Perception scoring methodology

Sentiment is normalized 0–1 with weighted drivers.

Last updated: 2026-05-05

Perception scoring methodology

Every attribute, driver, and overall score in Brand Perception is a number between 0 and 1, where 0 is uniformly negative and 1 is uniformly positive. This page explains exactly how those numbers are produced so you can defend them in stakeholder reports.

The pipeline

Source selection. Pranas pulls the most recent N AI Scan responses for the brand within the active scope (date range, country, models). Only Unaided + Problem-Solving + Feature prompts feed perception by default — Aided prompts are excluded because the prompt itself biases the answer.
Attribute extraction. An LLM reads each response and extracts descriptive phrases attached to your brand. Each phrase is normalised (lemmatised, lowercased) and clustered into canonical attributes (e.g., "very fast", "blazing speed", "low latency" → speed).
Polarity tagging. Each occurrence is tagged positive (+1), neutral (0), or negative (−1) based on the surrounding sentence, not just the adjective.
Per-attribute score. For each attribute:
```
score = (positives − negatives) / total_mentions × 0.5 + 0.5
```
This rescales the [-1, +1] polarity range to [0, 1]. An attribute with all positives scores 1.0; all negatives scores 0.0; balanced scores 0.5.
Visibility weighting. Attributes attached to brands with very low overall visibility are down-weighted so a single mention can't crown someone "the affordability leader". Narrative Drivers applies the related but distinct correction of multiplying each category score by that brand's coverage in the category.
Composite score. The overall sentiment score is the mention-count-weighted mean of all attribute scores, so frequently-mentioned attributes count more.

What counts as a "mention"

A mention is an attribute–brand pair within a single AI response. Repeated mentions of the same attribute within one response count once (per-response dedupe) to prevent verbose models from skewing the score.

Why scores can move without new content shipping

Model updates. When a model provider re-trains or fine-tunes, perception can shift overnight.
Competitor moves. Your absolute attributes don't change, but the relative ranking does.
Prompt-set changes. Adding new Problem-Solving prompts changes which response corpus the analysis sees.
Date-range changes. Looking at the last 7 vs last 30 days uses different scan data.

Always note the active filters when sharing a perception score.

Score interpretation reference

Score band	Polarity mix behind it	How to read it
1.00	All mentions positive	Uncontested strength
0.75	Mostly positive with some neutral or negative	Solid, worth reinforcing
0.50	Balanced positive and negative, or all neutral	No clear narrative yet
0.25	Mostly negative	Active perception problem
0.00	All mentions negative	Priority remediation

Scores below 5 total mentions are flagged low-confidence and should be treated as directional.

Confidence indicators

Each driver and attribute carries a sample size. Pranas flags any score computed from fewer than 5 mentions as low-confidence. Treat these as directional, not definitive.

Brand Perception overview
Gap analysis
Narrative Drivers
Dashboard filters