PAPER
AgenticGEO — what happens when strategy varies per document instead of following fixed rules
If you choose a different optimization strategy per piece of content instead of applying one rule to everything, does the result change?
AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization
- Authors
- Jiaqi Yuan · Jialu Wang · Zihan Wang · Qingyun Sun · Ruijie Wang (corresponding author) · Jianxin Li
- Affiliation
- Beihang University, School of Computer Science · with independent contributors
- Venue
- arXiv preprint (code released)
- Submitted
- 2026-03-02
- arXiv
- arXiv:2603.20213
- We verified
- 2026-08-22
WHAT THE PAPER SAYS
It did. The authors argue that the static heuristics proposed in earlier GEO work — add keywords, add citations, add statistics — fail to optimize roughly half of samples, and propose an agentic system that evolves a strategy per document. In their own evaluation they report gains of 26–28% in-domain and 37–70% out-of-domain over the previous best (AutoGEO).
Read the original on arXiv ↗METHOD
How it was measured
Read the conditions before the numbers. The same figure means something different under a different sample or environment.
- Datasets
- GEO-Bench 1,000 queries (5,000 documents) · MS MARCO 1,000 queries (5,000 documents) · e-commerce 416 queries (2,180 documents)
- Generation engines used for evaluation
- Qwen2.5-32B-Instruct · Llama-3.3-70B-Instruct (open models, not commercial engines)
- System
- Critic Qwen2.5-1.5B · evolver Qwen2.5-7B-Instruct · rewriter Qwen2.5-32B-Instruct
- Baselines
- 14 in total — static heuristics (keywords, authority, jargon, fluency, citations, quotations, statistics) and learned approaches such as AutoGEO
- Metrics
- Attributed word count · position-weighted citation order · a combined overall score
FINDINGS
What came out
- In-domain (GEO-Bench)
- 25.48 vs the best baseline 23.71 (AutoGEO) on Qwen2.5-32B — +26.08%. 24.52 vs 22.78 on Llama-3.3-70B — +27.71%
- Out-of-domain (MS MARCO)
- 34.10 vs 30.67 — +70.07% (Qwen), 33.50 vs 30.04 — +70.31% (Llama). Transfer to unseen domains is the paper's main claim
- Out-of-domain (e-commerce)
- 26.58 vs 21.18 — +47.58% (Qwen), 26.88 vs 21.50 — +36.59% (Llama)
- Feedback efficiency
- 98.1% of peak performance from 700 engine feedback samples while using only 41.2% of the supervision
- ⚠️ The conditions on these numbers
- The evaluation engines are open models (Qwen, Llama), not commercial answer engines like ChatGPT or Perplexity, and this is generation-stage evaluation over a supplied document set rather than a full pipeline with retrieval and reranking. It does not translate into citation rates in a commercial service
LIMITATIONS
Limitations the authors state themselves
Not our criticism — this is what the authors wrote in the paper.
- There is no dedicated Limitations section — the conclusion stresses the method's durability and ecosystem benefits without a systematic limitations analysis.
- Evaluation engines are open models (Qwen2.5-32B, Llama-3.3-70B) rather than commercial answer engines.
- Multi-turn rewriting is reported as optimal at three turns, with diminishing marginal benefit beyond that.
- The theoretical guarantee (regret bound O(√T)) rests on the assumption that the replay buffer grows linearly.
NAVIRANG'S READING — NOT THE PAPER'S CONCLUSION
In our claims table this paper holds the third position. The original GEO paper said visibility rises; SAGEO Arena said it fell in a realistic pipeline; this paper says that happened because the rules were static, and that varying the strategy per document raises it. The interesting observation is that existing heuristics fail on roughly half of samples — meaning the 'add citations, add statistics' checklist circulating in the industry does not work on every document. Moving it into practice is harder, though: evaluation ran on open models rather than commercial answer engines, and skips retrieval and reranking, which is exactly the segment SAGEO Arena took issue with. We read this only as evidence that the prescription has to differ per document, and we do not use its improvement figures to predict our own results.
IN COMPARISON
Where it diverges from other papers
The claims-table entries this paper appears in.
RELATED
Related reading
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.