AI analyst
An AI analyst that reads your analytics for you
MetricMind's AI analyst is built into the analytics, so it can query your traffic, contacts, searches, Web Vitals, errors and app crashes directly. Ask a question in plain words and you get an answer with the query behind every number. Every morning it checks what changed and investigates the changes that matter. It drafts goals, funnels and alerts, but nothing changes until you click.
Updated
Why did contacts drop on the Android app this week?
Worked through 5 steps and 7 queries
Contacts from the Android app fell 21% (212 → 167). Almost all of the drop is on version 2.4.1, released Oct 24: crash-free sessions there are 96.1%, against 99.4% on 2.4.0, and most crashes happen on the Sell screen (NullPointerException in SellViewModel.publish).
Crash-free sessions by release
- 2.3.2 (38)99.5%
- 2.4.0 (40)99.4%
- 2.4.1 (41)96.1%
People on 2.4.0 contacted sellers as often as last week, so the crash explains the drop. Next step: fix the publish crash and ship 2.4.2; I'll watch crash-free sessions on the new release.
Assumptions · “this week” is Oct 21–27, today left out · contacts = calls, WhatsApp and email
Questions you can ask
Ask the way you would ask a colleague who knows your data. The analyst picks the metrics, runs the queries and explains the result, with a chart when a trend reads better as a picture.
- Why did contacts drop on the Android app this week?
- Which campaign brought visitors who actually contacted us?
- Which page got slower since the last release?
- What do people search for that we don't have?
- Is the one-page form better than the old one, or is it chance?
- What changed since Monday?
How an answer is built
- It runs read-only queries over your own site's data: pages, sources, campaigns, countries, devices, contacts, searches, sessions, goals, funnels, Web Vitals, errors, app releases and revenue.
- For “why” questions it splits a change across platform, source, page, country, device, browser and app release, and ranks what drove it.
- For “is it better” questions it runs a significance test and says plainly whether the difference could be chance.
- Every answer shows the queries behind it, lists its assumptions (which days “this week” covers, what counted as a contact) and offers follow-up questions.
The morning briefing
At your site's local hour, statistics (not the model) compare yesterday with the same weekday and with the week before: visitors, pageviews, contacts, goals, errors, Web Vitals and frustrated visits. Normal swings and changes on tiny numbers stay quiet.
When a change is real, the analyst investigates it from facts worked out by code first: which source, page, country or device carries the change, the hour it started, and which error is behind a rise. The finding ends with a next step: act, ask, or keep watching. New sites get a first look on day one and a first-week report after seven days.
Safeguards that keep it honest
| Safeguard | What it does |
|---|---|
| Query receipts | Every number in an answer links to the query that produced it. |
| Assumptions | Each answer and finding lists the judgment calls behind it; any of them can become a team rule in one click. |
| Number check | Every number in a finding is checked against the query results before it's saved; invented numbers go back to the analyst. |
| Fix verification | A recommended fix comes with a measurable check, re-measured a few days later: fix confirmed, or still not fixed. |
| Read-only access | Queries can only read, and only your site's data. |
| Click to apply | Goals, funnels, alerts, timeline notes and feature flags it suggests are created only when you click. |
| Untrusted text stays data | Page titles, search terms, event names and feedback written by visitors are treated as data, never as instructions; the analyst's test suite plants such instructions and checks they are ignored. |
Where answers reach you
- Chat in the dashboard, with conversation history.
- Findings and alerts by email, Slack and signed webhooks.
- Scheduled questions answered every morning or every Monday, with a quiet mode that only writes when something needs action.
- Claude, Cursor and other MCP clients can query the same metrics and ask the analyst from your editor.
Questions
What data does the AI analyst see?
Your question, your business context and the results of the queries it runs, which are mostly aggregated figures. It never receives identities: emails and traits sent with identify() stay out of its reach. The AI model provider processes these results only to write the answer.
Can the analyst change my settings?
No. It can propose a goal, funnel, alert, timeline note or feature flag (created switched off), and nothing is created until you click.
What happens when the data can't explain a change?
The analyst says so, names what it ruled out, and suggests what to measure so the next change can be explained.
Can I teach it how my business works?
Yes. MetricMind reads your site to learn what you do, and you can add rules such as what counts as a lead or which traffic to ignore. Every answer follows them.
Can I use it from Claude or Cursor?
Yes. MetricMind runs an MCP server: connect with your account (OAuth) or an API key, choose which sites the app may read and whether it may ask the analyst.
Try MetricMind free during the beta.
Websites and apps, with the analyst. No card. Open the live demo
More from MetricMind
Mobile app analytics
Privacy-first app analytics for iOS, Android, React Native and Flutter: screens, events, crashes and release health beside your website. No IDFA or ATT prompt.
AI traffic and crawlers
See the visitors ChatGPT, Claude, Perplexity, Gemini and Copilot send you, whether they convert, and which AI crawlers (GPTBot, ClaudeBot…) read your pages.
Cookieless analytics
Analytics with no cookies, no fingerprinting and no stored IP addresses, hosted in the EU: traffic, contacts, search, Web Vitals and errors in one 9 KB script.