AI search research: primary source reviews, retrieval, and evidence

Executive conclusion and overview

According to C-SEO Bench, changing source position in model context mattered more than any fixed rewrite it tested [1]. Later research narrowed the interpretation: being crawled, retrieved, reranked, and cited are different events. A tactic that helps one stage is neutral or harmful at another when the measured outcome changes.

This is why aiseo-audit 2.0 reports four pipeline stages and labels every factor by evidence strength. The audit is a transparent, repeatable heuristic. It is not a fitted estimate of citation probability and it does not claim to observe a proprietary engine’s index or ranking system.

The pipeline works by separating access, retrieval, citation fitness, and provenance. This means evidence from one stage does not silently justify a claim about another. The reason for the evidence tiers is to preserve each paper's scope. As a consequence, unsupported observations remain visible without entering the score.

The evidence is a record of measured outcomes, limits, and tool changes from each primary source review.

What makes content citable in AI search?

Citable content is defined as source material that directly answers the target query, supplies verifiable evidence, and remains direct when lifted into an engine-generated response. What Gets Citedtested 252,000 controlled trials and reported that topic match and product completeness behaved as gatekeepers, while other signals acted as smaller differentiators [2].

Position bias in LLM citations

Position bias refers to a language model favoring a source because of where it appears in the supplied context. According to C-SEO Bench, this effect was stronger than the tested content rewrites. The finding supports a firm limit: page edits cannot guarantee retrieval order or placement inside a private model context.

Retrieval-augmented generation sources

Retrieval-augmented generation refers to selecting sources before a model writes its answer. According to SAGEO Arena, aligned structural fields improved retrieval while body-only changes affected the later generation stage differently [3]. This is why the audit reports retrieval alignment separately from citation fitness.

Current findings

These conclusions recur across the reviewed literature or have a strong controlled design within a clearly stated scope.

Evidence tier refers to the stated strength and scope behind a factor. Supported means that direct outcome evidence exists in a relevant setting. Conditional means that the effect depends on the tested regime. Heuristic is defined as a useful proxy without isolated causal validation. Diagnostic is a type of visible observation that stays outside the score. These labels prevent an observation from silently becoming a promise.

Ten peer-reviewed paper reviews

Each review was completed against the primary paper. It records the experimental regime, metric, results, contradictions, scope limits, and the specific changes the paper justified in the tool.

Key takeaways and how to use the evidence

1. The evidence is strongest when the paper directly measures the outcome named by the factor.
2. Start with eligibility. A page an engine cannot fetch or parse cannot benefit from downstream improvements.
3. Supply the real queries your audience uses. Query-blind checklists cannot distinguish a strong general page from the right page for a specific request.
4. Fix stage constraints before polishing low-confidence signals. Retrieval problems and citation-fitness problems need different interventions.
5. Preserve uncertainty. Treat conditional and heuristic factors as hypotheses to test on your pages, not universal rules.
6. Re-measure. Deployed engines and fetched web content both change, so a single observation is not a durable outcome.

Bottom line: the baseline is the repeatable page result, while observed engine citations remain a separate outcome.