Skip to main content

Analytics

The Analytics dashboard shows how your team uses ZenSearch — query volume across every surface, agent behavior and latency, AI model spend, search-widget engagement, and agent/automation reliability. It is organized into four tabs.

TabWhat it covers
Search AnalyticsQuery volume, surface breakdown, agent performance, satisfaction, and queries by API key
Model UsageRequests, tokens, cost, latency, prompt-cache hit rate, and a per-model / per-operation breakdown
Search WidgetSearches, clicks, click-through rate, zero-result rate, and top queries for the embeddable docs widget
ReliabilityAgent run cost (actual vs estimate), early-termination reasons, automation run status, and verification failures

The Search, Search Widget, and Reliability tabs offer 7-day / 30-day / 90-day windows. Model Usage additionally offers 24h and 1 year. Date ranges are computed in UTC so period boundaries line up with how the backend buckets events.

Search Analytics​

Summary stats​

A row of headline counts for the selected period:

  • Total Queries — every query event across all surfaces
  • Searches — keyword/semantic search queries
  • Auto — agentic chat turns on the default tier (a real agent loop ran)
  • Research — multi-step research-tier chat turns
  • Automations — scheduled or triggered automation runs
  • Satisfaction — thumbs-up rate over rated answers (shows -- until there are ratings)

Query volume over time​

A stacked area chart of Searches, Chat, and Agents over the period (daily for 7-/30-day, weekly for 90-day).

Surface breakdown​

A pie chart that splits every query into exactly one cell by surface: Web Search, Chat, Auto, Research, Automations, Widget, API, Slack, Teams, and Extension. Web traffic is further split by type (search vs chat vs auto vs research) so a single surface doesn't dominate the chart.

Agent performance​

This card spans every tier where an agent loop actually ran — default auto, research, and automations — not just research-tier chat. The headline number is Total Agent Turns, with these metrics underneath:

MetricMeaning
Avg IterationsMean iteration count over multi-step turns (more than one iteration)
P95 Iterations95th-percentile reasoning depth across agentic tiers
Avg Tool CallsMean tool invocations per agent turn
P50 LatencyMedian per-turn latency
P95 Latency95th-percentile per-turn latency
Auto / ResearchWeb turn counts split by tier

:::note Latency is full server wall-clock, not agent-loop-only The P50 / P95 Latency figures measure the full server turn — from request authentication through to the response (or stream) completing — which is the total user-perceived latency, not the time spent inside the agent loop alone. The percentiles are taken over agentic chat turns (auto + research) that recorded a duration; automation runs and legacy rows without a stamped duration are excluded so they can't skew the numbers. A dash (--) means no agentic turn recorded a duration in the window. :::

User satisfaction​

Thumbs-up rate, with a breakdown of positive ratings, negative ratings, and source clicks.

AI model usage (managed vs BYO)​

Splits queries by whether they ran on ZenSearch AI (managed models) or your own BYO models.

Queries by API key​

A bar chart of query counts per API key, covering requests authenticated with a key (the embeddable widget or direct API calls).

Model Usage​

This tab reports AI model consumption and cost across all operations (chat, search, embeddings, reranking, agents) — not only agent runs.

Summary cards​

  • Total Requests
  • Total Tokens
  • Input / Output token split
  • Total Cost (USD)

Performance​

Average and P95 latency, error rate, and error count for model calls in the period.

Prompt cache​

A cache hit rate over input tokens, plus cached-read and cache-write token counts. Long prompt prefixes are cached automatically by supported providers (Anthropic, OpenAI, Groq), so savings appear once traffic builds up.

Usage timeline, by model, and by operation​

A timeline bar chart of token usage, a per-model breakdown (cost, tokens, latency, cache hit rate, error rate, requests) with expand/collapse, and a per-operation view.

Search Widget​

Engagement analytics for the embeddable search widget, aggregated across all of a team's API keys:

  • Total searches, total clicks, click-through rate
  • Average response time
  • Unique queries and zero-result rate
  • Top queries, top clicked results, search trends over time, and the zero-result queries worth adding content for

Reliability​

Resilience signals for agent runs and automations. The tab is empty until chat agents or automations have executed.

Cost: actual vs estimate​

Total actual spend vs the pre-flight estimate, an estimate-accuracy percentage (100% means actual matched the estimate; above 100% means runs cost more than projected), plus counts of pre-flight blocks, mid-run cuts, and total runs. See Cost & Auto-Resume for how these ceilings are enforced.

Why runs stopped early​

A breakdown of early-termination reasons (for example budget exhaustion by iterations, tool calls, tokens, wall-clock, or cost).

Automation run status​

Run counts by lifecycle status, with a callout when runs breached their SLA targets.

Verification failures​

A breakdown of automation acceptance-criteria failure modes — for example insufficient sources or low confidence — for automations that define verification contracts.

  • Cost & Auto-Resume — the per-run and per-team dollar ceilings behind the Reliability tab
  • AI Agents — agent execution, the live cost meter, and the in-chat context-window meter
  • Evaluation — offline retrieval-quality scoring (NDCG, MRR, MAP)