Key Driver Analysis From Text, Not Just Survey Scores
Key driver analysis usually runs on survey scores. Here is how to run it on tickets and reviews, and how correlation, regression, Shapley and causal methods rank the same six drivers.
Key driver analysis finds which parts of the customer experience move an outcome such as satisfaction, NPS or retention, usually by regressing an overall survey score on customers' ratings of individual attributes. You can run the same analysis on support tickets, chats and reviews by first turning each record's text into fields (did it mention a billing error, was the agent rude, was the issue fixed) and then using those fields as the drivers.
Below: how the survey version works, how to build drivers from text, and a worked example on 9,186 gym reviews where five common methods produce five different rankings of the same six drivers.
What key driver analysis is
MeasuringU describes key driver analysis as a way to identify which features or aspects of an experience have the biggest impact on an outcome variable, such as likelihood to recommend. The classic setup is a survey with one overall question ("How satisfied are you overall?") and a battery of attribute questions ("How satisfied are you with the speed of service?"). The analyst regresses the overall score on the attributes, and the attributes with the largest weights become the drivers.
Two problems show up quickly. Attribute ratings tend to move together, so their regression weights are unstable, a problem known as multicollinearity. A common answer is Shapley value regression, which borrows a fair-division rule from game theory to split the explained variance among correlated drivers. The second problem is coverage. A survey can only find drivers it asked about, and it only hears from the people who answer it. In our study of about 9,000 support conversations at a consumer-goods company, satisfaction read from every conversation averaged 2.91 out of 5, against 3.62 among the customers who returned a survey.
Finding drivers in tickets and reviews
Text has the coverage a survey lacks: every customer who wrote in, in their own words, about whatever went wrong. It has no columns, so the first job is to make some.
Write the fields as questions you would ask of every record. "Does this review describe a billing or cancellation problem?" "Does it describe rude or absent staff?" Each answer is a typed field: yes or no, a fixed list of values, or a number. We call these fields Dimensions.
Let one record carry several drivers. A review can complain about the front desk and a surprise charge at once. In our gym data, 57 reviews named two or more different failures.
Extract every record. Sampling brings back the coverage problem.
Check the labels before trusting the drivers. A field that is wrong a third of the time produces a driver that is wrong too. How to measure that is in how to tell if your AI tagging is accurate.
Join an outcome. The star rating on a review, the CSAT on a ticket, a renewal or a refund from your own systems. Keep context fields too (location, channel, period, length), because they become your controls.
Then run the driver analysis. The fields behave like any other column.
A worked example: six drivers, five rankings
9,186 public Google reviews of 33 Planet Fitness clubs in metro Chicago, January 2025 to July 2026. In this analysis the six failure fields were extracted from the review text with patterns checked by hand on samples. The outcome is whether the review is one or two stars, which 7.5% are.
Dimension Labs analysis
Broken equipment ranks first by mentions and near the bottom once the drivers are compared fairly
Rank of each of six gym drivers under five methods, 1 is the strongest and "2=" a tie. Under each rank: mention count, correlation r, regression and causal effect in points, Shapley share.
| Driver | Count | r | Regr. | Shapley | Causal |
|---|---|---|---|---|---|
| Broken equipment | 1122 | 40.20 | 6+20 | 56% | 6+4.7 |
| Front desk, staff | 2=107 | 10.31 | 1+53 | 134% | 1+36.3 |
| Billing | 2=107 | 20.30 | 3+46 | 226% | 3+25.4 |
| Cleanliness | 2=107 | 30.27 | 4+41 | 322% | 4+25.1 |
| Safety, theft | 541 | 50.19 | 2+48 | 411% | 2+25.6 |
| Crowding | 633 | 60.09 | 5+22 | 62% | 5+6.7 |
Read across the broken-equipment row. It is the most-mentioned driver. A plain correlation puts it fourth. Once the other drivers and the controls are in the model, it falls to last on the regression, second to last on the Shapley share, and last on the causal estimate, at 4.7 points. Front desk and staff problems sit at the top of every column except the count.
Dimension Labs analysis
The most-mentioned gym driver has the smallest causal effect
The same six drivers, ranked two ways. Left: reviews that name each one. Right: its effect on the chance of a one- or two-star review.
The Shapley column answers a different question from the causal column. It measures how much of the variation in ratings across all 9,186 reviews each driver accounts for, so it rewards drivers that are both strong and common. Safety problems have a large effect per review (25.6 points) but appear in only 41 reviews, so their Shapley share is 11%. Both views are useful. The causal effect tells you how bad it is when it happens; the share tells you how much of the total problem it is.
The regression and causal numbers also differ in size. The regression here is a straight-line model of a yes-or-no outcome; the causal estimate comes from a logistic model averaged over every review (g-computation), with a bootstrap range and refutation tests. With rare events and a 7.5% base rate the two can differ a lot in size while agreeing on order.
From the report: Chicago gym report
The methods side by side
Methods compared
Key driver methods, from counting to cause
What each method answers, and what it can and cannot separate.
| Method | Question it answers | What it separates |
|---|---|---|
| Mention countUse it for: Workload and monitoring | How often does it come up? | Overlapping driversNot handledConfoundersNot handled |
| CorrelationUse it for: A first list of candidates | Does it travel with the outcome? | Overlapping driversNot handledConfoundersNot handled |
| Multiple regressionUse it for: Standard survey key driver analysis | What is its effect with the other drivers and controls in the model? | Overlapping driversPartly; weights get unstable when drivers are highly correlatedConfoundersOnly the controls you include |
| Shapley value or relative importanceUse it for: Share of the problem, when drivers overlap | How much of the variation in the outcome does it account for? | Overlapping driversOnly if you include them; the gym estimates here did not; shared variance is split by a fair-division ruleConfoundersOnly the controls you include |
| Causal estimateUse it for: Deciding what to fix and what the fix is worth | How much would the outcome change if this stopped happening? | Overlapping driversHandledConfoundersThe ones you measure, and the result is stress-tested |
For a survey with ten attribute ratings and one overall score, regression with Shapley weights is a reasonable default. For drivers built from text, two things change. The fields are often rare and yes-or-no, so the regression is fitting a handful of positives, and the records differ in ways a survey's fixed format avoids: review length, location, channel, and how angry the writer already was. That is the case for the causal estimate. Our post on correlation vs causation in CX data shows the length effect in the same gym reviews, and causal inference covers the estimators.
Practical notes
Report ranges, not just ranks. The staff and safety intervals in the gym data overlap with billing and cleanliness, so "first" is a soft claim.
Keep rare drivers, flag them. Crowding appeared in 33 reviews and its range starts below zero. It stays on the list as unproven.
Run it per segment. The staff effect was 24 points at inner-suburban clubs and 43 at outer-suburban ones.
Use more than one outcome. The drivers of a bad rating and of a stated intent to cancel were different lists in this data (why gym members say they will cancel).
FAQ
Can I do key driver analysis without a survey?
Yes. Any outcome attached to a record works: a review's star rating, a ticket's CSAT, a renewal or cancellation from your billing system. The drivers come from the text once it has been turned into fields.
Is Shapley value regression causal?
No. It is a fair way to split explained variance among correlated drivers. It does not separate a driver from something that travels with it, unless the controls in the model already do.
How many records do I need?
Enough that each driver appears a few dozen times. In the gym data the rarest driver tested appeared 33 times, and its range was too wide to call.
Dimension Labs reads every ticket, chat, call, email, review and survey response, builds the business-specific Dimensions for each record, joins them to CRM, billing and usage data, and shows which ones drive ratings, retention and cost, with confidence intervals on the Prove plan and up. Plans start at $299 a month with no sales call. See customer intelligence, or start with one file.