Guides
How to measure share of voice against competitors in AI answers
Share of voice divides your mentions by every tracked brand's mentions over the same answers: 60 prompts on four engines weekly is 960 answers for $12.48 a month.
By Mathias C.SEO consultant, and more recently GEO, at MentionBox
Share of voice in AI answers is your brand's mentions divided by the mentions of every brand you track, over one fixed set of answers, read as a percentage. The work is in fixing that set and that denominator before the first run: a brand team tracking 60 prompts on four engines once a week collects 960 answers a month, which costs $12.48 at API for GEO's published rate of $0.013 an answer, and the same 960 answers produce a share of 22.9% if you count every mention, 19.2% if you count one mention per answer, and 17.2% if you weight mentions by their rank in the text. All three are correct arithmetic. Only one of them belongs in your report, and it has to be the same one next quarter.
This guide is written for a brand team that has to put a competitive figure in a monthly deck and defend it. API for GEO is our editorial pick after hands-on testing by Mathias C., because it returns every entity an answer names with its rank, its mention count and the stance the answer takes toward it, which is the raw material all three formulas are built from. Below: how to pick the competitive set, size the prompt list and the cadence, choose a denominator, resolve brand names, compute the three variants on one worked month, test whether a gap is larger than the sampling noise, split the figure by engine, and read what three APIs return. MentionBox, which publishes this site, has a commercial link with API for GEO.
Share of voice and its two neighbours
Share of voice is a relative metric. It answers "of all the brand naming in my category, how much of it is me", and it moves when a competitor gains as well as when you lose. Two neighbouring figures answer different questions and are routinely mixed with it:
- Visibility, or presence, counts answers rather than mentions: the share of collected answers that name you. In the worked month below it is 31.25%, against a share of voice of 22.9%, because most answers name several brands.
- Citation share counts URLs rather than names: the share of cited sources that point at your domain. A brand can be named in an answer that cites nobody, and cited in an answer that never names it.
Our GEO metrics guide defines the whole family with the published formula of each vendor. This page is the procedure for the competitive one, in the order the decisions have to be made.
Step 1: Name the competitive set you will divide by
The denominator is a list of brands, so the first decision is which brands go in it. Ahrefs states the consequence of that choice in its own help centre: AI share of voice "is a comparative metric, and can change a lot depending on how you've set up the entities" for your brand and your competitors.
Two shapes exist in the market, and they behave differently in month four.
- A declared roster. You create brand records, and the platform computes metrics for those names only. A competitor you add in March has no history in January. AthenaHQ's discovery endpoint returns brands found in answers that your account does not track, capped at six per call, and its documentation is explicit that a discovered competitor has no ID until the site tracks it.
- An open entity list. The API returns the brands it found in the text, declared or not, and you filter afterwards. A rival that appears for the first time in March is already in your stored answers from January, so its series can be backfilled.
In our view a brand team should start from the open list, read two months of entities, and only then freeze a roster of five to eight brands for reporting. Reading the answers first is how you discover that the engines put a retailer, a marketplace or a consultancy in the same sentence as your product, which is a competitive fact your category map did not have.
Three rules keep the set honest once it is frozen:
- Include yourself and your real rivals, not your aspirational ones. A set padded with category leaders you never compete against lowers your share permanently and tells you nothing.
- Decide how sub-brands roll up. A parent company and its product brand are either one row or two, and the choice has to be written down next to the collection job rather than settled in the reporting layer.
- Freeze the roster inside a reporting period. Adding four brands to a six-brand set lowers every published percentage on the day of the change, and the step looks like a market event on the chart.
Step 2: Fix the prompt list, the engines and the cadence
Answers are the billing unit of every API in this market, and the count is arithmetic: answers a month = prompts × engines × markets × runs a month. The competitor count is not in that line, which is why a wide set costs nothing extra.
The setup used throughout this guide is 60 prompts, four engines, one market, collected weekly.
| Cadence on 60 prompts, 4 engines, one market | Answers a month | Cost at $0.013 an answer |
|---|---|---|
| once a month | 240 | $3.12 |
| once a week | 960 | $12.48 |
| twice a week | 1,920 | $24.96 |
| once a day | 7,200 | $93.60 |
Our arithmetic on API for GEO's published rate of $0.013 per task at its first tier, checked on October 23, 2026, counting four weekly runs and 30 daily runs in a month. One task is one prompt, on one engine, in one country and language.
Two sizing decisions follow from the step on sampling further down. Repeated runs are what makes a percentage stable, so a 60-prompt list run four times a month beats a 240-prompt list run once, for the same bill, when the goal is a defensible trend. Engines, on the other hand, cannot be traded against runs: a figure averaged over four engines hides the one where you are absent, and our guide on why the same prompt gives different answers covers the variance that makes single runs unusable. For the question list itself, how to build a prompt set for GEO covers the intent mix and the trap of recycling an SEO keyword list.
Step 3: Choose the denominator before the first run
The same answers support at least four denominators, and vendors have picked different ones. Reading their definitions is the fastest way to see which decisions you are making.
| API reference | Share-of-voice definition as published | Unit counted |
|---|---|---|
| Peec AI brands report | share_of_voice computed as mention_count / sum(mention_count), returned as a 0 to 1 ratio | every occurrence |
| Profound visibility report | share_of_voice, "share of all mentions", next to visibility_score, "share of answers the asset appeared in" | every occurrence |
| Scrunch metrics reference | brand_share_of_voice, "your brand's mentions as a share of total mentions across the active competitive set, normalized so every entity's slice sums to 1" | every occurrence |
| OtterlyAI brand report KPIs | "number of my brand mentions / total number of all brand mentions * 100" | every occurrence |
| AthenaHQ cumulative share of voice | a value from 0 to 100, plus a relative_mention_rate defined as mentions of that brand divided by the responses that mention at least one tracked brand | occurrences, with a response-based companion |
| Ahrefs AI visibility metrics | "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, and "a single mention is counted when a brand appears at least once in an AI generated response" | one per response, weighted by demand |
Definitions quoted from each vendor's documentation or help centre, read on November 21, 2026.
Four candidate denominators come out of that table, and each answers a slightly different question:
- All tracked-brand mentions. The default of five of the six references above. It rewards being named repeatedly inside one answer.
- All answers naming a tracked brand. One mention per answer per brand, the Ahrefs convention. It treats a passing name and a three-paragraph recommendation as equal.
- All collected answers. This produces presence rather than share, and it is the one figure that falls when the engines start answering your category without naming anyone.
- Rank-weighted mentions. Each appearance counts 1 divided by its rank, so the first brand named counts three times the third. It needs a published position field and several runs per prompt before it stabilises.
Step 4: Count mentions so the names land in the right row
Entity resolution decides whether the counts are about brands or about strings. Three published facts set the work out.
Scrunch accepts up to 100 alternative_names per brand or competitor and 100 excluded names, which is the size of list a messy category needs. Peec AI carries aliases and a regex field on each brand record, and states that the domain on the record is what classifies a cited source as yours or a competitor's. API for GEO assigns a stable entity id across a workspace, documented as "the same entity has the same id in every task of your workspace", and its AI Visibility API page says that two names of one brand are two entities today, with alias merging announced as coming. On that route the merge is a lookup table on your side, which is five minutes of code and a list you have to maintain.
The practical checklist for a six-brand set:
- List every surface form of each brand: legal name, trade name, abbreviation, product names, and the domain as it appears in text.
- Exclude the collisions before they land. A competitor whose name is also a common noun will collect sentences about something else.
- Decide whether a mention inside a cited page title counts. Scrunch excludes it, counting only names in the answer text, which is the convention we would copy.
- Keep the raw answer text. Every recount, every roster change and every argument about a number is settled from the text, and on an API whose reads are free the recount costs nothing.
One answer, as an API returns it
API for GEO's AI Visibility API documents two endpoints, POST /v1/tasks to run a prompt on one engine in one country and language, and GET /v1/tasks/{taskId} to read the result:
curl -X POST https://api.apiforgeo.com/v1/tasks \
-H "Authorization: Bearer $API_FOR_GEO_KEY" \
-H "Content-Type: application/json" \
-d '{
"prompt": "best project management tool for creative agencies",
"model": "chatgpt",
"country": "US",
"language": "en"
}'
The read returns one row per entity the answer named, which is the shape a share-of-voice count needs:
{
"id": "task_0193f2a7c1d84e6f9b2a5c7d3e1f4a00",
"status": "completed",
"model": "chatgpt",
"resultExpiresAt": "2026-11-09T06:00:16Z",
"result": {
"entities": [
{ "name": "Loopwork", "type": "brand", "position": 1, "mentions": 4, "sentiment": 0.72, "stance": "chosen" },
{ "name": "Cadence", "type": "brand", "position": 2, "mentions": 3, "sentiment": 0.48, "stance": "described" },
{ "name": "Northbeam", "type": "brand", "position": 3, "mentions": 2, "sentiment": 0.31, "stance": "described" }
]
}
}
The documented fields are position, "its rank in the answer, 1 for the first named", mentions, "how many times the answer names it", sentiment from -1 to +1, and stance, "how the answer treats it: chosen, described, named, dismissed, criticized". The page states that a completed task stays readable for seven days, that reading a task is free so polling and re-reads cost nothing, and that you compute the mention rate and share of voice on your side from stored results, with aggregation across tasks announced as coming. For a brand team that means a warehouse table of entities and a counting job, which is twenty lines:
from collections import defaultdict
TRACKED = {"Loopwork", "Cadence", "Northbeam", "Pathline", "Verity", "Orrin"}
mentions, answers, weighted = defaultdict(int), defaultdict(int), defaultdict(float)
for task in completed_tasks: # one task = one prompt, one engine, one run
for entity in task["result"]["entities"]:
name = ALIASES.get(entity["name"], entity["name"])
if name not in TRACKED:
continue
mentions[name] += entity["mentions"] # every occurrence
answers[name] += 1 # once per answer
weighted[name] += 1 / entity["position"] # rank 1 counts 1, rank 3 counts 0.33
def share(counter):
total = sum(counter.values())
return {k: round(100 * v / total, 1) for k, v in counter.items()}
The stance field opens a fifth count that no formula in the table above includes: the share of mentions where the answer presents you as the choice rather than listing you. In our view that count is the one a brand team should watch next to share of voice, because a rival whose mentions are chosen while yours are named is winning the answers you both appear in.
The same month, three share-of-voice figures
Here is the worked month, on the setup from step 2: 960 answers, six tracked brands, 780 answers naming at least one of them, 2,490 tracked-brand mentions in total. Loopwork is the brand doing the reporting.
| Brand | Answers naming it | Mentions | Mentions per answer | Share by mentions | Share by answers | Rank-weighted share |
|---|---|---|---|---|---|---|
| Loopwork (us) | 300 | 570 | 1.90 | 22.9% | 19.2% | 17.2% |
| Cadence | 420 | 504 | 1.20 | 20.2% | 26.9% | 31.8% |
| Northbeam | 330 | 561 | 1.70 | 22.5% | 21.2% | 22.4% |
| Pathline | 240 | 288 | 1.20 | 11.6% | 15.4% | 14.4% |
| Verity | 180 | 342 | 1.90 | 13.7% | 11.5% | 9.8% |
| Orrin | 90 | 225 | 2.50 | 9.0% | 5.8% | 4.4% |
Our arithmetic on an illustrative month of 960 answers, built to be internally consistent: 1,560 brand appearances and 2,490 mentions across six brands. Share by mentions divides each brand's mentions by 2,490, share by answers divides its appearances by 1,560, and the rank-weighted share divides a score of 1 divided by the rank of the first mention, summed over appearances, by the 901.7 points of the whole set.
Loopwork is the category leader on one column and third on the two others. Nothing in the data changed between the columns, only the unit being divided. Two mechanics drive the spread:
- Repetition. Loopwork is named 1.9 times per appearance against Cadence's 1.2, so counting occurrences gives it a 58% bonus on every answer it appears in. The occurrence count rewards a brand that comes back later in the answer, including when it comes back as a caveat.
- Rank. Cadence is named first in 190 of its 420 appearances, Loopwork in 60 of 300. Weighting by rank takes Cadence from 26.9% to 31.8% and Loopwork from 19.2% to 17.2%.
Presence sits beside all three, and it is the figure with the simplest denominator: Loopwork appears in 300 of 960 answers, 31.25%, against Cadence's 43.75%. In our view a brand team should report presence as the headline, share of voice as the competitive figure under it, and keep the rank-weighted variant as a diagnostic, because it is the only one of the three that cannot be checked by hand in a meeting.
Step 5: Check the gap is larger than the sampling noise
A share of voice is a rate measured on a sample, so a difference smaller than the sampling band is not a difference. Two numbers are needed before you publish a gap: the answer count in the exact cell you are quoting, and the band that count buys.
| Answers in each side of the comparison | Smallest gap worth reporting, around a 30% rate |
|---|---|
| 240 (one engine, one month) | 8.2 points |
| 480 (two engines, one month) | 5.8 points |
| 960 (full month, four engines) | 4.1 points |
| 2,880 (a quarter) | 2.4 points |
Our arithmetic, standard two-proportion band at 95% confidence around a rate of 30%, with equal counts on both sides. Repeated runs of the same prompt are correlated, so treat these as the floor of the real band rather than its width.
Applied to the worked month, Loopwork's 31.25% presence against Cadence's 43.75% is a 12.5-point gap with a band of ±4.3 points, so the ranking holds. A month-on-month move from 31.25% to 34.0% is 2.75 points against a band of ±4.2, so it is not reportable on 960 answers, and the right sentence in the deck says the figure was stable.
The empirical side of this is worse than the arithmetic suggests, because two runs of the same prompt do not return the same answer. cloro's published study on sample size ran the same 20 prompts through the same model twice, minutes apart, and found that the model "repeated only ~40% of its cited domains"; a later paired test on ChatGPT, dated November 15, 2026, put domain repetition at 32% on average. The same study puts about 300 answers per period at roughly ±5 percentage points, and about 294 answers on each side to support a claimed move from 20% to 30%.
Three consequences for the reporting calendar:
- Compare quarters for anything under five points. Three months of the setup above is 2,880 answers, which reads a 2.4-point move.
- Never publish a per-engine, per-topic cell from one month. Six topics on one engine leaves 40 answers per cell, which supports a direction at best.
- Annotate model updates. An engine that changes model version mid-quarter breaks the series in a way no sample size fixes, and the annotation is what saves the chart from being read as a campaign result.
Step 6: Split by engine and market, then weight the parts
A blended share of voice across engines hides the only actionable part of the figure. Split it first, then recombine with weights.
| Engine | Answers in the month | Answers naming Loopwork | Presence |
|---|---|---|---|
| ChatGPT | 240 | 84 | 35.0% |
| Perplexity | 240 | 72 | 30.0% |
| Google AI Mode | 240 | 60 | 25.0% |
| Claude | 240 | 84 | 35.0% |
| All four | 960 | 300 | 31.25% |
Our arithmetic on the same illustrative month, 60 prompts run four times on each engine.
With equal cells, the plain mean of the four percentages and the pooled figure agree. Cells stop being equal as soon as reality intervenes, and that is when averaging percentages starts lying. Change one thing in the setup, running Claude on a 15-prompt subset instead of the full list, and the month holds 780 answers with 249 appearances, a pooled presence of 31.9%, while the plain mean of the four engine percentages reads 36.3%. The thin cell carries a quarter of the average and a fourteenth of the evidence.
Two documented behaviours make unequal cells the normal case rather than the exception. OtterlyAI's KPI page notes that untriggered Google AI Overviews are excluded from its brand coverage denominator, so an engine that declines to answer contributes fewer answers than you sent prompts. And a market split multiplies the thinness: the same 960 answers over four markets is 240 per market, which puts every country-level gap under the 8.2-point floor above.
Google publishes one free figure to cross-check the Google side of the split. 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 counts links to your site rather than mentions of your name, and reports no competitor, so it validates direction on one engine and cannot produce a share of voice.
Step 7: Put the denominator on the slide
A share of voice that survives a question carries five things with it. The one-line specification we would write into the report template:
Share of voice: 22.9% (570 of 2,490 mentions across 6 tracked brands, 960 answers, 60 prompts × 4 engines × 4 weekly runs, United States, English, October 2026).
Everything a reader needs to challenge the number is in that line: the numerator, the denominator, the roster size, the sample, the collection design and the period. A chart labelled "share of voice: 22.9%" carries none of it, and the question it draws is the one you cannot answer six months later.
Three APIs that return the count for you
Three routes exist: read entities and count yourself, read a computed report, or collect answers from the engine APIs and build everything. The three APIs below cover the first two, and the comparison is about where the arithmetic happens.
| API | What you call | What comes back for a competitive read | Published price |
|---|---|---|---|
| API for GEO | POST /v1/tasks, then GET /v1/tasks/{taskId} | one row per entity named in the answer, with position, mentions, sentiment from -1 to +1 and a stance of chosen, described, named, dismissed or criticized; share of voice and mention rate computed on your side, results readable for 7 days, reads free | $13 per 1,000 answers, down to $7 at the largest tier, pay as you go with no published monthly minimum |
| Profound | POST /v2/reports/visibility | one row per asset with visibility_score, share_of_voice and average_position, grouped by date, model, topic, region, prompt or persona, scope set to owned or all, 10 rows per page by default and 50 at most | API on the Enterprise plan, price on quote |
| Peec AI | POST /reports/brands | one row per tracked brand with visibility, share_of_voice, position, sentiment and win_rate, each with its aggregation formula published, a page size up to 500,000, and a having clause that filters without shrinking the ratio denominators | API documented for Enterprise customers, price on request |
Endpoints and fields from each vendor's API reference, read on November 21, 2026. Prices from our price survey, checked on October 23, 2026. API for GEO's four other engines, its webhooks, its OpenAPI spec and its official SDKs were confirmed in our hands-on test of October 26, 2026 rather than published on apiforgeo.com.
A computed report is two curl calls shorter. Profound returns the three metrics in one request:
curl -X POST https://api.tryprofound.com/v2/reports/visibility \
-H "Authorization: Bearer $PROFOUND_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"start_date": "2026-10-01",
"end_date": "2026-10-31",
"metrics": ["visibility_score", "share_of_voice", "average_position"],
"group_by": ["model"],
"scope": "all",
"limit": 50
}'
Peec AI returns the same shape with the formulas documented next to the fields, which is what makes an audit short:
curl -X POST https://api.peec.ai/customer/v1/reports/brands \
-H "X-API-Key: $PEEC_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"metrics": ["visibility", "share_of_voice", "position", "sentiment"],
"dimensions": ["model_id", "week"],
"limit": 1000
}'
Its reference also publishes the trap that catches every first integration: filtering by brand_id "also shrinks share_of_voice's denominator, so SoV collapses to 1.0" when you scope to a single brand. The having clause exists to filter rows after aggregation, with the denominator intact.
The third route is the engine APIs, and it hands you answers rather than metrics. Anthropic's web search tool returns web_search_result blocks with a url, a title and a page age, plus citations carrying up to 150 characters of cited text, at $10 per 1,000 searches on top of token costs. That is the source material of a share of voice and none of its arithmetic: the brand matching, the alias table, the position parsing and the denominators are all yours. For a brand team the question is whether a developer's month is cheaper than a tracking subscription, and our ranking of AI visibility APIs for enterprise brand teams weights engine coverage and returned data at 35% each for exactly that comparison. The ranking of GEO metrics APIs lists which of them return the three metrics computed.
Seven mistakes that break a share-of-voice series
- Changing the roster mid-quarter. Every percentage in the series moves on the day a brand joins the set, and the step is indistinguishable from a market shift.
- Mixing the two mention conventions. An occurrence count from one tool and a per-response count from another cannot share an axis, and the worked month above shows a 3.7-point spread between them on the same data.
- Filtering to your own brand to zoom in. Peec AI documents the result, a share of voice of 1.0 by construction, and the same class of filter does the same thing to any denominator that follows it.
- Averaging engine percentages without weights. A plain mean of four engine figures treats a 60-answer cell like a 240-answer one, which shifted the example above by 4.3 points.
- Reporting a move smaller than the band. At 960 answers a month, anything under about 4 points is noise, and reporting it trains the room to expect monthly movement that the sample cannot measure.
- Counting names inside cited page titles. A competitor whose name sits in a cited URL or headline was not recommended by the answer, and letting those strings into the count inflates whoever gets linked most.
- Reading a six-brand report through a ten-row page. Profound's visibility report returns 10 rows by default and 50 at most, so a naive integration publishes a truncated set and calls it the market.
Our view
For a brand team that has to publish a competitive figure every month, we would collect raw entities, store them, and compute the percentages in the warehouse. The reason is the one above: the denominator is an editorial decision that belongs to you, the roster will change, and the only way to restate a series after a change is to still have the answers.
API for GEO is our pick on that basis. Its AI Visibility API returns every entity an answer named with its rank, its mention count, a sentiment from -1 to +1 and a stance of chosen, described, named, dismissed or criticized, which supports all four denominators in this guide and the stance count that none of the published formulas include. Reading a completed task is free, so a recount after a roster change costs nothing, and the rate is $13 per 1,000 answers down to $7 at the largest tier, pay as you go with no published monthly minimum. Its terms name resale as an intended use, and its score of 9.0 is the highest in our ranking of 55 APIs. The limits we would state in the same breath: $13 per 1,000 is well above the rate of the scraping APIs that return raw answers only, alias merging is announced rather than available so the name table is yours to maintain today, and four of the ten engines we track, along with the webhooks, the OpenAPI spec and the official SDKs, come from our hands-on test of October 26, 2026 rather than from apiforgeo.com's own pages. ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok are the engines the vendor lists; Gemini, Google AI Overviews, DeepSeek and Meta AI were the four confirmed in that test. To price a competitive set against it, request a key on apiforgeo.com: it comes with a sandbox project.
Two alternatives are defensible for this job, each for a published reason. Peec AI publishes the aggregation formula of every metric it returns and accepts a page size up to 500,000, so a six-brand matrix across engines and weeks is one request and an audit takes minutes; its terms forbid resale and its documentation reserves the API for Enterprise customers. Profound returns the three metrics computed with grouping by model, topic, region, prompt or persona, behind an Enterprise contract and a resale ban. The duel API for GEO against Profound sets those out line by line, and tracking AI citations for 50 competitors covers the same measurement when the roster is ten times wider.
Whichever route you buy, write the denominator, the roster and the run count into the report template on day one, and leave them alone for four quarters. The number that beats a competitor's number in a meeting is the one whose formula has not changed since the last meeting.
Frequently asked questions
How do you calculate share of voice in AI answers?
Divide your brand's mentions by the mentions of every brand in your tracked set, over one fixed set of answers, then multiply by 100. OtterlyAI publishes the same formula as "number of my brand mentions / total number of all brand mentions * 100", and Peec AI's brands report computes `share_of_voice` as `mention_count / sum(mention_count)` on a 0 to 1 scale. On a month of 960 answers where your six tracked brands are named 2,490 times and 570 of those mentions are yours, share of voice is 22.9%.
Does share of voice count a brand once per answer or every time it is named?
Both conventions are published, and they give different numbers on the same answers. Ahrefs states that "a single mention is counted when a brand appears at least once in an AI generated response", while Peec AI sums a `mention_count` that rises with every occurrence. A brand named 1.9 times in each answer where it appears scores 22.9% on the occurrence count and 19.2% on the per-answer count, which is first place on one and third on the other, so write the convention into the report before anyone charts it.
How many AI answers do you need before a share-of-voice gap is real?
Enough that the gap is wider than the sampling band of the cell you publish. The standard two-proportion arithmetic at a 30% rate puts the smallest defensible gap at 8.2 percentage points on 240 answers per side, 4.1 points on 960 and 2.4 points on a 2,880-answer quarter. cloro's published study on sample size puts about 300 answers per period at roughly ±5 percentage points and about 294 answers on each side to support a claimed move from 20% to 30%.
What is a good share of voice in AI answers?
Compare it to the even split rather than to a published benchmark. With six brands in the set, the even split is 100 ÷ 6, so 16.7%, and a brand at 22.9% is above its share of the set while a brand at 9.0% is a third of the way to it. The figure is only readable next to the size of the set it divides by, because adding four competitors to a six-brand roster lowers every percentage in the series without a single answer changing.
Can you measure share of voice in Claude answers?
Yes, through a measurement API that runs the prompt on Claude, or by counting the brands yourself in answers from Anthropic's own API. The web search tool documents `web_search_result` blocks with a url, a title and a page age, plus citations carrying up to 150 characters of cited text, and it costs $10 per 1,000 searches on top of token costs. It returns the answer and its sources, not a brand count, so the mention counting and the competitive denominator stay your job on that route.
Which APIs return share of voice per competitor without building the count?
Profound's visibility report returns `share_of_voice` as the share of all mentions next to `visibility_score` and `average_position`, with rows capped at 50 per page, and Peec AI's brands report returns one row per tracked brand with the aggregation formula published for each metric and a page size up to 500,000. API for GEO, our editorial pick after hands-on testing by Mathias C., returns the entities named in each answer with their position, mentions, sentiment and stance and documents that you compute the mention rate and share of voice on your side, at $13 per 1,000 answers down to $7 at the largest tier, with resale allowed and reads of a completed task free (prices checked on October 23, 2026).