Search Position Versus Citation Priority: Evidence for a Separate Re-Ranking Pass in AI Answer Generation
Abstract
This paper presents observational evidence that AI systems apply a separate evaluation pass between search retrieval and answer citation, with criteria that differ from search ranking factors. Through introspective analysis of a single LLM's citation process, we document five re-ranking criteria: source type credibility classification, consensus detection across multiple sources, evaluative depth weighting, self-ranking discount, and claim specificity preference. We observe that content can rank first in search results but fail to get cited in the AI answer because it lacks specificity, third-party validation, or evaluative depth. Conversely, a lower-ranked source with specific verifiable claims about third-party entities may receive preferential citation despite its lower search position. This challenges the prevailing GEO assumption that search rank equals citation probability and suggests that optimisation for search ranking and optimisation for citation probability require different strategies.
Keywords
search position, citation priority, re-ranking, AI answer generation, source evaluation, self-ranking discount, claim specificity, generative engine optimisation
1. Introduction
A foundational assumption in generative engine optimisation is that search ranking position correlates with citation probability: sources that rank higher in the AI system's search results are more likely to be cited in the generated answer. This assumption, if correct, would mean that GEO strategy could focus entirely on search ranking optimisation. However, if the AI system applies a separate evaluation pass with different criteria between retrieval and citation, then search ranking optimisation alone may be insufficient or even misaligned with citation optimisation.
This paper investigates whether such a separate re-ranking pass exists by examining the relationship between search position and actual citation behaviour in an LLM's answer generation process.
2. Methodology
The LLM was prompted with a competitive service query that produced 10 search results. Through introspective probing, the model was asked to describe its internal process for moving from retrieved search results to cited sources in its generated answer. The model was asked to identify: which sources it would cite, which it would not cite, and the specific criteria differentiating cited from uncited sources. All sources are anonymised in this report.
3. Results
3.1 Five Re-Ranking Criteria
| Step | Criterion | Effect on Citation |
|---|---|---|
| 1 | Source type credibility scan | Classifies each source as self-interested or third-party; directories receive higher trust than agency self-published content even when the latter ranked higher |
| 2 | Consensus detection | Same entity mentioned by multiple sources increases citation confidence for that entity |
| 3 | Evaluative depth weighting | Specific claims (named clients, founding dates, specific awards) preferred over generic descriptions |
| 4 | Self-ranking discount | Publisher ranking itself in first position heavily discounted; third-party evaluation of the same entity weighted more |
| 5 | Specificity preference | Citable content contains specific verifiable claims; non-citable content contains generic interchangeable descriptions |
3.2 Position-Citation Disconnect
The search result ranked at position 1 was a self-published listicle in which the publisher appeared prominently. Despite achieving the highest search ranking through content freshness and format-query alignment, this source received a self-ranking discount during the citation pass. Meanwhile, a lower-ranked source containing specific evaluative claims about third-party entities (named clients, specific project histories, award details) received preferential citation treatment despite its lower search position.
3.3 Citation Failure Patterns
Sources that failed the citation pass despite high search rankings shared common characteristics: generic descriptions interchangeable between any service provider, self-promotional language without external validation, absence of specific verifiable claims, and self-ranking by the publisher. Sources that passed the citation pass despite lower search rankings shared: specific third-party evaluative claims, named entities and verifiable facts, and non-self-interested positioning.
4. Discussion
The existence of a separate re-ranking pass between retrieval and citation has significant implications for GEO strategy. Currently, most GEO advice focuses on optimising for search retrieval: content freshness, format matching, keyword alignment, and domain authority. These factors determine search position but, based on our observations, they do not determine citation probability. A separate set of criteria -- source credibility, claim specificity, third-party validation, and self-interest assessment -- appears to govern the transition from retrieved source to cited source.
The self-ranking discount is particularly notable. A publisher that ranks itself in first position achieves maximum search visibility but may trigger maximum citation scepticism. This creates a strategic tension: the content format that optimises search ranking (self-published listicle with publisher prominently featured) may be the format that triggers the strongest citation discount.
For practitioners, this suggests that two distinct optimisation strategies may be needed: one for search retrieval (format match, freshness, keyword alignment) and one for citation probability (specific third-party claims, evaluative depth, non-self-interested positioning).
5. Limitations
- Introspective methodology: The re-ranking criteria are based on the model's self-report of its citation process.
- Single query, single session: The observation is based on a single competitive service query.
- Single model: Different AI systems may implement different re-ranking criteria.
- Correlation vs. causation: The position-citation disconnect may have explanations beyond a separate re-ranking pass.
6. Conclusions
We observe evidence that AI systems apply a separate evaluation pass between search retrieval and answer citation, with criteria that differ from search ranking factors. Content can rank first in search results but fail to get cited due to self-ranking discount, lack of specificity, or absence of third-party validation. This challenges the GEO assumption that search rank equals citation probability and suggests that search optimisation and citation optimisation may require different strategies.
Confidence: MODERATE for the existence of a separate pass. The specific re-ranking criteria are introspective self-report.
References
- The Scientific Institute for Generative Intelligence. "Query Framing Effects on AI Source Selection." SIGI-2026-055. generativeintelligence.institute, March 2026.
- The Scientific Institute for Generative Intelligence. "Content Extraction by Source Type." SIGI-2026-057. generativeintelligence.institute, March 2026.
- The Scientific Institute for Generative Intelligence. "Entity Density in Search Snippets." SIGI-2026-060. generativeintelligence.institute, March 2026.