Saltar al contenido principal

Analytics

The Analytics section helps you understand how users interact with Appilot and where your knowledge base needs improvement.

Key metrics​

Usage metrics​

  • Active users: How many people use the assistant daily/weekly/monthly
  • Questions asked: Total volume and trend over time
  • Conversations: Number of chat sessions started

Knowledge metrics​

  • Coverage: Percentage of active domains/pages with published knowledge
  • Request rate: How often users create knowledge requests (indicating gaps)
  • Resolution time: How quickly your team resolves requests

Impact metrics​

  • Top domains: Which web applications generate the most questions
  • Top pages: Which specific pages users need the most help on
  • Unanswered questions: Topics where knowledge is missing

Cost metrics​

  • Token usage: AI tokens consumed, broken down by source (extension, widget)
  • Model usage: Which AI models are being used and their cost

Using analytics effectively​

Prioritize knowledge creation​

Look at top pages with no knowledge — these are high-demand areas where users need help but your knowledge base has no answers. Creating content for these pages has the highest ROI.

Identify training gaps​

If a specific page has many questions despite having published knowledge, the content might be:

  • Unclear or incomplete
  • Outdated
  • Not matching the questions users actually ask

Review the questions and improve the content.

Measure ROI​

Track these before and after deploying Appilot:

  • Support ticket volume for covered domains
  • Time-to-complete for key workflows
  • Error rates on complex forms
  • User satisfaction scores

Monitor costs​

Token usage analytics help you:

  • Understand AI costs per domain or use case
  • Identify unusually high usage (possible misuse)
  • Plan capacity and budget

Conversations & Reports​

Each domain workspace in the Backoffice carries two side-by-side surfaces for inspecting what the agent did and how users felt about it:

  • AI Interactions lists every conversation the agent had on this domain. Filter by time range, response mode (🟢 / 🟡 / 🔵 / 🟠 / 🔴), feedback, or a free-text search. Click any row to open the conversation detail with the full inspector: prompts the agent rendered, the agent trace (iterations + tool calls), Knowledge chunks it received, Controls it considered, and the final response.
  • Reports is the inbox of complaints. It surfaces only turns where a user explicitly tapped 👎 and then "Send report"; silent 👎 ratings do not appear here. Click a row to open the same conversation detail, auto-scrolled to the reported turn so you can see exactly what the user objected to.

Both surfaces share one detail page, so you debug the same way regardless of how you arrived: from a triage scan of all conversations, or from the user-complaint queue.

Using Reports as the prioritization signal​

The Reports queue is the cheapest, highest-signal pointer at what to author next:

  1. Open Reports for the domain.
  2. Sort by recency. Read each report's description in the user's words.
  3. For each one, ask: did the agent emit 🔴 ("I don't have guidance"), or did it answer confidently but wrongly?
  4. 🔴 reports → add a Knowledge entry covering the missed topic.
  5. Confident-but-wrong reports → check whether existing Knowledge is misleading, or whether a missing Control / Form is blocking the agent from acting.

The deep inspector (Prompt Inspector, KB chunks, Controls considered) lets you confirm a hypothesis before authoring. See Authoring for the agent for the decision rules.