Key driver analysis: how it works and how to read the output

Key driver analysis is a statistical technique that ranks the themes in customer feedback by their measured relationship with a score such as NPS, CSAT or CES. It exists so the budget follows impact instead of volume. Every feedback dashboard can tell you what customers mention most. Almost none can tell you what customers will leave over, and the two lists are rarely the same. This guide explains how the analysis works without the maths degree, how to read its output as a quadrant, and why derived importance beats asking customers what matters. It ends with the pitfalls that produce false drivers. An interactive example sits a few paragraphs down; hover over or tap the bubbles.

Key takeaways:

• Frequency and importance are different measures. The topic mentioned most and the topic driving your detractors are usually different topics.
• Key driver analysis needs two inputs you may already have: scores per customer, and open feedback classified into stable topics.
• Read the output as a quadrant: high-impact problems to fix first, high-impact strengths to protect, low-impact noise to consciously park.
• Derived importance beats stated importance. People answer politely; their scores answer honestly.
• 75% of companies run an ongoing CX measurement programme, yet only 34% consistently follow up after a complaint. The ranking is the bridge between the two numbers.

The problem it solves: loud versus important

Key driver analysis solves the oldest problem in feedback programmes: the loudest topic and the most important topic are rarely the same. Most dashboards only show the loud one. The 2026 CX Maturity Study, which Loyalty Group ran with Hello Customer among 203 CX managers, describes the resulting state precisely: measurement is broad, action is inconsistent. 75% of companies run an ongoing measurement programme, but the insights are not systematically translated into prioritised improvements. Only 13% score as highly mature across all six dimensions the study rates: strategy, measurement, improvement, management ownership, management support for employees and documentation.

The root cause is almost always the same chart: topics ranked by how often they are mentioned. Volume feels like a safe proxy for importance, and it is a poor one. Customers readily mention what is easy to talk about: range, price, the website. What quietly decides their loyalty, whether the refund arrived or whether the stock was real, comes up far less often. A ranking by mentions sends the budget to the loudest topic; a key driver analysis sends it to the one statistically attached to your detractors and your churn.

This is what the output looks like:

Read it the way the analysis intends. Every bubble is a theme from open feedback. Its size is how often customers mention it, its horizontal position is how they feel about it, and its height is its measured impact on the score. The whole argument of this article is visible in one comparison. Billing draws a fifth of the mentions Price does and still sits in Fix this now. Store ambiance is mentioned more often than Billing and has half its impact. Volume decides which bubble you notice; height decides which bubble is costing you customers.

How key driver analysis works, without the maths degree

Key driver analysis needs two ingredients, per customer:

1. A score: NPS, CSAT or CES from your surveys, chosen per touchpoint as described in the three-layer measurement system.
2. Topics: the same customer's open feedback, classified into a stable set of themes. This is the output of your customer feedback analysis; at volume, an engine like ISAAC produces it continuously, in more than 30 languages.

The analysis then asks one question per topic. How do the scores of customers who mention this topic differ from the scores of those who do not, taking the other topics into account? A topic whose mentioners score sharply lower is a negative driver; one whose mentioners score higher is a positive driver. The output is an impact figure per topic. In the widget above it is the height of each bubble; hover one and the tooltip shows it as a score out of ten.

The statistical machinery behind "taking the other topics into account" matters, because feedback themes overlap: the customer angry about refunds often also mentions customer service. Naive correlation double-counts entangled topics. Modern implementations use methods built for exactly this. Relative weight analysis is what market research methodologists recommend as the practical choice for driver analysis when attributes are numerous and correlated. Shapley values, the game-theory approach that fairly attributes each factor's marginal contribution to an outcome, have become a foundation of explainable analytics. You do not need to run these by hand; you do want to know your tooling uses something of this class rather than raw correlations.

Reading the output: the four quadrants

The output of a key driver analysis reads as a quadrant once impact is paired with sentiment, and each quadrant carries its own instruction. They are the four labels on the widget above and on the key driver analysis map inside Hello Customer:

Bubble size is the third dimension. A large bubble in "Fix this now" is an expensive, visible problem that funds its own repair. A small bubble high in the same quadrant is the classic hidden driver: rarely mentioned, strongly tied to detractors, and the main reason key driver analysis exists. "Promote" earns equal standing with the fix list: a high-impact strength is a moat to protect, never a budget line to trim because "nobody complains about it".

One habit turns the quadrant from a chart into a routine: at every monthly review, name the topic that moved quadrants since last time. Quadrant migration is the earliest structural signal your feedback system produces.

Turn every driver into a sentence

A driver worth acting on should leave the analysis as a sentence, because the chart is for the analyst and the organisation acts on sentences. Translate it into one line with five parts: the topic, the direction, the magnitude, the segment, and one customer's words. For Delivery time, one of the four fix-this-now themes in the widget: "Delivery time is a top negative driver this quarter; customers who mention it score on average eleven points lower, concentrated in two regions after the carrier change; 'second week in a row my slot came and went.'"

That sentence does things a bubble cannot. It survives being forwarded without the chart. Forwarding is the normal case. 58% of the CX managers in the 2026 study name organisational silos as their main hurdle, so most drivers cross a department boundary before anyone can act on them. The sentence also names an owner implicitly, because whoever runs delivery knows this is theirs. And it carries its own evidence, so the first meeting is spent on the fix rather than on whether the problem is real. A chart is information; a sentence with a quote is an assignment.

Why not ask customers what matters?

Stated importance and derived importance disagree, reliably, which is why key driver analysis derives importance from scores instead of asking for it. Ask customers to rank what they care about and everything scores high: quality, price, service, all "very important". The polite answer carries almost no information. Derived importance skips the question entirely and reads the answer from behaviour: which topics, when they appear in a customer's own words, travel with low scores and departures.

The gap between the two is where the money hides. Nobody rates "stock accuracy" as a top-three priority in the abstract; the customers who drove twenty minutes for an item the website promised was in stock leave anyway. Their verbatims plus their scores tell you what their survey rankings never would.

This is also the honest limit of the method: derived importance needs the topic to appear in feedback at all. A friction so complete that customers leave without a word never enters the model. That is why the analysis lives inside a broader listening system, next to sentiment trends and behavioural data, rather than replacing it.

From driver to decision

A driver ranking earns its keep in the two weeks after it is produced:

1. Route the top negative driver to the department that owns it, with the impact figure, the affected segment and ten raw verbatims attached.
2. Cost the fix against the driver's value. The impact figure converts to money through the cohort method in the customer experience ROI guide: customers mentioning this topic churn at X% against baseline Y%, times customer value. The Verde Group's research offers an encouraging benchmark for the category. A 10% reduction in the most damaging customer friction associates with a 2 to 5% gain in revenue and profit.
3. Fix, then re-measure the same driver. The impact figure and the score at the affected touchpoint, before and after. A driver that stops driving is the cleanest proof a feedback programme can produce, and the moment-by-moment redesign that gets it there is what customer experience design covers.

Mikkel Korntved, CEO of Loyalty Group and co-author of the study, describes the moment most organisations stall at: "Believing in customer experience is free. Steering by it costs a seat at the executive table, and that's exactly where most companies hesitate." A driver ranking with a cost and an owner attached is what earns that seat.

What the 2026 CX Maturity Study says about measuring and acting

The 2026 CX Maturity Study explains why so few driver rankings exist in the first place: the inputs are unevenly collected. NPS is used by 74% of companies and CSAT by 72%, but sentiment analysis by only 33% and CES by 30%. 37% of CX managers name access to structured, relevant data as one of their biggest obstacles. A key driver analysis needs the score and the classified open feedback for the same customer; most programmes have the first and skip the second.

The study's broader finding is the one this article exists to fix. Companies measure broadly but do not always act on what they measure, and only 34% consistently follow up after a complaint. A driver ranking is the shortest route from the 75% who measure to the minority who act. It turns a pile of feedback into an ordered list with an owner per line. Our overview of the customer experience metrics worth tracking shows which scores make the best input. The full trends analysis of the study shows how far most companies still have to go.

The pitfalls that produce false drivers

Five conditions produce false drivers in a driver ranking, and each one is avoidable before the analysis runs:

• Unstable topics. If the taxonomy changed mid-period, the driver ranking compares apples to renamed apples. Freeze categories before trusting trends.
• Small samples. A topic mentioned nine times can top the impact chart by accident. Set a minimum mention count per topic per period, and treat anything below it as a hypothesis.
• Correlation read as causation. The analysis says a topic travels with low scores; whether it causes them is what the re-measurement after the fix establishes. Treat the ranking as the ordered list of what to investigate and fix first.
• One company-wide run. Drivers differ by journey and segment: what drives detractors in onboarding is different from what drives them at renewal. Run the analysis per journey, compare across segments, and let the differences inform the fix.
• Forgetting the positive side. Programmes that only model detractor drivers systematically underinvest in the strengths keeping promoters promoters.

Frequently asked questions

What is key driver analysis in customer experience?

Key driver analysis is a statistical technique that ranks the themes in customer feedback by their measured relationship with a target metric such as NPS or CSAT. It identifies which topics move the score, rather than which are mentioned most.

What data do you need for a key driver analysis?

For each customer you need a score and their open feedback classified into consistent topics, ideally with metadata such as segment and journey. A few hundred responses per touchpoint gives a workable first ranking; stability improves with volume.

How is key driver analysis different from a correlation analysis?

Feedback topics overlap heavily, and simple correlations double-count entangled themes. Proper key driver analysis uses attribution methods built for correlated inputs, such as relative weight analysis or Shapley values, so each topic's contribution is estimated fairly.

How often should you run key driver analysis?

Run the analysis continuously or monthly per journey, with a deliberate review each quarter. The valuable signals are ranking changes and quadrant migrations, which only exist if the analysis repeats on a stable taxonomy.

Does key driver analysis prove causation?

No. It produces a prioritised, evidence-based list of what to fix first. Causation is established the practical way: fix the top driver, re-measure the same metric at the same touchpoint, and compare.

Which metric should key driver analysis use as the target?

Use the metric you already measure at that touchpoint: NPS for relationship or transactional surveys, CSAT for service moments, CES for processes. Running the analysis against a score the touchpoint does not collect is the fastest way to a ranking nobody trusts.

The ranking is the beginning

Key driver analysis does one thing extremely well: it replaces the loudest-topic budget with an evidence-ordered one. Everything after the ranking is ordinary, disciplined work (routing, fixing, re-measuring), which is exactly why it succeeds. The organisations stuck between measuring and acting already have the data. What they are missing is an ordering they trust enough to act on.