Skip to content

Integration

How to track AI citations for 50 competitors with one API

Fifty competitors do not multiply the bill: 120 prompts on four engines, collected weekly, is 2,080 answers a month, $27.04 at $0.013, or $0.53 per brand.

A ring of white cubes surrounds one blue cube centered on a pale background

SEO consultant, and more recently GEO, at MentionBox

Fifty competitors cost no more to track than five, because an AI visibility API bills answers rather than brands. One prompt list of 120 questions, run on four AI engines once a week, is 2,080 answers a month, and every brand named in those answers can be counted from the same payload: $27.04 at $0.013 an answer, which is $0.53 per brand per month across your own brand and 50 rivals. What scales with the competitor count is the naming work, the share-of-voice denominator and the page size of the report you read back.

This guide is written for an agency or a brand team that already collects AI answers and now wants a competitive set wide enough to be a market view rather than a shortlist. API for GEO is our editorial pick after hands-on testing by Mathias C., because its per-prompt statistics return every brand and competitor named in the stored answers instead of only the ones you declared. Below: the arithmetic that stays flat, the two shapes an API can take, the only published ceilings on a competitor set, the requests that read 51 brands at once, the denominators that break when you filter, the citation ownership fields, and the sample size a 50-brand ranking needs. MentionBox, which publishes this site, has a commercial link with API for GEO.

What actually scales when you go from 5 competitors to 50

Three things grow with the competitor count, and the one most teams budget for is not among them.

  • The answer count does not grow. An answer is one prompt, one engine, one country, one run. Every brand named inside it is read from the same text, so brand 51 is free to count.
  • The naming work grows linearly. Each competitor needs its legal name, its trade names, its abbreviations, its product brands and its domains, or its mentions land in the wrong row. Fifty brands at four or five surface forms each is 200 to 250 strings to maintain.
  • The denominators change shape. Share of voice is a ratio over the whole competitive set, so adding thirty brands lowers every existing percentage. A series that spans a roster change is two series.
  • The read gets bigger. A monthly report over 51 brands, 4 engines and 6 topics is 1,224 rows before any date breakdown, which runs into page size caps that were sized for a shortlist.

Our guide to GEO metrics defines the metrics themselves, and how to build a prompt set for GEO covers the question list those 2,080 answers come from. This page is about the brand list next to it.

The bill follows answers, so the competitor count is free

Answers a month = prompts × engines × countries × runs a month. The competitor count appears nowhere in that line.

Cadence on 120 prompts, 4 engines, one countryAnswers a monthCost at $0.013 an answerPer brand, for 51 brands
once a month480$6.24$0.12
once a week2,080$27.04$0.53
twice a week4,160$54.08$1.06
once a day14,400$187.20$3.67

Our arithmetic on API for GEO's published rate of $0.013 per credit at the first tier, checked on October 23, 2026, counting 52 weeks across 12 months for weekly cadences and 30 days for daily ones. One credit is one prompt, on one model, in one locale.

Two consequences follow for a competitive study. The first is that widening the set is the cheapest thing you can do to a tracking setup, far cheaper than adding engines (each engine is another 25% of answers on a four-engine basket) or raising the cadence (daily collection is roughly seven times the weekly bill). The second is that a flat per-request price inverts the logic: SE Ranking's AI search leaderboard costs 7,500 credits per call and compares a primary domain against up to 10 competitors, so a 50-brand market there is five calls and 37,500 credits rather than one call.

On the other routes the plan floor decides instead of the volume. OtterlyAI sells API access from its $189 Standard plan, Trakkr from its $500 Scale plan, Peec AI on Pro, Advanced, Enterprise and agency plans, Profound and AthenaHQ on an Enterprise contract, with AthenaHQ's API billed as an add-on above its $295 Starter plan. Prices checked on October 23, 2026. For a brand-level comparison of what those plans deliver, our ranking of AI visibility APIs for agencies weights returned data at 35% and price at 30%.

Two shapes: a declared roster, or every entity named in the answer

This is the design decision that decides how a 50-brand study behaves, and vendors sit on both sides of it.

A declared roster means the platform matches answers against brand records you created, and computes metrics only for those. A brand you add in March exists from March. An open entity list means the API returns the brands it found in the text, whether or not you asked for them, and leaves the filtering to you.

APIHow competitors enter the dataWhat is published about it
API for GEOopen list, read from the stored answersentities[] on POST /v1/queries/stats is documented as every brand and competitor named in the answers, each with its own visibility and avgPosition, and entityType takes brand or competitor
Profoundboth, on two endpointsthe answers endpoint returns mentions as a flat deduped list in the order entities appear in the response; the visibility report is asset-scoped, with scope set to owned or all
Peec AIdeclared roster, with suggestionsGET /brands returns name, domains, aliases, regex, is_own and a source of manual, suggestion or onboarding; brands detected alongside yours are suggested when they appear at least twice
Scrunchdeclared rostercompetitor_id identifies a competitor tracked in your brand configuration, and share of voice is computed over your brand plus all active competitors
AthenaHQdeclared roster, with a capped discovery listcompetitors[] on a response carries id, name, url and position; discovered_competitors returns brands found in answers that the site does not track, and those have no ID
OtterlyAIboth, in one reportcompetitorBrandsAnalysis covers configured competitors, while detectedBrands and allBrandsAnalysis cover brands found automatically
Promptwatchdeclared roster, labelled per mentionbrandMentions[] carries a relation of SELF, DIRECT_COMPETITOR, OTHER or IGNORED next to visibilityScore, position and sentiment

Field names from each vendor's API reference, read on November 16, 2026, except Promptwatch's response schema, read on October 23, 2026.

The practical difference arrives in month four, when a brand you had never heard of starts appearing in answers. On an open list it is already in your data with a history, because the entity came out of the text. On a declared roster you add it and the series starts at zero, which is why AthenaHQ's documentation is worth quoting in full on the point: a discovered competitor "has no ID: it cannot be used in competitor_ids filters or the metrics endpoints until the website tracks it".

The only published ceilings on a competitor set

Competitor caps are the figure this market does not publish. We read the API reference of every vendor in this guide: three ceilings exist in writing, and the rest state nothing.

APIPublished ceilingWhat it applies to
API for GEOno competitor cap; the brand and alias list belongs to the project and apiforgeo.com publishes no endpoint to manage itproject setup, done in the application
SE Ranking Data API1 to 10 competitors, stated twice on the endpointone leaderboard request, at a flat 7,500 credits
OtterlyAI100 competitors per brand report, each with up to 100 name variations and 100 domain variations; 1,000 prompt IDs; suggestions capped at 20 per callthe brand report payload
AthenaHQ6 discovered competitors per call, highest mention rate firstthe discovery list, not the tracked set
Scrunch100 alternative names and 100 excluded names per brand or competitor, with a 422 above thatthe naming lists, not the roster
Profoundno competitor cap; visibility rows default to 10 per page and cap at 50, and an explicit asset selection returns all matches and ignores limitthe page size of the report
Peec AIno competitor cap; POST /brands creates one brand per call, with no bulk brand endpoint in the referencethe write path

Each vendor's API reference, read on November 16, 2026. An absence in this table means nothing was published, not that no limit exists in a contract.

Two readings are worth keeping. A cap on a request, like SE Ranking's ten, is an engineering problem you solve with more calls. A cap on a payload, like OtterlyAI's hundred, is a real ceiling for a market study, and a hundred is above the fifty this guide sizes. The absence of a published figure elsewhere is not a promise: ask for it in writing before you model a 200-brand category on it.

Set up 50 names so the counts hold

Entity resolution is where a 50-brand study fails, and the published tools for it are uneven.

  • List every surface form. Scrunch accepts up to 100 alternative_names per brand and competitor, and rejects the save with a 422 above it. OtterlyAI takes brandVariations and brandDomainVariations, 100 each, plus a brandDomainWildcard and a brandRegex of up to 1,000 characters. Peec AI carries aliases and a regex field on each brand record.
  • Exclude the collisions before they land. A competitor whose name is also a common noun will swallow unrelated sentences. Scrunch's excluded_names and excluded_websites lists, 100 entries each, exist for that case.
  • Attach a domain to every competitor record. Peec AI states plainly what the domain is for: it is "the website domain used to classify sources as You or Competitor". Without it, the citation side of your report has nothing to join on.
  • Collapse aliases server-side where the API does it. API for GEO's prompt tracking page documents aliases collapsing onto one entity, with the three forms of a brand name and its domain counting as the same entity in the mention positions.
  • Keep one domain per brand record. Scrunch rejects a repeated normalised domain across brand and competitors with a 409, which is the right behaviour and a nuisance for sub-brands sharing a site.
  • Freeze the roster inside a reporting period. Add brands at the quarter boundary, note the date under every chart, and keep the previous roster running for one period if you can afford the parallel read.

The requests that read 51 brands at once

Three shapes cover the market. The field names below come from each vendor's own reference, linked in each paragraph.

Read every entity out of the prompt statistics

API for GEO's prompt tracking page documents four endpoints: POST /v1/tags, POST /v1/queries, POST /v1/jobs/execute and POST /v1/queries/stats. The first three build and run the list, and the fourth is the competitive read:

curl -X POST https://api.apiforgeo.com/v1/queries/stats \
  -H "Authorization: Bearer $API_FOR_GEO_KEY" \
  -H "x-project-id: $PROJECT_ID" \
  -H "Content-Type: application/json" \
  -d '{
    "models": ["chatgpt", "perplexity", "google-ai-mode", "claude"],
    "country": "US"
  }'

The response carries one object per prompt, and the competitive layer sits inside it:

{
  "id": "q_3f9a1c20",
  "text": "best case management software for small law firms",
  "type": "commercial",
  "runs": 90,
  "visibility": 64.8,
  "avgPosition": 1.8,
  "entities": [
    { "name": "Loopwork", "entityType": "brand", "visibility": 64.8, "avgPosition": 1.8 },
    { "name": "Cadence", "entityType": "competitor", "visibility": 52.2, "avgPosition": 2.4 },
    { "name": "Northbeam", "entityType": "competitor", "visibility": 40.0, "avgPosition": 3.2 }
  ],
  "sources": [
    { "url": "https://review-site.com/pm-tools", "domain": "review-site.com", "category": "review", "mentionsBrand": true }
  ],
  "lastExecutedAt": "2026-09-02T06:00:14Z"
}

The documentation describes entities[] as every brand and competitor named in the answers with its own visibility and average position, visibility as responses mentioning the brand divided by runs, and avgPosition as the mean rank of the brand's first mention inside the answer. One pass over that array across 120 prompts gives you a 51-row leaderboard without a second call. The gap to know about: per-entity share of voice is not on this endpoint. It sits on POST /v1/models/platform-details, documented on the AI model comparison page, which takes an entities array and returns visibility, shareOfVoice, avgRanking, citationRate and totalResponses per model, one entity at a time. Fifty-one brands is fifty-one calls there, comfortable inside the 300 requests a minute published for the pay-as-you-go tier, and the reads themselves are not billed.

Ask a platform report for the whole asset set

Profound's visibility report is asset-scoped, so a competitive ranking is a matter of scope and assets:

curl -X POST https://api.tryprofound.com/v2/reports/visibility \
  -H "Authorization: Bearer $PROFOUND_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "category_id": "00000000-0000-0000-0000-000000000000",
    "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
  }'

The endpoint reference defines visibility_score as the share of answers the asset appeared in, share_of_voice as the share of all mentions and average_position as the average rank when mentioned, lower being better. scope takes owned or all, and limit is "Page size; default 10, max 50". Two details change how you page a 50-brand read: an explicit assets selection, passed as a list or as {"op": "in", "value": [...]}, "returns all matches and ignores limit", and the /v2/reports/visibility/stream variant takes the same body and ignores paging altogether. The rate limit is 600 requests an hour per key, so plan the backfill rather than looping over brands.

Ask for one row per brand, with the ratio intact

Peec AI's brands report returns a row per tracked brand and publishes the aggregation formula for every metric, which is rare enough to be worth building on:

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"],
    "limit": 1000
  }'

The brands report reference gives share_of_voice as mention_count / sum(mention_count), visibility as sum(visibility_count) / sum(visibility_total), position as sum(position_sum) / sum(position_count) and win_rate as sum(win_count) / sum(visibility_total), with visibility, share of voice and win rate returned as 0 to 1 ratios, sentiment as a 0 to 100 score and position as a rank where 1 is best. Its limit defaults to 1,000 and accepts up to 500,000, so a 51-brand matrix across models and weeks fits in one call, inside a rate limit of 200 requests a minute per project.

Scrunch's query API takes the same job from the other end, with brand_share_of_voice and competitor_share_of_voice as separate metrics and a breakdown list that includes competitor_id and competitor_name, so one request returns the full set. Its documentation states the identity that a market study needs: your brand's share plus the sum of every competitor's share equals 1.

Keep the denominator stable, or the leaderboard lies

Share of voice over 51 brands is a ratio with a moving bottom, and three published behaviours decide whether your numbers survive a filter.

APIPer-competitor fieldPublished definition
API for GEOshareOfVoice per modelentity mentions divided by all mentions within that model, requested one entity per call
Peec AIshare_of_voicemention_count / sum(mention_count), a 0 to 1 ratio
Scrunchbrand_share_of_voice, competitor_share_of_voicebrand plus all active competitors sums to 1; the denominator stays the full set even when you filter to some competitors
AthenaHQvalue, relative_mention_ratea 0 to 100 percentage, plus mentions of the brand divided by responses mentioning at least one tracked brand
Profoundshare_of_voiceshare of all mentions, with rows limited by scope
SE Ranking Data APIshare_of_voiceranked against up to 10 competitors, computed on brand presence and link presence

Definitions quoted from each vendor's API reference, read on November 16, 2026.

Three traps come out of that table.

  • Filtering to one brand can collapse its share to 100%. Peec AI documents the case and the fix: brand_id in filters shrinks the share-of-voice denominator too, so filtering to one brand collapses the ratio to 1.0, and you use having instead to keep the denominator whole.
  • Deactivating a competitor rewrites history. Scrunch states that deactivated competitors and archived sub-brands are excluded from numerators, share-of-voice denominators and breakdown values, so a cleanup changes last quarter's chart.
  • Scales differ by vendor. Peec AI and Scrunch return 0 to 1, AthenaHQ and API for GEO return 0 to 100. Joining two vendors in one dashboard without normalising gives a competitor a share of 0.52% instead of 52%.

Which pages the engines cite for each competitor

The citation side of a competitive study needs one field: the one that says whose page a cited URL belongs to.

APIFieldPublished values
API for GEOcategory and mentionsBrand on each entry of sources[]categories seen in the documented example include review and brand, with a boolean per source
Peec AIdomain_classification filter on the domains reportCORPORATE, EDITORIAL, INSTITUTIONAL, OTHER, REFERENCE, UGC, COMPETITOR, OWN, RELATED, plus custom classifications
Scrunchowner, with source_type kept for compatibilitybrand, competitor, other
AthenaHQdefault_source_typeowned, competitor, partner, third_party, plus custom types
Trakkrownership filter on its pages endpointowned, competitor, editorial, social, video, other
Profoundcitation_category inside citation_detailscategories not enumerated in the reference we read

Field names and value lists from each vendor's API reference, read on November 16, 2026, except Trakkr's filter, read on October 23, 2026.

Peec AI's domains report is the one that also hands you the denominator question on the citation side. It publishes citation_rate as sum(citation_count) / sum(retrieved_chat_count) and retrieval_rate as sum(retrieval_count) / sum(total_chat_count), which separates a competitor the engines never read from one they read and never cite. Those are two different competitive problems, and only the second is about content.

Two counts are worth building for a 50-brand set, both from the same payload. The first is the share of citations each competitor's own domain takes in the answers where your category comes up. The second is the list of third-party domains that cite competitors and not you, which is the shortlist a PR or partnership plan actually works from. Our guide to tracking AI citations of your own pages covers the first count on your own side, with the URL normalisation rules that keep it honest.

The sample size a 50-brand ranking needs

A wide roster makes the statistics harder, not easier, because the brands at the bottom appear in too few answers to rank. The published numbers come from cloro's study on sample size, dated August 14, 2026: a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model minutes apart, about 300 answers per period for a rate accurate to roughly ±5 percentage points, about 72 answers for ±10 points, and about 294 answers on each side to support a claimed move from 20% to 30%.

Apply that to the 2,080-answer month in the table above:

Cell you want to publishAnswers in the cellReading
whole month, all engines2,080safe for a full leaderboard
one engine520safe
one engine, one month, one topic of sixabout 87roughly ±10 points, a direction at best
one engine, one weekabout 120not reportable per brand

Our arithmetic on a 120-prompt, four-engine, weekly setup, against the precision bands published in cloro's study of August 14, 2026.

Three rules follow, and they are what keep a 50-row table defensible:

  • Rank the brands that clear the band, and report the rest as a pooled "others" row with its combined share. A leaderboard with 30 rows under 2% is a chart of noise.
  • Compare quarters rather than weeks for the tail. Three months of the setup above is 6,240 answers, which makes a 3% brand readable.
  • Publish the denominator next to the percentage. "Mentioned in 7% of 520 answers on ChatGPT" survives a client question in a way that "7% share of voice" does not.

Six mistakes on a wide competitive set

  1. Quoting a competitor count as a cost driver. The invoice follows prompts, engines, countries and runs. Fifty brands on the setup above is the same $27.04 as five.
  2. Adding brands mid-quarter without noting the date. Every share of voice in the series moves on the day the roster changes, and the step looks like a market event.
  3. Filtering to one brand to "zoom in". On Peec AI that collapses share of voice to 1.0 by design, and the same class of mistake hits any vendor whose denominator follows the filter.
  4. Letting the discovery list stand in for the roster. AthenaHQ returns at most 6 discovered competitors and gives them no ID until the account tracks them, so a discovery call is a prompt to act, not a data source.
  5. Skipping the domain on each competitor record. Without it, mentions and citations cannot be joined, and the citation half of the study quietly reports zeros.
  6. Reading a 51-brand report through a 10-row page. Profound's visibility report defaults to 10 rows and caps at 50 per page, so a naive integration publishes a top ten and calls it a market.

Our view

For a 50-brand competitive set, we would build on an API that returns the entities it found rather than only the ones we declared, and that does not bill us for reading the same month twice.

API for GEO is our pick on that basis: entities[] carries every brand and competitor named in the stored answers with its own visibility and average position, aliases collapse onto one entity, sources[] carries a category and a per-source brand flag, 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 and reads over stored runs not billed. It covers the ten engines we track: ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok are listed on apiforgeo.com, while Gemini, Google AI Overviews, DeepSeek and Meta AI were confirmed in our hands-on test of October 26, 2026. Its terms name resale as an intended use, and its score of 9.0 is the best in our ranking of 55 APIs. The limits we would state to a client: $13 per 1,000 answers is roughly eleven times DataForSEO's rate for raw answers, per-entity share of voice takes one call per brand, the brand and alias list is set on the project with no endpoint published to manage it, and four of those engines plus the webhooks, the OpenAPI spec and the official SDKs come from our own test rather than from the vendor's pages. To price a competitive set against it, request a key on apiforgeo.com: it comes with a sandbox project.

Three alternatives are defensible for this job, each for a published reason. Peec AI publishes the formula behind every metric and a 500,000-row page size, which makes a 51-brand matrix one request and an audit trivial, and its terms forbid resale. Scrunch is the only one that states the share-of-voice identity outright, with 100 alternative names per brand to resolve a messy roster, sold from a $500 Agency Core plan. Profound has the deepest answer-level schema, with citation positions in the text and an asset selection that returns all matches, behind an Enterprise contract and a resale ban. The duels API for GEO against Profound and Profound against Peec AI set those out line by line, and tracking ten clients without seat fees covers the same arithmetic when the 50 brands belong to 10 different clients.

Settle three things in writing before you widen the set: how a sub-brand rolls up, what the share-of-voice denominator contains, and whether a brand added next quarter can be counted over answers already collected. The first two are yours to decide. The third is the vendor's, and it is the one that cannot be fixed later.

Frequently asked questions

How many competitors can you track with one AI visibility API?

Most API references publish no ceiling at all, and the three that do apply it to a request or to a payload rather than to your account. SE Ranking's AI search leaderboard takes 1 to 10 competitors per call, OtterlyAI's brand report accepts at most 100 competitors with up to 100 name variations and 100 domain variations each, and AthenaHQ's discovery endpoint returns at most 6 suggested competitors. Profound, Peec AI, Scrunch and API for GEO state no maximum in the pages we read on November 16, 2026, so the practical limit there is the entity resolution work and the page size of the report you read.

Does tracking 50 competitors cost 50 times more?

No, because the unit you are billed for is an answer, and an answer is one prompt sent to one engine once. A 120-prompt list on four engines collected weekly is 2,080 answers a month, which costs $27.04 at API for GEO's $0.013 a credit whether you name 5 competitors or 50, so 51 tracked brands work out at $0.53 each per month. The exception is a request priced flat: SE Ranking's leaderboard costs 7,500 credits per call for up to 10 competitors, so a wider set means more calls.

Which API fields give each competitor its own share of voice?

Four vendors publish both the field and its formula. Peec AI's brands report computes `share_of_voice` as `mention_count / sum(mention_count)` and returns it as a 0 to 1 ratio. Scrunch returns `brand_share_of_voice` and `competitor_share_of_voice`, and states that your brand plus every active competitor sums to 1. AthenaHQ's cumulative share of voice returns a 0 to 100 `value` plus a `relative_mention_rate`, defined as mentions of that brand divided by the responses mentioning at least one tracked brand. API for GEO documents `shareOfVoice` per model as entity mentions divided by all mentions within that model, one entity per call.

Can an API show which pages AI answers cite for a competitor?

Yes, through a source ownership field, and five vendors publish the vocabulary. Peec AI filters its domains report on `domain_classification`, whose built-in values include COMPETITOR, OWN and RELATED. Scrunch's `owner` dimension takes `brand`, `competitor` or `other`. AthenaHQ's `default_source_type` takes owned, competitor, partner or third_party. Trakkr's pages endpoint filters `ownership` on owned, competitor, editorial, social, video and other. API for GEO returns a `category` and a `mentionsBrand` flag on each entry of `sources[]`.

How many answers before a competitor's share of voice can be reported?

Count the answers in the cell you want to publish, not the answers in the month. cloro's published study of August 14, 2026 puts about 300 answers per period at roughly ±5 percentage points and 72 answers at about ±10 points, and measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model. A 2,080-answer month split by four engines leaves 520 answers per engine, which clears that band, while the same month split by engine and by six topics leaves about 87 answers per cell, which does not.

Do competitors have to be declared before collection starts?

On a platform that computes metrics against a declared roster, yes, and that decides whether a brand added in March has any history in January. AthenaHQ is explicit about the consequence: a discovered competitor has no ID and "cannot be used in `competitor_ids` filters or the metrics endpoints until the website tracks it". Peec AI suggests brands detected alongside yours when they appear at least twice, and you accept them into the project. On an API that stores the raw answers and does not bill reads, a competitor added later can be counted over the runs already collected, which is the practical reason to keep the answer text.