Start with the question the dataset can answer
A spreadsheet can look self-explanatory: labels across the top, records underneath and perhaps a few numbers ready to chart. But a familiar layout does not tell you what the data means. Before calculating anything, write down the question you hope to answer. Then ask whether the dataset contains the people, places, time periods and measurements needed for that question. A file about purchases at one store, for example, cannot by itself describe every shopper in a city.
Make a copy of the original file before exploring it. Note its name and any accompanying description, then scan its size, headings and first few records. This first pass is not about finding a striking result. It is about learning what is present, what is absent and which details you need to clarify before interpreting the numbers.
Find out what one row represents
The meaning of a row is sometimes called the dataset’s grain. One row might represent a person, a purchase, a product, a store on a particular day or a survey response. Those are different units of observation, even if they share columns such as date and price. Look for an identifier and check whether it repeats. Repetition may be expected: a customer can make several purchases, and a store can appear once for every reporting day.
Getting the grain wrong can distort a simple count. If each row is a purchase, counting rows gives the number of recorded purchases, not the number of customers. If a table combines product details with purchase records, the same product may appear many times. State the grain in one sentence before making a summary. If you cannot do that confidently, inspect the documentation or ask how the file was assembled.
Read columns as definitions, not just labels
A column heading is a clue, not a complete definition. A field named “amount” might mean an item price, an order total, a refund or a value before discounts. Check a few entries for each column and note its apparent type: text, number, date, category or identifier. An identifier made of digits is still a label; averaging customer IDs would not produce a meaningful result.
Pay attention to missing values and codes. A blank cell might mean the information was not collected, did not apply or was removed. A zero could be a genuine measurement or a stand-in for missing information. Categories may also use abbreviations whose meanings are not obvious. Avoid replacing blanks or merging categories until you know what they signify. Small assumptions made at this stage can quietly shape every later calculation.
Check units, scale and time
Numbers need units to be interpretable. A distance could be recorded in miles or kilometers; a price may be per item, per order or per month. Percentages and proportions also differ in presentation: 0.25 and 25% can express the same share, while 25 without a label is ambiguous. Look for units in column names, notes or a separate data dictionary. Confirm whether totals and component values use compatible units before comparing or adding them.
Dates deserve similar care. A date might mark when an event happened, when it was recorded or the end of a reporting period. If rows summarize weeks or months, treating them as individual days will mislead. Check the time span and whether any periods are absent. When comparing groups, also look at the denominator: a count of purchases and a purchase rate per customer answer different questions.
Use metadata to set the limits of your conclusions
Metadata is information about the dataset. It may describe where the records came from, how fields were defined, which records were included, how missing entries were handled and whether the file was revised. It might appear in a data dictionary, a readme file, a separate worksheet or notes beside a download. Read it before treating an unusual pattern as a discovery: a change in collection or a filtered subset could help explain what you see.
Finish your first pass with a short record of what you know and what remains uncertain. Include the row grain, important column definitions, units, time coverage, missing-value conventions and any limits on who or what the data represents. Then try a few basic checks, such as counts by category and ranges for key numbers. These checks will not prove the dataset is complete or correct, but they can expose misunderstandings early and help you make claims that fit the evidence.
