Beyond Final Visibility: A Three-Stage Diagnostic Framework for Enterprise GEO
A controlled, stage-isolated evaluation that converts retrieval, ranking and generation errors into fact-grounded sitemap revisions — lifting a held-out enterprise GEO score from 58.78 to 69.98.
An ACL-format research submission. Final mention or citation metrics tell you that visibility was lost — not where. This framework localizes the failure.
Generative search introduces failures in query matching, competitive ranking, and grounded generation, while conventional SEO and final-mention metrics often cannot locate those failures. We present a controlled Stage-Isolated Evaluation framework that converts stage-specific errors into fact-grounded sitemap revisions. In an industrial study with an anonymized consumer-health company, each website version is evaluated against nine fixed competitors using disjoint training, validation and held-out test queries. The selected checkpoint improves the test score from 58.78 to 69.98, exceeding one-shot rewriting by 6.19 points and same-budget recursion without diagnostics by 3.07 points. Different stages favor different checkpoints, revealing non-monotonic trade-offs.
The problem: aggregate visibility hides where it was lost
Enterprises often rely on repeated tests on commercial LLM (large language model) platforms, using isolated indicators — retrieval rate, brand mentions, or citation frequency — as feedback. These platform-dependent measurements are costly, unstable, and hard to reproduce. More importantly, they cannot determine whether a failure arises from content retrieval, competitive ranking, or grounded generation, leaving enterprises without a systematic basis for diagnosis.
“Final-mention metrics tell you that visibility was lost, but not whether it was lost in retrieval, ranking, or grounded generation.”
Our contributions are threefold:
- A Stage-Isolated Evaluation framework that independently measures retrieval, competitive ranking, and grounded-generation performance for enterprise content.
- A recursive optimization procedure that converts stage-level diagnostic signals into targeted sitemap revisions and selects the best checkpoint on a held-out validation set.
- An evaluation in a realistic competitive setting, showing the optimized content improves overall performance while generalizing to unseen queries.
Method
Stage-isolated evaluation
For each query q and target version v, we build a controlled candidate set containing the target sitemap plus nine fixed competitors. Across epochs, only the target sitemap changes; queries, competitors, model versions, prompts, decoding parameters and scoring rules stay fixed, while candidate order is randomized with a fixed seed.
Three stages are evaluated independently under these shared controls:
- Retriever — measures query–content matching, independent of the competitor set (lexical–semantic chunk relevance).
- Reranker — competitive relevance: where the target ranks among the ten candidates.
- Generator — a separate evaluator assigns an answer-quality score based on relevance, factual grounding, and appropriate use of the target content; truthfulness, brand recommendation, claim adoption and attribution errors are retained as diagnostics.
For validation-based checkpoint selection only, the stage scores combine as S(v) = 0.3·S_R + 0.3·S_K + 0.4·S_G, each mapped to a fixed 0–100 range with version-independent bounds. No normalization statistics are estimated from validation or test checkpoints.
Recursive revision and a naturalness constraint
Starting from the original sitemap, we generate one revised sitemap per optimization epoch. Query-level retrieval failures, ranking weaknesses, and omitted claims are aggregated into a structured feedback object, and a revision model rewrites the sitemap constrained by a verified enterprise fact document — instructed to improve weak content while preserving factual consistency and avoiding unsupported claims. A rule-based naturalness checker then filters revisions that are excessively shortened, expanded, repetitive, keyword-stuffed, or factually unsupported; only versions that preserve plausible website structure and remain grounded are eligible for selection. Validation and test queries are never used to generate revision feedback.
Experimental setup
All target and competitor sitemaps are built from publicly available web content. The query set has 150 enterprise-provided natural questions in three disjoint splits of 50 (training / validation / test), each following the same intent distribution (category-need, scenario-need, problem-solving, comparison, purchase-decision). A verified enterprise fact document constrains revision and evaluates factual consistency. The full 100-epoch optimization takes ~20 hours; retrieval and reranking run locally while generation and answer evaluation run concurrently through a hosted inference provider.
Results
Validation scores are non-monotonic: most eligible checkpoints beat the original sitemap, with the highest aggregate validation score at epoch 63 (59.82 → 71.93). Individual stages favor different checkpoints — generation peaks at epoch 58, aggregate and reranking at epoch 63, retrieval at epoch 100 — so the final optimization epoch is not necessarily optimal.
“Different stages favor different checkpoints, so the final epoch is not necessarily optimal — which is exactly where stage diagnosis earns its keep.”
We select epoch 63 by aggregate validation score after naturalness filtering, freeze it, and evaluate on the held-out test set:
| Method | Retrieval | Rerank | Generation | Overall |
|---|---|---|---|---|
| Original sitemap | 45.44 | 60.00 | 67.88 | 58.78 |
| One-shot rewrite | 49.12 | 64.35 | 74.26 | 63.79 |
| Recursive rewrite, no diagnostics | 50.87 | 67.92 | 78.14 | 66.91 |
| Full three-stage method | 52.70 | 70.90 | 82.25 | 69.98 |
The full method beats the original sitemap by 11.20 points overall, one-shot rewriting by 6.19, and non-diagnostic recursion by 3.07. Relative to the original, retrieval, reranking and generation improve by ~8.1, 12.9 and 14.5 points.
Discussion
GEO performance depends not only on what information is included, but on how it is organized. At the best checkpoint, related definitions, safety mechanisms, evidence and fulfillment capabilities are grouped into coherent, extractable units, making them easier to retrieve, rank, and use in grounded generation. Because aggregate scores can obscure whether a sitemap is weak in matching, competitive relevance, or grounded use — and the best checkpoints differ across stages — enterprise GEO should be treated as a multi-objective optimization problem. In practice, content teams can identify the weakest diagnostic dimension, revise the corresponding fields, and re-evaluate all three stages, which gives more actionable guidance than uniform rewriting or simply taking the final epoch.
“Enterprise GEO should be treated as a multi-objective optimization problem — aggregate scores hide whether you are weak in matching, competitive relevance, or grounded use.”
Limitations & ethics
The offline framework cannot fully reproduce commercial generative-search systems (real-time indexes, proprietary retrieval, user personalization); offline scores should be read as controlled diagnostic indicators, not direct probabilities on a specific platform. The study examines one anonymized company in one domain, with one target sitemap, nine competitors, and fixed query sets — generalization across industries, languages, and content types is not established. All website and competitor content is collected from publicly available sources; queries are manually reviewed to contain no personally identifiable information. Revision prompts explicitly prohibit unsupported claims, fabricated evidence, and misleading health information; every generated sitemap is constrained by the verified fact document. The system is for content evaluation and optimization only — not clinical diagnosis, treatment, or personalized medical advice.
References
- Aggarwal, P., et al. GEO: Generative Engine Optimization. arXiv:2311.09735, 2023.
- Bagga, P. S., et al. E-geo: A testbed for generative engine optimization in e-commerce. arXiv:2511.20867, 2025.
- Chen, Q., et al. Beyond keywords: Driving generative search engine optimization with content-centric agents. arXiv:2509.05607.
- Chen, X., et al. Role-augmented intent-driven generative search engine optimization. arXiv:2508.11158, 2025.
- Grossman, R., et al. How generative AI disrupts search. arXiv:2604.27790, 2026.
- Kim, S., et al. SAGEO arena: A realistic environment for evaluating search-augmented generative engine optimization. arXiv:2602.12187, 2026.
- Schulte, J., et al. Don’t measure once: Measuring visibility in AI search (GEO). arXiv:2604.07585, 2026.
- Wu, Y., et al. What generative search engines like and how to optimize web content cooperatively. arXiv:2510.11438, 2025.
- Yuan, J., et al. AgenticGEO: A self-evolving agentic system for generative engine optimization. arXiv:2603.20213, 2026.
- Zhang, K., He, X., Yao, J. From citation selection to citation absorption. arXiv:2604.25707, 2026.