Quick answer
Root cause analysis (RCA) is a structured way to work back from a problem to the underlying reason it happened, so the fix removes the source instead of the symptom. For a metric drop, it means finding where and when the change started, then testing the candidate cause.
Formula: Contribution of a segment = segment change ÷ total change
What is root cause analysis?
Root cause analysis (RCA) is a structured way to work backwards from a problem to the underlying reason it happened, so that the fix removes the source instead of treating the symptom. The test of a root cause is practical: if you fix it, the problem should not come back. Google's site reliability engineering book describes postmortems as recording the root cause or causes of an incident and understanding all contributing causes well enough to put preventive actions in place.
In analytics, RCA usually starts with a metric that moved: sales down 22%, sign-ups flat, errors up. It sits at the end of diagnostic analytics. Diagnostic work breaks a change down to find where it is concentrated; RCA continues to why, then confirms the fix worked. The method names matter less than the discipline. Most wrong fixes come from skipping a step, not from choosing the wrong tool.
Symptom, proximate cause, root cause
Three levels are easy to confuse, and each needs a different action.
| Level | Question | Example (sample) | Action |
|---|---|---|---|
| Symptom | What did we notice? | Orders fell 22% week on week | Alert, triage |
| Proximate cause | What directly produced it? | The Pay button did nothing in mobile Safari | Rollback or hotfix |
| Root cause | Why was that possible? | Pre-release tests ran only on Chrome; the browser list was never updated | Change the process so it cannot recur |
The chain in this example, written as five whys: orders fell → payments failed in mobile Safari → the release used code the Safari versions in use could not run → no Safari device was in the pre-release tests → nobody owned the browser list. Stopping at the proximate cause is the commonest failure. Taiichi Ohno's well-known example is a stopped machine: replacing the blown fuse restores it, but the problem returns within months unless you ask why the fuse blew, and why that happened, until you reach something you can change.
Common RCA methods and when each fits
| Method | How it works | Fits | Watch out for |
|---|---|---|---|
| Five whys | Ask "why?" of each answer, roughly five times | A single chain of cause in a process | Different people give different chains; a problem can have more than one root cause (Institute for Healthcare Improvement) |
| Fishbone (Ishikawa) diagram | List candidate causes by category (people, process, tools, environment) | Brainstorming when the cause is unclear | Lists possibilities; it does not test them |
| Fault tree | Work from the failure down through AND / OR combinations of events | Engineering systems with redundancy | Needs a good model of the system |
| Change analysis | Ask what changed just before the problem began | Metric drops, outages, regressions | Misses causes that are not changes (slow drift) |
| Segment breakdown | Split the metric by time, channel, device, page, version | Product and marketing metrics | Slicing endlessly can find chance patterns |
For a metric drop, the last two do most of the work. Google's SRE book calls the recent-changes question "what touched it last": working systems tend to keep working until something external disturbs them, so look at deployments, configuration changes and shifts in load first, and annotate graphs with deployment times to see the link.
A data-first method for metric drops
- Pin the start. Look at hourly or daily data and find when it actually began, not when someone noticed. Compare with the same weekday and with seasonality.
- Check the data before the product. A tracking script that disappeared, a bot spike or a delayed import look exactly like a real drop.
- Split from the top down, in order. Break the change down by time, channel, device, browser or version, and page. At each level, keep the cell that explains most of the change, and continue inside it. Stop when the next split no longer concentrates anything.
- Match it to the change log. Releases, campaign changes, price changes, outages. This is hypothesis formation, not proof.
- Test the candidate. It must pass four checks: timing (it began when the cause arrived), size (it explains most of the drop), specificity (groups that did not get the change stayed flat) and reversal (undoing it brings the metric back).
- Fix the root, then verify. Roll back or fix the proximate cause now, change the process that allowed it, and watch the same breakdown to see that the cell recovers.
Google's SRE book describes the same loop: propose causes, test them by comparing what the theory predicts against what you see, and note that "correlation is not causation": two things can move together because they share a cause, or by coincidence.
Worked example: orders down 22%
Sample data: an online store. Last week: 40,000 sessions, 1,200 orders (3.00%). This week: 40,200 sessions, 936 orders (2.33%). Change: −264 orders.
Step 1: traffic or conversion? Sessions are flat (+0.5%), so the loss is in conversion, not in how many people came.
Step 2: split by device. Contribution = cell change ÷ total change.
| Device | Last week | This week | Change | Contribution |
|---|---|---|---|---|
| Desktop | 700 | 690 | −10 | 3.8% |
| Mobile | 500 | 246 | −254 | 96.2% |
| Total | 1,200 | 936 | −264 | 100.0% |
Step 3: split mobile by browser.
| Mobile browser | Last week | This week | Change | Contribution to total |
|---|---|---|---|---|
| Safari | 330 | 78 | −252 | 95.5% |
| Chrome | 170 | 168 | −2 | 0.8% |
Step 4: pin the start. Last week's mobile Safari orders averaged 47 a day (330 ÷ 7). This week: Monday 20, Tuesday 11, Wednesday 12, Thursday 10, Friday 9, Saturday 8, Sunday 8 (total 78). Assume that about 10 of a typical day's 47 mobile Safari orders arrive before 10:20. Ten of Monday's 20 orders did, in line with normal; the other 10 came in the remaining 13 hours 40 minutes, when a normal Monday adds about 37. The release log shows a release at 10:20 on Monday.
Step 5: look inside the funnel. Mobile Safari checkout starts barely moved (1,100 → 1,090), but payments completed fell from 330 to 78: 30.0% → 7.2% of starts. The loss sits at the payment step.
Step 6: test. Timing matches the release, size explains 95.5% of the drop, specificity holds (Chrome and desktop are flat), and in this sample rolling the release back on Tuesday morning returned Safari to about 44 orders a day, passing the reversal test. The root cause, in the five-whys chain from earlier, is that the release shipped code the Safari versions in use could not run and no test caught it.
Note what the contribution tables did not say: they located the problem to one browser and one hour. They did not say why. The "why" came from the change log and a test.
Common mistakes in root cause analysis
- Stopping at the first answer. "The button was broken" is the proximate cause. Ask why it was possible.
- Stopping at a person. "Someone made a mistake" is almost never a root cause; ask what let the mistake through.
- Assuming one cause. Several factors can combine. Name the main ones.
- Taking correlation for cause. A release and a drop at the same time is a hypothesis until it passes the four tests.
- Slicing until something appears. With enough dimensions, one cell will look odd by chance. Prefer splits you would have made anyway, and watch for mix shift, where each group improves while the total falls (see diagnostic analytics).
- Not verifying. A fix is only confirmed when the same cell recovers.
- Confusing normal swings with problems. Check anomaly detection and seasonality before starting an investigation.
For a traffic-specific checklist, see sudden drop in website traffic.
Root cause analysis in TapCub
Here is how the orders example looks in TapCub. TapCub helps you find where and when; it does not decide the cause for you.
- Catch it early. In All-platform analytics › Operations › Monitors, a monitor on a built-in measure (visitors, sessions, bounce rate, error rate, errors) or a saved metric (for example orders, once you define it) can use a period-change, forecast-band or anomaly condition, with a dry run that shows how many times it would have fired in the last 30 days. Add a collection-gap monitor so a broken script is caught within the hour. Monitors only notify; they do not act. See traffic alerts.
- Rule out the data. Check Web analytics › Bots for a bot spike, and the collection-gap alert for missing data, before blaming the product.
- Locate it. Open All-platform analytics › Operations › AI Assistant and the Why did it change tab. It compares against the previous period or the same period last year and splits by channel, referrer domain, country, device, entry page, platform and related dimensions, one or two levels deep, ranking each value by its contribution (that value's change ÷ the total change), the same calculation as the tables above. It reports where, not why, and it does not create monitors or send notifications.
- Go one level deeper. Web analytics › Audience lists browsers and operating systems and can expand into versions; click a browser to filter, then open sources or pages. Its version breakdown is meant for spotting one browser release, and Monitoring › Errors then shows errors for it.
- Match the change log. Annotations on the charts record releases and campaigns, and each new app or mini-program version gets a Release annotation automatically. Add a web release as an annotation so the step-4 question has an answer.
- Verify the fix. Watch the same breakdown the following week. The cell should return to its normal range.
See also: What is troubleshooting? Definition, six steps and an example.
Sources
- Beyer, B., Jones, C., Petoff, J. and Murphy, N. R. (eds.), Site Reliability Engineering, chapter "Effective Troubleshooting" (Google, 2016).
- Beyer, B. et al. (eds.), Site Reliability Engineering, chapter "Postmortem Culture: Learning from Failure" (Google, 2016).
- Institute for Healthcare Improvement, 5 Whys: Finding the Root Cause.
- Ohno, T. (1988). Toyota Production System: Beyond Large-Scale Production. Productivity Press.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Product Analyst
Focuses on event analysis, conversion funnels, and retention, using clear metric definitions and structured analysis to explain product performance.