Diagnosing and Repairing Citation Failures in Generative Engine Optimization

This paper introduces AgentGEO, a diagnostic framework that identifies specific citation failure modes in Generative Engine Optimization and applies targeted, iterative repairs to achieve over 40% relative improvement in citation rates while modifying minimal content.

The Gist

Imagine the internet is a massive, bustling library. In the old days, if you asked a librarian (a traditional search engine) a question, they would hand you a stack of index cards with the titles of the best books. You'd have to walk over, pick up the book, and read it yourself.

Today, we have Generative Engines (like advanced AI chatbots). Instead of handing you a stack of cards, the AI reads the books itself and writes a summary essay for you right there on the spot. It's incredibly convenient.

But here's the problem:
When the AI writes its essay, it often forgets to give credit to the specific books it used. If it doesn't write a footnote or a link (a citation), the original author of that book gets no traffic, no money, and no recognition. It's like the AI is the only one getting paid for the story.

This paper is about a new system called AgentGEO that helps authors get those missing footnotes back.

The Old Way: "Spray and Pray"

Before this paper, people tried to fix this by guessing. They thought, "Maybe if I just make my writing sound more professional, or add more statistics, the AI will like me."

They applied these generic rules to everything. It was like trying to fix a broken car engine by spraying WD-40 on the whole vehicle. Sometimes it helped, but often it made things worse because the problem wasn't "lack of polish"—it was something specific, like a flat tire or a disconnected wire.

The New Way: The "Detective Mechanic" (AgentGEO)

The authors of this paper realized that every time a webpage fails to get cited, it's for a different reason. They built a system called AgentGEO that acts like a detective mechanic.

Instead of guessing, AgentGEO follows a three-step process:

1. The Diagnosis (The Detective):
   When a webpage isn't getting cited, AgentGEO doesn't just rewrite the text. It compares the webpage to the ones that did get cited. It asks: "Why did the AI pick that other page and ignore mine?"

   • Analogy: Imagine you applied for a job and didn't get it. Instead of just rewriting your resume with fancy fonts, you ask the hiring manager, "Did I lack experience? Was my cover letter too long? Did I miss a specific skill?" AgentGEO finds the specific reason.

2. The Repair (The Mechanic):
   Once the problem is identified, AgentGEO has a toolbox of specific fixes. It doesn't use a sledgehammer; it uses a scalpel.

   • If the problem is technical: Maybe the website is hidden behind a login wall or the text is buried in messy code. AgentGEO fixes the code so the AI can actually "see" the content.
   • If the problem is intent: Maybe the user asked "How to buy a cake" but the page only talks about "The history of baking." AgentGEO rewrites the first sentence to say, "Here is where you can buy a cake," immediately.
   • If the problem is structure: Maybe the facts are there, but they are hidden in a giant wall of text. AgentGEO turns them into a neat bullet-point list or a table, making it easy for the AI to grab.

3. The Test (The Quality Control):
   After the fix, AgentGEO tests the page again. If it still doesn't get cited, it goes back to step 1, learns from the failure, and tries a different tool. It keeps looping until the page gets the citation.

Why This Matters: The "Long-Tail" Problem

The paper found something surprising. The old "generic" methods often hurt smaller, niche websites (the "long-tail"). They would try to force a medical advice page to sound like a news report, which made it look fake and untrustworthy to the AI.

AgentGEO is different because it respects the original content. It only changes what is broken.

• The Result: AgentGEO increased citation rates by 40% (a huge jump!) while only changing 5% of the actual text. The old methods changed 25% of the text and got much worse results.

The "MIMIQ" Benchmark

To make sure their system actually works, the authors built a new testing ground called MIMIQ.

• Old Tests: "Here is a webpage and one specific question. Can you make it rank for this question?" (This is like studying for a test by memorizing the answers to the practice questions).
• MIMIQ: "Here is a webpage. Now, imagine 60 different people asking 60 different versions of a question about it. Can your fix work for all of them?" (This is like learning the subject so well that you can answer any question, even the ones you haven't seen before).

The Bottom Line

The paper teaches us that diagnosis is better than guessing.

If you want your content to be seen by AI, you don't need to rewrite your whole life story to sound like a robot. You just need to find out exactly why the AI is ignoring you—whether it's a technical glitch, a confusing layout, or a mismatched topic—and fix that specific thing.

AgentGEO is the tool that does this detective work for you, ensuring that when the AI writes its summary, it gives credit where credit is due.

Technical Summary

Here is a detailed technical summary of the paper "Diagnosing and Repairing Citation Failures in Generative Engine Optimization."

1. Problem Statement

Generative Engine Optimization (GEO) aims to improve the visibility of web content within AI-generated responses (e.g., Google AI Overviews, Perplexity, ChatGPT). While existing GEO research focuses on contribution metrics (how much text from a source appears in the answer), the authors argue this is insufficient for content creators. The critical metric is citation (the binary event of a source being linked), as this is the primary pathway for traffic generation.

Current GEO methods suffer from two main limitations:

1. Generic Rewriting: Existing approaches apply uniform rules (e.g., "add statistics," "improve fluency") to all content. They fail to diagnose why a specific document was not cited.
2. Pipeline Blindness: Citation failures occur at various stages of the generative engine pipeline (fetching, parsing, generation). Generic rules cannot address upstream technical failures (e.g., JavaScript rendering issues, access blocking) or specific semantic mismatches.

The paper posits that citation failure is heterogeneous; a document might fail due to technical integrity issues, semantic misalignment, poor content quality, or systemic exclusion (e.g., competition from higher-authority sources).

2. Methodology: AgentGEO

The authors propose AgentGEO, an agentic framework that shifts the paradigm from "generic improvement" to "diagnostic repair." The system operates on a Diagnose-then-Repair loop.

A. Citation Failure Taxonomy

The framework is built upon the first systematic taxonomy of citation failures, categorized into four dimensions spanning the generative engine pipeline:

1. Technical Integrity (10.1%): Failures in fetching or parsing (e.g., access blocking, JS rendering failures, unparseable content, low signal-to-noise ratio due to boilerplate).
2. Semantic Alignment (62.2%): Content does not match query intent (e.g., intent divergence, missing entities, outdated info, localization mismatch).
3. Content Quality (27.1%): The topic is correct, but presentation is poor (e.g., information scarcity, fragmentation, excessive verbosity, unstructured layout).
4. Systemic Exclusion (0.6%): Structural disadvantages beyond content control (e.g., competitive redundancy with Wikipedia, context window truncation).

B. The AgentGEO Workflow

1. Diagnosis: For a target webpage failing to be cited, the agent simulates the GE pipeline. It compares the target against a cited competitor to identify the specific failure mode using the taxonomy.
2. Tool Selection: Based on the diagnosis and a query-specific memory (to avoid repeating failed attempts), the agent selects a specialized repair tool from a library.
   • Tools include: noise_isolation (wrapping ads/nav in semantic tags), intent_realignment (rewriting intros to match query), entity_injection (adding missing facts), data_serialization (converting text to tables), and bluf_optimization (putting key takeaways first).
3. Iterative Refinement: The agent applies the tool to a surrogate copy of the webpage and re-tests. If citation is achieved, the edit is recorded; if not, the failure is logged in memory, and the loop repeats.
4. Batch Aggregation: To prevent overfitting to specific queries, suggestions from multiple training queries are aggregated. Conflicts are resolved by prioritizing edits based on failure severity, resulting in a single, query-invariant optimization for the original webpage.
5. Localized Editing: Modifications are applied at the chunk level (HTML segments) rather than rewriting the whole page, preserving semantic integrity and fidelity to the original content.

3. Key Contributions

1. Diagnostic Approach (AgentGEO): A novel agentic system that iteratively diagnoses specific failure modes and applies targeted repairs, rather than applying generic rewriting rules.
2. Citation Failure Taxonomy: The first comprehensive classification of citation failures across the entire GE pipeline (fetching, parsing, generation), providing a foundation for future optimization tools.
3. MIMIQ Benchmark: A Multi-Intent Multi-Query document-centric benchmark. Unlike previous benchmarks that pair one document with one query, MIMIQ associates each document with 60 diverse queries (varying in intent, persona, and phrasing). It includes train/test splits to evaluate generalization and includes variants for Out-of-Distribution (OOD) and HTML structural robustness.
4. Tool Library: A suite of 9 specialized optimization tools addressing specific taxonomic failures (e.g., structural enhancement, information augmentation, persuasive refinement).

4. Experimental Results

The authors evaluated AgentGEO against baselines (Vanilla, GEO-Bench methods, AutoGEO) using LLMs (GPT-4.1-mini, Claude).

• Citation Rate Improvement: AgentGEO achieved a 79.52% citation rate in in-context generation, outperforming the best baseline (AutoGEO at 68.80%) by 10.72 percentage points. This represents a >40% relative improvement over baselines.
• Content Efficiency: AgentGEO modified only 5% of the original content on average to achieve these gains, whereas baselines altered ~25%. This proves that citation failures are often local issues requiring targeted fixes, not global rewrites.
• Generalization: On the MIMIQ benchmark, AgentGEO generalized well to held-out queries and OOD personas, whereas generic methods often overfit or degraded performance on long-tail content.
• Faithfulness: AgentGEO maintained high fidelity to the original content (TF-IDF similarity >94%), ensuring the optimization did not hallucinate or distort facts.
• Failure Analysis: The study revealed that for some documents (e.g., those facing dominant competitors like Wikipedia), optimization alone cannot overcome systemic exclusion, highlighting a limit of content-side interventions.

5. Significance and Implications

• Shift in Optimization Strategy: The paper demonstrates that "diagnosis beats generic rules." Understanding the specific failure mode is crucial for effective GEO.
• Equity in Visibility: The findings suggest that generic optimization rules can inadvertently harm long-tail or specialized content. Diagnostic optimization offers a more equitable path to visibility by addressing specific structural or semantic deficits.
• Systemic Limits: The research identifies that not all citation failures are recoverable. If a document is systematically disadvantaged by the engine's bias toward high-authority sources, content optimization alone cannot ensure visibility. This calls for broader scrutiny of how citation mechanisms function in AI-mediated information access.
• Technical Robustness: By addressing HTML parsing and fetching issues (not just text), AgentGEO offers a more holistic solution to the challenges posed by modern generative engines.

In conclusion, AgentGEO provides a scalable, diagnostic framework that significantly improves citation rates with minimal content alteration, offering a new standard for optimizing web content in the era of generative AI.