How to Brief an Original Research Report: Write the Headline Before the Survey
By Jascha Kaykas-Wolff, former CMO of Mozilla and BitTorrent ·
Brief an original research report by writing down the finding you hope to publish, as the sentence a reporter would quote, before anyone drafts a survey question. Then set the sample size that sentence needs, and decide in writing what you will publish if the number comes back dull. For most small teams a survey of their own customers or an analysis of their own product data is the right format, because nobody else holds that data and nobody can copy it next week.
A payroll software company with about 3,000 small-business customers wants a report that will earn links in January. Someone drafts 34 questions in Typeform over a weekend, the customer list gets two emails, and 287 people answer. In week three the writer finds the report's best number: 61% of restaurant owners said they had paid a late-payroll penalty in the past year. It is a great headline. It is also 23 people out of 38 restaurant owners in the sample, and on a group that small the margin of error runs about 16 points in either direction. The report ships with the number in its title. A trade reporter asks for the sample size, gets it, and doesn't write the story.
Thirty-four questions, and nobody had decided which one mattered.
Most guides to original research start at the questionnaire. Our view is that the questionnaire is the last thing to write. A research brief starts from the one or two findings worth publishing, works backward to the sample those findings need, and only then lets someone open the survey tool. It also says what happens if the answer is boring, before the data arrives, because the urge to go digging through crosstabs for a better number is strong and the number you dig up that way is usually noise.
What a research brief decides that a normal brief does not
A normal content brief assumes the facts already exist and the writer's job is to find and arrange them. A research brief commissions the facts. That moves two decisions from the writer to the editor: what the data has to be able to prove, and how much of it you need before you are allowed to say so. Get those wrong and no amount of good writing will rescue the report.
Original research formats compared
Pick the format from the data you can actually reach. Teams usually pick the paid panel because big-brand reports use one, then discover the per-response price once they screen for a job title.
| Format | Best for | Sample to brief for | Time and cost | Key trade-off |
|---|---|---|---|---|
| Customer survey | Companies with a few thousand customers they can email | 300 or more completed responses | A survey tool and two emails. Three to six weeks. | Your customers are not the market. The methodology note has to say so, every time. |
| Paid panel survey | Claims about a whole profession or a whole country | About 1,000 completed responses | Priced per completed response by the panel vendor. Screening for a narrow job title raises it fast. | Anyone with a budget can buy the same panel, and low-effort answers have to be filtered out. |
| Product data analysis | Software companies with usage data across many accounts | Hundreds of accounts, reported only in aggregate | An analyst plus a privacy and terms review. Two to four weeks. | The hardest data for a competitor to copy, and the slowest to get past legal. |
| Public dataset analysis | Small teams with no list and no budget | Whatever the agency published, usually a BLS or Census Bureau table | One person with a spreadsheet. One to two weeks. | Cheap and easy to check. Anyone can run the same analysis next week. |
| Expert interview roundup | New topics where nobody has numbers yet | 10 to 20 interviews | Mostly scheduling. Two to four weeks. | No percentages to quote, so fewer links. The quotes age better than survey numbers do. |
Where each format fits (and where it does not)
Customer survey. The default for a company with a list. Leave it alone if you have fewer than about 2,000 reachable customers, because response rates on customer surveys are low and you will struggle to reach 300. Never title the report as if it describes an industry. “Small restaurants on our platform” is honest, and reporters will quote it.
Paid panel survey. Use it when the claim is about everyone (“US accountants”) and you can afford roughly 1,000 completes after screening. Skip it for niche roles. A panel that has to find 1,000 hospital procurement managers will either cost a great deal or quietly fill the quota with people who are not.
Product data analysis. Our favorite, and the one we would brief first if the data exists. A finding shaped like “Across 1,400 accounts, payroll runs submitted after 4pm on a Friday were three times as likely to be late” is one no competitor can write. Budget for the privacy review. Your terms of service decide what you can publish, and legal will want to read every number, in context, before launch.
Public dataset analysis. The right choice for a team with no list. The Bureau of Labor Statistics publishes more than most marketers ever open. The weakness is that the source is public, so the value has to come from the cut you make and the explanation you write.
Expert interview roundup. Useful when the topic is too new for numbers. It earns fewer links than a statistic and holds up longer. Brief each interview the way our guide to interviewing a subject matter expert describes, with the same three or four questions put to everyone so the answers can be compared.
And one honest concession. If you have no customer list and nobody who can read a crosstab, a research report is probably the wrong project this quarter. Refreshing the three pages that already rank on page two will usually do more, and our guide to briefing a content refresh covers that work.
How big the sample has to be
The table below shows the classical margin of error at 95% confidence for a simple random sample, at a 50% result, which is where the margin is widest. Customer surveys and panels are opt-in, so they are not random, and the real error is larger than these figures. The American Association for Public Opinion Research advises against reporting a classical margin of error for opt-in samples at all. We use the table only to plan, as a floor.
| Completed responses | Margin of error (95%, random sample) | What it can support |
|---|---|---|
| 100 | ±9.8 points | Direction only. Fine for one topline number, too thin for any subgroup. |
| 300 | ±5.7 points | A topline headline, plus one or two large subgroups. |
| 400 | ±4.9 points | Toplines and a few subgroup splits of 100 or more each. |
| 1,000 | ±3.1 points | The size national polls tend to use. Most subgroup comparisons. |
| 2,400 | ±2.0 points | Small differences between groups, such as 41% vs 46%. |
The number that catches teams out is the subgroup. A 300-person survey has a decent topline, but split it by industry and every slice is back near the top row of that table.
Fields a research brief needs
- The headline finding, as a sentence with a blank in it. “___% of small restaurant owners paid a late-payroll penalty in 2025.” Write two at most. If you cannot write one, you are not ready to field a survey.
- The population, and who is excluded. “Owners or managers of US businesses with 5 to 50 employees, currently customers.” Former customers, partners and your own staff are out, and the brief says so.
- Every subgroup you plan to report, with its minimum size. Our rule: no percentage for any group under 100 responses. Below that, report the count (“23 of 38”) and let the reader judge it.
- A question budget. We cap ours at 12. Every question has to map to a finding in the brief or it gets cut, and the cap does most of the editing for you.
- The dull-result plan. Decide now whether you publish the topline whatever it says, reframe around a second finding you also named in advance, or shelve the report. Write it down, with a name next to it.
- The methodology note. Dates fielded, completed responses, how people were recruited, who was screened out, and the exact wording of every question behind a quoted number. Put it on the report page, visible, with no form in front of it.
- Data owner and sign-off. One person who holds the raw file and answers reporters' questions about it, and the name of whoever in legal approves product data.
- The success measure. Referring domains to the report page 90 days after launch, read in Ahrefs or a similar backlink tool. Page views will flatter a report that nobody cited.
Check what is already published before you field anything
Search the phrase a reporter would use to find your number, such as “late payroll penalty statistics,” and read what comes back. The ContentBrief.io brief generator does this in about 30 seconds and for free. It reads the top 10 results and returns their headings and word counts, the People Also Ask questions and the related searches. If the results already cite a fresh government figure, your 287-person survey will lose to it, and you should either ask a different question or analyze the government data yourself. If the results are listicles quoting a 2019 study nobody can find, there is room for a new number. Our guide to SERP analysis for a brief covers what to look for.
That is where our tool stops. It cannot design a questionnaire, clean survey data or tell you whether a difference between two groups is real. You need a survey tool and a spreadsheet for that, and ideally someone who has taken a statistics course since college.
Writing the report from the data
Give the writer a fact sheet, never the raw export: each number they may quote, its exact question wording, its base (“of 287 respondents”) and the subgroup it came from. A writer handed a 34-column spreadsheet will find a striking number in a 12-person cell and build a section around it. A language model will do the same thing faster. The approach in our guide to a content brief for AI writing applies here without changes.
Before launch, the editor checks every number in the draft against the fact sheet, one by one, as part of the pre-publish QA. Rounding drifts. A 48.6% becomes “nearly half” in one section and “more than half” in the social post.
Next year's edition
If the report earns links, someone will ask for it again next year, so archive the questionnaire exactly as it ran, awkward wording included, because changing a question breaks the comparison with last year and the year-over-year change is the part reporters come back for.