List Magnitude and Entity Anchoring: How Request Size Affects Which Service Providers LLMs Recommend

Introduction

When users ask an LLM to recommend service providers, the phrasing of the request — particularly the number of recommendations sought — is typically treated as a neutral formatting parameter. A user asking for “the top 3” versus “the top 10” expects the same entities to appear in both lists, with the longer list simply adding more entries. This study tests whether this assumption holds by systematically varying the requested list magnitude and tracking which entities appear at each level.

Entity anchoring — the phenomenon where certain entities are recalled with high consistency regardless of context parameters — has implications for competitive visibility in AI-mediated discovery. If some entities are anchored in the model’s recall patterns while others are magnitude-dependent, this creates a structural advantage that is independent of the entities’ actual quality or relevance. All entity names in this publication have been anonymised (Firm Alpha, Firm Beta, Firm Gamma, etc.).

Methodology

The magnitude probe (Probe ID: M01–M10) used the following prompts: “Name the single best brand design agency in [Country A]” for N=1, and “Name the top {N} brand design agencies in [Country A]” for N=2 through N=10. All tests were conducted on 24 March 2026 using a ChatGPT-class LLM. Token input remained nearly constant (17–19 tokens), confirming minimal prompt variation. The entities named in each response were catalogued and tracked across all magnitudes.

Results

Entity Appearance Across Magnitudes

Response Scaling with Magnitude

Entity Anchoring Analysis

Two entities demonstrated strong anchoring, appearing in 7 or more of the 9 list sizes where multiple entities were requested (N=2 through N=10): Firm Alpha appeared in 8 of 9 lists (absent only at N=4), and Firm Beta appeared in 7 of 9 lists (absent at N=3 and N=4). These anchored entities represent the model’s strongest recall associations for the tested category and geographic market.

The N=4 gap is particularly notable: at this specific magnitude, none of the primarily tracked entities appeared. The model populated the list with entities (Firm Delta, Firm Epsilon, Firm Zeta, Firm Eta) that appear at other magnitudes but are not among the most consistently anchored. This suggests that the specific number “4” may trigger a different recall pathway, though this is a single observation requiring replication.

Discussion

The entity anchoring phenomenon has significant implications for competitive visibility in AI-mediated discovery. Anchored entities benefit from a structural advantage: they appear in recommendations regardless of how the query is phrased, while non-anchored entities are excluded from shorter lists entirely. For service providers, this creates a binary visibility distinction — those whose brand signals are strong enough to achieve anchoring in the model’s recall patterns, and those who appear only when the query parameters happen to create sufficient “slots” for additional entities.

The linear relationship between N and response word count (121 to 184 words, approximately 7 additional words per additional entity requested) suggests that the LLM allocates roughly equal descriptive space to each entity. The sub-linear token scaling (discussed further in SIGI-2026-016) indicates that each additional entity receives less information than the previous one, creating a positional advantage for entities listed earlier.

Limitations

This study tests a single geographic market and industry category. Entity anchoring patterns may differ substantially across markets and service categories. The N=4 gap is a single observation that may not replicate. The M01 prompt (“Name the single best”) differs structurally from M02–M10 (“Name the top N”), creating a minor confound at N=1. All tests used a single LLM at a single time point.

Conclusions

List magnitude is a non-neutral parameter in LLM service provider recommendations. Entity anchoring creates structural visibility advantages for certain providers, while the linear word-count scaling and the N=4 recall gap reveal that the model’s entity retrieval mechanisms are more complex than simple ranked-list generation. For GEO strategy, achieving anchored status — consistent inclusion regardless of query magnitude — represents a qualitatively different and more valuable visibility outcome than occasional appearance in longer lists.

Confidence Statement: HIGH for the anchoring pattern. The N=4 gap is a single observation requiring replication. Single-model, single-market, single-time-point limitation acknowledged.