When a number is strong enough to show.
A pricing study is only as good as the sample under it. These are the rules Albi applies to every Van Westendorp, Gabor-Granger, MaxDiff and conjoint result: how many respondents each method needs, whose answers count, and what happens when there aren't enough.
The thresholds below come from the same code that grades your report, not from a separate policy document.
A floor for each method.
A Van Westendorp crossing can be moved by a handful of respondents. MaxDiff and conjoint estimate many parameters at once, so they need more. The floors follow published guidance: 100 completes for a firm price-sensitivity read, and Sawtooth's 300 per study and 200 per reportable subgroup for choice models.
| Method | Full from | Small sample from | Too thin from |
|---|---|---|---|
| Van Westendorp | n = 100 | n = 50 | n = 30 |
| Gabor-Granger | n = 100 | n = 20 | n = 10 |
| MaxDiff | n = 200 | n = 50 | n = 30 |
| Conjoint | n = 300 | n = 60 | n = 30 |
| Importance × willingness-to-pay | n = 200 | n = 60 | n = 40 |
| Scale | n = 100 | n = 30 | n = 15 |
| Choice | n = 100 | n = 30 | n = 15 |
| Performance rating | n = 100 | n = 30 | n = 15 |
| Open text | n = 80 | n = 30 | n = 15 |
Below the “too thin” floor a cut is not offered at all. Importance × willingness-to-pay joins a MaxDiff to a price question, so it is graded on the respondents who answered both. Multiple-choice and ranking questions have no published floor and are not graded.
Your study's own result is never hidden
Hiding the only result a study produced helps nobody. Below the floor, your headline chart still renders, marked as a small sample. The full ladder applies to cuts: a segment, a country, a filter.
Conjoint: a minimum before launch
A conjoint design sets its own minimum sample by Orme's rule, n·t·a / c ≥ 500: respondents × tasks × options per task, over the most levels any attribute has. A plan below that minimum cannot be launched.
Four verdicts, applied to every cut.
Three confident-looking country charts built on 30 respondents each is how a reader gets misled. So a cut is offered only when it can carry a chart.
Full
At or above the method's floor. The chart renders with no caveat.
Small sample
Below the floor but still a reading. The chart is marked “Too few to trust”, and the ranges, not the single numbers, are the finding.
Too thin
A cut this small cannot carry a chart. Albi withholds it and shows the numbers only, marked “too thin”.
Not offered
Below the lowest floor, the cut is not reportable and does not appear as a filter at all.
Whose answers count, and why.
Every real respondent is checked when they submit. A check raises a flag; flags are evidence, kept on record with when and why they were raised. Whether a flag removes someone from the analysis is a rule, and nobody is dropped silently: the analysis shows how many were excluded, and why, beside the n it used.
Never counted
- Respondents who were screened out, rejected, or didn't finish.
- Duplicate: the same person (the same panel ID or the same customer link) answering a second time. The first answer counts.
- Answered under 70% of the required questions they were shown. Questions that logic hid from them don't count against them.
Flags mark, you decide
Speeding or straight-lining correlates with noise but doesn't prove it: a decisive respondent can answer a short survey fast and honestly. So every flag in the table below only marks a respondent for you to review, however many they have. You decide whether to exclude them.
Notes, not quality flags
- Same answer at every price (from 3 prices asked): “I wouldn't pay any of these” is a real answer.
- Left questions unanswered: an optional or later-added question skipped.
- No start time recorded: a missing timestamp, not a quality finding.
| Flag | Raised when |
|---|---|
| Finished too fast | Finished in under a third of the median time of the other completes in the same run. Until a run has 10 completes, a third of the survey's estimated time. Checked once, at submit. |
| Same answer throughout | Gave the same answer to every scale question, with three or more scale questions. |
| Answer not valid | Sent an answer the question cannot take; the answer is dropped. |
| Choices look random | In a conjoint fit, their choices look no better than random. |
| Inconsistent answers | Answers that contradict each other, flagged by an analysis or a reviewer. |
| Prices out of order | Van Westendorp prices not in order (too cheap ≤ cheap ≤ expensive ≤ too expensive). |
| Same price four times | The same price for all four Van Westendorp questions. |
| Accepted a higher price after refusing a lower one | Gabor-Granger answers that contradict each other. |
| Always picked the same position | MaxDiff: picked the same position on every screen (checked from 4 screens). |
| Always chose the option in the same place | Conjoint: chose the option in the same place every time (checked from 4–8 tasks, depending on options per screen). |
You can overrule any of it
On the Responses view you can exclude or include anyone by hand, or dismiss a flag you judge a false positive. The flag stays on record, marked dismissed. Your decision changes the analysis only.
What you pay for
A duplicate is not charged. A respondent left out for answering under 70% of the questions, or excluded by you, finished the survey and was paid on the panel, so they are charged.
Synthetic respondents
AI-generated respondents don't pass through the submission checks above. Their answers count once generated, unless screened out or under 70% complete.
No attention checks yet
We don't use attention checks yet, and a skipped question is never treated as a failed one. We plan to add them in the formats Prolific allows. Until then we reject no one on Prolific: every participant who isn't a duplicate or screened out is approved and paid.
Want to see a report graded this way?
Design partners run their first study with us from October 20.