What Is Share of Voice in AI Search?
Share of voice in AI answers is a proportion within your own prompt set and your own competitor list. Here is the exact formula and what it cannot mean.
Share of voice is the most useful comparative metric in AI answer monitoring and the easiest one to overstate. It compares how often a model names you against how often it names a set of competitors — but only the competitors you listed, only on the questions you asked, and only under the conditions you recorded. Get those three qualifiers into the sentence and the number is genuinely informative. Drop them and it becomes a market-share claim nobody can support.
The definition
Share of voice in SiteRank AI is a proportion within a user-defined tracked set of entities, computed over a user-defined prompt set.
share_of_voice(X) = observations mentioning entity X
──────────────────────────────────────────
observations mentioning any tracked entityBoth parts of that definition are user-supplied, and that is the whole point of the metric's honesty. The numerator depends on which aliases you declared for entity X. The denominator depends on which competitors you decided to track. Neither is discovered by the tool, and both are shown next to the figure.
The observations counted are valid observations only — provider calls that completed with an ok status. Errors and timeouts are excluded from both parts, and the excluded count is reported. A failed call is not evidence that a model preferred a competitor.
Why the denominator is "any tracked entity", not "all observations"
There is an obvious alternative: divide by all valid observations. That produces a different quantity, and a less useful one.
Suppose 40 valid observations, of which 12 name nobody in your tracked set at all — the model answered generically, or named companies you never listed. Dividing by 40 mixes two separate facts into one number: how often the model discusses this category in terms of named vendors, and how you fare among the vendors it names. Dividing by the 28 observations that mention someone tracked isolates the second.
The first fact is still worth having, and it is exactly what mention rate reports against the full valid denominator. Keeping the two separate means a drop in share of voice can be distinguished from a drop in whether the category is being answered with vendor names at all.
A worked example
Ten unbranded category prompts, sampled four times each, against one model with web search enabled, over one calendar month. Four entities in the tracked set: your site plus three named competitors.
| Quantity | Value |
|---|---|
| Attempted calls | 45 |
Excluded (non-ok) |
5 |
| Valid observations | 40 |
| Observations mentioning at least one tracked entity | 28 |
| Entity | Observations mentioning | Share of voice | 95% CI |
|---|---|---|---|
| Your site | 9 | 32.1% | 17.9% – 50.7% |
| Competitor A | 12 | 42.9% | 26.4% – 60.9% |
| Competitor B | 9 | 32.1% | 17.9% – 50.7% |
| Competitor C | 3 | 10.7% | 3.7% – 27.2% |
Two things to notice immediately.
The shares sum to more than 100%. 32.1 + 42.9 + 32.1 + 10.7 = 117.8%. That is not an error. A single answer frequently names several vendors, so one observation can appear in several numerators while counting once in the shared denominator. Share of voice here is the proportion of vendor-naming answers in which you appear, not a slice of a fixed pie. Any tool whose shares always total exactly 100% has either forced them into a pie or is counting something else — most likely first mention only, which measures sentence order rather than inclusion.
The intervals overlap almost completely. Your 32.1% and Competitor A's 42.9% are statistically indistinguishable at n=28. Reporting "we are second" from this data would be a fabrication dressed as a finding. What the data supports is: at this sample size, four vendors are present in this category's answers, one appears somewhat less often than the rest, and the top three cannot be separated.
Citation share is a different metric with a different unit
Alongside share of voice, SiteRank AI reports citation share, and the two are frequently confused because both are "our slice of the tracked set".
citation_share(D) = citations to domain D
──────────────────────────────
citations to any tracked domain| Metric | Unit counted | Denominator |
|---|---|---|
| Share of voice | Observations | Observations mentioning any tracked entity |
| Citation share | Individual citations | Citations to any tracked domain |
The unit is the difference. Share of voice counts answers; citation share counts links. An answer that cites your domain three times contributes three to a citation-share numerator and one to a share-of-voice numerator. The two numbers are therefore not comparable to each other, and a dashboard that puts them side by side without labelling the unit invites exactly the wrong comparison.
Continuing the same window: 210 citations to tracked domains in total, of which 24 point to your domain.
citation_share = 24 / 210 = 11.4% (95% CI 7.8% – 16.4%)One statistical caveat worth stating in the interface as well as the article: citations within a single answer are not independent draws. Six citations from one response reflect one retrieval event, not six. The binomial interval above therefore understates the true uncertainty, and should be read as a floor. The same is true, more mildly, of repeated samples of the same prompt — which is why volatility is reported alongside.
What share of voice is not
It is not market share
Nothing in this calculation observes a market. It observes a set of model responses to a set of questions someone chose. A company with 60% share of voice on ten prompts may have a rounding error's worth of customers; a market leader absent from your prompt set has 0% share of voice and is still the market leader. The two quantities are unrelated, and presenting one as evidence of the other is the single most common misuse of this metric.
It is not stable when the inputs change
Both denominators move when you change your inputs:
| Change | Effect |
|---|---|
| Add a competitor to the tracked set | The denominator grows; your share falls, with no change on your site |
| Remove a competitor | Your share rises, again with no change on your site |
| Add or reword a prompt | The observation population changes; comparability to prior periods breaks |
| Change the model or the web-search setting | Different instrument; the series should show a break, not a continuous line |
This is why the tracked set and the prompt set are versioned, and why a change to either marks the chart. A share-of-voice series with a silent competitor addition in the middle is two series drawn as one, and it will read as a decline that never happened. Deciding who belongs in the set in the first place is its own problem, covered in choosing competitors for LLM visibility monitoring.
It is not a ranking
Ordering the four entities in the table above by their point estimates produces a list, and a list looks like a ranking. It is not one. The ordering is unstable at these sample sizes — the overlapping intervals say so — and there is no ordered result set inside a generative answer to rank against in the first place. SiteRank AI does not use the word "rank" for generative answers, and neither should a report built on this data.
It is not causal
If your share rises from 32% to 45% across two months, the candidate explanations include: you published better content; a competitor's content was deindexed or deprioritised; the provider changed the served model; the retrieval index refreshed; or you sampled a favourable set of draws from an unchanged process. A monitoring series cannot separate these. Report the change, the interval, and the volatility, and stop there. The mechanics of that discipline are set out in how LLM visibility monitoring actually works.
Reading the number well
A few habits make share of voice useful rather than decorative.
Read the denominator first. "32.1% of 28" and "32.1% of 4" are the same percentage and completely different evidence.
Read it against mention rate. Share of voice can rise while mention rate falls, if the model started naming fewer vendors overall and you survived the cut. Both numbers together tell you whether the category is contracting or you are gaining.
Segment by prompt type. Branded prompts almost guarantee a mention and will inflate your share against competitors who are not named in the prompt. Unbranded category prompts are the comparison that means something.
Watch volatility before celebrating a delta. Volatility is the proportion of samples that disagree with the majority for the same prompt under identical conditions — same prompt id and version, provider, returned model, web-search flag and locale. High volatility means single observations are close to coin flips, and a month-over-month move of a few points is consistent with nothing having changed.
Never quote it without its scope. The defensible sentence names the prompt set, the tracked set, the model, the web-search setting, the window, and both denominators. Anything shorter is a slogan.
Evidence tiers
| Statement | Tier |
|---|---|
| Share of voice as defined here is a proportion computable from stored observations | ESTABLISHED STANDARD — arithmetic over recorded data |
| Excluding failed calls from both numerator and denominator | ESTABLISHED STANDARD |
| Reporting a 95% Wilson interval at the actual n | ESTABLISHED STANDARD |
| Unbranded prompts give a more comparable share than branded prompts | EMERGING PRACTICE |
| Share of voice in model answers relates to commercial outcomes | HYPOTHESIS — plausible, undemonstrated, and not scored as a deficiency anywhere in SiteRank AI |
The last row is the one to keep in view. There is no published evidence linking share of voice in generated answers to revenue, and no provider documents how entities are selected for inclusion in an answer. Google states that there are no additional requirements or special optimisations needed to appear in its AI features beyond standard search practice, and that no special structured data is required for them — which is a useful corrective to any product that presents share of voice as a lever with a documented mechanism behind it.
Key takeaways
- Share of voice = observations mentioning entity X ÷ observations mentioning any tracked entity, over valid observations only.
- Both denominators are user-defined, both are shown, and changing either changes the number without anything changing on your site.
- Shares can sum to more than 100%, because one answer can name several vendors. That is correct behaviour, not a bug.
- Citation share counts citations, not observations. Different unit, not comparable to share of voice.
- Overlapping confidence intervals mean no ordering, and no ordering means no ranking.
- It is not market share, and no published evidence connects it to commercial outcomes.
Official sources & further reading
- AI features and your website — Google Search Central
- Web search tool — OpenAI
- Overview of OpenAI crawlers — OpenAI
- Does Anthropic crawl data from the web, and how can site owners block the crawler? — Anthropic
- Perplexity crawlers — Perplexity
Related reading
Frequently asked questions
Are we counted if a model names us but does not link to us?
Yes — share of voice counts observations in which the entity is mentioned, whether or not a link accompanies it. Linking is counted separately by citation share, whose unit is the individual citation and whose denominator is citations to any tracked domain. A site can therefore hold a reasonable share of voice alongside a low citation share, which usually means the model is willing to name you but is sourcing the answer somewhere else.
Should I put our brand name in the prompts to improve the number?
Not if you want the number to mean anything. A prompt containing your brand name nearly guarantees a mention, so branded prompts inflate your share against competitors the prompt never names. Keep branded and unbranded prompts in separate segments and treat the unbranded category prompts as the comparison that carries information. Because the prompt set is versioned, adding branded prompts mid-series also breaks comparability with earlier periods.