Skip to content

AI Analysis

SurveyThis uses advanced AI to transform raw survey responses into actionable insights — executive summaries, sentiment analysis, theme extraction, and more.

How It Works

When you run an AI analysis, SurveyThis processes your survey responses through a multi-stage LLM pipeline:

  1. Data Collection — All completed responses are gathered, including per-question analytics (averages, distributions, counts).
  2. Text Processing (Map) — Open-text responses are split into batches of 50 and processed in parallel. Each batch extracts themes, sentiments, notable quotes, issues, and suggestions.
  3. Merging (Reduce) — Results from all batches are intelligently merged — themes are deduplicated by name, sentiments resolved by majority, and quotes deduplicated.
  4. Analysis — The merged data is combined with quantitative analytics and sent to GPT-4o-mini with a specialized prompt for your chosen analysis type.
  5. Results — A structured JSON report is generated with findings, recommendations, and chart data.

Analysis Goal: If you set an analysis goal when creating the survey (e.g., "Understand why Q4 retention dropped"), the AI uses this context to focus its analysis on what matters most to you.

Running an Analysis

  1. Navigate to Dashboard → AI Insights.
  2. Select a survey with completed responses.
  3. Choose an analysis type.
  4. For Custom analysis, enter a specific question like "What do customers under 30 care about most?"
  5. Click Analyze. Processing time depends on response volume:
    • Up to 100 responses: ~15–30 seconds
    • 101–500 responses: ~1–2 minutes
    • 501–1,500 responses: ~2–4 minutes
    • 1,500+ responses: ~4–6 minutes

Results are stored permanently and can be viewed anytime from the AI Insights section.

Monthly Limits

AI analyses are counted per calendar month (resets on the 1st) across all your surveys:

  • Free — 1 analysis per month
  • Pro — 30 analyses per month
  • Business — Unlimited analyses

Each analysis request counts as 1 toward your monthly limit, regardless of the analysis type. Failed analyses (e.g., no completed responses) do not count.

Requirements

  • The survey must have at least one completed response.
  • More responses produce better analysis — we recommend at least 20–30 responses for meaningful AI insights.
  • Including open-text questions significantly improves analysis quality, especially for sentiment analysis and theme extraction.

Technical Details

  • Model: GPT-4o-mini (temperature 0.3 for consistency)
  • Text batching: 50 responses per chunk, up to 5 chunks processed in parallel
  • Output: Structured JSON with type-specific fields, chart data, and recommendations
  • Fallback: If LLM text processing fails for a question, a local lexicon-based sentiment analysis runs as a backup
  • Token tracking: Each analysis records tokens used for transparency

See Analysis Types for details on each analysis type and its output structure.

Responses & AnalyticsAnalysis Types