In The B2B Numbers Census, a website-content audit of 614 B2B companies observed in September 2026 and published in October 2026 by ClaimTheStory Research, 1 in 3 companies never published a number from their own data.
If you hold a dataset and want something worth publishing from it (or want help choosing which internal records to pull, covered in proprietary data for content), turning those rows into a published piece takes eight steps in order: ask a question your audience cares about, compute the finding, check it before drafting, pick an honest angle, structure the piece, build one clear chart, decide what you can responsibly publish, and write a methods note outside readers can inspect.
Data is not content until it answers a question
A table of raw numbers is not content on its own. Numbers become useful to an outside reader only when you add context: who was measured, what they are compared against, and why the difference matters to a named audience.
Bare number
42% of construction firms’ invoices due in 2025 were not paid in full by their due date. Is that normal for construction, or a shift?
Finding
That share rose 11 points in one year, from 31% for invoices due in 2024 to 42% for invoices due in 2025.
Ask questions your data can answer
Start with a question your audience wants answered, not with a chart you want to draw. Before looking for patterns, list what your extract contains: the unit being counted, the time span and the segments available. Then pair columns to see which questions the table can answer, and which it can’t.
In our illustrative invoicing table, pairing industry, due-date year and late share answers three questions:
- Did the share of invoices not paid in full by their due date rise in every industry between 2024 and 2025, or only in some?
- Which industries had the highest share of invoices not paid in full by their due date in 2025?
- Which industries issued the most invoices through the platform in 2025?
It can’t answer whether invoices sent to larger clients were paid late more often, or how many days past the due date those invoices were settled. Those questions need client-size fields and actual payment dates, and neither is in this table. Name the extra fields a question needs before promising an answer.
Can answer
- Did late shares rise in every industry, or only in some?
- Which industries had the highest late share in 2025?
- Which industries issued the most invoices in 2025?
Can’t answer
- Were invoices to larger clients paid late more often? It needs client size.
- How many days late were invoices settled? It needs payment dates.
The first question, whether late shares rose everywhere or only in some industries, is the strongest choice. It has a year-over-year comparison and a clear audience: finance leads in the industries concerned. In an account of its own client work, the agency Digital Third Coast described filtering US Census data to answer one narrow question for a client’s audience.
Find the finding and show the arithmetic
Once you’ve picked the question, run the arithmetic across every row and sort the results before you write a sentence. Keep every calculation in a working file; you’ll publish it later with your methods note.
Check the finding before you write a word
Before drafting anything, test your finding with four checks:
- Base size. Set a minimum group size before you look at the numbers, so small groups don’t produce wild swings. In the invoicing example, the team sets an illustrative floor of 50 accounts per industry. Education is the smallest group, at 90 accounts, so all 10 industries stay in the table.
- Fair comparison. Make sure a change over time is a real shift, not a change in which accounts are counted. Keeping only the 3,150 accounts active in both 2024 and 2025 stops new sign-ups and departed customers from skewing the comparison.
- Other explanations. Test whether a few high-volume accounts drove an industry’s rise. In Construction, check the share of individual accounts whose late rate went up, not only the invoice share. That tells you whether the shift was broad or concentrated in a few firms.
- Limits and contrary rows. State what the records can’t show, and keep rows that don’t fit the main pattern. The invoicing data covers one software platform’s customers, not all US businesses, and it measures how often invoices missed their due date, not how many days late they were paid. Software’s 1-point drop stays in the table.
If a check exposes a flaw, narrow the finding to what the data supports. Don’t bury the caveat.
Choose the angle your audience cares about
One checked table can support several headlines, but the rows limit which angles are honest. Test an angle with one rule: can you state it in a single sentence using only numbers from your table, without inferring causes the data doesn’t measure? (If you’re shaping a checked finding for journalists, see our guide to data-led PR.)
Structure the narrative: context, change, meaning
A common structure for data articles has four parts: context, change, meaning and limits. A slide deck for your own team usually ends with an action item. A published article ends with what the reader can do with the finding instead, and keeps its limits in the text, not in fine print.
- Context: who was measured, over what period, and what the baseline looked like.
- Change: where the numbers moved, and where they stayed flat.
- Meaning: what people in the affected groups should pay attention to.
- Limits: what the data doesn’t cover and what it can’t explain.
- Context 1. What we measured and for whom
- Change 2. The two industries where the share not paid in full by the due date rose. 3. The eight industries where it barely changed, including Software’s 1-point fall
- Limits 4. What this does and does not show
- Meaning 5. What finance leads in construction and logistics can watch
- For checking 6. Methods note and downloadable table
One chart, explained in words
Pick one chart that answers your main question, and let the text around it explain what it shows. To compare changes across categories, a horizontal bar chart is easier to read than a crowded line graph with many series.
In the invoicing piece, the chart shows horizontal bars of the percentage-point change from 2024 to 2025 for all 10 industries, sorted from largest rise to fall, with Construction (+11) and Logistics (+7) in a highlight color and the other eight in gray. The chart title states the finding: “More construction and logistics invoices due in 2025 were not paid in full by their due date; elsewhere little changed.” The 2025 invoice counts stay off the chart, so the bars stay clean, and go in the downloadable table below it. For finished examples of how published pieces pair text and visuals, see data storytelling examples.
Decide what you can responsibly publish
Before drafting the final copy, sort your data into three release levels with the person who owns it:
- Publish as rows: category summaries that meet your base-size rule.
- Publish only as combined totals: groups below your threshold. Keep them out entirely if a grand total would let readers subtract the visible rows and work out the hidden group.
- Keep private: customer names, account IDs, raw transaction logs and any field that could expose an individual business.
In the invoicing example, the illustrative release policy publishes industry percentages and aggregate counts only, removes all customer names, and leaves out any industry with fewer than 50 accounts (a rule chosen for this dataset, not a general standard). A minimum account threshold does not make data anonymous on its own: if one large customer issues most of the invoices in a narrow industry, an insider might still recognize it. Always assess disclosure risk with the data owner before publishing.
Publish as rows
Industry percentages and counts for all 10 industries; each has at least 50 accounts.
Combined totals only
None needed; no industry falls below the 50-account rule.
Keep private
Customer names, account IDs and the raw invoice records.
Write the methods note journalists can inspect
A clear methods note lets reporters, analysts and skeptical readers check how you counted before they use your numbers. Every methods note needs nine items (our guide to writing up survey and research findings covers methods notes for survey data in full):
- Source and owner: where the records came from and who holds them.
- Collection method: how records were extracted and filtered.
- Dates and definitions: exact date boundaries, what puts a record in a given period, and how terms like “late” or “closed” are defined.
- Exclusions: how canceled, incomplete or missing records were handled.
- Sample and base sizes: the total accounts or records behind every row.
- Segment definitions: whether industry or company size was self-reported or verified.
- Limits: who the data covers and what it can’t show.
- How to cite: the exact source credit you want writers to use.
- Data access: where readers can download the summary table or ask questions about it.
- Source and owner Invoices issued through [Company]’s platform by the 3,150 business accounts active in both 2024 and 2025.
- Collection Extracted in January 2026; disclosure risk assessed with the data owner before release.
- Dates and definitions Each invoice counts in the year of its due date (1 January 2024 to 31 December 2025). “Late” means not paid in full by the due date; invoices unpaid at extraction count as late.
- Exclusions Canceled and credited invoices are left out.
- Base sizes 3,150 accounts; a minimum of 50 accounts per industry, and all ten meet it.
- Segments Industry is self-reported by the issuing account.
- Limits Our customers only. Shares of invoices, not accounts; how often invoices were late, not how many days. Does not show why.
- How to cite [Company], Invoice Payment Data 2025.
- Data access The full arithmetic table, available for download.
If you have data and don’t know what to do with it, we find a question it can answer that your customers would care about. Book a demo