Integration
AI citation tracking API: see which pages AI answers cite
Seven steps to log every URL an AI answer cites for your prompts, with the endpoint and fields each API publishes, from $1.20 per 1,000 collected answers.

By Mathias C.SEO consultant, and more recently GEO, at MentionBox
To track citations of your site in AI answers with an API, you send your prompt list to an API that runs those prompts on the AI engines, then read the list of URLs each answer cited and keep the ones on your own domain. Published rates for 1,000 collected answers run from $1.20 on a scraping API to $15.83 on a platform API, while the engines' own APIs bill the search instead: $10 per 1,000 searches on OpenAI and on Anthropic, $14 per 1,000 on Gemini after a free monthly allowance.
This guide is written for a brand team that wants a monthly citation figure it can defend, and for the developer who will write the collection job. API for GEO is our editorial pick after hands-on testing by Mathias C., because its response carries the cited URLs with a category and a brand flag next to the per-brand metrics, so the script stores rather than computes. Steps 1 and 2 settle what you are counting and where, steps 3 and 4 choose the API and show the requests, steps 5 to 7 cover sample size, the four figures and the bill.
What an AI citation tracking API returns
Three layers come back, and vendors expose them at different depths:
- The answer. The text the engine produced for your prompt, which you need in order to tell a citation apart from a brand mention in prose.
- The source list. One entry per cited URL, with the domain and usually the rank of the citation inside the answer. This is the citation layer.
- The classification. A label on each source, so you can separate your own pages from a competitor's, an editorial site, a forum or a reference work.
A citation and a brand mention are two separate counts. An answer can name your brand and cite nobody, cite your page without naming you, or do both. The four figures in step 6 keep those cases apart. Our companion guide on tracking a brand in ChatGPT with an API covers the mention side, and query fan-out covers the searches an engine runs before it picks any source at all.
1. Decide what counts as a citation
Write the rule down before the first request, because changing it later invalidates the trend line.
Four decisions settle it:
- URL or domain. Count a cited URL once per answer, and aggregate to the domain separately. A single answer that cites three pages of your site is one cited answer and three citations.
- Retrieval or citation. A retrieval is a page the engine read; a citation is a page it showed. Peec AI's domains report publishes both and divides them:
citation_rateis defined assum(citation_count) / sum(retrieval_count), withretrieval_countandretrieval_ratereturned beside it. OpenAI's web search guide draws the same line inside its own response, where thesourcesfield "returns the complete list of URLs the model consulted" while inline citations "show only the most relevant references". - URL identity. Decide upfront that a tracking parameter, a trailing slash and a redirect target are the same page. Profound ships a
clean_urlfield next tourlin its citation details, which tells you the problem is real enough for a vendor to solve it in the schema. - Ownership. Fix the list of domains that count as yours, including a documentation subdomain, a help centre on a vendor's domain and any syndicated copy. A citation of your help centre on a third-party platform is not a citation of your site, and you will want both counts.
2. Choose where you measure, the engines or your own logs
Two surfaces answer two different questions, and a brand team needs both.
| Surface | What it tells you | What it cannot tell you |
|---|---|---|
| Prompts run through an API | Which URLs the engines cite for the questions you chose, per engine, per country, over time | What real users actually asked |
| Your own server logs | Which AI agents fetched which pages, with the user agent the vendor documents | Whether the answer cited you, or which brand it named instead |
| Search Console and analytics | Clicks and impressions, with AI features folded into the Web search type | The citation itself, when nobody clicked |
Each row from the vendor documentation linked in this section, read on October 31, 2026.
The prompt-driven surface is the one this guide is about. The log surface is free, and the engines document the user agents you should look for. OpenAI separates three: OAI-SearchBot, which it says "is used to surface websites in search results in ChatGPT's search features"; ChatGPT-User, used "for certain user actions in ChatGPT and Custom GPTs"; and GPTBot, used "to crawl content that may be used in training our generative AI foundation models", on its bots page. Perplexity documents two: PerplexityBot/1.0, "designed to surface and link websites in search results on Perplexity", and Perplexity-User/1.0, which "might visit a web page to help provide an accurate answer and include a link to the page in its response". Anthropic documents three: ClaudeBot for training, Claude-User for user questions and Claude-SearchBot, which "navigates the web to improve search result quality for users".
A hit from Perplexity-User or ChatGPT-User is the closest free signal to a real citation event, since both are described as visits made while answering someone. It still stops short of a citation, because the engine may read a page and leave it out of the answer. OtterlyAI is one of the vendors that joins the two: its API introduction lists "agent analytics (AI crawler traffic on your own domains)" next to brand reports, prompts and citations.
Google is the exception on the analytics side. Its documentation on AI features and your website states that sites appearing in AI features "are included in the overall search traffic in Search Console" and are reported "in the Performance report, within the 'Web' search type", so the AI share cannot be split out.
3. Pick an API that publishes citation fields
Two families sell this. Scraping APIs return the answer and its source list at the lowest price per answer, and leave every count to you. Platform APIs return a citation report with the shares already computed, and sell access inside a subscription. Our ranking of GEO metrics APIs scores the second family, and the ranking for enterprise brand teams weights engine coverage and report-ready metrics the way a brand team does.
| API | Price per 1,000 answers, entry ticket | Citation data published | Engines, of the 10 we track | Main limit |
|---|---|---|---|---|
| API for GEO | $13, down to $7 at the largest tier; pay as you go, no minimum published | sources[] per prompt: url, domain, category, mentionsBrand, plus YouTube citations | 10 | highest price per answer of the scraping-capable options; four engines, plus webhooks, SDKs and the OpenAPI spec, confirmed in our hands-on test of October 26, 2026 and not yet documented on apiforgeo.com |
| Profound | not published, Enterprise plan on quote | citations and citation_details: url, clean_url, title, hostname, path, positions, groups, citation_category, first_cited_at | 9 | API on the Enterprise plan only; terms forbid reselling the data; sync only |
| Peec AI | not published | domains report with citation_count, retrieval_count, citation_rate; domain_classification and url_classification filters | 9 | price on request; collection method not named for several engines |
| Scrunch | not published, Agency Core from $500 a month | citations[] per response: url, domain, source_type, title, snippet | 9 | collection method not published; no price per answer |
| Trakkr | not published, API on the $500 Scale plan | /get-citations with six views, source_type, mentions_brand, domain_rating, and a citation_gained webhook | 8 | no answer text in the endpoints we read; API gated at $500 a month |
| OtterlyAI | $15.75, Standard plan at $189 a month | citation statistics endpoint: citationShare, domainRank, mostCitedUrls[], competitor shares | 7 | no source category in the citation statistics schema; 2,000 API requests a month on Standard |
| cloro | $4, Lite plan at $30 a month | aioverview.sources[]: position, url, label, description, plus videos and ads | 6 | no share of voice, position or sentiment; no Claude |
| DataForSEO | $1.20 on the standard queue, $50 minimum payment | cited URLs in the scraped answer, plus an LLM Mentions database at $0.001 per row | 5 | no computed metrics; interface scraping for ChatGPT and Gemini only |
Prices, engine counts and plan floors from each vendor's pricing page and API reference, checked on October 23, 2026. Field names from the API reference pages linked in step 4, read on October 31, 2026.
API for GEO bills one credit per prompt, per model and per run, from $0.013 down to $0.007, and its AI Visibility API page states that "the AI Visibility API does not bill reads. Credits are consumed by executions, when a prompt runs against a model." On a citation job that has a direct consequence: re-aggregating three months of stored source lists under a new ownership rule costs nothing. It covers the ten engines we track: ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok are listed on its own pages, while Gemini, Google AI Overviews, DeepSeek and Meta AI were confirmed in our hands-on test of October 26, 2026. Its terms allow resale, and its score of 9.0 is the best of the 55 APIs this site rates. The honest trade-off is the price per answer, roughly ten times DataForSEO's. MentionBox, which publishes this site, has a commercial link with API for GEO.
The competing strengths are published too, and they are real:
- Profound has the deepest citation schema of the eight, with the character positions of each citation in the answer text and a
first_cited_attimestamp per URL, plus Python and TypeScript SDKs and a hosted MCP server. - Peec AI is the only one that publishes the retrieval denominator, and its nine built-in domain classes (
CORPORATE,EDITORIAL,INSTITUTIONAL,OTHER,REFERENCE,UGC,COMPETITOR,OWN,RELATED) save you a classification project. Its API introduction now states that access "is available on Pro, Advanced, Enterprise, and agency plans", a change our guide on API for GEO against Peec AI follows. - Trakkr is the only one in the list with a
citation_gainedwebhook event, signed with HMAC-SHA256, which turns citation tracking into an alert rather than a monthly read. - OtterlyAI computes the share for you, and joins citation data to crawler traffic on your own domains in the same API.
- cloro documents its collection method, runs real browser sessions in 250 countries and gives 500 free credits a month, which is enough to prototype the whole pipeline.
- DataForSEO is the cheapest route at volume, with SDKs in four languages, webhooks, up to 100 tasks per call and a separate LLM Mentions API billed at $0.001 per row.
4. Read the fields, with the requests that return them
Field names decide how much code you write. Three requests show the three shapes.
API for GEO returns the citation layer inside the per-prompt statistics. Its prompt tracking page documents POST /v1/queries/stats, whose example response carries runs, visibility, avgPosition, an entities[] array, the raw responses[] and the source list:
{
"runs": 20,
"visibility": 35.0,
"avgPosition": 2.4,
"sources": [
{
"url": "https://review-site.com/best-tools",
"domain": "review-site.com",
"category": "review",
"mentionsBrand": true
}
]
}
One pass over sources[], filtered on your own domains, gives you the cited-answer count. The category field is what lets you report where the engine goes when it does not come to you, and mentionsBrand tells you whether the cited page names the brand.
Profound asks for the citation layer explicitly through its include array. Its query answers endpoint takes a category, a date window and a page size up to 200:
curl -X POST https://api.tryprofound.com/v2/prompts/answers \
-H "Authorization: Bearer $PROFOUND_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"category_id": "00000000-0000-0000-0000-000000000000",
"start_date": "2026-10-01",
"end_date": "2026-10-31",
"include": ["date", "model", "response", "citations", "citation_details"],
"limit": 200
}'
The documentation describes citation_details as "citation metadata", where positions identify citation locations in the answer text and each groups entry "represents a rendered citation pill". That last detail is the one most pages miss: an engine can group several URLs behind one visible source pill, so the number of pills a reader sees is lower than the number of cited URLs you count.
OtterlyAI skips the raw layer and returns the shares. Its citation statistics endpoint is a read over a brand report:
curl -G https://data.otterly.ai/v1/reports/brand/$REPORT_ID/citations/stats \
-H "Authorization: Bearer $OTTERLY_KEY" \
--data-urlencode "startDate=2026-10-01" \
--data-urlencode "endDate=2026-10-31" \
--data-urlencode "country=us" \
--data-urlencode "domainLimit=50"
It answers with domainRank, a domainCitations object holding current, total, citationShare and mostCitedUrls[], a citations[] array and a competitors[] array with each competitor's domain share. Nothing is left to compute, and nothing can be recomputed either, which is the trade both platform APIs make.
5. Set the sample size and the schedule
A citation list is a draw, not a fact. cloro's published study on sample size, dated August 14, 2026, measured the overlap between the cited domains of two runs of the same prompt on the same model, minutes apart: "the mean overlap was 0.40: the identical model answering the identical prompt repeated only about 40% of its cited domains on the second run", with per-prompt overlap from 0.06 to 1.00.
The same study gives the volumes. About 300 answers in a period resolves a rate to roughly ±5 percentage points, 72 answers carry a 95% interval of about ±10 points, and a claimed move from 20% to 30% needs about 294 answers on each side. Its own illustration puts 50 prompts across 6 engines sampled daily at 2,100 answers a week, and concludes that "weekly reads are sound and daily reads are noise".
Three rules follow for a citation job:
- Run each prompt at least four times a month and compare quarters, not weeks. A 60-prompt list on four engines at four runs is 960 answers a month, which clears the ±5-point band on a quarterly comparison.
- Keep the engine, the country, the language and the day of the week fixed. A change in any of them moves the source list on its own.
- Store the raw source lists, not only the aggregates. The 0.40 overlap figure means your month-to-month deltas will include real churn, and you can only separate churn from a collection change if the URLs are still on disk.
6. Compute four citation figures
Four counts cover most reporting, and each answers a question a brand team actually gets asked.
| Figure | Formula | Example on 60 prompts × 4 engines × 4 runs = 960 answers |
|---|---|---|
| Cited-answer rate | answers citing at least one of your URLs ÷ all answers | 134 answers: 14.0% |
| Citation share | your citations ÷ all citations in those answers | 210 of 4,800: 4.4% |
| Page concentration | citations on your top three URLs ÷ all your citations | 148 of 210: 70.5% |
| Cited without mention | answers citing you but not naming your brand ÷ cited answers | 31 of 134: 23.1% |
Formulas ours, worked on an illustrative 960-answer month. The field names needed for each one are in step 4.
The first figure is the headline. The second puts you next to everyone else the engines read on the same questions, and it is the one that moves when a competitor publishes a comparison page. The third is the one that drives the content plan, because a 70% concentration on three URLs means the rest of the site is invisible as a source. The fourth is the gap worth a meeting: the engine trusted your page enough to cite it and still did not name you, which is usually a page that answers the question without stating who wrote it.
Two of these need the answer text as well as the source list, which rules out any API that returns citations only. Trakkr is the case to watch there: the endpoints we read publish no answer-text field, and its prompts page states that "the tracking collection asks for a ranked list of brands for the question, rather than reproducing a full consumer chat".
7. Put a price on the job
The volume is prompts × engines × runs × countries × frequency. For 60 prompts, four engines, four runs a month, one country:
| Route | Monthly cost for 960 collected answers |
|---|---|
| API for GEO | $12.48 at $0.013 a credit, categories and per-brand metrics included |
| cloro, Lite plan | $30, the plan floor; that volume uses about 4,800 of its 37,500 credits |
| DataForSEO LLM Scraper, standard queue | about $1.15 of usage after a $50 minimum payment, for the two engines it scrapes from the interface |
| OtterlyAI, Standard plan | $189, the first plan with API access, capped at 2,000 API requests a month |
| Trakkr, Scale plan | $500, the plan its documentation requires for API access |
| Profound | on quote, Enterprise plan only |
Our arithmetic on the rates each vendor published, checked on October 23, 2026: 60 prompts × 4 engines × 4 runs = 960 answers.
Three things move that total before you commit. Speed is billed: DataForSEO charges $1.20 per 1,000 on the standard queue, $2.40 on the priority queue and $4 live, and a cloro request in sync mode costs 7 credits instead of 5. Each extra country multiplies the whole line, and citation lists differ by country more than brand mentions do. And a plan floor does not scale down, so a platform API at $189 or $500 a month costs the same whether you collect 960 answers or 9.
Build it on the engine APIs instead
If you only need a handful of engines and have someone to write the counting, the official APIs return citations directly and bill the search rather than the answer.
| Engine API | Citation fields | Search price |
|---|---|---|
| OpenAI web search tool | url_citation annotations with "the URL, title and location of the cited source", a web_search_call item, and sources with "the complete list of URLs the model consulted" | $10 per 1,000 calls plus the tokens the search feeds the model |
| Anthropic web search tool | web_search_result with url, title, page_age, and web_search_result_location citations with url, title, cited_text up to 150 characters | $10 per 1,000 searches plus token costs |
Gemini google_search tool | google_search_call, google_search_result and annotations | $14 per 1,000 search requests, after 5,000 free a month shared across Gemini 3 and newer models |
| Perplexity Sonar | search_results with title, url, date, last_updated, snippet, source | included in the request price |
From each provider's own documentation: OpenAI's web search guide, Anthropic's web search tool, Gemini's Google Search tool and Perplexity's API reference, read on October 31, 2026.
Anthropic's citation block is the most precise of the four, since cited_text gives you the sentence the model took from your page, and the documentation notes that cited_text, title and url "do not count toward input or output token usage". The limit of this route is the same one that applies to brand tracking: you are measuring a model you configured, not the assistant your buyer uses, and the model decides whether to search at all. Our piece on the web search tool against ChatGPT answers sets out that gap, and the same caution applies to Google, where no AI Mode API exists.
Mistakes that cost a quarter of data
- Counting domains when you meant answers. Ten citations in one answer is not ten cited answers, and a domain-level count hides whether your citations come from one lucky prompt or thirty.
- Letting URL variants split a page in two. A redirect, a
utm_sourceparameter and an AMP copy become three cited pages in a naive count. Normalise at ingestion, not in the report. - Mixing the engine's API with its interface in one series. The two surfaces cite differently, and a month that mixes them measures two populations. Log the collection route with every answer.
- Reporting a citation rate without the retrieval rate. Peec AI's two counters exist because they lead to different work: never retrieved means a crawling or indexing problem, retrieved and never cited is a content problem.
- Ignoring grouped citations. Profound's
groupsfield exists because engines render several URLs behind one pill. A count of pills and a count of URLs are both defensible, and mixing them inside a quarter is not. - Treating one run as a measurement. With a mean overlap of 0.40 between two runs of the same prompt, a single run is a sample of one, and most of the month-on-month movement a one-run setup reports is noise.
Our view
Which API to build on depends on what you will do with the source list:
- Lowest price per answer, developers on hand. DataForSEO at $1.20 per 1,000 on the standard queue, with your script doing the normalisation and the counting. Our score: 7.7 out of 10. The catch is engine coverage, five of the ten we track, with interface scraping on two.
- Citation shares already computed. OtterlyAI at $15.75 per 1,000 inside a $189 plan, with
citationShare,mostCitedUrlsand competitor shares returned, and crawler traffic on your own domains in the same API. Our score: 6.7 out of 10. - Deepest citation schema. Profound, for a brand team that can buy an Enterprise contract and wants citation positions, categories and a first-cited date per URL. Our score: 6.5 out of 10, held down by the undisclosed price and the resale ban.
- Raw source lists and the computed layer in one response. API for GEO, our editorial pick, at $13 per 1,000 answers down to $7 at its largest tier, pay as you go with no published minimum, reads over stored runs not billed, resale allowed, and 9.0 out of 10. Its price per answer is about ten times DataForSEO's, and four of its ten engines plus its webhooks, OpenAPI spec and official SDKs were confirmed in our test of October 26, 2026 rather than documented on apiforgeo.com.
In our view the decision turns on one question: whether you will ever want to recount. A platform API that returns shares is cheaper to integrate and impossible to re-aggregate when the ownership rule changes, while an API that hands back the URLs leaves you a month of work and a dataset you own.
Frequently asked questions
What is an AI citation tracking API?
It is an API that runs a prompt on an AI engine and returns the list of URLs that answer cited, usually with the domain, the rank of the citation and a category for the source. You give it the prompts your buyers type, it gives you back the sources the engine put under its answer, and you count the ones on your own domain. Prices published on October 23, 2026 run from $1.20 per 1,000 collected answers on DataForSEO's LLM Scraper to $15.83 on Promptwatch's Essential plan, with Profound, Peec AI, Scrunch and Trakkr selling API access inside a subscription or on quote.
Which APIs return the exact URLs an AI answer cited?
All eight APIs in this guide do, at different levels of detail. API for GEO, our pick after hands-on testing by Mathias C., returns `sources[]` with `url`, `domain`, `category` and a `mentionsBrand` boolean per source, alongside share of voice, position and sentiment, at $13 per 1,000 answers down to $7 at its largest tier, and bills one credit per prompt, per model and per run. Profound returns `citations` plus a `citation_details` array carrying `url`, `clean_url`, `title`, `hostname`, `path`, `citation_category`, `first_cited_at` and the character positions of each citation in the answer; OtterlyAI returns a citation statistics endpoint with `citationShare` and `mostCitedUrls`; Trakkr exposes a `/get-citations` endpoint with six view modes.
How much does it cost to track citations for 60 prompts a month?
Sixty prompts on four engines, run four times a month, is 960 collected answers. At the rates published on October 23, 2026: about $12.48 with API for GEO at $0.013 a credit, with the citation categories and the per-brand metrics included; about $1.15 of usage with DataForSEO's LLM Scraper on the standard queue, after its $50 minimum payment, for the two engines it scrapes from the interface; $30 a month on cloro's Lite plan, where that volume uses about 4,800 of its 37,500 credits. Platform APIs are sold by the plan rather than by the answer: OtterlyAI's first plan with API access is $189 a month, and Trakkr's documentation states that API access requires its $500 Scale plan.
Can Google Search Console show which AI answers cited my pages?
No. Google's documentation on AI features states that sites appearing in AI features such as AI Overviews and AI Mode "are included in the overall search traffic in Search Console" and are reported "in the Performance report, within the 'Web' search type", with no separate line. The Search Analytics API documents six search types, `discover`, `googleNews`, `news`, `image`, `video` and `web`, and none of them isolates a generative AI answer. Search Console also counts clicks and impressions, never the URL an answer cited when nobody clicked.
What is the difference between a retrieval and a citation?
A retrieval is a page the engine read while building its answer; a citation is a page it showed as a source. Peec AI's domains report publishes both counts and divides them: its `citation_rate` is defined as `sum(citation_count) / sum(retrieval_count)`, with `retrieval_count` and `retrieval_rate` returned next to it. OpenAI's web search guide draws the same line in its own response, where the `sources` field "returns the complete list of URLs the model consulted" while inline citations "show only the most relevant references". A page that is read and never cited is a different problem from a page that is never read.
How many runs per prompt before a citation figure is stable?
More than one, and the floor is published. cloro's 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, with per-prompt overlap from 0.06 to 1.00. It 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%. Four runs per prompt per month on a 60-prompt list reaches that range on a quarterly comparison.