Skip to content

Guides

GEO metrics explained: share of voice, visibility, position

Share of voice, visibility and average position are computed by the vendor over stored answers, and five API references publish five different denominators.

Three transparent cylinders holding stacked cubes, the center one filled with blue blocks, outer ones mostly empty.

SEO consultant, and more recently GEO, at MentionBox

Share of voice, visibility and average position are computed by the measurement vendor, not by the AI engine. Each one is a count over a stored set of answers, and the formula sits in the API reference: visibility is the share of answers that name your brand, share of voice is your mentions divided by every brand's mentions, and average position is the mean rank of your brand inside the answers that name it. The denominators differ from one vendor to the next, which is why the same 200 collected answers can be reported as 34.0% visibility by one API and 40.0% by another, with both figures correct under their own definition.

This page is written for whoever reads one of these numbers in a monthly report, and for whoever writes the code that produces it. It goes through one real API response field by field, then the published formula of each metric, the arithmetic on a 200-answer month, the free figure Google now publishes, and the mistakes that make two tools disagree about the same brand.

Three numbers on one API response

API for GEO publishes the fields with their definitions on its AI Visibility API page, which makes it a usable worked example. Its POST /v1/visibility/entities returns one row per tracked entity plus a time series:

{
  "results": [
    {
      "name": "Loopwork",
      "entityType": "brand",
      "visibility": 68.4,
      "shareOfVoice": 24.1,
      "avgPosition": 1.8,
      "mentions": 137,
      "sentiment": { "score": 76, "label": "positive" }
    }
  ],
  "timeseries": [
    { "bucket": "2026-08-10", "visibility": 64.2, "avgPosition": 2.1 }
  ]
}
FieldDefinition published next to it
visibility"Percentage of in-scope responses mentioning the entity."
shareOfVoice"Entity mentions ÷ all entity mentions in scope."
avgPosition"Mean rank of the first mention inside the answer. Lower is better."
mentions"Raw mention count across in-scope responses."
sentiment"Score from 0 to 100 and its derived label."

Field definitions from the API for GEO AI Visibility API page, read on October 31, 2026. MentionBox, which publishes this site, has a commercial link with API for GEO.

The load-bearing words in that table are "in scope". The filters sent with the request define the set of stored answers the four numbers are computed over, so a date range, an engine, a country or a prompt tag changes all of them at once. The same vendor's model comparison endpoint narrows the scope to one engine and describes share of voice there as "entity mentions ÷ all mentions within this model". A per-engine share of voice and an all-engine share of voice are two different quantities, and the field name says nothing about which one is on screen.

What the denominator changes, in numbers

Take one month on one engine: 200 stored answers for your prompt list, 170 of which name at least one brand. Your brand appears in 68 answers, with 78 mentions, against 540 brand mentions in total. Its first mention ranks first in 20 answers, second in 24 and third in 24.

FigureFormulaResult on these 200 answers
Visibility, all answers in the denominatoranswers naming you ÷ all collected answers68 ÷ 200 = 34.0%
Visibility, brand-bearing answers onlyanswers naming you ÷ answers naming at least one brand68 ÷ 170 = 40.0%
Share of voiceyour mentions ÷ all brand mentions78 ÷ 540 = 14.4%
Average positionsum of your first-mention ranks ÷ answers naming you140 ÷ 68 = 2.1
Prominence score, Scrunch scaletop of response 100, middle 50, bottom 0, averaged3,200 ÷ 68 = 47

Formulas quoted from the API references linked in this article, read on October 31, 2026. The counts are an illustration, chosen to be arithmetically consistent.

Six percentage points of visibility separate the first two rows, and no collection difference produced them. The gap grows with the number of answers that name nobody, which is exactly the population that moves when an engine starts answering a category generically.

The same names in five other API references

API referenceVisibilityShare of voicePosition
Peec AI brands reportsum(visibility_count) / sum(visibility_total), documented elsewhere as responses mentioning your brand ÷ total responses × 100mention_count / sum(mention_count), returned as a 0 to 1 ratiosum(position_sum) / sum(position_count), 1 is best
Profound glossary"The percentage of responses that mention your brand, out of all responses that mention at least one brand""How often your brand is mentioned compared with all brands""The average position of your brand relative to competitors across responses"
OtterlyAI brand report KPIsbrand coverage: prompts mentioning the brand ÷ all prompts in the windowmy brand mentions ÷ total brand mentions × 100sum of brand mention positions ÷ number of prompts where the brand appeared
Scrunch metrics referencebrand_presence_percentage, "share of responses whose text names your brand, regardless of tone"brand_share_of_voice, mentions as a share of the active competitive set, "normalized so every entity's slice sums to 1"brand_position_score 0 to 100: "top-of-response = 100, middle = 50, bottom = 0"
Ahrefs AI visibility metricsimpressions, "calculated by summing up search volumes (based on Google) of prompts where your brand appears as an AI Answer""a brand's percentage share of impressions compared to other tracked brands", a weighted average across platformsno position metric among the AI visibility metrics documented

Definitions from each vendor's own documentation or help centre, read on October 31, 2026.

Four details in that table change what a report means:

  • Ahrefs weights by demand. Its share of voice divides impressions rather than mentions, and impressions sum the Google search volume behind each prompt, so appearing in a high-volume prompt counts for more than appearing in a niche one. Its POST /v3/brand-radar/sov-history returns one share_of_voice value per brand per date, described as an estimate. A brand can gain mentions and lose share of voice if the gains land on low-volume prompts.
  • Scrunch measures where you sit in the text, not your rank among competitors. A single 0 to 100 score built from top, middle and bottom of the answer answers a different question than "second out of five brands", and the two cannot be charted on the same axis.
  • Peec AI returns the numerator and the denominator separately. visibility_count and visibility_total come back as counts, which is what lets a percentage be re-aggregated across weeks, engines or countries without averaging averages. Its visibility page states the formula as "(Number of responses mentioning your brand / Total responses) × 100".
  • One vendor can describe its own metric two ways. Profound's glossary defines the visibility score as a percentage of responses, while its MCP analytics documentation describes the same visibility_score as a measure of "how often and how prominently a brand appears in AI answers for a category over a date range", and warns that the field is "a raw decimal. Multiply it by 100 to match the percentage the Profound platform shows, so 0.42 is 42%".

Average position improves when visibility collapses

Average position is averaged over the answers where your brand appears, and over nothing else. OtterlyAI publishes that denominator as "number of prompts where brand appeared", and Peec AI's position_count carries the same restriction.

The consequence is counter-intuitive on a bad month. In the 200-answer example, losing the 24 answers where the brand ranked third takes visibility from 34.0% to 22.0% and the average position from 2.1 to 1.5. The brand looks better ranked and is named in a third fewer answers. Any chart of average position alone, without the appearance count beside it, can report an improvement that is a loss.

Sentiment and citation rate are no steadier

The two metrics usually shown next to the first three are built on scales that differ as much:

  • Sentiment scale. Peec AI returns ((sum(sentiment_sum) / sum(sentiment_count)) / 2 + 0.5) * 100, which maps an underlying -1 to +1 average onto 0 to 100, so a perfectly neutral month reads as 50. Scrunch builds its brand_sentiment_score by scoring "positive = 100, mixed = 50, negative = 0" per response and averaging, so a set of half positive and half negative answers also reads as 50. API for GEO publishes a 0 to 100 score with a label per brand per response. The three numbers share a range and not a method.
  • Citation rate. API for GEO's model comparison endpoint defines citationRate as the "share of this model's answers that cite at least one source", a property of the engine. Scrunch's brand_citation_rate is the "share of all responses (whether they cite anything or not) that cite at least one URL you own", a property of your site. Peec AI's domains report divides again differently, with citation_rate defined as sum(citation_count) / sum(retrieval_count), citations over pages the engine read. Three fields, one name, three questions answered. Our guide on tracking AI citations of your pages works through that family on its own.

The one figure you can get without a paid API

Google now publishes a number of its own. Search Console's Generative AI performance report covers AI Overviews and AI Mode, defines impressions as "how many times links to your site were shown to a user in a generative AI feature on Google Search", and was rolled out "to all websites worldwide" on August 31, 2026. It groups data by page, by country, by device and by text or multimodal search, and excludes Search Labs experiments.

It measures a link to your site rather than a mention of your brand, and it stops there: no clicks, no position, no competitor, no answer text. A site can be named in thousands of AI Overviews without a single link, and the report would show nothing. It answers "are AI surfaces showing my pages", and it cannot produce a share of voice.

The engine APIs are the other cheap route, and they return raw material rather than metrics. OpenAI's web search guide documents url_citation annotations and a sources field that "returns the complete list of URLs the model consulted"; Anthropic's web search tool returns web_search_result blocks and citations carrying up to 150 characters of cited text; Perplexity's API returns search_results with title, URL and date. Every share, rank and sentiment figure in a GEO report is computed on top of outputs like these, by the vendor or by you.

Computing the three metrics yourself

From raw answers, the counting job is short, and four decisions decide whether the series holds up:

  1. Fix the denominator in writing. All collected answers, or only answers naming a brand. Store both counts so either percentage can be produced later, the way Peec AI's two fields do.
  2. Collapse aliases before counting. "Loopwork", "Loop Work" and "loopwork.io" are one entity, and a brand whose aliases are missed loses mentions to nobody. API for GEO states that it resolves aliases onto one entity; a hand-rolled counter needs the same list, maintained.
  3. Define rank once. Order of first mention among tracked brands, with ties resolved by position in the text, is the definition most vendors publish. A prominence band inside the answer is a second, defensible choice, and the two must never share a column.
  4. Keep counters, not ratios. Storing 68 and 200 lets you re-aggregate by week, engine or country. Storing 34.0% leaves you averaging averages, which silently overweights the engines that returned fewer answers.

The sample size settles the rest. cloro's published study of August 14, 2026 measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model, minutes apart, and puts about 300 answers per period at roughly ±5 percentage points, with about 294 answers on each side to support a claimed move from 20% to 30%. A one-run month produces a number with no error bar, whatever the field name says.

Five mistakes that make two tools disagree

  • Comparing percentages across vendors. Two visibility scores for the same brand in the same month rest on different denominators, prompt lists, engines, countries and run counts, which makes them two measurements of two populations.
  • Filtering to your own brand before aggregating. Peec AI's documentation spells out the trap: filtering by brand_id in filters "also shrinks share_of_voice's denominator, so SoV collapses to 1.0 when scoping to a single brand". Its having clause exists for that reason. Any share metric computed after a single-entity filter is 100% by construction.
  • Charting average position alone. It is averaged over appearances, so it improves as appearances disappear.
  • Averaging engine percentages without weights. Four engines with 400, 300, 200 and 50 answers do not contribute equally to a brand's visibility, and a plain mean of four percentages pretends they do. Ahrefs makes its weighting explicit; a spreadsheet usually does not.
  • Changing the competitor set mid-quarter. Share of voice divides by the mentions of the brands you track, so adding two competitors lowers your share without anything changing in the answers. Ahrefs' own help page notes that AI share of voice "can change a lot depending on how you've set up the entities for your brand and competitor brands".

Our view on which number to report

In our view visibility is the figure to put at the top of a monthly report, because its denominator can be stated in one line and defended in a meeting. Share of voice is the figure that moves when the category moves, and it belongs next to the list of brands it divides by, never alone. Average position is a diagnostic, useful beside the appearance count and misleading without it.

Frequently asked questions

What is share of voice in AI search?

Share of voice is your brand's mentions divided by every tracked brand's mentions across the same set of AI answers, usually shown as a percentage. Profound's glossary defines it as "how often your brand is mentioned compared with all brands", OtterlyAI publishes the formula as my brand mentions ÷ total brand mentions × 100, and Peec AI returns it as `mention_count / sum(mention_count)` on a 0 to 1 scale. Ahrefs computes a different quantity under the same name: its AI share of voice is "a brand's percentage share of impressions compared to other tracked brands", where impressions sum the Google search volumes of the prompts where the brand appears.

What is the difference between visibility and share of voice?

Visibility counts answers, share of voice counts mentions. If your brand appears in 68 of 200 collected answers, visibility is 34.0%; if those answers hold 78 of the 540 brand mentions in the set, share of voice is 14.4%. Visibility rises when you appear in more answers, share of voice falls when competitors are named more often inside the answers where you already appear, so the two can move in opposite directions in the same month.

How is average position calculated in an AI answer?

Most vendors take the rank of your brand's first mention among the brands named in the answer, then average it over the answers where your brand appears. OtterlyAI publishes the denominator explicitly as the number of prompts where the brand appeared, Peec AI returns `sum(position_sum) / sum(position_count)` with 1 as best, and Profound's glossary gives a worked case: a brand listed first in half the responses that mention it and second in the other half has an average position of 1.5. Scrunch measures prominence instead, scoring top of response 100, middle 50 and bottom 0 and averaging that into a `brand_position_score`.

Can I compare a visibility score from two different tools?

No, unless both publish the same denominator. Profound's glossary divides by "all responses that mention at least one brand", which drops every answer that names nobody, while Peec AI's documentation divides by total responses analysed and Scrunch's `brand_presence_percentage` is the "share of responses whose text names your brand". On 200 answers of which 170 name a brand, the same 68 appearances read as 40.0% or 34.0%. Prompt lists, engines, countries and run counts differ on top of that.

Does Google report my visibility in AI answers?

Partly, and only as impressions. Search Console's Generative AI performance report covers AI Overviews and AI Mode, defines impressions as "how many times links to your site were shown to a user in a generative AI feature on Google Search", and was rolled out to all websites worldwide on August 31, 2026. It groups data by page, country, device and text or multimodal search, and it reports no clicks, no position and no brand mention, since it measures links to your site rather than the words of the answer.

How many answers do I need before a share of voice is stable?

Several hundred per period. cloro's published study of August 14, 2026 measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model minutes apart, puts about 300 answers per period at roughly ±5 percentage points, and needs about 294 answers on each side to support a claimed move from 20% to 30%. A 50-prompt list on four engines run four times a month reaches 800 answers, which is enough for a quarterly comparison and not for a weekly one.