Every report opens with what it cannot tell you

Not an appendix, not a footnote, not a file sent afterwards on request — the first thing on the results page, before any chart. Six things are stated every time, because a number without them can be read to mean almost anything.

• Where the respondents came from.

  A closed group — your own customers, employees or patients, uploaded by you — or an open group answering a public link. The two support completely different claims, so the report names which one it is before it shows a single number.

• When fieldwork actually ran.

  Start and end dates, because a study run across a festival week, a fee deadline or an outage is not comparable with one that was not, and the reader is entitled to spot that themselves.

• Contacted, valid, rejected — and why the rejected were rejected.

  Three counts and a breakdown of the reasons. A study that contacted four thousand people to get three hundred usable answers is a different study from one that contacted four hundred, and only the first number makes that visible.

• Language mix, channel mix and completion rate.

  Which languages people actually answered in, how they answered, and how many who started finished. Language mix in particular is a finding in itself: a study answered almost entirely in English by a workforce that is not is telling you who did not participate.

• Known bias, written by the researcher in plain words.

  A free-text field for what this study cannot tell you — the list you uploaded skewed to recent customers, the branch that did not send its data, the shift that has no smartphones. It is the most valuable box on the page and the easiest to leave empty.

• How much of it was inferred rather than asked.

  A conversation can establish an answer without ever putting the question. That is useful and it is not the same as asking, so the share of answers derived rather than asked directly is reported as its own number.

Almost every study run here is a closed group: a client’s own customers, employees or patients. That is exactly the right sample for “what do our people think”, and the wrong one for “what does the country think”. The reporting is built so that distinction survives the trip from the dashboard into somebody’s deck.

• The standing line, on every report.

  Findings are directional only unless quota controls were applied and met. It appears whether or not anyone wants it there, because the alternative is a chart of your own customers being read in a board meeting as a measure of a market.

• Confidence intervals appear only when they mean something.

  A margin of error is shown only for a study that was quota-controlled, whose quotas were met, and that has at least a hundred valid responses. Below that, no interval is drawn at all — an interval printed on a convenience sample is a decoration that reads as a guarantee.

• Quota control is a claim the researcher makes, and the block says so.

  The cells a study was designed around are declared and shown, and whether the study was actually quota-controlled is recorded as a separate, explicit statement. Quotas are often managed outside the platform, so the two are kept apart rather than one being inferred from the other.

• The counts are of valid responses, not of attempts.

  A fixed set of rules — consent given, enough time on the conversation, enough of the mandatory questions actually covered, quality floors met, no duplicate — decides what counts. The same gate decides what you are billed for, so nobody has an incentive to be generous with it.

Break a report down far enough and a group becomes one person. The floor is the control for that — and how it behaves matters more than the number you choose.

• Off by default, and that is deliberate.

  The floor starts at one, which hides nothing, and you raise it if your own policy wants group-level hiding. A platform-imposed threshold would blank a small study entirely, and it is your study about your own people — the honest promise is the one made to the respondent before they answered, not a number applied to your dashboard afterwards.

• A hidden group is shown as withheld, never quietly dropped.

  When you do raise the floor, a group below it is replaced by a visible “withheld” marker rather than removed. A chart that silently omits a slice misreports its own composition, which is worse than showing nothing.

• The total is never hidden, and the arithmetic is closed.

  Totals always show. And if exactly one group falls below the floor, the next-smallest is withheld with it — otherwise anyone could subtract the visible groups from the total and recover the hidden one, which would make the whole exercise theatre.

On a study where respondents were told their answers are named, the floor does not apply: they already know the organisation can see who said what, and hiding a group there protects nobody. Access to those identities is controlled by permission instead, and every look is logged.

A margin of error on a sample that cannot carry one. A comparison against a benchmark we cannot show you the source of. And a projection from your customers to a population you did not sample. Any of the three would make a report look more authoritative and make it worth less, and a research buyer catches all three in the first five minutes.

Where a study genuinely is quota-controlled and large enough, the interval is calculated and shown. The rule is about what the sample supports, not about being cautious for its own sake.

A spreadsheet or a document leaves the platform carrying the same methodology block and obeying the same reporting floor as the dashboard it came from. An export that dropped either would become the version that gets forwarded — which is precisely why it is the version that must not be looser.