
This turns a metric into a to-do list: your own pages to strengthen, third-party pages to earn presence on, and inaccurate descriptions to correct.
Common mistakes
- Comparing across scopes. Different questions or competitors make numbers incomparable.
- Counting aliases inconsistently. Missing a brand variant undercounts that brand.
- Mixing citations and mentions. They signal different things.
- Over-reading single runs. Variation is normal.
- Tracking only giants. Include the competitors you actually meet, not only the category leaders.
Checklist
- Scope written down (questions, brands, assistants, market, period)
- Alias list per brand
- Clear rule for citations vs mentions
- One counting rule (per answer or per occurrence)
- Documented weighting, kept stable
- Sample size reported with every figure
- Breakdown by assistant, topic, intent and market
- Sources reviewed for each weak topic group
How SearchFirefly helps
SearchFirefly asks your tracked questions across ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode, in each market’s language, and calculates your AI visibility and share of voice against the competitors you track. For every question you can see the answer, the brands it names and the sources it cites, market by market, with prioritized recommendations for the gaps that matter most.