Market research tool
MaxDiff design generator
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Pair balance of the sample design
17–18times each pair meets
Every claim shown 40 times, 4 per respondent
10 items · 5 per set · 8 sets × 10 versions = 80 sets · pair SD 0.42
A MaxDiff design decides which items appear together in each best-worst question. A good design shows every item equally often (one-way balance), lets every pair of items meet about equally often (two-way balance), spreads items across positions, and never shows an item twice in a set. This generator builds such designs in your browser, reports the balance, and exports a data template the MaxDiff analysis tool opens directly.
Generate a MaxDiff design
Paste the items, choose how many appear per set, how many sets each respondent answers and how many versions to rotate, then generate. The same seed always gives the same design. "Use sample data" fills in the ten snack claims used by the MaxDiff analysis tool's sample. Nothing is uploaded.
Items and design settings
Paste the items to test, one per line, choose how many appear in each set, then generate. Everything runs in this browser tab.
One pair per line, by item text or number, separated by | ; , or a tab. These two items never share a set.
3–5 is typical; at most 8.
Different set lists, assigned to respondents in rotation.
The same seed and settings give the same design.
List the items to test, or use the sample of ten snack claims.
How the generator works
Each set is filled greedily: every slot takes the eligible item shown least often so far, breaking ties by how often it has already appeared with the items in the set, then at random. Local search then swaps items between sets of the same version, and replaces items, whenever that lowers — in this order — the spread of item counts within each version, across the design, of pair counts across the design, and of pair counts within each version. The search restarts from new random constructions within a fixed time budget and keeps the best design; it stops early when every criterion is at its theoretical minimum. Finally, the order inside each set is chosen so that each item appears in each position about equally often.
When a balanced incomplete block design exists for your settings — for example 7 items, 3 per set and 7 sets, where every pair meets exactly once — the generator finds it: the test suite checks this for several known designs. Prohibited pairs (two items that must never be compared, such as two prices) are excluded from every set and left out of the pair statistics.
The method
Formula
Exposures per respondent = S · k / n; mean pair count = V · S · k(k − 1)/2 ÷ [n(n − 1)/2]
where
- n
- number of items
- k
- items per set
- S
- sets per respondent (one version)
- V
- number of versions
- Balance
- the generator minimises Σ count² for item, pair and position counts; with fixed totals this minimises their variance
Worked example
The sample has 10 snack claims shown 5 per set in 8 sets, so each respondent sees each claim 4 times — above the usual minimum of 3. With 10 versions the design has 80 sets in all.
Every claim appears exactly 40 times across the design (SD 0.00) and 4 times within every version. There are 45 pairs of claims and 800 pair appearances, an average of 17.78 per pair; the generator gives every pair 17 or 18 appearances (SD 0.42), as even as whole numbers allow. Each claim sits in each of the 5 positions exactly 8 times.
Download the data template, collect the best and worst picks, and open the filled file in the MaxDiff analysis tool: it recognises the respondent, task, item and best/worst columns without any setup.
Using it well
When to use it
Use it when
- Use: Before fieldwork, to program a MaxDiff (best-worst scaling) question in any survey platform.
- Use: You want each respondent to see every item about three or more times, for individual-level scores.
- Use: Some items must never appear together, or you want several versions to spread order effects.
Not the right tool when
- Avoid: Items are full product profiles with attributes and prices — that needs a conjoint design.
- Avoid: You only need a ranking of a few items: with fewer than about six items, a ranking question is simpler.
- Avoid: Your survey platform builds and randomises MaxDiff sets itself and you cannot import a design.
How to interpret it
Rule of thumb
Check one-way balance first: within a version every item should appear the same number of times (±1). Pair counts that differ by one are normal; a pair that never meets is fine in a short single-version design, but more versions spread the pairs across the sample. Aim for at least three exposures per item per respondent when you plan to estimate individual (HB) scores, and keep sets to three to five items so each choice stays easy.
Questions
Frequently asked questions
How many sets does each respondent need?
A common rule of thumb is that each respondent should see each item at least three times, which needs at least 3 × items ÷ items-per-set sets. With 12 items in sets of 4 that is 9 sets. The tool suggests this number when the field is left blank and warns when the design falls short.
What is a balanced incomplete block design?
A design in which every set has the same size, every item appears equally often and every pair of items appears together equally often. It exists only for some combinations of items, set size and number of sets — 7 items in 7 sets of 3 is the classic example (the Fano plane). When one exists, the generator finds it; otherwise it gets as close as whole numbers allow.
Why use several versions?
Respondents are assigned to versions in rotation. Different versions put items in different company and positions, so order and context effects average out across the sample, and pairs that never meet in one version meet in another.
How do I analyse the results?
Download the data template, fill in the best and worst pick of each set with the item text or its position (1 = first shown), and open the file in the MaxDiff analysis tool. It maps respondent_id, task, item_1… and best/worst automatically and estimates counts, aggregate logit and HB scores.
Is the design random or fixed?
Both: it is built with random tie-breaks and restarts, but from a seeded random number generator. The same items, settings and seed always give the same design, so you can regenerate it later; change the seed to get a different, equally balanced design.
Sources and next steps
References & further reading
Try it on your own data — free, no signup
Upload a CSV, Excel, SPSS, Stata, SAS, R, Parquet or JSON file. Everything runs in your browser; your file never leaves your device.
Open the workspace →