The Data Analysis Process in 6 Phases | Free Data Analytics Course

What you'll learn

The six phases of a data analysis: ask, prepare, process, analyze, share and act.
How to turn a vague request into a specific, measurable business question.
How to judge whether a data source is trustworthy before you use it.
How the whole process plays out on a real question: "Why was December so big?"
The process mistakes that waste the most time, and how to avoid them.

📥 Practice this lesson

Follow the December example with the practice file. The Summary sheet has revenue by month and by channel, so you can check every number in this lesson yourself.

sales-practice-data.xlsx · sheets Sales Clean and Summary

Good analysis isn't a burst of clever formulas; it's a repeatable process. Analysts who skip steps end up answering the wrong question, trusting broken data or producing a beautiful chart nobody acts on. In this lesson you'll learn a six-phase framework that's widely used in analytics training (ask, prepare, process, analyze, share, act) and you'll run it end to end on a question from the course dataset, with real numbers at every step.

Ask
Prepare
Process
Analyze
Share
Act

The phases are presented in order, but real projects loop back. You might discover during analysis that you need another data source (back to Prepare) or that the question was wrong (back to Ask). That's normal. The framework keeps you from forgetting a phase, not from ever going backward.

The scenario: a VP gets excited about December

Imagine you're the analyst at the retailer in our dataset. In early January the VP of Sales sends you a message:

"December was huge! I want to double the holiday advertising budget next year. Can you pull something together to support that?"

It's a typical request: enthusiastic, vague and already leaning toward an answer. Let's walk it through the six phases.

Phase 1: Ask, and define the real question

The Ask phase is about understanding the decision behind the request. Here the decision is "should we spend more on holiday ads?" The VP's assumption is that December's spike came from holiday shoppers, the people ads can reach. That's the assumption to test.

A useful check is to make the question SMART, an analytics version of the classic goal-setting acronym: specific, measurable, action-oriented, relevant and time-bound. Compare:

Vague request	SMART analysis question
"Pull something together on December."	"How much of December 2025 revenue came from consumer channels (online and retail) that holiday advertising can influence, and how did that compare with November and other months of 2025?"

The SMART version names the time period, the metric (revenue), the comparison (November and other months) and the decision it supports (ad budget). Before you touch any data, confirm it with the stakeholder: "Is this what you need to decide on the budget?" Five minutes here can save five days later.

Also identify your stakeholders: the VP (decision-maker), the marketing manager (who owns the ad budget) and maybe the wholesale sales lead. Each one cares about a different slice of the answer.

1. Which of these is the best-defined analysis question?

"How are sales doing?"
"Are customers happy with our products?"
"Which product category had the largest revenue drop between Q3 and Q4 2025 in the Northeast?"

The third question is specific (one region, category level), measurable (revenue), time-bound (Q3 vs. Q4 2025) and clearly tied to an action, such as investigating that category. The others are too broad to answer with a single analysis.

Phase 2: Prepare, and find trustworthy data

Now you decide what data you need and where it lives. For this question you need every 2025 order with its date, channel and revenue. In our case that's the sales export from the order system; in a real company it might be a table in a SQL database or a data warehouse.

Before you trust a source, run a quick credibility check. A popular checklist in analytics training is ROCCC:

Reliable: is it accurate and complete, or known to have gaps?
Original: is it the system of record, or a copy of a copy someone edited by hand?
Comprehensive: does it include everything you need (all channels, all 12 months)?
Current: is it up to date for the period you're analyzing?
Cited: do you know who produced it and how?

Prepare is also where you think about privacy. Sales data often includes customer names, emails or addresses. If you don't need them for the question, don't pull them. Our dataset has no personal information at all, which is the safest setup.

Phase 3: Process, and clean before you count

Raw data is almost never ready to analyze. The raw sheet in our practice file has 203 rows, but only 200 real orders: three rows are duplicates. It also has four blank revenue cells, seven state names typed inconsistently (like "CA", "california" and "Texas " with a trailing space) and one date in a different format.

Does that matter? Add up the raw revenue column as is, and you get $90,231. The correct total after cleaning is $89,735. The duplicates inflate revenue by $853, and the blanks hide $357. Small here, but in a real company those errors can be thousands of rows, and they're exactly the kind of mistake that destroys trust in an analyst. You'll learn the full cleaning workflow in Lesson 5.

Keep the raw data untouched

Never clean the only copy. Duplicate the raw sheet (or file), clean the copy and write down every change you make in a short cleaning log. If someone questions a number, you can show exactly how you got it.

2. The raw revenue column adds up to $90,231, but the clean total is $89,735. If 3 duplicate rows add $853, how much revenue was hidden in the 4 blank cells?

$496
$357
$853

Raw total = clean total + duplicates − blanks, so blanks = $89,735 + $853 − $90,231 = $357. The two errors partly cancel each other out, which is why a total that "looks about right" isn't proof the data is clean.

Phase 4: Analyze, and let the data push back

With clean data, start with the headline number. December revenue was $18,319, about 2.8 times November's $6,588. The VP is right that December was huge. The next step is to break it down by channel:

Wholesale (81.7%) $14,961Only 6 orders, from businesses buying in bulk.
Online (10.0%) $1,83212 orders from individual shoppers.
Retail (8.3%) $1,5268 in-store orders.

Six wholesale orders produced more than four-fifths of December's revenue. The three largest alone (an Arizona office-chair order of $5,229, a California dumbbell order of $3,887 and a New Jersey smart-watch order of $2,587) add up to $11,703.

This is where the data pushes back on the assumption. Holiday advertising targets consumers, so look at the consumer channels only (online plus retail):

Month (2025)	Online + retail revenue
July	$3,298
September	$3,465
November	$2,065
December	$3,358

Consumer revenue did rise from November to December, by about 63% ($3,358 ÷ $2,065 ≈ 1.63). But December's consumer revenue was roughly the same as July's and slightly below September's. The giant spike came from wholesale buyers, who don't respond to holiday ads the way shoppers do.

Good analysis also asks what could make this conclusion wrong. With only one year of data, you can't tell whether these wholesale orders repeat every December or were a one-time event. That limitation belongs in your final message.

Sharing means communicating the result so the decision-maker understands it in under a minute. Lead with the conclusion, back it with one chart and state the limitation:

Example summary for the VP

December revenue ($18,319) was 2.8× November, but 82% of it came from six wholesale orders. Online and retail sales in December ($3,358) were similar to July and September. Doubling holiday ads would target the smallest part of the spike. Recommendation: test a modest ad increase first and talk to the wholesale team about whether these bulk orders will repeat. Limitation: we only have one year of data.

That's five sentences. The details, tables and methods go in an appendix for anyone who wants them. You'll practice this skill in Lesson 12.

Phase 6: Act, and close the loop

The analysis is only finished when someone decides something. In this case the VP might approve a smaller ad increase run as an A/B test (you'll learn how to read one in Lesson 13) and ask sales to forecast wholesale orders. Your job in the Act phase is to agree on how success will be measured, then come back with the results. That follow-up builds your credibility faster than any dashboard.

Here's the whole process on one page:

Phase	Key question	Typical output
Ask	What decision are we supporting?	A written SMART question confirmed with the stakeholder
Prepare	What data do we need, and can we trust it?	List of sources, access, privacy check
Process	Is the data clean and complete?	Clean dataset and cleaning log
Analyze	What does the data say, and what could prove us wrong?	Tables, calculations, findings
Share	What should the audience remember?	Summary, chart, dashboard or slides
Act	What will we do, and how will we measure it?	Decision, next steps, success metric

3. During analysis you discover the dataset has no customer IDs, so you can't measure repeat purchases, which the stakeholder cares about. What should you do?

Go back to the Prepare phase: look for another source or tell the stakeholder what can and can't be answered
Estimate repeat purchases by guessing from order dates
Skip that part and don't mention it

The process loops. When the data can't answer part of the question, you either find better data or clearly state the limitation. Guessing or quietly dropping the topic undermines trust.

Common process mistakes

Starting with the data instead of the question.

Write the decision and the question first; then pick only the data you need.

Trying to "prove" what the stakeholder already believes.

Treat the stakeholder's idea as a hypothesis to test. Report what the data says, even when it disagrees.

Analyzing raw data without checking it.

Count rows, check duplicates and blanks, and reconcile totals before any chart.

Sending every chart you made.

Share the one or two that answer the question. Put the rest in an appendix.

Disappearing after the presentation.

Agree on a success metric and follow up with the results.

4. In which phase would you check whether a spreadsheet you received is the original system export or a copy someone edited by hand?

Prepare
Analyze
Share

Judging the credibility of a data source (is it original, reliable, current?) belongs in the Prepare phase, before you invest time cleaning and analyzing it.

Practice exercise

The regional director for the Midwest writes: "Our region is always last. Fix it." Using the practice file (sheets Sales Clean and Summary):

Rewrite the request as a SMART analysis question.
Find the Midwest's revenue, number of orders, average order and share of total revenue.
Find how much of the Midwest's revenue came from wholesale.
Write a two-sentence finding for the director.

See solution

1. SMART question: "Why did the Midwest generate the lowest revenue of the four regions in 2025: fewer orders, smaller orders or a different channel mix? And which of those can we act on in 2026?"

2. Key numbers: the Midwest had $13,639 in revenue (15.2% of $89,735), 32 orders and an average order of $426.22 (13,639 ÷ 32). Its average order is close to the company average of $448.68.

3. Wholesale: 6 wholesale orders brought in $9,467, which is 69.4% of the Midwest's revenue (9,467 ÷ 13,639). That's in line with the company-wide pattern, where wholesale is 68.5% of revenue.

4. Finding: "The Midwest's orders are about as large as everyone else's ($426 vs. $449 on average); the gap is volume, with 32 orders vs. 45 to 64 in other regions. Growing the number of orders, especially wholesale accounts, matters more than raising order size."

Summary

Analysis follows six phases: ask, prepare, process, analyze, share and act, and it's fine to loop back.
Spend real time on Ask: a SMART question tied to a decision prevents wasted work.
Check sources for credibility (ROCCC) and privacy before using them.
Clean before you count: errors in our raw file shift revenue from $89,735 to $90,231.
Let the data challenge assumptions: December's spike was 82% wholesale, not holiday shoppers.
Lead with the answer when sharing, and follow up after the decision.

Next lesson: Types of Data and How to Structure a Table. You'll learn to classify every column in the dataset and build tables that are easy to analyze.