A survey with a dozen rated attributes produces a dozen scores, and it's tempting to just fix whichever ones scored lowest. That instinct is more wrong than it feels like it should be - a low score on something nobody actually cares much about isn't a priority, and a merely-decent score on something that matters enormously might deserve more attention than an attribute scoring worse but mattering less. Importance-performance analysis, a technique developed by researchers John Martilla and John James back in 1977 and still widely used in marketing, hospitality, and customer experience research today, exists specifically to sort this out - by asking not just how something performed, but how much it matters in the first place, and looking at the two together.
Table of Contents¶
- The Two Questions It Combines
- Building the Chart
- Reading Each Quadrant
- Common Pitfalls in Practice
- Running It Per Segment
- A Worked Example
- FAQ
The Two Questions It Combines¶
The method starts from asking respondents two separate things about each attribute you care about, rather than just one. First, performance: how well are we actually doing on this - the standard satisfaction-style rating most surveys already collect. Second, importance: how much does this actually matter to you - a question most surveys skip, assuming importance can just be inferred from context, when in practice it often can't be. Asking both explicitly, for the same set of attributes, is what makes the rest of the analysis possible.
Building the Chart¶
Once you have both scores for every attribute, plot them on a simple two-axis chart: importance along one axis, performance along the other, with each attribute placed as a single point based on its two scores. Draw a line down the middle of each axis - often the overall average importance and average performance across all attributes, though a fixed midpoint on the scale works too - and you end up with four quadrants, each one representing a genuinely different kind of finding. Reports supports scatter plot charts, which is exactly the chart type this technique needs if you want to build it directly from survey data rather than a separate spreadsheet exercise.
Reading Each Quadrant¶
Concentrate here (high importance, low performance) is the quadrant that deserves the most attention - attributes people say matter a lot, that you're currently not doing well on. This is where the clearest, most defensible priorities live.
Keep up the good work (high importance, high performance) covers the things that matter and that you're already doing well - not a place to redirect resources away from, since these are likely a real part of why people are satisfied at all, but also not where new investment needs to go.
Low priority (low importance, low performance) is exactly what it sounds like - things people don't rate as mattering much, that also aren't performing well. It's tempting to want to fix everything with a low score, but this quadrant is a legitimate place to consciously deprioritize, since the effort spent here won't move the outcomes people actually care about.
Possible overkill (low importance, high performance) covers attributes performing well that respondents don't consider especially important - not a problem exactly, but worth a second look if significant resources are being poured into maintaining something nobody's asking for at that level.
Common Pitfalls in Practice¶
Where you draw the dividing lines matters more than it looks like it should. Using the average importance and average performance across your specific attribute list, as most implementations do by default, means the quadrant boundaries shift depending on which attributes you happened to include - add a couple of universally beloved attributes to the list and the performance average climbs, quietly reclassifying attributes that scored the same as before into a less flattering quadrant purely because the average moved around them. A fixed midpoint on the rating scale itself (the middle of a 1-10 scale, for instance) avoids that instability, at the cost of being a less locally-calibrated dividing line - which approach to use is worth deciding deliberately rather than defaulting to whichever one the analysis tool happens to use.
Small sample sizes per attribute are a second common pitfall - if only a handful of respondents rated a particular attribute's importance, its position on the chart is much less stable than an attribute rated by your full sample, and a single outlier respondent can visibly shift a low-volume attribute's plotted position in a way that looks meaningful but is really just noise. It's worth checking the underlying response count behind any attribute sitting close to a quadrant boundary before treating its placement as decisive.
Running It Per Segment¶
The blended, whole-audience version of the chart is a useful starting point, and it can hide real disagreement between groups who experience the same attributes very differently. Running the same importance-performance analysis separately for two meaningfully different segments - new customers versus established ones, say, or two different customer tiers - sometimes reveals an attribute sitting in "concentrate here" for one group while sitting comfortably in "keep up the good work" for another, a distinction the combined chart averages away into a single, less actionable middle position. This is worth doing whenever you already suspect two segments have genuinely different priorities, rather than as a routine step for every analysis, since splitting the sample also means each segment's chart is built on a smaller, less stable set of responses per attribute.
A Worked Example¶
A hotel chain asks guests to rate both importance and performance on ten attributes: room cleanliness, check-in speed, wifi quality, breakfast variety, and others. The results plot wifi quality squarely in "concentrate here" - guests consistently rate it as one of the most important attributes to their stay, and performance scores are middling at best. Breakfast variety, meanwhile, lands in "possible overkill" - performance scores are excellent, reflecting real investment in an elaborate breakfast program, but importance ratings are only moderate; guests like it, but it's not what's driving their satisfaction with the stay overall. The chain redirects some budget from expanding the breakfast program toward a wifi infrastructure upgrade instead - a reallocation the raw performance scores alone would never have suggested, since breakfast variety and wifi quality had similar performance ratings before importance was factored in at all.
FAQ¶
Do I need a separate importance question for every attribute I'm rating?
Yes, ideally - asking importance for each attribute individually gives you a cleaner result than trying to infer importance from a single ranked list, though a ranking exercise can work as a lighter-weight substitute if survey length is a real constraint.
What if two attributes land right on the dividing line between quadrants?
Treat the placement as approximate rather than exact - the real value of the technique is the overall pattern across attributes, not a precise binary classification for every single point near a boundary.
Is importance-performance analysis the same as key driver analysis?
Related, but not identical. Key driver analysis infers importance statistically from how attribute ratings move together with an overall outcome; importance-performance analysis asks respondents directly how important each thing is. Both are useful, and they can validate each other when they broadly agree.
Can I use this for something other than customer experience?
Yes - the method works for any set of rated attributes paired with an importance rating, employee engagement dimensions and product features included, not just traditional customer satisfaction attributes.
Should I use average scores or a fixed scale midpoint to divide the quadrants?
Either is defensible, but be consistent about which one you use wave over wave - average-based dividing lines shift depending on the specific attribute list, while a fixed midpoint stays stable across waves and attribute lists, which matters more if you're tracking quadrant placement over time.
Is it worth running a separate chart for each customer segment?
When you already suspect two segments prioritize things differently, yes - a blended chart can average away a real disagreement between groups. For a first pass, or when segments are small, the combined chart is usually the more stable starting point.
For more on prioritizing what to act on from survey data, see Key Driver Analysis and Beyond Averages: The Professional's Guide to Survey Analysis.