Historical client work | Product-focused text analytics, NLP & voice of the customer
The product question behind the customer language
On Beat Digital’s work for Nike used text analysis to investigate products, consumer needs, purchasing decisions, and emerging demand. The focus went beyond measuring whether a conversation sounded positive or negative: it connected the product being discussed to the issue, the surrounding language, and the source comment.
Chat and survey feedback provided a way to investigate what people wanted, what was going wrong, and which signals deserved a closer look. This historical work predates today’s mainstream generative AI tools. Its value came from the analytical design, product context, and traceable evidence behind a finding.
From product term to a specific issue
The cleat example shows two complementary views: related tokens and dependency words. Related terms establish the context around the product, while dependency words help identify actions and problems associated with it. In the displayed output, that language included returns, breakage, products falling apart, and customization.
Those views were connected to sentence-level evidence. An analyst could inspect the original wording, its sentiment score, and the field from which it came. Each visual pattern remained connected to the original language and the context needed to investigate it.
Cleat durability and customization were different problems
The source comments distinguish a physical product issue from an ordering experience. A faulty cleat and a pair falling apart after one season point toward durability and returns. Another comment describes a customized cleat order failing while the order succeeded without customization. Treating all three as one negative product category would hide the difference.
| Signal | Evidence in the historical output | Score shown | Source field |
|---|---|---|---|
| Faulty cleat / return | “return faulty cleat” | -2.00 | reason_for_contacting_nike |
| Durability after use | “a problem with my daughter’s cleats falling apart after one season of use.” | -2.00 | reason_for_contacting_nike |
| Customized ordering failure | The comment describes an unsuccessful customized cleat order that went through without customization. | -1.35 | overall_comments |
The sentiment scores are reproduced from the historical output. The important diagnostic chain is product → related language → failure mode → source sentence. The resulting questions differ: investigate a possible durability pattern, review returns, or examine the customization ordering journey.
Product-focused text analysis and visual exploration
- Prepare and tokenize customer comments while preserving product and source context.
- Examine related terms, dependency words, and recurring phrases.
- Score and review sentiment at the sentence level.
- Use coxcomb and radial visualizations to explore product-language relationships.
- Return to the underlying comments to assess what a pattern actually means.
The visual analysis supported a close reading of the product evidence. Counts and scores helped surface a pattern; the original sentences helped explain it. Product teams and analysts could investigate an issue without confusing a purchase-process complaint with a defect in the physical item.
Purchasing drivers and product familiarity
The historical Nike presentation also examined parental purchasing behavior. Comfort, fit, durability, and price were recurring considerations, and product familiarity mattered in family purchases. The work considered the relationship between consumer conversations, social activity, athlete associations, and expressed purchasing needs.
Those findings belonged to the analyzed population and period. Their broader value is the analytical capability: identify the audience, distinguish what people discuss from what they intend to buy, and investigate the product attributes that matter to their decisions.
Emerging trends and product demand
The work investigated retro-product requests and emerging interest in non-traditional sports such as dance. On Beat’s historical product-trend analysis also included identifying leggings as an emerging category through coxcomb visualizations of consumer language.
A changing pattern in text can create an early product hypothesis. The next step is to examine its frequency, context, audience, and supporting evidence. The strength of the approach was connecting an early signal to the product, audience, and source language behind it.
Sentiment methods were tested, not assumed
Related Nike project work compared switch-polarity and shift-polarity methods on survey comments containing negation words such as “not,” “no,” and “never.” The evaluation concerned a deliberately difficult subset of comments, with further score audits planned against consumer chat data.
That methodological work is part of the capability: test how language changes a score, inspect errors, and define the population on which an evaluation was performed. A historical experiment on selected negation comments should not be presented as a current benchmark for all text analysis.
What this work demonstrates
Nike’s case illustrates product-level issue identification, dependency and token analysis, contextual sentiment, audience and purchasing-driver analysis, and emerging-trend investigation. It combines visual exploration, product context, and source-level evidence to support a specific investigation.
These capabilities sit alongside On Beat Digital’s broader customer analytics practice: CSAT key driver analysis, NPS loyalty matrices, neural networks, principal components analysis, statistical comparisons, and logistic regression. Those services have their own dedicated pages so their methods and applications can be assessed in detail.