AI Q&A (BETA)
Ask your data in natural language: which models, what it can query, and its limits — plus BETA notes.
What it is#
AI Q&A lets you ask questions in natural language; the system queries your GA4 / GSC data within a controlled scope and answers, listing the sources it used. Entry: Console → AI Q&A (marked BETA).
- Good for: quick lookups, spotting anomalies, comparing two dimensions
- Not for: exact financial reconciliation (answers are model-generated — trust the reports)
Models#
By default it runs on Cloudflare Workers AI — no keys for you to configure:
| Model | Notes |
|---|---|
| llama-3.3-70b (default) | Supports tool calling; best overall |
| llama-4-scout | Supports tool calling |
| llama-3.1-8b | Faster and cheaper, but no tool calling |
Note: When no model is configured the endpoint returns AI_NOT_CONFIGURED (503); the page then falls back to demo data (marked demo) — do not treat it as real.
What it can query#
The model can only fetch data through three controlled tools — it has no free database access:
| Tool | Purpose |
|---|---|
| query_ga4 | Query GA4 dimensions and metrics |
| query_gsc | Query Search Console dimensions |
| list_notes | Read site timeline notes |
Available fields (whitelist):
- GA4 dimensions: date, pagePath, pageTitle, sessionDefaultChannelGroup, sessionSource, sessionMedium, country, city, deviceCategory, operatingSystem, browser, eventName
- GA4 metrics: sessions, activeUsers, newUsers, totalUsers, screenPageViews, averageSessionDuration, engagementRate, bounceRate, eventCount, conversions, totalRevenue
- GSC dimensions: query, page, country, device
- At most 25 rows per query (10 by default)
Note: The page subtitle mentions PageSpeed / AdSense, but the tool layer currently covers only GA4, GSC and timeline notes — not PSI / AdSense queries.
Limits#
- At most 4 tool-calling rounds per question; history keeps the last 12 messages, each up to 4000 characters
- Answers are not streamed (returned whole); up to about 800 tokens per answer
- Conversation storage limit 50 and monthly ask limit 500 (shown in the UI for quota hints)
- Citations in answers are currently an empty array
BETA notes#
- Marked BETA; features and prompts are still changing
- The backend does not enforce a plan check yet (the UI copy calls it a Pro feature)
- On request failure it generates a demo answer marked demo — do not base decisions on it
Related docs#
- AI traffic (/ai-traffic)
- Overview & visitors metrics (/overview-metrics)
- Data latency & metric definitions (/data-latency)