Muon

Diagnose what changed

Muon does not stop at "conversion dropped" — it checks whether the change is real, what kind of explanation fits, what has been ruled out, and what to do next.

What is Muon's investigation engine?

When Muon detects a Finding, it first checks the data and whether the change is real. It then tests technical, traffic and product explanations, ruling out the ones that do not fit.

The result shows what remains, what was excluded and what the data is still too weak to answer — so you start with a clear diagnosis instead of an empty dashboard.

Example

One possible diagnosis: a version regression.

In this example, a global conversion drop is concentrated in one environment and coincides with a release and error increase. Other changes can lead to different diagnoses — or show that the measurement itself is broken.

Signup conversion ↓
Safari62% of drop
iOSconfidence 0.93
/signupimpact high
Release 1.8.4timing match
Browser errors +320%related

Each step is chosen because it concentrates the change more than any alternative at that level. Muon measured that Safari accounts for 62% of the total drop, narrowed within it to iOS and the /signup page, then found the change began right after release 1.8.4 — coinciding with a 320% spike in browser errors.

This is offered as a likely, ranked path to investigate — correlation, not proven causation. The 0.93 is statistical confidence in the affected segment, not a claim that the release broke the flow. What triggers this →

What Muon checks

Different explanations need different checks.

Muon starts with the data itself, then checks technical conditions, traffic composition and product behavior. Dimensions such as browser, country, campaign and release are used only when they help distinguish the explanations.

BrowserOSDevicePageCountryReferrerUTM campaignReleaseError type

Audience & environment

Browser, OS and device isolate rendering and platform-specific regressions — a change that only hits Safari on iOS looks nothing like one that hits everyone.

Context & acquisition

Page, country, referrer and UTM campaign locate the change in the product surface and traffic source — a broken landing page or a single campaign cohort.

Release & errors

Release markers and error type connect the change to what shipped and what broke — the two signals most often behind a sudden product move.

Browser Health →
How diagnosis works

Check, rule out, then explain.

Checks run in order. Broken measurement stops the investigation; a product explanation is only shown after the data, technical and traffic alternatives have been considered.

1 · Check the data

Muon looks for missing events, contradictions and tracking changes. If measurement is broken, interpretation stops until it is fixed.

2 · Check the change

Historical comparisons and uncertainty tests distinguish a meaningful change from normal variation.

3 · Rule out explanations

Technical failures, traffic composition, automation, seasonality and product behavior are checked in order, using the data available for this case.

4 · Show what fits

Muon shows the best-fitting explanation, the checks behind it, what was ruled out and what remains unknown.

5 · Recommend the next step

The result ends with an action: fix measurement, filter traffic, inspect a product step, pause a rollout, collect missing data or simply check again later.

What changedCheck the dataRule out explanationsShow what fitsNext step

Every step before the last is deterministic. The explanation is the only stage that uses an LLM — and it can only describe what the statistics already found.

Why this matters

Most teams lose hours manually slicing dashboards — filtering by browser, then OS, then page, then release, trying to guess where a metric actually moved. Muon does that first round of investigation automatically, the moment a Finding appears.

You start from a narrowed, ranked hypothesis instead of a blank filter. Muon describes what it finds as likely and related — it points you at the segment and the correlated release or error to check, and leaves the final call to you. It surfaces likely causes; it never claims proven causation.

What triggers investigation

Investigation runs on every Finding — a structured signal that a change is real and meaningful, detected by statistics before anyone asks. Findings engine →

Browser Health signals

Frontend error spikes are one of the strongest related signals in investigation. Muon connects them to the affected metric, release and page. Browser Health →

FAQ

Questions, answered directly.

How does Muon find the affected segment?
Muon uses segment contribution to measure how much each dimension value accounts for the change, information gain to find which dimension best separates the affected traffic from the rest, and correlation scoring to link the change to releases and errors. The result is a ranked list of likely segments, not a guess.
Does Muon prove causation?
No. Muon surfaces likely, ranked candidates with a confidence score, describing relationships as likely or related — never as proven causation. It narrows where a change is concentrated and what it correlates with, so a human can decide.
Which dimensions does it check?
Muon investigates across browser, OS, device, page, country, referrer, UTM campaign, release and error type, looking for the dimension and value where the change is most concentrated.
Does an LLM do the investigation?
No. Deterministic checks decide whether the data is correct, whether the change is real and which explanations fit. An LLM can only phrase those checked results in plain language.
Can it correlate with releases and errors?
Yes. Release markers and browser error spikes are first-class signals in investigation. Muon checks whether a change began right after a release and whether it lines up with an error spike, offering these as related signals to investigate.
Open source · self-hosted

Let Muon run the first round of investigation.

The investigation engine ships in the open-source core. Spin up Muon with Docker and let it narrow the cause before you open a dashboard.