Numbers you can put your name on
100% from n=2 is the most confident-looking number in any report — and the least supportable. Gotopico is built to refuse to show you that number. This page is the policy, with the real thresholds.
Thin bases don't get reported
One policy decides whether a percentage may exist, everywhere in the product — the analysis layer and the pre-field checks quote the same thresholds, so a study is never told two different stories about the same base.
Operational rates vs. reportable proportions
Whole-sample rates — survey completion, per-question drop-off — are a different animal: hiding them would obscure the fielding status of a freshly launched survey without preventing any misreading. So they are annotated instead of suppressed: the rate is always shown next to its base, tagged when it sits below the reporting minimum, with its margin of error making a thin base look thin.
No precision for withheld figures
A margin of error is never quoted for a suppressed percentage. Quoting a precision for a figure withheld as unreliable invites reporting the precision instead — so the policy withholds both.
Drag the sample size and watch the verdict change. These are the exact thresholds the product applies — suppress below n=15, flag low-base below n=30 — not a simplified demo.
62% ± 8.7 pts
95% interval: 53.3% – 70.7% · n = 120
Margin of error is the 95% normal-approximation half-width, the same formula the reporting layer uses. Small bases make wide intervals: at n=30 a 50% result is ±17.9 points.
Every base travels with its chart
Illustration of the dashboard's chart header.
Charts in the analytics dashboard carry their base in the header — an n=412 badge next to the title, not a footnote in an appendix. Model-fit banners state their sample the same way.
Significance testing done properly
“Is this difference real?” gets a real answer: the appropriate test for the data, an effect size next to the p-value, and intervals chosen for coverage rather than convenience.
NPS confidence intervals are built from Wilson intervals on the promoter and detractor shares, combined conservatively — the interval is honest about the score being a difference of two noisy percentages.
Comparing NPS across segments runs a significance test on the promoter rates and states the verdict in plain language, p-value included — whether or not the difference holds up.
Cross-tabulations mark individually significant cells using adjusted residuals — but only inside a table whose overall chi-square test passed first. Cell marks are never inferred from eyeballing percentages.
The output states the exact residual threshold a cell had to clear, and how many cells cleared it. Zero marked cells in a significant table is reported as a real result, not papered over.
Guardrails before you spend a cent on fieldwork
Every planned pricing or choice study is evaluated before it fields — and re-validated when you hit publish. Each check returns a verdict of ok, warning, or blocking, with a plain-language message and a concrete remedy. A design that would produce statistically meaningless results cannot be fielded.
A Van Westendorp plan under 15 respondents is blocked outright — the curve intersections would be driven by single respondents.
A MaxDiff plan below the Sawtooth/Orme rule of thumb (n ≥ 500 × items ÷ (tasks × shown)) gets a warning with the exact number to aim for.
A conjoint design whose respondents can't mathematically cover its utility parameters — tasks × (concepts − 1) below the parameter count — cannot be fielded at all.
More than 15 choice tasks per respondent triggers a fatigue warning: past that point, speeding and straight-lining degrade the data you paid for.
Segment-level estimates with fewer than 100 respondents per segment are blocked; below 200 per segment you get a warning before you commit.
And after fielding, the same rule of thumb is run against what you actually collected: choice-model results carry an achieved-sample verdict — high, medium, or low confidence, with directional-only results labeled as exactly that.
Plenty of tools claim a “colorblind-friendly” palette. Ours is validated by an automated check that parses the actual hex values out of the chart theme source — so the claim can never drift from the colors the app really draws with — and fails the change if any check breaks.
- Adjacent series stay distinguishable under simulated protanopia, deuteranopia, and tritanopia, with a minimum perceptual-difference floor
- A separate, higher separation floor for full-color vision
- Every hue kept in a usable lightness band, and above a chroma floor, against both the light and dark card surfaces
- 3:1 contrast against the chart surface — or the check demands visible labels instead
The check covers exactly the number of series the app is allowed to draw at once — beyond that, extra categories fold into “Other” rather than dipping into unvalidated colors.
We never paywall honesty
Some analyses are plan features. Honesty about your numbers never is. On every plan, every reported percentage passes the same base-size policy, low bases stay flagged, margins of error stay attached, and the pre-field guardrails stand between every user and a statistically invalid study. Upgrading buys you more analysis — it never buys anyone the ability to hide a caveat.
Keep digging
Choose the right research method
MaxDiff, conjoint, Van Westendorp, and more — organized by the decision you need to make.
Data quality you can defend
Multi-signal fraud scoring, method-aware quality checks, and a flag-don't-delete posture.
How we stack up
See how Gotopico compares to other survey platforms on methods, analytics, and rigor.
Run research you can defend
Field a study with the guardrails on, and report numbers that carry their own evidence.