A traffic chart can show a peak. A useful report explains what the numbers support, what remains uncertain, and what you can check next. Nanolytica's Reports tab does that with a language model running on your device.
Try it with real public analytics
On iPhone and iPad, start with the suggested smallest model. All three choices remain available. Memory pressure can reload the page during loading or generation, including in CPU mode; a desktop browser may have more capacity.
No account is needed for this demo. It opens the Reports tab on a public dashboard for the last 30 days. Choose a model, press Load model, then generate a report or ask a question.
Try the local AI demo →Model downloads start only after you press Load model. GPU downloads are approximately 400 MB, 1 GB or 2.3 GB; the small CPU fallback is about 640 MB. Browser support, storage and memory affect availability.
One question, a bounded snapshot
- Select a site, date range, timezone and filters in the dashboard.
- The browser fetches aggregate metrics from Nanolytica, including available page, source, goal and traffic-timing breakdowns.
- A local worker receives a compact evidence snapshot and privacy-focused instructions. WebGPU runs the model on a supported GPU; the smallest model also has a CPU option.
- The report or answer streams into the page. Report PDFs are created locally; chat is separate and stays on the current page.
Questions worth asking
When does activity concentrate?
Ask: “Which hours and weekdays had the most pageviews? Explain any coverage limits.” Look for counts and the selected timezone, not assumptions about people's routines.
What should we improve next?
Ask: “Which pages deserve a closer look, and what aggregate evidence supports that?” An answer should separate recorded observations from possible explanations.
A small example
This is an illustrative calculation, not live demo data or a model-generated claim: if a period records 500 pageviews and /pricing accounts for 100, its pageview share is 20%. If source and conversion counts come from separate lists, that does not tell us which source converted on /pricing.
The same boundary applies to busy dates. A spike can suggest checking a release or campaign calendar; the data cannot establish that a holiday or launch caused it without supporting evidence.
What stays local—and what does not
Your analytics are collected and stored by Nanolytica. “Local AI” refers to inference: the selected aggregate snapshot, question and answer are processed in your browser, rather than sent to an external AI service. Model assets download from third-party hosts, which receive ordinary download-request metadata. Browser caches can retain those assets.
Changing the site, period, timezone, filters or model clears chat context. A small amount of recent conversation supports follow-up questions; it is not a persistent account chat history.
If loading or generation fails, Reports offers a copyable diagnostic with the model, task stage, browser capabilities and error category. It contains no analytics or chat text and is not sent automatically. A hard browser shutdown can only be detected later as an interrupted task, not as a confirmed crash.
Check the answer against the dashboard
Small models can misread numbers or make unsupported statements. Tables can be truncated to fit the model context. Missing data is not zero, anonymous visitor estimates are not verified people, and measured page engagement is not session duration. Treat recommendations as proposals to validate with aggregate evidence.
Read model setup and troubleshooting and the AI data boundary.