Given a spreadsheet with a dozen numeric columns, "what's interesting in here?" isn't a question one chart can answer. The efficient order is broad first, then specific: see everything at once, decide what's worth a closer look, then zoom into exactly that.
The workflow
1. Start with the Heatmap Generator in Correlation mode. Pick every numeric column that's plausibly related, and it computes every pairwise Pearson correlation at once, color-coded on a diverging scale — strong positive correlations in blue, strong negative in red, near-zero in white. This is the "which pairs are worth investigating" step, and it replaces manually eyeballing dozens of possible column combinations.
2. Zoom into a specific pair with the Scatter Plot Generator. A correlation matrix cell gives you a number; the scatter plot shows you the actual shape of the relationship — and that distinction matters. The Pearson correlation coefficient only captures linear relationships. Two columns can show a weak correlation in the heatmap while still having a real, strong, curved relationship the number doesn't capture — the scatter plot (with its computed trend line) is what reveals whether the relationship is genuinely linear, curved, or driven by a handful of outlier points pulling the number in one direction.
3. Check for outliers with the Box Plot Generator. A single extreme value can distort both a correlation coefficient and a trend line — a box plot flags outliers explicitly, using the standard Tukey fence (any value more than 1.5× the interquartile range beyond Q1 or Q3), rather than leaving them invisible inside an average or a correlation number. Run this on the columns involved in any correlation that looked surprising in step 2, before trusting it.
Why the order matters
Running these in the other direction — picking a scatter plot first, on a guess — means checking pairs one at a time and likely missing a real relationship you didn't think to check. Starting with the heatmap's full pairwise view first means nothing plausible gets skipped, and the more detailed tools only get used where they're actually needed.
A concrete example
A dataset with temperature, ice_cream_sales, and umbrella_sales: the heatmap's correlation matrix immediately shows temperature strongly positively correlated with ice cream sales, and strongly negatively correlated with umbrella sales — visible in one grid, no manual comparison needed. Zooming into temperature vs. ice cream sales with the scatter plot confirms a clean, strongly linear relationship with a tight trend line. A box plot on each column separately would catch it if, say, one unusually cold day with abnormally high ice cream sales were quietly distorting that correlation — a data-entry error or a genuinely unusual outlier worth investigating either way, not just averaging over.
One caution that applies at every step
None of these three tools can tell you why two variables are related — only that they appear to be, and how strongly. A strong correlation between ice cream sales and umbrella sales isn't one causing the other; both are driven by a third factor (weather) neither chart can see directly. Treat every result from this workflow as a lead to investigate, not a conclusion to report.