Machine Learning · Employee Voice · Employer Brand
This project analyzes 604,320 public workplace reviews to examine what predicts whether employees recommend their employer, how much information is actually contained in familiar workplace ratings, what employees’ written comments add to those scores, and why negative employee voice appears to relate differently to recommendation across employers.
Glassdoor Employee Reviews: What They Reveal About Employee Voice and Employer Brand
The original machine-learning project has now been translated into a practitioner-facing research guide examining the broader organizational implications of the findings: why several workplace ratings can tell much of the same underlying story, why narrative employee voice adds context that scores cannot, what the results mean for eNPS and employee listening, and why firm-level differences raise a new question about employer-brand resilience.
Read the full guide →The research question
Employee-review platforms create an unusual form of organizational data because they combine structured ratings with unsolicited narrative employee voice. The project began with a relatively direct machine-learning question: how accurately can an employee’s recommendation of an employer be predicted from the information contained in a public workplace review?
The modeling results answered that question, but they also exposed a more interesting measurement problem. If ratings for senior management, culture, career opportunities, compensation and work-life balance are strongly interrelated, how much should managers treat movement in each score as an independent diagnosis? And if open-ended employee language adds information beyond those ratings, what is lost when listening systems reduce workplace experience to a dashboard?
What the analysis found
Ratings of management, culture, career opportunity, compensation and work-life balance were strongly interrelated, suggesting that they partly express a broader evaluation of the employer.
Replacing the five structured ratings with a single latent workplace-evaluation factor reduced ROC-AUC only from .929 to .928 — virtually no loss in predictive performance.
Text describing what employees disliked about work was more discriminating than positive commentary in distinguishing whether reviewers recommended their employer.
Firm-level models showed meaningful variation in how strongly the tone of negative employee commentary was associated with recommendation, suggesting that organizational context matters.
From employee voice to employer-brand resilience
The firm-level analysis extended the project beyond prediction. Across the employers examined, the relationship between negative employee commentary and recommendation varied substantially. Some firms appeared relatively less sensitive to changes in negative employee voice, while others appeared more sensitive.
Salesforce and McDonald’s provide useful examples from opposite ends of that broader historical continuum: negative employee commentary had a weaker relationship with recommendation at Salesforce and a stronger relationship at McDonald’s. The comparison is not a ranking of either employer and should not be interpreted as a statement about their current employee experience.
Instead, the result raises a more interesting question: why does negative employee voice appear to carry different implications for advocacy depending on the employer?
One possible explanation is employer-brand resilience: the idea that accumulated trust, familiarity, identification or reputational strength may give some employers greater capacity to absorb negative employee experiences without seeing the same relationship with advocacy. Employer-brand resilience was not directly measured in this project and remains a hypothesis for future research.
What this means for employee listening
The findings do not imply that organizations should stop measuring employee experience. They suggest that leaders should be careful about mistaking the measurement system for the phenomenon being measured. A recommendation score can efficiently summarize broad sentiment, and structured ratings can help organize workplace feedback, but neither necessarily explains why employees feel the way they do.
That has particular implications for employee Net Promoter Score (eNPS). A recommendation question can function as a useful pulse of generalized employee sentiment, but the project reinforces the case for interpreting that signal alongside richer survey items, operational data and narrative employee voice.
60-second explainer
5 ratings. 1 bigger story.
This short explainer summarizes the central findings: five workplace ratings share a powerful underlying signal, negative employee voice adds information the ratings miss, and the relationship between negative voice and recommendation varies across employers.
Methods
The project used a historical public Glassdoor dataset and retained 604,320 reviews with clearly positive or negative recommendation outcomes. An 80/20 stratified train-test split was used for predictive model evaluation.
Limitations
Public workplace reviews are self-selected and should not be treated as a representative sample of employees. The data are historical, so firm-level results are not current employer assessments. The structured ratings and recommendation outcome were also collected within the same review event, making common evaluative tone and common-method effects important interpretive considerations.
The analysis is predictive and associative rather than causal. Sentiment and TF-IDF models simplify language, and the exploratory multilevel model generated a convergence warning. The employer-level results therefore motivate further research rather than establishing a completed theory of employer-brand resilience.
Research note: This project was completed in 2026 in connection with graduate coursework in machine learning and predictive analytics at New Mexico State University. It has not been peer reviewed. Glassdoor is a trademark of Glassdoor LLC. This independent project is not affiliated with, sponsored by, or endorsed by Glassdoor.