MaxDiff: what it is and how to use it

Cloud4Feed Research Team ·Updated 22 Oct 2026 ·8 min read

MaxDiff, short for Maximum Difference Scaling and also called best–worst scaling, is a survey method for finding out what matters most to people. Instead of rating each item on its own, respondents see small sets of four or five items, all of them reasonable, and pick the most important and the least important one in each set. They cannot say “I want them all”.

The method was developed by Jordan Louviere in the late 1980s and first published in 1992. It is widely used in market research because it fixes a common problem with rating scales: when customers rate items one by one, most items land between 4 and 5, and the list no longer shows a priority. People do not judge options in isolation; they compare them. MaxDiff measures exactly that comparison.

1. 1Design the experiment. A balanced design, usually based on a balanced incomplete block design (BIBD), places every item in many different sets so each item appears about equally often and competes with every other item.
2. 2Collect the choices. Each respondent completes a series of sets, choosing the most and least important item in each.
3. 3Model the choices. A choice model, multinomial logit (MNL) for the total sample or Hierarchical Bayes (HB) for each respondent, estimates how likely each item is to be chosen over the others.
4. 4Report utilities and shares. The model gives each item a utility score on a common scale, often converted into preference shares that add up to 100.

Enter each item with how often it was shown, chosen as most important and chosen as least important. The ranking updates as you type.

Rating scales ask how good or important each item is on its own. MaxDiff asks which item wins when items compete. The difference shows up in how clear the result is.

Each item should describe one idea, at the same level of detail as the others, in plain words customers use. Avoid items that overlap (“fast delivery” and “quick shipping”) or that bundle two things (“fast and cheap delivery”).

Most studies use 8 to 30 items with 4 or 5 items per set. Aim to show each item at least three times to every respondent. A quick way to estimate the number of sets: three times the number of items, divided by the items per set. With 20 items and 5 per set, that is 12 sets.

Decide before fieldwork which groups you want to compare, such as customer type, region or value tier, and make sure each group has enough respondents to report on its own.

Utility scores are the model’s raw output. They are usually centred on zero, so positive items are above average and negative items below. Differences between utilities are meaningful, so you can say how far apart two items are.

Preference shares convert utilities into the probability of each item being chosen, scaled so that all items add up to 100. They can be read as ratios: an item with a share of 20 is twice as important as an item with a share of 10. This is the version most teams present.

Segment results come from Hierarchical Bayes, which estimates scores for every respondent. The same product can be judged by very different values: one segment may put speed first, another reliability, another price transparency. That is where measurement turns into design.

Like every survey method, MaxDiff is only as good as its items and sample. Check that the items you care about were in the list, and that the respondents represent the customers you want to serve.

The real value of MaxDiff is that it tells you what to leave out. Items at the bottom of the ranking are the work you can deliberately stop or postpone, so resources go to the few things that move customers. A roadmap is a strategy only when it also says no.

MaxDiff shows what matters; CSAT, CES or item-level ratings show how well you deliver it. Put the two together and start where importance is high and performance is low.

When segments disagree, one average priority list serves nobody well. Use the segment results to tailor journeys, messages and service levels to what each group values most.