Skip to content

AI engines

How to track your brand in Perplexity with an API in seven steps

Seven steps to measure how often Perplexity names your brand: route, prompt set, fields to read, runs per prompt and a bill from about $2 per 1,000 answers.

Six ring-shaped markers on stems, one filled solid blue among five outlined white ones

SEO consultant, and more recently GEO, at MentionBox

To track your brand in Perplexity with an API, you send a fixed set of buyer prompts to an API that answers them with Perplexity's search attached, then count three things in every answer: whether your brand is named, whether one of your pages was retrieved, and which competitors appear beside you. The seven steps below take a brand team from an empty spreadsheet to a weekly figure. The official route costs about $1.80 per 1,000 cited answers at Perplexity's published rates, read on November 15, 2026, and an API that computes the per-brand metrics for you costs $13 per 1,000.

This guide is for a brand team that wants its own numbers rather than a dashboard seat, and for the developer who will run the job. API for GEO is our editorial pick after hands-on testing, because one call returns the computed layer per brand and Perplexity sits among the engines it runs natively; the official Agent API is cheaper and hands you raw material instead. Steps 1 to 3 settle what you are measuring, steps 4 and 5 collect it, steps 6 and 7 turn answers into numbers, and the cost section at the end prices a month of it. Teams comparing platforms rather than endpoints will also want our ranking of AI visibility APIs for enterprise brand teams.

1. Decide which Perplexity you are measuring

Perplexity sells one brand and three different things to measure, and a report that mixes them is unreadable. The consumer product at perplexity.ai answers a logged-in user with its own model selection and its own tools. The Agent API answers the configuration you send. The Search API returns a ranked list of pages with no model in the loop at all.

Perplexity states the gap between the first two in its developer FAQ: "The API uses the same search system as the UI with differences in configuration, so their outputs may differ", and, on the same page, that the two "share search infrastructure, but their tools, configuration, and models differ, so results can differ".

SurfaceWhat it answersWhat a measurement of it supports
perplexity.ai, logged inPerplexity's own model, tools and personalisation"this is close to what a buyer sees", if someone can collect it
Agent API with the web_search toolthe model, effort level, filters and location you seta stable series against a frozen configuration
Search APIPerplexity's ranked retrieval set, no generated answer"which pages does Perplexity's index put forward for this query"

The honest sentence for an API series is the second one, and it remains the most useful of the three, because it is the only surface you can hold still for twelve months. Our explainer on what AI visibility measures and what it misses sets out where that claim stops.

2. Pick the route

Six routes return Perplexity answers or Perplexity sources, at prices that span a factor of thirteen. The choice turns on who computes the metrics: the cheap routes hand over raw answers for you to count, the expensive ones return per-brand figures ready to put in a slide.

RouteWhat it readsWhat comes backPrice per 1,000 answers
API for GEOPerplexity run natively, in the country and language asked forraw answer, cited URLs by category, brand mentions, position, share of voice, sentiment, themes, fan-out questions$13, down to $7 at its largest tier
Perplexity Agent API, web_search toolthe model you chose, with Perplexity's search attachedsearch_results with queries and results, inline [web:ID] markers, a usage.cost breakdownabout $1.80 at low effort on openai/gpt-6-luna
Perplexity Search APIPerplexity's ranked index, no generated answerresults with title, url, snippet, date, last_updated$5.00, or $1.00 with Fast Search
DataForSEO AI Optimization, Perplexity LLM Responsesthe official Perplexity modelsanswer text, annotations, search_results, a web_search flag, money_spent$0.60 of platform fees plus the model's tokens, after a $50 minimum payment
Bright Data AI Scrapersthe perplexity.ai interfaceprompt, answer in text, HTML and markdown, sources, answer URL$1.50
Profoundanswers captured in the browserbrand mentions, position, share of voice, sentiment, source categories, fan-out queriesnot published, Enterprise plan on quote

API for GEO, DataForSEO, Bright Data and Profound figures from our comparator, checked on October 23, 2026; Perplexity rates from its pricing page and DataForSEO's LLM Responses rates from its own pricing page, both read on November 15, 2026.

API for GEO bills one credit per prompt, per engine and per run, from $0.013 down to $0.007 per credit, with no upfront commitment published and analytics reads over stored runs unmetered, so re-reading a year of history 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, along with its webhooks, OpenAPI spec and official SDKs. Its terms allow resale, which decides the question for anyone reporting to clients under their own brand. The trade-off sits in the same table: $13 per 1,000 against $0.60 to $1.80 for the routes that hand you raw answers. A key is requested from its pricing page, and it comes with a sandbox project. MentionBox, which publishes this site, has a commercial link with API for GEO.

The competing strengths are published too, and they are real. DataForSEO is the cheapest metered route to official Perplexity answers, with SDKs in four languages, webhooks and an MCP server. Bright Data reads the interface and includes proxies, unblocking and parsing in its $1.50 per 1,000 records. Profound has the widest published engine list of the platforms we checked and documents its collection method plainly. The duels API for GEO vs Profound and API for GEO vs DataForSEO take those differences line by line, and our ranking of AI answer scraping APIs covers the cheap end.

3. Write a prompt set that can name your brand

A measurement is only as good as the questions behind it, and Perplexity rewards natural phrasing. A buyer asks "which project management tool do creative agencies actually use?", not "project management tool". Sales calls, support tickets and your site's internal search give you that wording for free.

Sort the list into two kinds, because they answer different questions:

  • Comparative prompts ask for a choice ("best", "alternatives to", "X or Y"). These are the answers where brands get named and ordered, so they carry your share of voice.
  • Informative prompts ask how something works. Fewer brands appear, and the value sits in the retrieval set: the domains Perplexity pulled to answer are the pages you need to be among.

Tag every prompt with a product line, a persona and a country before the first run, so figures can be split later without re-collecting. Fifty prompts per product line is enough to see a trend. Then freeze the wording: a prompt edited in month three breaks the series it belongs to, so a changed prompt becomes a new prompt with its own history. Our guide on building a prompt set for GEO covers the sampling in more detail.

4. Freeze the configuration that decides the source list

On the Agent API the source list is a function of your settings as much as of your prompt, so the settings are part of the measurement. Four of them move it.

The first is the effort level, which Perplexity documents as a bundle of model, step budget and search budget. From its effort page, read on November 15, 2026:

EffortModelMax stepsSearch results per callSearch type
fastopenai/gpt-6-luna110fast
lowopenai/gpt-6-luna515as set on the tool
mediumopenai/gpt-6-luna1515as set on the tool
highopenai/gpt-6.1-terra4015as set on the tool
xhighanthropic/claude-opus-5-530015as set on the tool

Bundles from Perplexity's Agent API effort page, read on November 15, 2026. An ultra level exists with no frozen configuration, which makes it unsuitable for a series.

From low upward the search budget is fixed at 15 results, 4,000 tokens of search context and 4,000 tokens per page. The step budget is the part that bills: a fast run makes at most one tool call, while an xhigh run has 300 steps available and no published ceiling on how many searches it will choose to make, each one metered.

The other three settings live on the web_search tool:

  • search_type, web for the standard index or fast for the cheaper, lower-latency one. The two need not return the same pages, so pick one and never change it mid-series.
  • filters, with search_domain_filter taking up to 20 domains as an allowlist or, with a minus prefix, as a denylist, plus search_recency_filter (hour, day, week, month, year) and four date filters in MM/DD/YYYY form that separate publication date from last update.
  • user_location, an object with country as an ISO 3166-1 alpha-2 code, plus region, city, latitude and longitude, the documentation noting that latitude and longitude must be provided together with country. A bare call has no country. Your buyer does.

5. Read the fields that carry the answer and its sources

A searched call declares the tool and nothing else. The endpoint is POST https://api.perplexity.ai/v1/agent, and its OpenAI compatibility page accepts POST /v1/responses as an alias:

curl https://api.perplexity.ai/v1/agent \
  -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "effort": "low",
    "input": "best project management tool for creative agencies",
    "tools": [{
      "type": "web_search",
      "search_type": "fast",
      "max_results": 15,
      "user_location": {"country": "US"}
    }]
  }'

The response is a typed output array rather than a choices array, and the sources sit in their own item. The shape, from Perplexity's Agent API quickstart, with our prompt in place of its example:

{
  "id": "resp_example",
  "status": "completed",
  "output": [
    {
      "type": "search_results",
      "queries": ["best project management tool creative agencies"],
      "results": [
        {
          "id": 1,
          "title": "Page title",
          "url": "https://example.com/page",
          "snippet": "Result preview text",
          "source": "web",
          "date": "2026-01-01"
        }
      ]
    },
    {
      "type": "message",
      "content": [{ "type": "output_text", "text": "Answer text with [web:1] markers" }]
    }
  ]
}

Five fields do the work of a brand report:

What you needWhere to read itWhat it gives you
the answer textoutput_text on the message itembrand mentions and their order inside the answer
the pages retrievedresults on the search_results item, with id, url, title, snippet, source, date, last_updatedwhether your domain was pulled at all
the pages actually usedthe [web:ID] markers in output_text, resolved against results[].idwhether your page was cited rather than merely retrieved
the searches that ranqueries on the search_results itemthe fan-out behind one prompt, and the invocation count you are billed for
the bill for that runusage, with input_tokens, output_tokens, tool_calls_details and a cost objectper-run cost, split by tokens and by tool

Field names from Perplexity's Agent API quickstart, its web search tool page and the Create Agent Response reference, read on November 15, 2026.

The distinction in the third row is the one most Perplexity reports miss. Retrieval and citation are two different events: a page can sit in results and never be referenced by a [web:ID] marker, which means Perplexity read it and did not use it. Reported as one number they look like progress; separated, they tell you whether to work on indexing or on the page itself. Our piece on how ChatGPT, Perplexity and Claude choose sources covers what drives the second step.

Two storage rules follow from the field shapes. The id on each result is an index scoped to one response and reused on the next call, so key your store on the URL and keep the id only to resolve markers inside that same answer. And the absence of a search_results item means no search happened and the answer came from the model's weights, which is a column in your table rather than a zero.

The same pipeline on DataForSEO's Perplexity endpoint reads different names for the same things. Its Live Perplexity LLM Responses endpoint takes a user_prompt of up to 500 characters and a model_name, offers web_search_country_iso_code for localisation on Sonar-family models, and returns annotations with the sources, a search_results array, a web_search boolean saying whether search was active, and money_spent for the token cost. Its documentation is explicit that "Perplexity uses web_search in all sonar-family models by default, but it's not guaranteed to work with every request", which is the same caveat in different words: store the flag with every answer.

6. Decide how many runs per prompt, and how to queue them

Perplexity does not answer twice the same way, so one run is one draw. cloro's study on sample size, published on August 14, 2026, measured a mean overlap of 0.40 between two runs of the same prompt on the same model: "only about 40% of its cited domains" came back the second time. The same study puts the statistical floor at "about 294 answers on each side" to state that visibility moved from 20% to 30%, and notes that at roughly 300 answers a visibility rate near 25% carries a margin of about five points.

Two designs reach that floor. Repetition runs the same prompt several times, which is the only way to learn whether a single prompt is stable. Breadth runs more prompts once each, which reaches the same brand-level total for the same money while telling you nothing about variance. Most brand teams want both: five runs on a core of twenty prompts, one run on a long tail of thirty.

Volume then runs into two published ceilings. Perplexity's rate limits page, read on November 15, 2026, sets the Agent API at 1 query per second on Tier 0, rising to 3 at Tier 1 after $50 of cumulative purchases, 8 at Tier 2 after $250 and 33 at Tier 4 after $1,000, while the Search API holds 50 query units per second at every tier, independent of spend. A thousand answers at 1 QPS is a sixteen-minute job, which is fine weekly and painful if you want everything inside one minute.

For anything beyond a handful of prompts, queue rather than block. The Agent API reference documents a background field that runs the response asynchronously: with stream: false the request "returns immediately with status: 'queued'", and you poll by id. A store field controls whether a response can be retrieved later, which is worth setting deliberately if your pipeline re-reads results rather than writing them straight to your own database.

7. Turn answers into five numbers

Five figures cover what a brand team is asked for, and the last two are specific to a citation engine like Perplexity.

MetricFormulaExample on 50 prompts × 5 runs = 250 answers
Mention rate (visibility)answers naming the brand ÷ all answersnamed in 55 answers: 22%
Share of voiceyour mentions ÷ all brand mentions in the same answers55 of 380 mentions: 14%
Average positionmean rank of the brand where it appears2.4, lower being better
Retrieved rateanswers whose search_results hold one of your URLs ÷ all answers90 answers: 36%
Citation rateanswers where one of your URLs carries a [web:ID] marker ÷ all answers40 answers: 16%

Worked example on a 250-answer month, to show the arithmetic. The formulas are set out in full in our piece on GEO metrics.

Read the gap between the last two rows before anything else. In the example above Perplexity retrieved one of the brand's pages in 36% of answers and used one in 16%, so the retrieval is working and the selection is not. That points at the pages, their structure and their freshness. A low retrieved rate with a high conversion to citations points somewhere else entirely, at indexing and at whether PerplexityBot can reach the content at all. Perplexity's crawler page describes PerplexityBot as the agent that surfaces and links websites in search results and states that it "is not used to crawl content for AI foundation models", with a published IP list for verification alongside the user-agent string.

Counting is where a month of collection gets lost, so three rules are worth fixing in writing before the first report:

  • Match each brand through a list of variants: the legal name, the product names and the domain. Count one answer once, even when the brand appears four times in it.
  • Decide whether a brand name inside a cited URL counts as a mention, as a citation, or as both, then never revisit the decision mid-series.
  • Fix the competitor list per topic. Share of voice moves on its own whenever a new brand enters the denominator, and a shifting denominator makes a flat series look like a decline.

An API that computes the layer skips this section. API for GEO's stats endpoint, POST /v1/queries/stats, answers with runs, visibility, avgPosition, entities[] carrying the per-brand metrics, sources[] with the cited URLs and their categories, and responses[] with the raw answer text, and its job endpoint, POST /v1/jobs/execute, takes models[], a country and an executionsPerQuery count, returning creditsCharged so the bill is readable per job. Profound answers with visibility_score, share_of_voice and average_position, grouped by date, model, topic, region, prompt or persona. Either way the definitions above are the ones to check against your own, because two vendors rarely compute share of voice over the same denominator.

What a month of Perplexity tracking costs

The volume is prompts × runs × frequency. For 50 prompts, five runs a week, Perplexity only, one country: 50 × 5 × 52 ÷ 12, about 1,083 answers a month.

RouteMonthly cost for about 1,083 Perplexity answers
API for GEOabout $14, at $0.013 per credit, metrics included
Agent API, low effort, one Fast Search per runabout $2 of usage, and you write the counting
Agent API, low effort, standard web searchabout $3 of usage, same counting
Perplexity Search API with Fast Search, no answerabout $1, ranked sources only
DataForSEO Perplexity LLM Responses, Liveabout $0.65 of platform fees plus the model's tokens, after a $50 minimum payment
Bright Data AI Scrapersabout $1.60, interface records, resale not allowed
Profoundon quote, Enterprise plan

Perplexity and DataForSEO rates from their own pricing pages, read on November 15, 2026; API for GEO, Bright Data and Profound from our comparator, checked on October 23, 2026. Token volumes behind the Agent API lines are a working assumption of about 4,500 input and 700 output tokens per answer, since Perplexity publishes a search budget rather than an average.

Two things outside that table decide the real bill. The first is the invocation count: from low effort upward several searches per run are normal, and at $0.0025 each a run that searches five times costs more in tool calls than in tokens. Read tool_calls_details in usage on a sample of fifty runs before you commit to a volume. The second is breadth: each extra engine or country multiplies the whole total, which is the budget decision that actually gets made. Our costing piece on tracking 100 prompts across five AI engines runs that arithmetic across the engines.

Retention deserves a line in the same conversation. Perplexity's developer FAQ states that "retention depends on the API, model, and account configuration, and Zero Data Retention (ZDR) is not the default", so if your prompt set encodes anything confidential, settle that before the first scheduled job rather than after.

Where Perplexity reports go wrong

  1. Reading citations off a chat completion. Sonar Chat Completions support ended on September 27, 2026 and the sources moved into a typed output item. Code that indexes choices[0] returns nothing rather than failing loudly. Our piece on Perplexity Sonar API citations and the Agent API covers the full field mapping.
  2. Letting the effort level float. Switching from low to high changes the model and the step budget, so the month that follows is a different experiment. Pin the effort level in your job definition and version it.
  3. Counting retrieval as citation. A URL in results was read; a [web:ID] marker means it was used. Merged into one metric, they hide which half of the problem you have.
  4. One run per prompt. A single draw against a mean overlap of 0.40 produces movement that is noise, and a quarterly report built on it will contradict itself.
  5. Comparing a logged-out scrape with an API series. They measure different surfaces under different configurations, and after cloro's November 12, 2026 entry the logged-out route is no longer available from every vendor who used to sell it.

A fair summary is that Perplexity gives you more structure than most engines and fewer brand figures than any dashboard. The retrieval set, the searches that ran, the citation markers and the per-tool cost all arrive in one object, which is more than ChatGPT's API hands over. The counting still has to happen somewhere, and the decision is whether you write it or buy it. Our companion guide on tracking your brand in ChatGPT with an API runs the same pipeline on the other large engine.

Frequently asked questions

Can I track my brand in Perplexity with the official Perplexity API?

Yes for the raw material, no for the metrics. A call to `POST https://api.perplexity.ai/v1/agent` with the `web_search` tool attached returns the answer text plus an `output` item of type `search_results`, holding the searches that ran and the pages retrieved with their `url`, `title`, `snippet`, `date` and `last_updated`. Nothing in that response is a brand figure: mention rate, share of voice, position and citation rate are counts you write yourself over stored answers. The search is billed at $0.0025 per invocation, or $0.001 with Fast Search, on top of the model's tokens.

Does the Perplexity API return the same sources as perplexity.ai?

Not reliably, and Perplexity says so. Its developer FAQ states that "the API uses the same search system as the UI with differences in configuration, so their outputs may differ", and that the two "share search infrastructure, but their tools, configuration, and models differ, so results can differ". On the Agent API you also choose the model, the effort level, the search filters and the location, four variables a perplexity.ai reader never touches. An API series is a measurement of your own configuration over time, which is still useful as long as the configuration never moves.

How much does it cost to track 50 prompts in Perplexity every week?

Fifty prompts run five times a week make about 1,083 answers a month. At the rates published on Perplexity's pricing page, read on November 15, 2026, that is about $2 of Agent API usage at `low` effort on `openai/gpt-6-luna` with one Fast Search per run, and about $1 on the Search API with Fast Search if a ranked source list without an answer is enough. With metrics included it is about $14 with API for GEO at $0.013 per credit, at the prices checked on October 23, 2026, and on quote with Profound, whose API comes with the Enterprise plan only.

Which API returns share of voice for Perplexity?

API for GEO and Profound both compute it. API for GEO lists Perplexity among the engines it runs natively and returns `visibility`, `avgPosition`, `entities[]` with their per-brand metrics, `sources[]` with categories and the raw `responses[]`, at $13 per 1,000 answers down to $7 at its largest tier, with analytics reads over stored runs unmetered and resale allowed. Profound returns `visibility_score`, `share_of_voice` and `average_position`, grouped by date, model, topic, region, prompt or persona, on an Enterprise contract whose terms forbid reselling the data. The official Perplexity endpoints and the scraping APIs return answers and sources, and leave the arithmetic to you.

How many times should I run each prompt in Perplexity?

Three times at least, five if the budget allows, on a fixed weekday. cloro's published study on sample size measured a mean overlap of 0.40 between two runs of the same prompt on the same model, meaning "only about 40% of its cited domains" came back the second time, and puts the volume needed to confirm a move from 20% to 30% visibility at "about 294 answers on each side". At around 300 answers per period the margin on a visibility rate near 25% tightens to roughly five points.

Can I scrape perplexity.ai instead of calling the API?

It is getting harder, and one vendor has already stopped. cloro's changelog entry of November 12, 2026 records that it discontinued its Perplexity endpoint because "Perplexity blocks access for logged-out sessions", and that the endpoint now returns `400 (MODEL_UNAVAILABLE)` at no credit cost. Bright Data still publishes a Perplexity scraper at $1.50 per 1,000 records, returning the prompt, the answer in text, HTML and markdown, the sources and the answer URL, and its terms do not allow reselling the data. Test any interface route on your own prompts before you build a year of reporting on it.