---
title: "Driver Analysis: Which CX Drivers Actually Move the Metric"
date: "2026-07-28"
description: "Driver analysis is a statistical method that identifies which underlying factors — support responsiveness, onboarding, product reliability, pricing — correlate most strongly with a customer experience outcome metric such as Net Promoter Score, CSAT, or renewal likelihood."
keywords: ["driver analysis", "cx driver analysis", "key driver analysis", "customer experience drivers", "nps driver analysis"]
author: "Perspective AI Team"
category: "AI Conversations at Scale"
slug: "driver-analysis-cx-which-drivers-move-the-metric"
excerpt: "Driver analysis is a statistical method that identifies which underlying factors — support responsiveness, onboarding, product reliability, pricing — correlate…"
image: "https://getperspective.agency/assets/b06d1eee-3ca2-4fc7-8192-2e90a02d30b1"
tags: ["driver analysis", "cx driver analysis", "key driver analysis", "customer experience drivers", "nps driver analysis"]
lastModified: "2026-07-28"
definition: "Driver analysis is a statistical method that identifies which underlying factors — support responsiveness, onboarding, product reliability, pricing — correlate most strongly with a customer experience outcome metric such as Net Promoter Score, CSAT, or renewal likelihood. In CX practice it is usually called key driver analysis, and its output is a ranked list of drivers showing which attributes appear to move the metric and roughly how much each one contributes."
faqs: [{"question": "What is key driver analysis in customer experience?", "answer": "Key driver analysis is a statistical technique that ranks which attributes — support, onboarding, pricing, reliability — correlate most strongly with a customer experience outcome metric like NPS or CSAT. It answers \"which factors move together with our score, and by how much.\" It's a prioritization tool: it converts a long list of survey attributes into a short shortlist of things that appear to matter most, so teams can focus improvement effort rather than guessing."}, {"question": "What statistical methods are used for CX driver analysis?", "answer": "The three common methods are bivariate correlation, multiple regression, and relative importance analysis. Bivariate correlation ranks each driver against the outcome one at a time but overstates importance because drivers are intercorrelated. Multiple regression weights all drivers together but grows unstable under multicollinearity. Relative importance methods — Johnson's relative weights and Shapley value regression — split explained variance fairly across correlated drivers and are the most defensible choice for typical CX survey data."}, {"question": "Does driver analysis prove what causes customer satisfaction?", "answer": "No. Driver analysis identifies correlation, not causation, so it cannot prove that any factor causes a change in your metric. Halo effects, confounding variables, reverse causality, and drivers you never measured all break the link between correlation and cause. Driver analysis is best treated as a way to generate strong hypotheses about what matters, which you then confirm through follow-up conversations, controlled changes, or experiments."}, {"question": "Why does driver analysis tell you which factors matter but not why?", "answer": "Driver analysis operates entirely on compressed data — a numeric score and a fixed set of attribute ratings — so it can detect statistical association but has no access to the customer's reasoning. Knowing that \"onboarding\" ranks as a top driver doesn't reveal what about onboarding mattered: the speed, a missing integration, or a confusing first session. Recovering that \"why\" requires asking the customer directly, which is what conversational follow-up adds on top of the model."}, {"question": "How do you combine driver analysis with customer interviews?", "answer": "Use driver analysis to triage and interviews to diagnose. First, run the model to find the two or three highest-importance drivers. Then route the relevant customers into short, AI-moderated interviews that ask why that specific factor shaped their rating and probe vague answers. Finally, theme the conversation verbatims and feed the causal patterns back as sharper candidate drivers for the next cycle, closing the loop a static survey leaves open."}]
---

## What is driver analysis in CX?

Driver analysis is a statistical method that identifies which underlying factors — support responsiveness, onboarding, product reliability, pricing — correlate most strongly with a customer experience outcome metric such as [Net Promoter Score](https://hbr.org/2003/12/the-one-number-you-need-to-grow), CSAT, or renewal likelihood. In CX practice it is usually called key driver analysis, and its output is a ranked list of drivers showing which attributes appear to move the metric and roughly how much each one contributes.

The critical thing to understand before you run one: driver analysis tells you *which* factors move together with your score. It never tells you *why*. Correlation is a direction to look, not a reason — and the distance between those two things is where most measurement programs quietly lose the plot. Done well, driver analysis is one of the most useful techniques in the [customer experience analytics](/blog/customer-experience-analytics-from-dashboards-to-the-why-behind-the-numbers) toolkit; done in isolation, it produces a confident-looking chart that points at a problem no one on the team actually understands yet.

## How driver analysis works: the three common methods

Driver analysis works by quantifying the statistical relationship between a set of candidate drivers — the attributes you measured on a survey — and a single outcome metric you care about. Three families of methods dominate CX practice, in ascending order of rigor.

### Bivariate correlation

Bivariate correlation measures the relationship between each driver and the outcome one factor at a time, using a Pearson (linear) or Spearman (rank) coefficient. It is fast, intuitive, and the method behind most quick "top drivers" charts. Its fatal flaw is that it ignores the fact that drivers are correlated with each other. When a customer is happy, they tend to rate *everything* higher — support, product, value alike — so measuring each attribute against the outcome in isolation systematically overstates how many things "drive" the score. It's a reasonable first screen and a poor final answer.

### Multiple regression

Multiple linear regression models all of the drivers simultaneously and reports a standardized coefficient (a beta weight) for each one, representing its association with the outcome while holding the other drivers constant. That "holding the others constant" is a genuine improvement over one-at-a-time correlation, but regression on CX data has a well-known failure mode: multicollinearity. When predictors are highly intercorrelated — routine in a survey battery where every attribute is a flavor of "how good was your experience" — coefficients become unstable, standard errors inflate, and a driver's sign can even flip from positive to negative between two samples of the same population. A common diagnostic is the variance inflation factor (VIF); values above 5 (some analysts use 10) flag a driver whose weight you should not trust. Regression also assumes the relationships are linear and additive, which real customer perceptions frequently are not.

### Relative importance analysis

Relative importance analysis was designed specifically to divide credit fairly among correlated predictors, which is exactly the problem CX data creates. Techniques including Johnson's relative weights and Shapley value regression (also called the LMG method, or dominance analysis) estimate each driver's average contribution to the model's explained variance across every possible ordering of the predictors, rather than a single fragile coefficient. The output is easy to communicate: each driver's importance is expressed as a share of the variance the model explains, and the shares sum to 100%. For the correlated attribute batteries typical of survey-based CX programs, this is the most defensible of the three approaches — and increasingly the default in serious [NPS and CSAT measurement](/blog/how-to-measure-customer-experience-2026) work. At national scale, the [American Customer Satisfaction Index](https://www.theacsi.org) goes further still, using structural-equation cause-and-effect models to link drivers to satisfaction and loyalty — a useful reminder that even the most rigorous driver modeling is still modeling association within the attributes it chose to measure.

| Method | What it does | Strength | Weakness |
|---|---|---|---|
| Bivariate correlation | Ranks each driver against the outcome individually | Fast, intuitive, easy to explain | Ignores correlation between drivers; overstates importance |
| Multiple regression | Weights all drivers together, controlling for the rest | Isolates each driver's marginal association | Unstable under multicollinearity; assumes linear, additive effects |
| Relative importance (relative weights / Shapley) | Splits explained variance fairly across correlated drivers | Stable ranking; sums to 100% of R² | More complex; still only measures association, not cause |

Whichever method you pick, all three share one trait: they operate on a set of attributes you chose in advance and a score you already collected. That constraint matters more than the choice of algorithm — and it's the root of the limitation that follows.

## What driver analysis tells you — and what it can't

Driver analysis tells you where to look; it does not tell you what you will find when you get there. Used well, it converts a wall of undifferentiated survey attributes into a short, ranked shortlist — which is genuinely valuable when you have twenty candidate drivers and budget to fix two.

The highest-leverage way to use that shortlist is importance-performance analysis (IPA). Plot each driver's importance (from your model) against how well you currently perform on it (from the survey), and you get four quadrants: high-importance, low-performance drivers are your priorities; high-importance, high-performance ones are strengths to protect; low-importance drivers are distractions for now. This quadrant map is the most defensible artifact driver analysis produces, and it pairs naturally with the broader set of [CX metrics that actually matter](/blog/customer-experience-metrics-in-2026-the-8-that-matter-nps-csat-ces-clv-and-more) rather than fixating on one headline number.

It's also worth knowing that *derived* importance (what the statistics infer) and *stated* importance (what customers say matters when you ask them directly) routinely disagree — customers under-report hygiene factors they take for granted. Neither number is "true" alone; the gap between them is itself a signal worth investigating.

Here is what driver analysis structurally cannot deliver, no matter how clean the model:

- **Causation.** A correlation between "onboarding" and NPS does not establish that better onboarding causes higher NPS.
- **Mechanism.** Even if the effect is causal, the model can't tell you *what about* onboarding matters — the speed, the human contact, a specific missing integration.
- **Unmeasured drivers.** The model can only rank the attributes you put on the survey. The real driver of this quarter's churn may be a factor no one thought to ask about.
- **Segment differences.** A driver that dominates for enterprise accounts can be irrelevant for self-serve users; aggregate models average that away.

## The 'why' gap: correlation is not a reason

The central limitation of driver analysis is that a correlation is never, by itself, a reason. This is not a nitpick — it's the difference between a chart that says "fix onboarding" and a team that knows what to change on Monday. As the statisticians' warning goes, [correlation does not imply causation](https://en.wikipedia.org/wiki/Correlation_does_not_imply_causation), and CX data violates the assumption in several specific ways.

**The halo effect and reverse causality.** Overall satisfaction bleeds into how customers rate individual attributes. Someone who loves your product rates its support 9/10 partly *because* they love the product, not only because support was excellent. That halo inflates the correlation between nearly every attribute and the outcome, which is a large part of why bivariate methods overstate driver counts.

**Confounding and omitted variables.** Two drivers can both correlate with the outcome because a third, unmeasured factor moves all three. Attribute the effect to the visible driver and you can spend a quarter optimizing something that was never the lever.

**You can only rank what you measured.** No statistical sophistication escapes this. If the survey battery was written last year and the real reason accounts are leaving is a competitor's new feature, driver analysis will confidently rank last year's attributes and never mention the actual cause. The strongest driver is invisible because nobody asked about it.

**Aggregation hides the segments that matter.** A single model over your whole base produces a single ranking, but the "why" behind a detractor score is frequently segment-specific — and the [common CX mistakes](/blog/customer-experience-mistakes-2026) that hurt most tend to hide inside averages that look fine at the top line.

Each of these is a variant of the same problem: the score is a compressed number, and driver analysis works entirely inside that compression. It can point at a factor. It cannot reconstruct the human reasoning that the number flattened in the first place — which is exactly the reasoning a survey discards and a [voice-of-customer program](/blog/the-complete-guide-to-voice-of-customer-programs-in-2026) exists to recover.

## Combining driver analysis with conversational follow-up

The fix is not to abandon driver analysis — it's to pair it with a method that supplies the why it can only point at. Driver analysis is excellent at narrowing a wide field to the two or three factors worth investigating; a conversation is what actually investigates them. Used together, the quantitative model does the triage and the qualitative follow-up does the diagnosis.

This is the [listening half of customer experience](/blog/ai-for-customer-experience-the-listening-half-of-ai-cx) that most measurement programs skip, and it's where AI-moderated interviews change the economics. Historically, following up on a driver meant either another survey (more scores) or a handful of manual interviews (which don't scale). Perspective AI runs the interview at survey scale: after driver analysis flags a top driver, you route the relevant customers into a short conversation that asks *why* that factor shaped their score, follows up on vague answers, and probes for the specific incident behind the rating — the context a dropdown can never hold.

A defensible workflow looks like this:

1. **Run driver analysis** to identify the two or three drivers with the highest relative importance for your target metric.
2. **Route by response.** Send customers whose ratings load on a top driver into a brief AI-led interview that opens with an open question about that specific factor.
3. **Analyze the verbatims at scale.** Use [text analytics for customer feedback](/blog/text-analytics-for-customer-feedback-2026) to theme the conversations, so you get the causal patterns behind the driver, not just anecdotes.
4. **Feed the "why" back into the model.** The themes you surface become better candidate drivers for the next round — closing the loop that a static attribute battery leaves open.

That sequence is also how [customer sentiment analysis](/blog/customer-sentiment-analysis-in-2026-methods-tools-and-the-conversational-edge) earns its keep: sentiment scored from real conversations carries its reasoning attached, where sentiment scored from a survey field is just another number waiting for a driver analysis to guess at its cause.

## How to run a defensible driver analysis

A defensible driver analysis is less about the algorithm and more about the discipline around it. Follow these steps to produce a ranking you can actually stake a roadmap on.

**Step 1: Define the outcome metric precisely.** Decide up front whether you're modeling relationship NPS, transactional CSAT, or a behavioral outcome like renewal — the drivers differ by metric. If you're still choosing, start from a clear view of [how to measure customer experience](/blog/how-to-measure-customer-experience-2026) across relational, transactional, and qualitative layers.

**Step 2: Ground your candidate drivers in the journey, not the survey template.** The most common failure is inheriting last year's attribute list. Derive candidate drivers from recent open-ended feedback first, so the battery reflects what customers actually talk about — the single biggest lever on whether the analysis finds the real driver.

**Step 3: Check you have enough data.** A common rule of thumb is 10–20 observations per predictor for a stable regression; with fifteen drivers, that's a floor of roughly 150–300 clean responses, and more if you plan to segment.

**Step 4: Screen for multicollinearity.** Inspect the correlation matrix and VIF values before you trust any coefficients. If drivers are heavily intercorrelated, prefer a relative-importance method over raw regression betas.

**Step 5: Use relative-importance methods, and segment before you aggregate.** Report relative weights or Shapley values rather than bivariate correlations, and run the model separately for segments that plausibly differ (enterprise vs. self-serve, new vs. tenured) before you pool them.

**Step 6: Validate stability.** Bootstrap the importance estimates or hold out a sample. If a driver's rank swings wildly across resamples, treat it as a hypothesis, not a finding.

**Step 7: Close the why gap.** Take your top two or three drivers into conversations, as above. A driver analysis that ends at the chart is a to-do list of unanswered questions; a driver analysis that ends at a set of interviews is a plan.

Teams that build this into a repeatable rhythm rather than a one-off study climb the [customer experience maturity model](/blog/customer-experience-maturity-model-2026) faster, because they stop relitigating what the drivers are and start acting on why they move. It's a core habit in any serious [customer experience strategy](/blog/how-to-build-a-customer-experience-strategy) and a recurring theme in the [2026 CX improvement playbook](/blog/how-to-improve-customer-experience-2026-playbook).

## Frequently Asked Questions

### What is key driver analysis in customer experience?

Key driver analysis is a statistical technique that ranks which attributes — support, onboarding, pricing, reliability — correlate most strongly with a customer experience outcome metric like NPS or CSAT. It answers "which factors move together with our score, and by how much." It's a prioritization tool: it converts a long list of survey attributes into a short shortlist of things that appear to matter most, so teams can focus improvement effort rather than guessing.

### What statistical methods are used for CX driver analysis?

The three common methods are bivariate correlation, multiple regression, and relative importance analysis. Bivariate correlation ranks each driver against the outcome one at a time but overstates importance because drivers are intercorrelated. Multiple regression weights all drivers together but grows unstable under multicollinearity. Relative importance methods — Johnson's relative weights and Shapley value regression — split explained variance fairly across correlated drivers and are the most defensible choice for typical CX survey data.

### Does driver analysis prove what causes customer satisfaction?

No. Driver analysis identifies correlation, not causation, so it cannot prove that any factor causes a change in your metric. Halo effects, confounding variables, reverse causality, and drivers you never measured all break the link between correlation and cause. Driver analysis is best treated as a way to generate strong hypotheses about what matters, which you then confirm through follow-up conversations, controlled changes, or experiments.

### Why does driver analysis tell you which factors matter but not why?

Driver analysis operates entirely on compressed data — a numeric score and a fixed set of attribute ratings — so it can detect statistical association but has no access to the customer's reasoning. Knowing that "onboarding" ranks as a top driver doesn't reveal what about onboarding mattered: the speed, a missing integration, or a confusing first session. Recovering that "why" requires asking the customer directly, which is what conversational follow-up adds on top of the model.

### How do you combine driver analysis with customer interviews?

Use driver analysis to triage and interviews to diagnose. First, run the model to find the two or three highest-importance drivers. Then route the relevant customers into short, AI-moderated interviews that ask why that specific factor shaped their rating and probe vague answers. Finally, theme the conversation verbatims and feed the causal patterns back as sharper candidate drivers for the next cycle, closing the loop a static survey leaves open.

## Conclusion

Driver analysis remains one of the most useful techniques in CX measurement — but only when you respect the boundary of what it can do. It tells you *which* factors correlate with your metric, ranked and prioritized; it never tells you *why*, and treating a ranked correlation as an explanation is how teams end up optimizing the wrong lever with total confidence. The methods matter (prefer relative-importance approaches over raw correlation), the discipline matters more (segment, validate, ground your drivers in real feedback), but the decisive move is pairing the quantitative model with the qualitative follow-up that supplies the missing reason.

The shift underneath the numbers is simple: the score points at a driver, and a conversation explains it. Perspective AI runs those conversations at the scale your survey already reaches, turning the top drivers your analysis flags into interviews that surface the causal why. If you want your next driver analysis to end in a plan instead of a chart, [start with a conversation-first study](/research/new), see [how research teams are already doing it](/studies), or explore what it looks like [built for CX teams](/roles/cx-teams).