Query Framing Effects on AI Source Selection: How Single-Word Changes Alter Source Pools

1. Introduction

Generative engine optimisation currently operates on the assumption that semantically similar queries produce functionally equivalent search results. Practitioners optimise content for a canonical query (e.g., a superlative form) and assume that alternative framings of the same question will surface similar source pools. This assumption, inherited from traditional SEO where synonym handling is well-developed, may not hold for AI systems that interpret queries through intent classification rather than keyword matching.

If AI systems classify different query framings as different intents -- even when the underlying information need is identical -- they may route those queries to systematically different source types. This would mean that a source optimised for one framing could be invisible for alternative framings of the same question, with significant implications for GEO strategy.

2. Methodology

Two query variants targeting the same competitive service category in the same geographic market were submitted to an LLM with web search enabled. Variant A used a superlative framing (seeking the top-ranked providers). Variant B used an advisory framing (seeking guidance on how to evaluate providers). Each variant returned 10 search results. Results were compared on: total unique domains, source type distribution, specific domains present or absent, and position stability of sources appearing in both variants.

3. Results

3.1 Source Pool Differences

3.2 Source Substitution Patterns

Two agency homepages present in the superlative framing dropped out entirely in the advisory framing. They were replaced by: one additional directory source and one design-focused advisory source. This substitution pattern suggests that the advisory framing actively suppresses individual provider sites in favour of evaluative and comparison-oriented content.

3.3 Format-Query Alignment

One listicle source maintained the first search position across both query variants, despite the different framings. This suggests that content formatted as a comprehensive evaluative list achieves strong format-query alignment that transcends specific intent variation. The listicle format appears to satisfy both the superlative intent (ranking) and the advisory intent (evaluation guidance).

4. Discussion

The observed differences between source pools suggest that AI systems do not treat superlative and advisory framings as equivalent, even when they target the same information need. The advisory framing appears to trigger a source preference shift toward evaluation-oriented content (directories, comparison sites, advisory guides) and away from individual provider sites. This is consistent with the hypothesis that AI systems classify query intent before selecting sources, and different intent classifications route to different source type preferences.

The practical implication for GEO is that content must be optimised for multiple query framings, not just the canonical superlative form. A provider that appears in superlative query results but not advisory query results is invisible to a significant portion of searchers who use alternative framings. The common GEO practice of targeting only superlative queries may miss advisory-framing traffic entirely.

5. Limitations

• N=2 query variants: This is an exploratory observation based on only two framing variants. Additional framings (comparison, review, pricing) were not tested.
• Single model, single session: Results may vary across models, sessions, and time points.
• Single query domain: The framing effect may not generalise to all query categories.
• No statistical significance testing: With N=2 variants, formal statistical testing is not applicable.

6. Conclusions

We observe that single-word changes in query framing produce different source pools, suggesting AI systems route queries to different source types based on perceived intent. Advisory framings appear to suppress individual provider sites while elevating directory and evaluation sources. Content formatted as comprehensive evaluative lists may achieve format-query alignment that is robust across multiple framings. These exploratory findings, while limited in scope (N=2), identify a potentially significant gap in current GEO practice.

Confidence: LOW. N=2 query variants, single model, single session. Exploratory observation only.