The Price-Credibility U-Curve: How Service Pricing Magnitude Affects Large Language Model Sentiment Assessment

1. Introduction

Pricing is among the most psychologically complex signals in service evaluation. In human decision-making, the relationship between price and perceived quality is well-documented as non-linear, with both excessively low and excessively high prices triggering different forms of skepticism. Whether large language models, trained on vast corpora of human-generated evaluative text, reproduce similar non-linear pricing heuristics is an open empirical question.

This study directly tests the relationship between service pricing magnitude and LLM sentiment assessment through a controlled single-variable experiment. By presenting 18 price points spanning three orders of magnitude ($500 to $500,000 AUD) while holding all other variables constant, we isolate the independent effect of pricing on AI-generated evaluative sentiment.

The results reveal a pattern of remarkable complexity -- 9 threshold transitions compared to just 2 for the ratings probe (SIGI-2026-001) -- suggesting that pricing is a fundamentally more volatile signal for LLM evaluation than star ratings. This volatility has immediate practical implications for service providers whose pricing information is available to AI systems generating recommendations.

2. Methodology

2.1 Probe Design

The pricing probe (PQ01) consisted of 18 prompt variations (PQ01_v00 through PQ01_v17) presented to a single LLM system. Each prompt asked the model to assess quality expectations for a brand design agency at a specified price point. The independent variable was the price (ranging from $500 to $500,000 AUD), while the service description remained constant across all variations as a "brand identity project."

2.2 Price Points Tested

The 18 price points were selected to provide coverage across three orders of magnitude: $500, $1,000, $1,500, $2,000, $3,000, $5,000, $7,000, $10,000, $15,000, $20,000, $25,000, $30,000, $50,000, $75,000, $100,000, $150,000, $250,000, and $500,000 AUD. This non-uniform spacing reflects the logarithmic nature of price perception.

2.3 Variable Isolation

Token input ranged narrowly from 41-43 across all variations, confirming prompt-level isolation. All tests were conducted sequentially on 24 March 2026 within a single session.

2.4 Evidence Level

This study is classified as Evidence Level 4 (Controlled Experiment). All findings are currency-specific (AUD) and cross-currency replication is required for broader applicability.

3. Results

3.1 Complete Sentiment Trajectory

3.2 Threshold Transitions

3.3 Sentiment Zones

3.4 The Instability Zone

The $20,000-$25,000-$30,000 range exhibited rapid sentiment oscillation: neutral at $20,000, negative at $25,000, then positive at $30,000. This instability zone represents three consecutive threshold transitions across just three data points, warranting targeted replication with finer price granularity.

3.5 Response Stability

Despite the sentiment volatility, response word counts remained stable at 181-213 across all price points (mean: 198.3). Tokens in ranged narrowly from 41-43, and tokens out from 302-319. Mean elapsed time was 9.65 seconds (range: 8.5-10.7s).

4. Discussion

The U-curve pattern revealed by this experiment has several striking characteristics. The "valley of suspicion" at $1,000-$5,000 suggests that under these controlled conditions, the LLM has internalised a price-quality heuristic where this range is perceived as insufficient for professional brand identity work. The $5,000 point produced the highest negative mention count (3 negative markers), indicating peak skepticism.

The premium credibility zone ($30,000-$250,000) is notable for its consistency: 7 out of 8 price points in this range produced positive sentiment with exactly 3 positive mentions per response. The sole exception ($100,000) produced 3 positive mentions with zero negative mentions -- the cleanest positive signal in the entire probe.

The ceiling effect at $500,000 is qualitatively distinct from the valley of suspicion. At $500,000, the LLM produced 2 positive and 0 negative mentions but classified as neutral overall -- suggesting that extreme pricing produces cautious evaluation rather than the active criticism seen in the valley of suspicion. The LLM appears to retreat to a more measured stance when confronted with pricing that may exceed its training data's range of normal professional services.

The 9 threshold transitions across 18 data points make pricing the most volatile signal tested in the research programme, with a transition density of 0.50 per data point compared to 0.11 for ratings (2 transitions across 19 points). This volatility is itself an important finding: it suggests that LLMs process pricing through more complex and less stable heuristics than they apply to star ratings.

5. Limitations

• Currency specificity: All prices are in AUD. The threshold values are likely currency-dependent and may not transfer directly to USD, EUR, GBP, or other currencies without cross-currency replication.
• Service type specificity: The probe was framed as a "brand identity project." Different service categories (web development, marketing strategy, consulting) may produce different pricing zones.
• Single-model limitation: Findings are from a single LLM system and may not generalise to other models.
• The instability zone ($20,000-$25,000): The rapid oscillation in this range may reflect genuine model uncertainty or could be an artifact of the specific price points chosen. Finer granularity testing is needed.

6. Conclusions

Under these controlled conditions, pricing below $1,000 and above $15,000 AUD produces positive AI sentiment, while the $1,000-$5,000 range produces predominantly negative sentiment. The non-linear U-curve pattern with 9 threshold transitions establishes pricing as the most volatile single variable tested in the SIGI research programme.

The premium credibility zone ($30,000-$250,000) demonstrates that under these specific conditions, higher pricing correlates with more positive AI sentiment up to a ceiling effect at $500,000. For service providers, these findings suggest that pricing strategy may have direct implications for AI-generated evaluative sentiment, though generalisation beyond these specific experimental conditions requires cross-currency and cross-model replication.

Confidence: HIGH. All 7 logic gates passed. Currency-specific (AUD); cross-currency replication needed for broader applicability.