Citation Wars: Can GEO Optimization Make Content Worse?

What the citation-wars paper actually tested

Chen Xu, Zitian Guo and Chenyan Xiong model GEO as a repeated Stackelberg game with partial monitoring. Publishers adapt documents after observing citation outcomes, while an answer engine chooses sources. The authors test a defense called Verifiable Content Reward, or VCR, that rewards factual substance that can be checked rather than surface tactics alone.

Across three benchmarks, the paper reports a 12.1 percentage-point improvement in Net Defense Utility over its strongest baseline. That is a result inside the paper’s experimental setup. It is not a measured uplift for an ordinary publisher and should not be converted into a promise about traffic, rankings or revenue.

Use the public repository without overstating it

The authors link a public reference implementation. Its README describes E-commerce, GEO-Bench and Researchy-GEO datasets, a five-round configuration, Gemini 2.5 Flash-Lite as the default answer engine, and GPT-4o mini for extraction or judging roles. The example commands describe 1,000 test examples and 3,000 training examples.

Those are repository implementation details, not universal properties of every result or production answer engine. The repository also offers a deterministic offline demo, which is useful for checking workflow plumbing without an API key. Reproduction still requires recording the commit, configuration, datasets, provider state and deviations from the authors’ setup.

Recognize citation over-optimization in an edit

• Certainty rises while evidence stays flat: “may help” becomes “will rank” without a new source.
• Caveats disappear: scope, date, population or failure conditions are removed to make a sentence easier to quote.
• Source density becomes decoration: links multiply, but the cited pages do not support the nearby claims.
• The user job fragments: one useful page becomes several thin URLs because each sentence is treated as a citation target.
• Measurement collapses: a mention, citation, click and conversion are reported as one visibility score.

The danger is not concise writing. The danger is optimizing the visible signal while discarding the substance that made the answer worth citing.

A worked repair: from quotable certainty to bounded evidence

Illustrative unsafe rewrite: “Adding statistics and schema makes any article more likely to be cited by AI engines.” It sounds decisive, but it invents a universal effect, merges two treatments and supplies no denominator or platform scope.

Evidence-first repair: “For this 20-page test, we added one source-backed statistic to ten comparable pages and left ten unchanged. We will record citation presence for a frozen prompt set over four weekly observations. The test cannot isolate effects on platforms outside the recorded set.”

The second version is less dramatic but more useful: it names the intervention, comparison, sample, observation window and limit. This example is an editorial demonstration, not a result from the paper or from a SearchEngineAnswer experiment.

Run a quality-preserving AEO test

1. Define the user job: write the decision or task the page must support before changing it.
2. Freeze the treatment: change one meaningful element, such as claim-level sourcing or a worked example;not length, schema, headings and links together.
3. Declare the quality floor: require zero unsupported material claims, intact caveats, accessible references and a subject-aware reviewer.
4. Freeze the prompt set: preserve wording, platform, mode, locale, account context and observation schedule.
5. Record the full denominator: keep uncited, unavailable and contradictory responses instead of saving only favorable examples.
6. Use a rollback rule: revert or repair when usefulness or factual support declines, even if citation presence rises.

Pair the claim-level citation workflow with the small SEO experiment method. The first protects the passage; the second protects the inference.

Score the release gate without pretending it is universal

A practical editorial gate can require 100% of material factual claims to map to evidence, zero unsupported material claims, preserved limitations and a named reviewer. Those thresholds are publishing controls, not ranking factors. Add job-specific checks for legal, medical, financial or safety-sensitive content.

Download the GEO rewrite quality-gate worksheet. Its rows are marked EXAMPLE-REMOVE so examples cannot silently become reported evidence.

Google’s people-first content guidance asks whether a page provides original information, substantial value, clear sourcing and a satisfying experience. It also says Google has no preferred word count. That is why “make it longer” is not a recovery method; stronger evidence, examples, decisions and limits are.

The decision rule

Keep an AEO edit only when it improves or preserves the reader’s ability to verify and act. If a variant wins more citations but loses a condition, weakens a source, invents certainty or fragments the task, it has failed. Our AI search ranking evidence register helps separate documented requirements, local observations and unknown selection mechanics.

Verification note: we rechecked the preprint and its linked repository on September 3, 2026. The paper was submitted on August 11, 2026 and had not been treated here as peer-reviewed production-platform evidence.

Primary documentation and reproducibility

• Xu, Guo and Xiong: The Citation Wars preprint
• Authors’ public reference implementation
• Google: creating helpful, reliable, people-first content