Digital Crystal Ball?

Voters as Consumers

Modern political campaigns treat candidates as “Brands” and voters as a “Market.” To win, candidates need Electoral Research.

With 143 million internet users in Indonesia—74% of whom use search engines—Google Trends captures the “Curiosity” of nearly 40% of the population. But curiosity doesn’t always equal a vote.

The Statistical Litmus Test

To compare Google’s “Search Interest” (Small Data) with the KPU’s “Real Count” (Total Population), we used the Predictive Accuracy (A) measure by Martin, Traugott, and Kennedy.

• Accuracy: Did Google predict the correct winner in a province?
• Precision: How close was the search ratio to the actual vote ratio?

We analyzed “Topic Queries” (Broad Match) rather than “Search Terms” to capture the full spectrum of intent for both candidates across all 34 provinces.

The Geographical Mismatch

At first glance, the maps look similar. But a deeper look reveals a contradiction.

Google Trends showed Candidate 02 (Prabowo) dominating search interest in 20 provinces. However, the Real Count showed Candidate 01 (Jokowi) winning in 21 provinces. Search volume was leaning one way, while the actual votes were leaning the other.

A High Rate of Error

The statistical value ‘A’ should ideally be 0 (Perfect match).

In reality, Google Trends only correctly predicted the winner in 13 out of 34 provinces. This means the search-based model was wrong more than 60% of the time when applied to provincial outcomes.

The Consistency Gap

Precision measures reliability. Our analysis found that the ‘A’ values varied wildly—from -0.029 (high precision in Jakarta) to -0.955 (extremely low precision in Bali).

Because the deviation was so inconsistent across the archipelago, Google Trends proved to be an unreliable tool for forecasting a national election in its current form.

Curiosity is not Support

Why did the data fail? Google Trends measures Volume, not Sentiment.

A voter might search for a candidate because they support them, or because they are looking for a scandal to criticize them. Without a way to distinguish between “Love-searching” and “Hate-searching,” search volume remains a blunt instrument.

Potential for Evolution

While it failed as a standalone predictor in 2019, Google Trends is not useless. It remains a powerful tool for measuring Issue Salience—what the public is thinking about right now.

To become a true predictive tool, the platform needs:

1. Sentiment Analysis integration.
2. User-Level Weighting to prevent “Power Users” from skewing provincial data.
3. Hybrid Modeling combining search data with traditional survey metrics.