Core Concepts API reference - LLM Pulse

Filters & common parameters

• project_id required for all project-scoped endpoints.
• metrics or metric: CSV of mentions, citations, responses, mention_rate (or visibility), citation_rate, ai_visibility_score, avg_position, avg_mention_position, sentiment_very_positive, sentiment_positive, sentiment_neutral, sentiment_negative, sentiment_very_negative, net_sentiment
• granularity: day|week|month (weeks start Monday)
• Time window: either range (days) orfrom & to (ISO8601). If you send to, from is required. Defaults with from/to: missing from = 30 days ago (start of day); missing to = now (end of day).
• Filters: model, collection_id (tag id), prompt (prompt ID), country_code, language_code, competitors (CSV IDs)
• include_project (default true): set false to exclude your project (return only competitors).

Metric definitions

These semantics are aligned 1:1 with the Overview UI.

• mentions: daily sums of mentions_count. For week/month we compute the PERIOD TOTAL (Mon-Sun for weeks) and repeat that same total for every day in the period. If a period’s total is 0, we “carry” the last non-zero period total (sticky).
• citations: response-level count that includes visible citations and background source references. Each actor counts at most once per response. Uses the same period-total and sticky-carry behavior as mentions.
• responses: total number of prompt executions (API calls to LLMs). Uses project-level execution count regardless of actor. • day: daily sum. • week/month: period total (no sticky carry, days with 0 executions show 0).
• Mention Ratemention_rate (also accepted as visibility): percentage per actor, using the project-level executions as denominator under the same filters (mentions_count / prompt_executions_count × 100). • day: compute the daily ratio. • week/month: compute the ratio-of-sums for the period (Σ mentions / Σ executions) and repeat that value for every day in the period. If the period's denominator is 0, carry forward the last non-null value.
• citation_rate: percentage of responses that include a visible or background citation for the actor (citations_count / prompt_executions_count × 100). • day: daily ratio. • week/month: ratio-of-period-sums with sticky carry, same as visibility.
• AI Visibility Scoreai_visibility_score: position-weighted visibility metric that accounts for mention prominence. Uses reciprocal position weighting (position 1 = 100%, position 2 = 50%, position 3 = 33%, etc.). Higher scores indicate both more mentions AND better positioning. • day: compute daily weighted score / executions. • week/month: ratio-of-period-sums with sticky carry, same as visibility.
• avg_position: average visible citation position. Background citations have no visible position and are excluded. • day: daily average excluding nulls. • week/month: average within the period (ignoring nulls) and then propagate (carry) the last known value to days without data, matching the Overview chart smoothing.
• avg_mention_position: average position of the brand when mentioned in AI responses (position 1 = first mentioned, position 2 = second, etc.). Lower is better. • day: daily average excluding nulls. • week/month: average within the period with sticky carry.
• Sentiment metricssentiment_very_positive,sentiment_positive,sentiment_neutral,sentiment_negative,sentiment_very_negative: for each actor and day, percentage of sentiments in that bucket over all sentiments for that actor. • day: daily ratio (bucket_count / total_sentiments * 100). Days with no sentiments return null. • week/month: we compute a ratio-of-period-sums (Σ bucket / Σ total * 100) and repeat that percentage for every day in the period. If the period’s denominator is 0, we carry forward the last non-null value (sticky), matching the Overview smoothing.
• Net Sentiment Scorenet_sentiment: for each actor and day, we calculate a score in [-100, 100]: ( (very_positive + positive) − (negative + very_negative) ) / total_sentiments × 100. • day: using the daily percentages of each bucket (the metrics sentiment_*) we calculate net = (pos+very_pos) − (neg+very_neg). • week/month: the score inherits the same aggregation/smoothing semantics as the sentiment buckets because it is calculated from those series.

In /metrics/summary, total is a sum for count metrics and an average for avg_position, avg_mention_position, and net_sentiment.

Output Formats for BI Tools

Read endpoints return nested JSON by default, which is convenient for code but unreadable for business intelligence tools. Pass the optional output parameter to get the same data as a rectangular table that Tableau, Excel, Google Sheets or a warehouse loader can consume directly.

• output omitted (default): the nested JSON documented for each endpoint. Unchanged, so existing integrations are unaffected.
• output=flat: the same metadata plus columns (ordered column names), rows (one flat object per row) and row_count.
• output=csv: the same rows as text/csv, with columns as the header row.

Supported on /metrics/timeseries, /metrics/summary, /metrics/sov, /metrics/prompt_summary, /metrics/top_sources, the /search_console/* series and the list-returning /dimensions/* endpoints. Endpoints whose payload is not a single table (/metrics/agent_traffic, /metrics/ai_traffic, /search_console/summary, /dimensions/models, /dimensions/locales) reject output instead of ignoring it.

/metrics/sov also accepts view=over_time (default), view=current or view=breakdown: its payload holds several different shapes and a table can only hold one at a time.

Columns by endpoint

Example

GET /api/v1/metrics/timeseries?project_id=1&range=30&metrics=mentions&output=csv

date,actor_type,actor_id,actor_name,actor_domain,metric,value
2025-01-01,project,1,My Brand,mybrand.com,mentions,10
2025-01-01,competitor,2,Competitor A,competitor.com,mentions,5

Notes:

• Errors are always returned as JSON, never as CSV, so a failure is never mistaken for data.
• Percentages stay on the raw 0-100 scale, exactly as in the JSON response.
• Text cells that begin with an equals, plus, minus or at sign are escaped so spreadsheets do not execute them as formulas.
• Paginated endpoints also return their paging metadata in the X-Total-Count, X-Page and X-Per-Page response headers, which a CSV body cannot carry.
• Flat output is capped at 200,000 rows. Above that, narrow the date range, request fewer metrics, or use granularity=week.
• Endpoints without a flat projection reject output with ERR_INVALID_PARAM instead of ignoring it.

Cache & Conditional GET

Metrics endpoints (timeseries, summary, sov, top_sources) support conditional GET. We compute an ETag from the project version and a hash of your query params, and use the project’s updated_at as Last-Modified.

Send If-None-Match or If-Modified-Since to receive 304 Not Modified when nothing changed.

Errors

Errors use a consistent JSON structure. Branch on the code field. The message field contains error details when available; otherwise, it contains a default token such as missing_authorization:

Error body shape:

Typical 422 cases:

• invalid metrics: only mentions, citations, responses, mention_rate (or visibility), citation_rate, ai_visibility_score, avg_position, avg_mention_position, sentiment_very_positive, sentiment_positive, sentiment_neutral, sentiment_negative, sentiment_very_negative, net_sentiment are allowed.
• invalid granularity: must be day|week|month.
• from/to must be ISO8601, or from is required with to.
• range is required when neither from nor to are sent.
• unknown competitor ids: ...: IDs not found under the project.

Versioning & Rate limits

Current version: v1. Future breaking changes will bump the path (e.g. /api/v2).

Rate limiting: 300 requests per minute per API key. Contact us for higher quotas.

Metrics endpoints support conditional GET (ETag/Last-Modified); see Cache & Conditional GET.