Does Clearer Writing Win Money Back? A Study of 94,335 Complaints

When a bank charges you a fee you never agreed to, or a credit card company refuses a refund, your next move is often to file a complaint with a government agency. That complaint is just text: a few hundred words describing what went wrong. A new study asks a deceptively simple question about those words. Does the way a consumer writes a complaint matter for whether the company pays up? The answer, drawn from nearly one hundred thousand real disputes, is a cautious and fascinating maybe.

The research, published in Frontiers in Behavioral Economics, analyzed 94,335 public complaints filed with the Consumer Financial Protection Bureau between 2023 and 2024. The complaints covered checking and savings accounts, credit cards, and prepaid cards. Fangda Cui, an independent researcher based in Toronto, wanted to know whether two qualities of writing, clarity and coherence, were associated with the outcome that matters most to consumers: monetary relief, meaning the company reported closing the complaint with a refund, fee reversal, or credit. In the sample, 15.7 percent of complaints ended in monetary relief, 14.7 percent in non-monetary relief, and a striking 69.6 percent got nothing more than an explanation.

Measuring something as subjective as clarity across tens of thousands of texts is a serious technical challenge, and the study tackled it with a two-stage pipeline. First, two human raters, working independently and without seeing how companies had responded, scored 1,500 complaints on seven-point scales for clarity and coherence. Clarity asked whether the intended meaning was easy to grasp; coherence asked whether the narrative hung together as a logically connected account. The raters agreed remarkably well, with exact agreement above 78 percent and intraclass correlations exceeding 0.93 for both dimensions.

Those human judgments then became training data. A machine-learning model, a Huber-loss gradient-boosting regressor, learned to predict the average human rating from 19 structural features of each text: word and sentence counts, Flesch reading ease, sentence-length variation, paragraphing, connective density, first-person pronoun usage, punctuation patterns, and more. Crucially, the model never saw the company responses or relief outcomes, so it could not accidentally learn to predict the answer. On 300 held-out complaints, the predicted scores correlated with human ratings at 0.928 for clarity and 0.923 for coherence, far above benchmarks built from readability formulas alone, which topped out around 0.71.

With presentation scores in hand, the study turned to the outcome models. Using linear probability models packed with controls, including the product, issue, company, state, and month of each complaint, plus 50 semantic components capturing what the text was about, six factual indicators such as whether a dollar amount was mentioned, and the median income of the complainant’s neighborhood, the analysis found that a one-point increase in predicted clarity was associated with a 0.394 percentage-point rise in the probability of monetary relief. For coherence, the estimate was 0.437 percentage points. Scaled to the spread of scores actually observed, moving from a typical low-clarity to a typical high-clarity complaint corresponded to roughly 0.43 to 0.48 percentage points, or about 4 to 5 more monetary-relief outcomes per 1,000 complaints.

Those numbers are small in absolute terms, and the study is refreshingly honest about that. The presentation scores added less than 0.0001 to the model’s explanatory power beyond all the controls. More troubling, when the analysis swapped the standard length adjustment for flexible splines over word count, sentence count, and mean sentence length, alongside finer complaint categories, the estimates dropped to essentially zero, with confidence intervals straddling no effect. Because the clarity and coherence scores are themselves built partly from length-related features, that richer adjustment changes what the contrast means. The headline associations, in other words, are fragile, and the author says so plainly.

There were genuinely intriguing patterns buried in the sensitivity analyses. Complaints that mentioned a specific dollar amount showed a notably stronger association between presentation and relief, with pooled interaction estimates above one percentage point, hinting at a possible complementarity: a clear narrative may help a complaint handler interpret concrete financial details, turning raw information into an actionable case. Meanwhile, a small population of complaints with identical repeated text, including one group of 1,163 complaints sharing the same 27-word message, had a monetary-relief rate of just 0.11 percent, compared with 17.33 percent for unique narratives. Templated mass complaints, it seems, do not fare well.

The study also probes a question of fairness. If well-written complaints win more relief, does that advantage differ by neighborhood income? The answer, at least in this data, was no clear evidence of moderation: neither the clarity-by-income nor the coherence-by-income interaction was distinguishable from zero, though the confidence intervals were wide enough to leave room for differences that matter. The author is careful to note that ZIP-code-level median income describes neighborhoods, not individuals, and that absence of evidence here is not evidence of equal treatment.

Why would presentation matter at all? The theoretical scaffolding comes from behavioral economics. Processing fluency research suggests that information that is easier to read is judged more favorably, often without conscious awareness. Complaint handlers at financial firms face heavy caseloads and limited attention, conditions under which people lean on peripheral cues like the quality of presentation. A clear, coherent account might simply be easier to evaluate, or it might make the complainant seem more credible. But the data cannot distinguish these mechanisms, and the study is explicit that handler effort, perceived credibility, and judgments of claim merit were never measured.

The deeper limitation is confounding by substance. A complaint that reads clearly may also be a complaint with better documentation, a more meritorious claim, or a more persistent consumer behind it. The study’s audit of its own factual indicators revealed a weak spot: the automated rule meant to detect transaction-specific details had specificity of only about 26 percent, meaning it frequently flagged generic financial language as specific. That leaves substantial room for factual differences to masquerade as presentation effects. The outcome itself is company-reported, not an independent judgment of whether the harm was real or the compensation adequate.

What the study does offer is a concrete policy idea worth testing. Complaint systems could offer structured prompts, asking separately about the disputed event, the timing, the amount, prior contacts, supporting records, and the desired remedy, while keeping a free-text field for anything else. Administrative-burden research shows that the cognitive and psychological costs of dealing with institutions fall unevenly across citizens, and the format of a request can shape who succeeds. But the author stresses that this observational design cannot establish that improving presentation would change outcomes, and that any such intervention should be evaluated experimentally, including its effects on participation and access.

In an era when large language models can polish anyone’s prose, the question of whether writing quality influences institutional outcomes has taken on new urgency. This study stops short of answering whether AI-assisted complaints fare better, and it explicitly does not estimate the effect of LLM use. What it delivers instead is a rigorous, unusually transparent measurement exercise: human-calibrated scores of how nearly 100,000 Americans actually write their grievances, and an honest accounting of how much, or how little, that presentation seems to matter once you control for everything else. The modest, fragile positive associations it finds are less a verdict than an invitation, a well-documented reason to run the experiments that could finally tell us whether the pen, or at least the well-organized complaint, is mightier than the fee.

Subject of Research: The association between the clarity and coherence of consumer complaint narratives and company-reported monetary relief in CFPB complaints

Article Title: Predicted clarity, coherence, and monetary relief in 94,335 Consumer Financial Protection Bureau complaints

Article References: Cui, F. (2026). Predicted clarity, coherence, and monetary relief in 94,335 Consumer Financial Protection Bureau complaints. Frontiers in Behavioral Economics, 5, Article 1981505. https://doi.org/10.3389/frbhe.2026.1981505

Image Credits: AI Generated

DOI: 10.3389/frbhe.2026.1981505

Keywords: consumer complaints, CFPB, behavioral economics, administrative burden, text clarity, processing fluency, monetary relief, machine learning, natural language processing, consumer finance, public policy, observational study