Marketing
Data Analysis: Turn a Business Question Into an Answer
Learn data analysis with an original inquiry dataset, a transparent cleaning log and a decision brief that separates evidence from assumptions.

What you will learn
- Frame a business question and define the unit of analysis.
- Clean a small dataset without silently replacing unknowns.
- Calculate a defensible comparison and explain its limits.
Data analysis turns information into an answer to a defined question. Start with the decision, identify what each record represents, document the cleaning steps and calculate a comparison another person can reproduce. The useful output is a traceable finding with a clear next question, not simply a more polished chart.
Start with the question and unit
Snowflake’s data analysis guide describes preparing and interpreting data in the context of a question. A practical analysis should make its unit clear: people, organizations, events, invoices or opportunities. Counting two events from one person is not the same as counting two people.
Compare “Why is marketing not working?” with “For eligible inquiries first received in this period, how many reached an accepted conversation within 14 days, by recorded source?” The second question defines a population, event, time window and comparison. It also exposes what data you need before answering.
This lesson uses a fictional service business. Its small dataset teaches cleaning and denominator choices, not statistical significance or a recommended marketing channel. No records refer to actual Finta users.
Inspect an original small dataset
Fictional inquiry extract before cleaning
| Row | Inquiry ID | Source | Received | Accepted within 14 days |
|---|---|---|---|---|
| 1 | A01 | Search | 2026-10-01 | Yes |
| 2 | A02 | Event | 2026-10-02 | Yes |
| 3 | A02 | Event | 2026-10-02 | Yes |
| 4 | A03 | Search | 2026-10-03 | No |
| 5 | A04 | Unknown | 2026-10-04 | No |
| 6 | A05 | Event | 10/05/2026 | No |
| 7 | A06 | Search | 2026-10-06 | Yes |
| 8 | A07 | Event | 2026-10-07 | Yes |
| 9 | A08 | Search | 2026-10-08 | No |
| 10 | A09 | Event | 2026-10-09 | No |
| 11 | A10 | Search | 2026-10-10 | Yes |
| 12 | TEST01 | Search | 2026-10-10 | Yes |
The extract has 12 rows, but not 12 eligible inquiries. A02 appears twice. TEST01 is a documented internal test. One source is unknown, and one date uses a different display format. The outcome window is complete for every eligible inquiry in this teaching example. A real analysis must check that condition, rather than labeling an immature record “No.”
Keep a cleaning log
Original cleaning decisions
| Issue | Evidence and treatment | Effect |
|---|---|---|
| Duplicate A02 | The source owner confirms rows 2 and 3 are an identical export duplication; retain one inquiry | Remove one duplicate row, not both |
| Internal test | TEST01 is explicitly documented as an internal test under the eligibility rule | Exclude one test row |
| Mixed date format | The source specification confirms US month/day/year, so 10/05/2026 means October 5 | Normalize display to 2026-10-05; preserve the raw value |
| Unknown source | No verified acquisition source exists for A04 | Keep Unknown as a separate group; do not guess Search |
| Outcome timing | All eligible inquiries have completed the 14-day observation window | A recorded No is eligible for this comparison |
Preserve the original extract and record the transformation version. If the duplicate had conflicting outcomes, resolve the conflict or hold the record; do not choose whichever result improves the chart. If the date convention were unknown, it would remain unresolved rather than being normalized by intuition.
The stocks-versus-flows lesson helps distinguish units and movements. A count of inquiries received during a period is different from all open inquiries at the end of it.
Calculate the comparison with explicit denominators
After the two exclusions, there are 10 eligible unique inquiries. Search has five inquiries and three accepted conversations. Event has four inquiries and two accepted conversations. Unknown has one inquiry and no accepted conversation.
Cleaned fictional inquiry comparison
| Recorded source | Eligible inquiries | Accepted conversations | Rate |
|---|---|---|---|
| Search | 5 | 3 | 60% |
| Event | 4 | 2 | 50% |
| Unknown | 1 | 0 | 0% |
| Total | 10 | 5 | 50% |
Search’s observed rate is 3 ÷ 5 = 60%. Event’s is 2 ÷ 4 = 50%. Overall it is 5 ÷ 10 = 50%. Do not average 60%, 50% and 0% as if the groups were the same size. The overall rate uses the total numerator and total denominator.
The difference is descriptive. It does not prove that switching budget from events to search will improve results. The inquiries may differ in suitability, timing, cost or how source was recorded. With such small counts, one additional outcome can change the apparent comparison substantially.
How do you write data analysis?
A defensible brief says: “Among 10 eligible unique inquiries with a complete 14-day window, five reached an accepted conversation. Search contributed three of five; events two of four; one source was unknown. This small observational comparison does not establish a channel effect. We should review fit, source capture and costs before changing allocation.”
Use the small-sample lesson when the evidence is sparse. If your actual question concerns content acquisition, the content measurement lesson keeps clicks, visitors, signups, activation and paid outcomes distinct.
Before sharing the brief, ask a teammate to reproduce its totals from the preserved extract and cleaning log. They should be able to explain which rows were excluded, why the denominators differ and where the unknown source remains. A result that cannot survive that handoff needs clearer working before it needs a better presentation.
Choose a chart only after the unit and decision are clear. A simple table may be enough. Label the period, group sizes and rate definition. Do not hide Unknown because it makes the presentation untidy. List the next data collection step and the person who can resolve it.
Copyable AI prompt
Try this with Aurora AI in Finta, Codex, or Claude. Share only material you are authorized to use, remove unnecessary personal information and review the draft.
Using [authorized dataset], [data dictionary], [eligibility rule], [observation window] and [decision question], propose and document a reproducible cleaning plan. Preserve raw values, identify duplicates and conflicts, and keep unknowns distinct from zero. Calculate group counts and rates with explicit denominators. Return a decision brief separating descriptive findings, assumptions, missing evidence and causal claims that cannot be supported. Do not invent records, change source systems or make external decisions.Keep the next investigation bounded
Choose the unresolved question that could change the decision and assign it to someone who can obtain the evidence. Explore Finta’s Metric Change Investigator for preparing a review from authorized material and consistent definitions. Keep the raw extract, transformations and calculation method available so the next discussion can build on the finding.
Sources and method. By Finta Editorial Team. Sources checked November 15, 2026. Worked examples and worksheets are original fictional teaching materials, not customer results.