Why Can GEO Become a Citation War? The Risk of Repeated Competition

One rewrite may look harmless. Five rounds of editing around “what gets cited” can produce a very different result.

The paper *Mechanism Design for Generative Engines: From Exploitation toward Win-Win Outcomes*, posted on August 11, 2026, treats GEO as a repeated game between content suppliers and a platform rather than a one-time copywriting exercise.

Its central risk is a citation war: each side adapts to the previous round until visibility competition displaces genuine content improvement.

Why a one-shot uplift test misses the problem

Many GEO experiments compare citation visibility before and after a rewrite. In a live ecosystem, platforms can add warnings, filters, or reranking. Suppliers then learn to adapt to those defenses.

The study simulates five rounds of this interaction. In an e-commerce benchmark, the supplier uses AutoGEO and the platform applies a conventional prompt-warning defense. Mild early edits evolve into authority framing, unsupported quantitative claims, and fabricated detail. Both rewritten-document quality and user-side answer utility fall below the no-exploitation reference.

This is a controlled simulation, not evidence that a named commercial engine follows the same trajectory. It exposes a dynamic risk that a single before-and-after test cannot measure.

Why penalty-only defenses can stall

A platform that penalizes manipulation-looking features faces a tradeoff. Clear structure and fact density can appear in genuinely helpful content too. An aggressive penalty suppresses useful evidence; a weak one becomes easy to adapt around.

Suppliers can also hide optimization intent in more natural prose. The paper’s local model describes an “inert stationary outcome”: defense effectiveness fades, suppliers stop producing real quality gains, and further rounds generate little joint value.

Signs of a citation arms race inside a content program

Pause and review when several of these signals appear:

• each revision is driven mainly by phrases seen on competing pages;
• unsupported terms such as “leading,” “expert recommended,” or “industry standard” keep accumulating;
• new sites are created chiefly to cite one another without clear editorial ownership;
• reporting rewards mentions but does not track factual drift or answer quality;
• repeated edits lack version diffs, evidence owners, and approval records;
• a short-term gain on one engine is presented as a durable cross-platform rule.

Change the objective to verifiable factual gain

A safer iteration should answer four questions: What useful fact was added or clarified? Which source supports it? What scope and limitation apply? Did the change improve a user’s ability to decide?

Teams can establish stop conditions. If a revision adds only promotional language, creates no factual gain, or causes AI answers to lose important qualifications, it should not proceed. Visibility, factual integrity, and user utility need to pass together.

GEO Radar supports cross-platform brand and competitor monitoring at [https://www.georadar.top](https://www.georadar.top). Its results are best used to identify changes that deserve investigation, not to drive endless adversarial rewriting. Page versions and evidence records remain essential.

Keep the research boundary visible

The study uses a local quadratic model, one platform and creator population, finite rounds, and benchmark environments. It does not establish a global equilibrium for real platforms, and it leaves paid placement and multi-platform competition for future work. “Citation war” is a governance model, not a factual description of a particular engine’s algorithm.

Sources for this article

• arXiv, August 11, 2026, *Mechanism Design for Generative Engines: From Exploitation toward Win-Win Outcomes*: https://arxiv.org/abs/2608.11390
• arXiv HTML paper, including the repeated game, five-round experiments, citation competition, and limitations: https://arxiv.org/html/2608.11390v1
• Authors’ public code and experiment materials: https://github.com/cxcscmu/GameTheory-GEO