Skip to content

PAPER

When only one of two documents is cited, what decided it

When only one of two candidate documents is cited, what difference decides it?

What Gets Cited: Competitive GEO in AI Answer Engines

Authors
Rahul Vishwakarma · Shushant Kumar · Ratnesh Jamidar
Affiliation
Sprinklr (Gurugram, India · Dubai, UAE) — a commercial AI/CX platform
Venue
arXiv preprint
Submitted
2026-05-25
arXiv
arXiv:2605.25517
We verified
2026-08-24

WHAT THE PAPER SAYS

Topical match and list position dominated, with pricing information and a recent timestamp next. The authors ran 252,000 pairwise comparisons injecting two candidate documents that differed in **exactly one factor**. Topical mismatch, missing pricing, an old timestamp and second position in a list were conditions that lost essentially always, and the gap between assertive and hedged phrasing was also large.

Read the original on arXiv ↗

⚠️ THIS PAPER HAS A DISCLOSED CONFLICT OF INTEREST

All authors are affiliated with Sprinklr, a commercial AI/CX platform, and the paper mentions an early pilot at Sprinklr, but **there is no conflict-of-interest declaration**. This differs from the Ranqo paper in the same library (arXiv:2606.20065), whose author stated his vendor position directly.

METHOD

How it was measured

Read the conditions before the numbers. The same figure means something different under a different sample or environment.

Trials
252,000 pairwise comparisons
Models
6 — Gemini-2.5-Flash · GPT-5-Nano · GPT-5-Mini · GPT-5.2 · Claude-3.5-Sonnet · Kimi-K2-Thinking
Factors
18 content factors
Design
Two candidate documents are injected directly, differing in exactly one factor, and the cited one is recorded. Documents are supplied directly rather than passing through a retrieval stage

FINDINGS

What came out

Conditions that lost essentially always
Topical mismatch · missing pricing · old timestamp · second position in a list — all four at odds ratios >>10,000
Phrasing
Assertive vs hedged phrasing, odds ratio 599 — the same content hedged was heavily disadvantaged
Others
Missing specification information odds ratio 8.63 · clear value proposition 5.22
⚠️ How to read an odds ratio >>10,000
It does not mean a ten-thousand-fold effect. It means one side won so consistently in pairwise comparison that the outcome separated statistically. 'This condition was effectively decisive' is the right reading
⚠️ The conditions of the experiment
This is a controlled environment with only two documents injected. The authors themselves note that "real RAG typically retrieves five to ten or more, so the real citation pool is larger than our testbed"

LIMITATIONS

Limitations the authors state themselves

Not our criticism — this is what the authors wrote in the paper.

  • Real services typically retrieve five to ten or more pages in RAG, so the actual citation pool is larger than this testbed — a limitation the authors state directly.
  • Documents were injected directly without passing through retrieval or reranking, so the study cannot see what gets filtered out during retrieval.
  • There is no conflict-of-interest declaration and all authors work at a commercial platform.

NAVIRANG'S READING — NOT THE PAPER'S CONCLUSION

Two items transfer straight into practice. **Missing pricing information** was effectively decisive against a page — which supports what we have told B2B and SaaS clients: even if you cannot publish a figure, publish the billing unit and the variables. And **hedged phrasing lost to assertive phrasing at an odds ratio of 599** — though this one needs care. It does not mean 'state things you are unsure of assertively'. We use it only to mean: do not hedge a sentence you have evidence for. Documents that assert without evidence become the raw material for the citation distortion another paper (2605.28565) demonstrated. Above all, this experiment skipped retrieval and injected documents directly. The very segment SAGEO Arena took issue with is missing here too, so we read the result as strictly a story about what happens after retrieval has been passed.

IN COMPARISON

Where it diverges from other papers

The claims-table entries this paper appears in.

If you want your own brand's numbers rather than a paper's

Every figure here came from someone else's sample. Send us a URL and we ask all 7 answer engines directly and measure yours.

We reply within one business day.

Free audit Call Email Blog