From data to insight: an analysis workflow

0:00 From data to insight

Sessions rose twelve percent. Is that an insight? No. It's an observation. Here's an insight: sessions rose twelve percent because a creator's video drove new mobile users, but they converted at half the site average because the landing page is slow on mobile, so fixing speed could unlock meaningful revenue. See the difference? It explains, it quantifies, and it points to an action. In this lecture you'll learn a question-first workflow, the decomposition method for diagnosing changes, how to separate signal from noise, how to use AI assistants safely for hypotheses, and how to present insights people act on.

0:43 Why it matters

Why is this the skill that matters most? Because stakeholders don't need more numbers. They need to know what's happening, why, and what to do. Analysts who deliver insights get invited into decisions. Analysts who deliver screenshots get asked for more screenshots. And with AI tools now summarizing reports in seconds, the value of a human analyst is exactly this: framing the right question, testing explanations against evidence, and making a recommendation someone can own.

1:16 Question-first workflow

Here's the question-first workflow. Question: what decision will this inform, who makes it, by when? Context: what changed, like campaigns, releases, seasonality, tracking? Hypotheses: list two to four plausible explanations before opening G A four. Evidence: which report or segment would confirm or refute each? Analysis: run the checks and note data quality caveats. So what: quantify the impact and recommend an action. Follow-up: how will we know if it worked? Writing hypotheses first protects you from confirmation bias, hunting until something supports what you already believed.

1:54 Decomposition

Now the decomposition method, my favorite diagnostic tool. When a KPI moves, break it into factors. Revenue equals sessions, times conversion rate, times average order value. Leads equal sessions times lead rate. Ask which factor moved, and within that factor, which segment explains most of the change: channel, device, country, landing page, new or returning. Think of it like a doctor checking vital signs one at a time instead of saying the patient seems unwell. A fifteen percent revenue drop could be one channel's traffic, or a mobile conversion problem. Very different fixes.

2:34 Signal versus noise

Separating signal from noise. Small numbers swing randomly. Before reacting, compare against a longer baseline, like the same weekday average over several weeks. Check volume: a conversion rate from thirty sessions is unreliable. Check tracking changes and annotations: did measurement change rather than behavior? And consider seasonality and events: paydays, public holidays, school terms, religious festivals, sales events. Then, where a decision is significant and two things moved together, remember correlation isn't causation. Propose an experiment to confirm.

3:08 Example 1: the revenue dip

First example, with illustrative numbers. An online homeware store compares two equivalent weeks. Sessions went from fifty thousand to forty-nine thousand, minus two percent. Conversion rate from two percent to one point six, minus twenty percent. Average order value sixty to sixty-one, up two percent. Revenue down about twenty percent. So conversion rate explains the drop. By device, desktop is unchanged but mobile fell sharply, starting the day of a site release with a new address lookup. Funnels confirm drop-off at add shipping info on mobile. Recommendation: roll back or fix the component.

3:48 Example 2: Riyadh travel leads

Second example, a business case. A Riyadh travel agency's leads fall thirty percent in a week, illustratively. Decomposition shows sessions stable and the lead rate down only on Arabic-language pages. A path exploration shows Arabic users reaching the form and leaving. The cause: the form's Arabic validation message broke after a translation update, blocking submission. Fixing it restored the lead rate. The written insight took five lines, and the recommendation had an owner and a deadline.

4:21 Watch me do it, part 1

Watch me decompose in pandas. I export two comparable weeks with sessions, orders and revenue by device. The lesson's script groups by week, computes conversion rate as orders over sessions, and average order value as revenue over orders, and prints them. Instantly I see conversion rate is the factor that moved. Then it pivots by device and prints the change in conversion rate per device, sorted. Mobile sits at the top of the list. Ten lines of code, and I've done in a minute what used to take twenty clicks.

5:00 Watch me do it, part 2

Then I use AI for hypotheses, carefully. I give an assistant the context, the aggregated table with no personal data, and the known events: a release on the fourteenth, no campaign changes. I ask for three or four hypotheses ranked by fit, the G A four report or exploration that would confirm or refute each, and a what, so what, now what draft using only the numbers provided, with estimates marked illustrative. Inside G A four, Ask Advisor can do similar work from your property. Either way, I treat the answer as a list of hypotheses, and I verify each one myself.

5:44 Presenting

Now presenting. Use what, so what, now what. What: mobile conversion rate fell from two point one percent to one point three percent since the fourteenth. So what: illustratively, about one fifth of weekly revenue is at risk, and desktop is unaffected. Now what: roll back the address lookup, the developers, today; re-test the funnel, analytics, tomorrow. Lead with the conclusion. Put charts after the story, and only the charts that support it.

6:16 Beyond GA4 and timeboxing

Some questions can't be answered in G A four alone. Profit needs cost of goods. Lifetime value across offline channels needs your C R M. Lead quality needs sales outcomes. Good analysts know when to pull in other sources, like C R M exports, ad cost data and BigQuery joins, and they state clearly which numbers came from where. And they timebox. Agree the time with the decision-maker up front. A good-enough answer before the decision beats a perfect one after it.

6:52 Common mistakes

Common mistakes. Opening G A four without a question and looking for insights. Reporting every metric that changed instead of the ones that matter. Explaining random noise as meaningful trends. Recommendations without owners or deadlines. Forgetting to check whether tracking changed. Presenting AI-generated explanations as findings. And false precision, like revenue impact to the last unit when the inputs are estimates.

7:19 Recap

Recap. An insight explains, quantifies and points to action. Write hypotheses before opening reports. Decompose K P Is to find the factor and segment that moved. Separate signal from noise. Use AI and Ask Advisor to generate hypotheses, then verify everything. Present as what, so what, now what, conclusion first, with owners and dates.

7:42 Try this now

Try this now. Take a real K P I change from a site you know. Decompose it into factors and find the segment that moved most. Write three hypotheses before you look any further. Then write a what, so what, now what summary with an owner and a date for the action.