Importance-Performance Analysis (IPA) Guide — Prioritizing Improvements in 4 Quadrants

You've finished the satisfaction survey, and now ten attribute scores sit in a row: Price 3.8, Support 4.1, Ease of use 3.5, Feature breadth 4.3… The moment you put that slide up in a leadership meeting, someone asks: "So where do we start?"

You don't have the resources to lift every satisfaction score at once. You need priorities. But "fix the lowest score first" gets it wrong, because fixing something the customer doesn't care about — even if its score is low — does nothing. The return on improvement lives where importance is high but satisfaction is low. The tool that puts that on a single scatter plot is Importance-Performance Analysis (IPA). In this guide we work through how to read the four quadrants, how to measure importance (this is the biggest fork in the road), where to draw the axes, and the trap that's specific to satisfaction data — all with a practitioner's feel for the work.

1. Why a "list of satisfaction scores" can't drive a decision

A report that just lines up attribute-level satisfaction scores in a bar chart is one of the most mass-produced outputs in research practice — and one of the least used to make decisions. The reason is simple: "how low the score is" and "what to improve first" don't line up.

A low score ≠ a high improvement priority

Suppose satisfaction with "the PDF design of your invoices" is a low 3.2. If customers don't weigh that heavily, improving it will barely move loyalty. Conversely, even if "how fast support first responds" sits at a middling 3.9, if customers value that most of all, lifting it to 4.3 has an outsized effect.

A decision needs two axes — "how much the customer weighs this item (importance)" and "how satisfied they are with it today (performance)." Plotting each attribute on those two axes and splitting them into four quadrants is the core idea of IPA. Ever since Martilla & James proposed it in 1977 in Importance-Performance Analysis (Journal of Marketing), it has remained a long-lived framework, used across CX, service quality, tourism, healthcare, and more.

How you measure satisfaction itself (5-point vs. 10-point, the Top 2 Box idea) is covered in the CSAT survey design guide. This article is one step past that — how to convert the satisfaction you measured into improvement actions.

2. The four quadrants of IPA — a map of where to start

Put importance on the vertical axis and performance on the horizontal axis, split each at its mean (or median), and every attribute lands in one of these four quadrants.

Principles for reading it

• In practice, look only at "Concentrate Here": the 2-3 items that land here become next quarter's improvement themes
• For "Possible Overkill," reallocate rather than cut: if you let quality slip just because importance is low today, it'll hurt when importance eventually rises. Start by reviewing how resources are allocated
• Resist the urge to fix "Low Priority": getting pulled toward items with low absolute scores is the biggest trap. Always read the importance axis alongside

3. How to measure importance — direct questions vs. statistical derivation

How you measure importance decides 80% of whether IPA succeeds. Do this carelessly and the vertical axis becomes untrustworthy, and the whole analysis collapses. There are two broad approaches to measuring importance, and each has the opposite weakness.

Method A: Direct questions (stated importance)

You ask, separately from satisfaction, "how important is each of the following to you?" It's simple, but it has a fatal quirk.

Respondents rate almost everything "important." Asked "Is support important?", nobody answers "I couldn't care less." As a result, importance scores all cluster at 4.0 to 4.5 (a ceiling effect), and there's barely any spread between items. The vertical axis collapses and the four quadrants stop working.

Method B: Statistical derivation (derived importance)

You treat the correlation or regression coefficient between each attribute's satisfaction and "overall satisfaction (or repurchase intent / NPS)" as its "importance." The logic is "when support satisfaction moves, overall satisfaction also moves a lot → support is important," and you never ask respondents about importance directly.

This avoids the ceiling effect of direct questions, but it has weaknesses of its own: multicollinearity (attributes correlate with each other, making coefficients unstable) and confusing correlation with causation. The practical mechanics of the correlation and regression that derivation rests on are covered in the survey aggregation and significance testing guide.

The practical conclusion: measure both and read the "gap"

Both academically and in practice, the best approach is to measure both stated and derived importance and read the gap between them.

• Items where customers say "important" when asked directly, but the impact on overall satisfaction is small → the customer's stated position. Loud voices that don't change behavior
• Items customers shrug off when asked directly, but that strongly move overall satisfaction → hidden drivers the customers themselves aren't aware of. This is the gold mine

Derived importance is continuous with "key driver analysis" (covered next), and IPA sits at the exit, visualizing those results in four quadrants.

4. Building the scatter plot — the unglamorous but consequential choice of split point

Where do you draw the crosshair that divides the four quadrants? This sways the conclusion more than it looks. There are mainly three options.

The pitfall of a mean split — "adding an item changes the conclusion"

The most-used mean split has an easily overlooked weakness. Because the dividing line is the mean of all items, adding or removing a single item shifts the line, and an attribute can jump from Concentrate Here into Keep Up the Good Work — with no actual change to the attribute.

"Support, which was in Concentrate Here in last month's report, is in Keep Up the Good Work this month — and we did nothing." It's not uncommon that nothing improved; the set of items being aggregated changed, and the mean line moved. If you use the mean as your split point, fix the item set and always annotate where the line sits.

How to build the scatter plot itself (a two-axis plot in Excel, axis labels, quadrant guide lines) is covered in the survey data visualization guide and the Excel survey aggregation guide.

5. The ceiling effect in satisfaction data — IPA's biggest trap

The thing that most often breaks an IPA is the "ceiling effect," where satisfaction data skews toward the high end.

Customer satisfaction comes out high by structure. On a 5-point scale it clusters at a mean of 4.0 to 4.5, and items rarely fall below 3.0 (a sampling bias is also at work: dissatisfied customers have already churned and don't respond in the first place). On the scatter plot, every item then crowds onto the right side (the high-performance side), and the horizontal split stops working. Everything lands in "Keep Up the Good Work," and nothing gets decided.

Remedies

• Split on the median: more robust to a skewed distribution than the mean. It can still bisect the items even when they bunch to the right
• Standardize performance (convert to z-scores): transform each item into its "relative position within the full set" before plotting. You're no longer dragged around by the high absolute values, and the relative differences between items become visible
• Use "dissatisfaction rate" or "gap from Top Box" instead of raw satisfaction: magnify the small differences within the high-score zone

Leaving the ceiling effect unaddressed and reading "everything's in the upper right, so we're in great shape" is the single most typical misreading of IPA. Once you've built the scatter plot, your first move is to suspect that the points have clumped together.

6. Improved IPA — the diagonal approach and the "asymmetry of satisfaction"

The classic four quadrants have drawn criticism, and several improved versions have been proposed. Two are worth knowing in practice.

The diagonal approach (Bacon 2003)

Instead of dividing at the vertical and horizontal means, you draw the 45-degree "importance = performance" line and measure priority by the distance from it. Bacon (2003) pointed out that four-quadrant classification becomes unstable when attributes cluster near a quadrant boundary, and showed that evaluating priority continuously by deviation from the diagonal (how far importance exceeds performance) is more robust. The advantage is that you avoid the binary "which quadrant does it fall into?"

The asymmetry of satisfaction — connecting IPA to Kano (Matzler 2004)

Classic IPA assumes "importance is a fixed value independent of performance," but Matzler et al. (2004) showed that an attribute's importance changes with the level of performance (it's asymmetric and nonlinear).

• Must-be quality: expected by default. Meeting it doesn't raise satisfaction, but missing it causes strong dissatisfaction (e.g., being able to log in, accurate billing)
• Attractive quality: its absence causes no dissatisfaction, but its presence makes satisfaction jump (e.g., unexpectedly proactive support)

In other words, "importance to a dissatisfied customer" and "importance to a satisfied customer" are different things, and computing derived importance separately for the high-satisfaction and low-satisfaction groups lets you separate must-be quality from attractive quality. Whether something in Concentrate Here is actually must-be quality (fixing it won't raise satisfaction, only resolve dissatisfaction) or attractive quality (fixing it makes satisfaction jump) changes what the investment means.

7. The editorial take — five things you must not do with IPA

From a vantage point of continuously following industry cases and the voices of hands-on practitioners, here are five accidents that happen with IPA again and again. None of them are "the analysis was wrong" — they're all "the reading missed."

1. Taking direct-question importance at face value

The most frequent accident. You ask about importance, everything comes back at 4.2, and the vertical axis collapses. Run IPA on direct questions alone and it will almost certainly break down under the ceiling effect. Always pair it with derived importance (correlation with overall satisfaction). If you can only capture one, choose derivation.

2. Ignoring the performance ceiling effect and relaxing because "everything's in the upper right"

Once you've built the scatter plot, your first move is to see whether the points have clumped together. If they crowd to the right, try a median split or z-score standardization. An IPA where "every item is in Keep Up the Good Work" has almost certainly missed a ceiling effect.

3. Not understanding the instability of the mean split, so the conclusion shifts every month

Every time you add or drop an item, the dividing line moves, and quadrants change with no connection to your initiatives. Fix the item set and always annotate the value of the dividing line. If you're tracking over time, the month-over-month comparison is meaningless unless you record where the line sat too.

4. Mistaking "Possible Overkill" for a strength and defending it forever

You convince yourself a low-importance × high-performance item is "our strength," keep pouring resources into it, and never get around to Concentrate Here. Possible Overkill is not something to take pride in — it's a candidate for reallocation. That said, check once from the asymmetry angle (is it must-be quality?) whether the "low importance" is only temporary.

5. Fixing a quadrant on the satisfaction of an attribute with small N

If your design has "only those it applies to" answer per attribute (e.g., only support users rate support satisfaction), N varies widely by attribute. Deciding a quadrant on the mean satisfaction of an N=15 attribute and acting because "it's Concentrate Here" is dangerous. Always show N per attribute, and withhold judgment on attributes with small N. The thinking on sample size is in how to determine survey sample size.

8. Running IPA with the Kicue survey tool

IPA splits into a "measure satisfaction and importance with questions" phase and an "plot it on a scatter chart and read the four quadrants" analysis phase. Kicue handles the former; the latter is done in combination with external tools.

• Measuring satisfaction and stated importance: supports question designs that measure each attribute's satisfaction and its "importance" on 5-point / 7-point Likert scales (question types, Likert scale design guide)
• Placing an overall-satisfaction question alongside: you can place the "overall satisfaction" and "repurchase intent" questions needed to compute derived importance within the same form
• CSV export with respondent IDs: outputs attribute-level satisfaction and overall satisfaction in a one-row-per-response structure, usable for the correlation calculations of derived importance and for splitting into high- and low-satisfaction groups
• GT and cross-tabulation: checking each attribute's mean and distribution is possible right on the dashboard

⚠️ What Kicue does not cover

• No IPA scatter-plot or four-quadrant plotting: the workflow is to export the CSV and build the plot and quadrant split in Excel (scatter plot) / R / Python / SPSS / JASP
• No derived-importance (correlation/regression) computation: statistical analysis is done in external tools. Kicue itself does not carry statistical-analysis features
• No z-score standardization or median-split preprocessing: both are done after export in a spreadsheet or statistics package

As related reading, working through the CSAT survey design guide, the CX metrics comparison guide, the survey aggregation and significance testing guide, the survey data visualization guide, and the VoC program design guide brings the full arc into view: "measure satisfaction → decide improvement priorities → run it operationally."

Summary — six points for using IPA to drive decisions

1. In practice, look only at "Concentrate Here" — the 2-3 items in high importance × low performance are the next improvement themes
2. A low score ≠ a priority — always read the importance axis alongside. Resist the urge to fix Low Priority
3. Lead with derived importance (correlation), follow with direct questions — direct questions alone almost always break under the ceiling effect
4. Suspect the performance ceiling effect — if the scatter plot bunches to the right, use a median split or z-score standardization
5. For the mean split, fix the item set and annotate where the line sits — otherwise month-over-month comparison is meaningless
6. Show N per attribute and withhold judgment on small N — attributes answered only by those they apply to get split thin

IPA is not "hard because the analysis is advanced." As long as you don't miss two things — how you measure importance, and the skew in satisfaction data — it's a high-value, cost-effective framework that can move a leadership-meeting decision forward with a single Excel scatter plot.

If you want to design a survey that measures satisfaction and importance, why not try the free survey tool Kicue? From Likert-scale design for attribute-level and overall satisfaction, to CSV export with respondent IDs, to GT and cross-tabulation, you can start building the input data for IPA in a single account (scatter-plot creation, derived-importance correlation, and z-score standardization are done in combination with Excel / R / Python / SPSS / JASP).