Client Portfolio Size and AI Credibility Assessment: An Oscillating Signal with No Positive Ceiling

1. Introduction

Client portfolio size is commonly presented as a credibility indicator by service providers. The implicit assumption is that more clients equals more experience, and more experience equals higher quality. However, this assumption may not hold in AI evaluation contexts, where LLMs may apply more nuanced heuristics that weigh specialisation against breadth, or quality against quantity.

This study tests the relationship between claimed client count and LLM sentiment assessment through a controlled single-variable experiment, varying client count from 1 to 1,000 while holding all other variables constant. The results reveal that client count is fundamentally limited as a positive credibility signal under these conditions.

2. Methodology

2.1 Probe Design

The client count probe (CC01) consisted of 19 variations (CC01_v00 through CC01_v18). Each prompt stated that a design agency had worked with a specified number of clients and asked the LLM to assess experience level. The independent variable was the client count, tested at: 1, 3, 5, 8, 10, 15, 20, 25, 30, 40, 50, 75, 80, 100, 150, 200, 300, 500, and 1,000.

2.2 Variable Isolation

Token input ranged from 31-32 across all variations (the minor variation due to digit count differences). All tests were conducted on 24 March 2026.

2.3 Evidence Level

Evidence Level 4 (Controlled Experiment). MODERATE confidence due to signal volatility suggesting less stable LLM evaluation of this variable.

3. Results

3.1 Complete Sentiment Trajectory

3.2 Threshold Transitions

3.3 Stable Neutral Bands

3.4 Response Metrics

Word counts ranged from 147 to 194 (mean: 176.7). Mean elapsed time was 8.3 seconds (range: 7.2-9.3s). Tokens out ranged from 217 to 304. The shortest responses were produced at the extreme low end (n=3, 147 words), consistent with a brief negative assessment of insufficient experience.

4. Discussion

The complete absence of positive sentiment across all 19 variations is the most significant finding of this probe. Unlike the awards probe (SIGI-2026-005), which achieved one positive result at n=200, or the ratings probe (SIGI-2026-001), which achieved consistent positive results above 4.7 stars, client count under these controlled conditions cannot generate positive AI endorsement at any magnitude.

The oscillating pattern between neutral and negative through the mid-range (10-100 clients) suggests that the LLM's evaluation of client count is less stable than its evaluation of ratings or even awards. Each threshold transition between neutral and negative in this range may reflect competing heuristics: more clients as evidence of experience (pushing toward neutral) versus more clients as evidence of commoditisation or lack of specialisation (pushing toward negative).

The return to negative at high counts (500-1,000) is consistent with a specialisation-versus-breadth heuristic. Qualitative analysis of responses at these magnitudes indicated that the LLM framed very large client portfolios as suggestive of a volume-focused business model rather than a quality-focused one.

The two stable neutral bands (25-40 and 150-300) represent practical ranges where client count claims avoid negative sentiment. However, the inability to achieve positive sentiment at any count means that client count, under these conditions, functions as a damage-limitation signal rather than a credibility-building one.

5. Limitations

• MODERATE confidence: The 8 threshold transitions across 19 data points and the complete absence of positive sentiment suggest either genuine evaluative instability or a variable that the LLM processes with less certainty than other signals.
• No context factors: Real-world client count claims typically include information about agency size, years in operation, or project types. These contextual factors may significantly modify the assessment.
• Industry specificity: Results are specific to the design agency context.
• Single-model limitation: Other LLM systems may evaluate client count differently.

6. Conclusions

Under controlled conditions, no client count produces positive LLM sentiment. The signal oscillates between neutral and negative, with stable neutral bands at 25-40 and 150-300 clients. High counts (500-1,000) return to negative, suggesting the LLM perceives very large portfolios as indicating commoditisation rather than expertise. Client count functions as a damage-limitation signal rather than a credibility-building one in AI evaluation contexts.

Confidence: MODERATE. All 7 logic gates passed, but signal volatility suggests the LLM's evaluation of this variable is less stable than other probes in the programme.