Data come in kinds, and the kind rules the analysis
Before any chart or test, name the kind of data in each column. The classification decides which summaries exist: a mean of categories is nonsense, a median of unordered labels is nonsense, and software will happily compute both if you let it. Two families, four kinds in total.
Categorical (qualitative) data
Nominal: labels with no order: blood group, city, favourite colour, pass/fail. Summaries that exist: counts, proportions, the mode. Ordinal: labels with order but no fixed spacing: satisfaction ratings (poor, fair, good, excellent), education level, race finishing positions. Medians and percentiles join the toolkit; means remain dubious, because the gap between "fair" and "good" is not a known number.
Numeric (quantitative) data
Discrete: counts: children per family, defects per batch, goals per match. Values are whole steps with gaps. Continuous: measurements: height, time, temperature, weight, limited only by instrument precision. Both support means and standard deviations; continuous data add the full modelling toolkit (regression, normal models), while low-count discrete data often need count-specific models (binomial, Poisson).
One survey, four columns. "City" (nominal): report the modal city and the share per city. "Satisfaction 1 to 5" (ordinal): report the median (say 4) and the share at each level; a mean of 3.7 is common practice and contested, since the steps are not known to be equal. "Visits last month" (discrete): mean 2.3 visits is fine, fractions of a visit exist only in averages. "Session length in minutes" (continuous): the full summary, mean 28.15 with SD 4.04 style, applies.
Numbers that are secretly labels: jersey numbers, postcodes, product codes, Likert items stored as digits. They sit in numeric columns and average beautifully into garbage (the mean postcode of a customer base describes nowhere). If the digits name rather than measure, treat the column as nominal whatever its storage type.
Why the kind matters downstream
Every later choice keys off this page: bar and pie charts for categories, histograms for numbers; chi-square tests for categorical associations, t tests for numeric means; correlation for two numeric variables, and rank methods when one side is ordinal. Naming the kinds at the start is the cheapest correctness there is.
Frequently asked questions
Is a 1-to-5 rating scale nominal or ordinal?
Ordinal: the labels are ordered, but the spacing between a 3 and a 4 is not guaranteed to equal the spacing between a 4 and a 5. That is why medians and distributions of ratings are safer than bare means.
Are percentages categorical or numeric?
Numeric (continuous, bounded): a percentage is a measurement derived from counts, and means of percentages can be meaningful, though averaging percentages from different-sized groups needs weighting by the group sizes.
Can data change kind?
Yes, by processing: continuous ages become ordinal age bands when grouped; counts become nominal when only presence/absence is kept. Grouping always discards information, so keep the raw form when you can.