Why Sentiment Analysis Isn't Enough for Urban Planning

Sentiment analysis has become a buzzword in smart city circles. But for urban planning, knowing that residents feel “negative” about transport is about as useful as a doctor knowing a patient feels “unwell.” Without magnitude, threshold, and specificity, sentiment alone cannot drive action.

The Problem with Binary Sentiment

Traditional sentiment analysis classifies text into positive, negative, or neutral categories. For brand monitoring or political polling, this may suffice. For urban planning, it creates three fundamental problems.

First, it tells you direction but not magnitude. Consider three statements about traffic: “the traffic is unbearable,” “the traffic is frustrating,” and “the traffic could be better.” All three register as negative. But they represent vastly different levels of urgency. The first demands immediate intervention. The second warrants investigation. The third suggests monitoring a trend. Binary sentiment treats them identically.

Second, it provides no actionable threshold. When does negative sentiment become serious enough to warrant spending public money? Without a calibrated scale, every piece of negative feedback carries equal weight, which means none of it carries enough weight to trigger a decision.

Third, it cannot be compared across time or geography. A “negative” rating in one district cannot meaningfully be compared with a “negative” rating in another, because the underlying intensity may differ completely.

What Urban Planning Actually Needs

Effective urban decision-making requires three things from sentiment data that binary analysis cannot provide:

• Granularity to detect emerging issues before they become crises, with comparable scores across topics, locations, and time periods
• Clear intervention thresholds tied to action, so that planners know when to celebrate, when to monitor, and when to act
• Alignment with established measurement frameworks so that AI-derived insights can sit alongside traditional planning evidence

The shift required is from sentiment to satisfaction. Sentiment tells you direction. Satisfaction tells you magnitude and urgency.

Adapting Net Promoter Score for Urban Contexts

The Net Promoter Score was introduced by Fred Reichheld in Harvard Business Review in 2003 and has since been validated across more than 400 companies by Bain & Company. Its core insight is that a 0–10 scale, segmented into Detractors (0–6), Passives (7–8), and Promoters (9–10), correlates strongly with real-world outcomes like customer retention and revenue growth.

Adapting this for urban planning required several modifications. The NPS thresholds were recalibrated through validation against survey data across 50+ GCC projects. The binary NPS segments were expanded into an 8-tier scale that maps sentiment intensity to specific intervention responses. And the scoring was calibrated against Parasuraman et al.'s (1988) SERVQUAL methodology for service quality measurement, grounding it in established social science.

The 8-tier intervention scale

The critical distinction: the same input — say, a negative comment about traffic — now produces different outputs depending on intensity. A score of 1.8 triggers immediate intervention. A score of 4.2 prompts root cause investigation. A score of 5.6 calls for trend monitoring. Same topic, same direction, fundamentally different responses.

Validation: Does It Actually Work?

Any AI-derived methodology needs rigorous validation. The satisfaction scoring approach has been tested through multiple channels.

Addressing Known Limitations

No methodology is without limitations, and transparency about these is essential for credibility.

Representation bias: Social media users skew younger, more urban, and higher income. This is mitigated by weighting against known demographic distributions and triangulating with review data, which draws from a broader user base.

Platform bias: Different social platforms attract different demographics and discourse styles. Multi-platform ingestion and platform-specific sentiment calibration address this.

Linguistic and cultural context: Sarcasm, dialect, and cultural expression vary interpretation. Arabic-native models trained on GCC social posts, combined with human-in-the-loop validation samples, mitigate this risk.

Ecological validity: What people say online may not match what they do in practice. Cross-referencing with behavioural data — mobility patterns, activity data — via RAG validation helps bridge this gap.

The Arabic Language Challenge

For practitioners working in the GCC, language is not a minor technical detail. Generic global AI models miss critical nuance in Arabic expression because Gulf dialects are fundamentally different from Modern Standard Arabic. The same concept is expressed with entirely different words across Egyptian, Gulf, Levantine, and Maghrebi dialects.

Beyond vocabulary, cultural context shapes expression. Religious references, social norms, and regional idioms all affect how satisfaction and dissatisfaction are communicated. A model that cannot parse Khaleeji sarcasm or recognise Najdi idiom will systematically misclassify 30–40% of sentiment in GCC social data.

This is why Arabic-native models, trained specifically on millions of GCC social posts and calibrated for cultural and religious context, are not optional but essential for any credible urban sentiment analysis in the region.

From Methodology to Practice

The satisfaction scoring framework described here is not theoretical. It has been applied at scale to measure pilgrim satisfaction during Hajj — 1.67 million pilgrims across 171 countries — and to inform architectural housing guidelines for the city of Al Ain using 2.7 million geotagged data points. Both case studies are covered in companion articles in this series.

For planners and policymakers considering these approaches, the key takeaway is this: sentiment analysis is a starting point, not a destination. The value lies in transforming unstructured human expression into calibrated, spatial, actionable satisfaction intelligence — intelligence that tells you not just that people are unhappy, but how unhappy, where, about what, and what to do about it.