TL;DR
This paper introduces AgentGEO, an agentic framework that moves Generative Engine Optimization (GEO) away from "one-size-fits-all" rewriting toward a surgical, diagnostic approach. By identifying exactly why a webpage fails to be cited by AI engines—whether due to technical parsing errors or semantic gaps—and applying targeted repairs, AgentGEO achieves a 40% relative improvement in citation rates while modifying only 5% of the original content.
What Problem Does This Paper Address?
As users shift from traditional search engines to "generative engines" (like Perplexity, ChatGPT Search, and Google AI Overviews), content creators face a crisis: AI models often synthesize answers without providing the citations that drive traffic back to the source.
Existing GEO methods attempt to fix this by applying generic rules—like "add more statistics" or "use an authoritative tone"—across all content. The authors argue this is fundamentally flawed for two reasons:
- Contribution vs. Citation: Current metrics often measure how much a document influenced a response, but for a creator, the only thing that matters is the citation (the clickable link).
- The "Shotgun" Approach: Generic rewriting doesn't account for the fact that different pages fail for different reasons. One page might be ignored because its HTML is messy (a technical failure), while another might be ignored because it doesn't directly answer the user's specific intent (a semantic failure).
Methods
The researchers developed a three-part solution to make GEO more precise and effective:
- A Taxonomy of Citation Failures: They categorized 949 "failure cases" into four dimensions: Technical Integrity (fetching/parsing issues), Semantic Alignment (relevance to query), Content Quality (how the info is presented), and Systemic Exclusion (being overshadowed by high-authority sites like Wikipedia).
- AgentGEO: This is an agentic system that operates in a "diagnose-then-repair" loop. For a given document, the agent simulates the AI engine's pipeline, identifies the specific failure mode, selects a specialized tool from a library (e.g., "Intent Realignment" or "Data Serialization"), and iterates until the document is successfully cited.
- Localized Editing: Unlike baseline methods that rewrite entire pages, AgentGEO uses a "chunk-level" editing mechanism. It only modifies the specific parts of the HTML or text necessary to trigger a citation, preserving the original meaning and "fidelity" of the creator's work.
Datasets
The authors introduced a new benchmark called MIMIQ (Multi-Intent Multi-Query).
- Why it's different: Most GEO benchmarks pair one document with one query. MIMIQ is document-centric, pairing each of its 204 webpages with 60 diverse queries spanning different personas and intents. This tests whether an optimization actually makes a page more "citable" in general, or if it's just overfitting to a single search phrase.
- Variants: They also created MIMIQ-OOD (Out-of-Distribution) to test generalization to unseen user personas and MIMIQ-HTML to test robustness against messy, real-world web code.
Key Insights and Ideas
- Diagnosis Beats Generic Rules: The "aha moment" here is that citation failure is rarely a global quality problem. Most pages are "almost there" and just need a small, targeted fix.
- Generic Rules Can Harm "Long-Tail" Content: The study found that applying standard GEO "best practices" (like those used by baselines) can actually decrease citation rates for niche or specialized topics. Generic rules are derived from aggregate patterns that don't always apply to the "long tail" of the internet.
- The "Wikipedia" Wall: The authors admit that some failures are "unrecoverable." If a generative engine is hard-coded or biased to prefer high-authority sources for a specific fact, no amount of content optimization will force it to cite a smaller creator.
Key Results
- Higher Success, Less Change: AgentGEO achieved a 79.52% citation rate (using GPT-4.1-mini), a significant jump over the best baseline (AutoGEO at 68.80%). Remarkably, it did this while changing only 5% of the content, whereas baselines typically mangled 25% of the page.
- Robustness: The system showed strong "cross-method" transferability. If AgentGEO optimized a page to be cited by one type of AI engine, those improvements generally held up when tested on a completely different model architecture.
- Efficiency in Length: AgentGEO was particularly effective on "Very Long" documents (>10k words), where its ability to target specific chunks prevented the "lost in the middle" phenomenon common in LLMs.
Limitations and Future Work
- Cost and Latency: Because AgentGEO relies on an iterative agentic loop with multiple LLM calls, it is currently more computationally expensive than simple rewriting scripts.
- Black-Box Engines: The study was conducted on a controlled, simulated RAG (Retrieval-Augmented Generation) pipeline. While this allows for deep diagnosis, the authors note that commercial engines (like Google's) are "black boxes" with proprietary ranking signals that may require further reverse-engineering.
- Systemic Bias: The paper highlights that GEO alone cannot ensure "equitable visibility" if the underlying AI engines have deep-seated biases toward specific dominant platforms.
Summary updated