What Evidence Do Language Models Find Convincing? (Wan et al. 2024)

1. What this paper is — and the reversal it forces

Most GEO advice rests on a quiet assumption: if you make a page look credible — author byline, citations, neutral tone, schema markup, statistics — an LLM choosing between sources is more likely to side with you. What Evidence Do Language Models Find Convincing? (Wan, Wallace, Klein — UC Berkeley, ACL 2024) is the first study to put a number on that assumption and the first to break it.

Given two real web pages that disagree on a contested question, the LLMs they tested chose between them mainly on topical relevance — how directly and specifically the page addresses the query — and were nearly insensitive to the stylistic credibility signals humans say they value most: scientific references, a neutral tone, formal citation style.

The result is uncomfortable because the dominant practice points the other direction. The honest framing is also narrower than the headline: this is a counterfactual study on five 2023–24 LLMs against the ConflictingQA dataset, not a sweep of every 2026 production engine.

2. The problem it formalizes

Retrieval-augmented generation hands a model multiple sources for the same query. On factual queries with one correct answer, the literature on retrieval-augmented LLMs already had a story: the model often defers to whichever source aligns with its parametric prior, or to the one cited more prominently in the prompt. On contested queries — is aspartame linked to cancer, is daycare bad for kids — there is no single correct answer to ground on, which is exactly the regime where credibility signals should matter most.

Wan et al.’s structural move is to take that regime and turn it into a paired comparison:

• Pick a contested query with no settled answer.
• Show the model two real web-page paragraphs that take opposing positions.
• Ask: which one are you convinced by?

That setup isolates “what makes evidence convincing” from “what is the true answer”. The same setup also lets the authors run counterfactual edits — strip the scientific references from one paragraph, rewrite another in a neutral tone, prepend a question-restatement prefix — and measure how the model’s preference shifts. The grounding mechanism the experiment sits on top of — retrieval, prompting, citation — is the answer loop; this paper takes that mechanism as given and asks the preference question on top of it.

Worth noting up front: this paper is about the preference step, not the retrieval step. Citability covers the structural question of whether a passage can be lifted into an answer at all; citation vs mention vs link covers how that lift gets credited. Wan asks the more upstream question: given the model has two retrieved candidates in front of it, which does it pick?

3. Methodology — ConflictingQA and the feature space

The dataset, ConflictingQA, is constructed from contested queries scraped across medical, political, and social topics — the kind of question where retrieving the top-N web pages reliably surfaces opposing views:

The five LLMs evaluated span open-source 13B and closed-source frontier classes — important for §6 critique:

The feature space the paper interrogates splits into two camps:

For each pair, the protocol records (a) a human preference label, (b) the model’s preference, and (c) how the model’s preference shifts when one paragraph is counterfactually edited along a single feature axis. The result is a per-feature win-rate curve, reported as bar charts in Figure 2 and as scatter plots in Figure 4 of the paper.

4. The reversal in detail — what moved preferences, what did not

The single most-quotable artifact from the paper is the asymmetry of the per-feature win-rate effects:

The phrasing the abstract uses is the cleanest one-liner the paper produces: models “rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone”.

Why this lands uncomfortably for GEO: a non-trivial share of practitioner advice on citability and E-E-A-T optimization treats stylistic credibility signals as the primary lever on whether the model “trusts” a source. Wan et al. is the cleanest counterfactual evidence to date that — for the preference step, in the configurations they tested — that lever is faint and topical relevance is the heavy one.

5. Key findings

The headline, written in a way that survives extraction:

One nuance the paper is careful with and that practitioners routinely lose: the results are direction and rank-ordering of effects, not point-estimate percentages with confidence intervals. The bars in Figure 2 and scatters in Figure 4 are the primary artifact — a citation that says “scientific references give X% lift” is not a quotable number from this paper, and any number quoted at that precision is being read off a chart, not from a published estimate.

6. GEO Wiki critique

Steel-man first. Three contributions are load-bearing and have held up:

1. First decomposition of ‘convincingness’ into named features. Prior work on retrieval-augmented trust mostly measured whether the model deferred to retrieved sources, not which feature of a source drove the deferral. Wan et al. is the first paper to interrogate this with counterfactual perturbation.
2. ConflictingQA is a reusable public asset. 238 contested questions × 2,208 paragraphs is a non-trivial dataset to construct, and the MIT license + GitHub release have already enabled follow-on work — Retrieval-Augmented Generation with Conflicting Evidence (Wang et al., arXiv 2504.13079, April 2025) builds directly on the conflict regime Wan opened.
3. Direct implication for corpus governance. If preference is driven by topical relevance and not by surface credibility signals, then the leverage for high-stakes RAG sits upstream — in which sources enter the candidate pool — not in how the candidates are styled. That is the same direction C-SEO Bench points: many on-page conversational-SEO rewrites do less than the field assumes.

Bounded critique — four points:

1. External validity is tight. All five evaluated models predate mid-2024. GPT-4 is gpt-4-1106-preview, Claude is Instant v1, and three of the five are 13B open models from the same Llama-2 lineage. There is no Claude 3.5 / 4, no GPT-4o, no Gemini, no Mistral, no LLaMA-3. A 2026 reader cannot quote this paper as evidence about today’s ChatGPT Search, Gemini, Perplexity, or Google AI Overviews and be on safe ground.
2. Feature coverage is a subset, not the whole surface. The “credibility signals” tested are scientific references, neutral tone, formal citations, “add more info”, readability, and unique-token count. The rest of the E-E-A-T surface — verified author identity, sameAs graphs, source-domain authority, statistic density, structured-data presence — was not perturbed. The honest read of the paper is “four named stylistic features barely move five 2023–24 models”, not “credibility is irrelevant to LLMs”.
3. The apparent contradiction with Aggarwal dissolves under measurement. Aggarwal et al. (KDD 2024) reports Cite Sources lifting Position-Adjusted Word Count by up to 30% and Quotation Addition by up to 41%. Wan reports adding scientific references barely moves preference. Both can be true simultaneously: Aggarwal measures whether a source gets pulled into a synthesized answer at all (an inclusion outcome), Wan measures which of two retrieved sources the model prefers in a head-to-head (a preference outcome). The two papers operate at different points in the answer loop. Practitioners who collapse them into one number lose this distinction.
4. The ‘why’ is undertheorized. The paper documents the phenomenon — relevance dominates, stylistic credibility does not — but does not strongly distinguish between candidate causes: a pretraining-objective effect, an RLHF artifact, or a prompt-format dependency in how the comparisons were framed. The follow-up question why matters because it determines whether the result is durable across new training regimes or a transient property of 2023–24 models.

Position. Foundational is not turnkey. Take the direction — topical relevance and specificity dominate the preference decision, surface credibility signals barely move it — and discard any cross-engine or cross-time extrapolation as a planning input.

7. Reproducibility

Verified at draft time (2026-05-28), not assumed:

Field value: code-and-data. Both the experimental code and the ConflictingQA dataset are openly available, so the headline experiments are independently reproducible — a meaningful strength relative to most papers in this space.

8. What it means for practitioners

Read this paper as a priority-shift signal, not as permission to abandon credibility work:

• Shift forward in the funnel. Spend more effort on whether a passage directly and specifically answers the query — query-shaped phrasing, concrete specificity, a one-line restatement of what the section is about. Wan et al. shows this is the lever that actually moved preference in the configurations they tested.
• Do not abandon surface credibility — relocate its job. Author bylines, sameAs identity graphs, schema markup, and source-domain authority were not in Wan’s perturbation set. They may still affect whether your page is retrieved at all (the upstream half of citability) and whether retrieved content survives a E-E-A-T trust filter. Wan only tested the preference step once two retrieved sources are in front of the model.
• The right way to read Wan alongside Aggarwal. Aggarwal et al.’s Cite Sources / Statistics / Quotation rewrites work in part because they also raise topical relevance — adding a statistic to a passage about a numerical question makes the passage more directly responsive to that question. Read the two findings as compatible: Wan tells you what kind of content edit moves the preference step (relevance, not style); Aggarwal tells you that the substance-additions kind of credibility edits do lift inclusion, partly via the same relevance channel.
• Do not over-extrapolate to 2026 production engines. The five models tested were 2023–24 snapshots. Read writing-for-ai-citation for the operational version of the priority shift — but treat any specific “X is dead, Y is the new lever” framing built on Wan alone as moving past what the paper actually demonstrates.

9. Further reading

• Aggarwal et al. 2024 — GEO: Generative Engine Optimization — the inclusion-side study; read alongside §6 critique pt. 3 for the measurement-level reconciliation.
• Puerto et al. 2025 — C-SEO Bench — independent replication-counterweight on the GEO methods literature; converges on the same “corpus governance > on-page styling” direction.
• Generative Engine Optimization — how Wan’s priority-shift signal fits into the broader GEO discipline.
• E-E-A-T — the wider credibility surface Wan only sampled a corner of; useful to read after §6 critique pt. 2.
• Citability — the retrieval-side preconditions that Wan’s preference step assumes are already met.

References

1. Wan, Wallace, Klein — What Evidence Do Language Models Find Convincing?, ACL 2024 Main. arXiv:2402.11782 · ACL Anthology 2024.acl-long.403 · DOI:10.18653/v1/2024.acl-long.403
2. Code and ConflictingQA dataset — github.com/AlexWan0/rag-convincingness (MIT)
3. Third-party dataset mirror — huggingface.co/datasets/kortukov/ConflictingQA
4. Aggarwal et al. — GEO: Generative Engine Optimization, KDD 2024. arXiv:2311.09735
5. Puerto et al. — C-SEO Bench: Does Conversational SEO Work?, NeurIPS 2025 D&B. arXiv:2506.11097
6. Liu, Zhang, Liang — Evaluating Verifiability in Generative Search Engines, Findings of EMNLP 2023. arXiv:2304.09848
7. Wang, Prasad, Stengel-Eskin, Bansal — Retrieval-Augmented Generation with Conflicting Evidence, 2025. arXiv:2504.13079