Ask the AI assistant about your data

Where: Product analytics › Intelligence › AI assistant. Three tabs: Ask, Change attribution and MCP.

How it answers

The assistant doesn't make numbers up. It turns your question into an analysis query — metric, time range, comparison, breakdown, filters — runs it on your project's real data and returns the result along with the query. Every answer is reproducible: Open in analysis shows exactly the same query.

Simple, clear questions are parsed by rules without calling a language model. Only vague questions, or follow-ups that need context, go to the model. If the project has no model configured, the top right says "rules only", and only phrasings the rules recognise can be answered.

Things you can ask

• Visitors over the last 7 days compared with the previous period
• Which channel brought the most sign-ups last month
• Bounce rate on mobile vs desktop
• Which pages visitors from ChatGPT read this week
• Sign-up conversion rate by country

Then keep going — "split that by device", "only the US". Follow-ups modify the previous query.

Getting sharper answers

• Name the metric. Visitors, sessions, pageviews and sign-ups are different numbers; "traffic" is vague.
• Name the time range. "Last 30 days" beats "recently".
• Use your own event names. If your sign-up event is sign_up, saying "sign_up events" is the most precise.
• If the answer misses the point, pick one of the suggested phrasings or rephrase.

The assistant doesn't save anything or create monitors on its own. If a result is useful, click Save as analysis to put it on a dashboard or turn it into a monitor later.

Why did it change?

When a number moves, the first question is which part moved. Click Why did it change? in an answer, or open the Change attribution tab:

1. Pick a metric and a comparison (previous period or same period last year).
2. Pick dimensions to split by — channel, referring site, country, device, landing page, platform and more — one or two levels deep.
3. Results are sorted by contribution: the change in that value ÷ the total change.

Say visitors fell by 1,200 last week. Change attribution might show that 900 of them were organic search from one engine, mostly on five blog posts. That gives you somewhere to look, which the total alone doesn't.

Connecting your own AI tool (MCP)

If you use an AI tool that supports the MCP protocol, you can connect your project and ask from there:

1. Create an API key under Configuration › Settings › API keys with the Read reports permission.
2. Copy the endpoint and configuration example from the MCP tab into your tool.
3. The tool gets a handful of read-only queries: list projects, overview, list metrics and events, run an analysis, run a saved analysis.

The connection is read-only — it can't change settings or delete data.

Usage

Model calls have a daily limit per site, shown at the top of the page. Questions the rules can answer on their own don't count against it.