2026 GEO Research Explained (II): Why AI Answers Cite a Competitor Before You

An AI answer citing you does not mean you have won that comparison. Whom it cites first often determines who first defines the user's decision framework.

The paper *What Gets Cited: Competitive GEO in AI Answer Engines*, made public on May 25, 2026, narrows the question to a specific competitive scenario: when two sources both enter a model's context, what makes one more likely to become the first citation in the answer?

The paper does not study how to manipulate a system. In a controlled setting, it isolates content signals. Its most valuable contribution for companies is to break down whether a page is effective into source elements that can be checked.

How the study put two sources into the same competition

The authors started with 100 anonymized product-review articles from 50 categories and constructed pairwise source pages. In each pair, only one factor differed between the two versions, such as topical fit, information completeness, price, date, trust signals, format, or list position.

In the experiment, the system provided the model with only these two candidate sources and controlled for position bias by swapping their order. The authors also anonymized brands and publishers to reduce familiarity effects from the model's pretraining memory as much as possible.

The study covered 6 LLMs, 18 content factors, and 252,000 trials, measuring which source the first citation marker in the answer pointed to. This design cannot reproduce the full complexity of the real internet, but it can answer relatively clearly how a single content difference affects preference for the first citation.

Which factors most affect the first citation?

The paper's conclusion is measured, and practical.

Topical match and retrieval-list position were the largest influences on the first citation. Clear price information and a more recent timestamp also provided consistent benefits. Completeness and trust signals produced relatively smaller gains, while changes to formatting alone had limited effects.

This means that when a company finds that a competitor is consistently cited first by AI, its first response should not be to add more adjectives, bold more paragraphs, or make the entire page sound more "expert." Check four more fundamental issues first:

1. Does the page directly answer the current question, rather than merely describe the product itself?
2. Can the page reliably enter the retrieval candidate pool, and what is its position in that pool?
3. Are the price, specifications, conditions of use, and comparison dimensions clear, verifiable, and current?
4. Does the page provide sufficiently complete conclusions, limitations, and evidence so that AI does not need a competitor's page to complete the answer?

Why price and date can become content signals

In real procurement, shopping, and service comparisons, users care not only about what something is, but also about how much it costs, whether it fits, and whether the information is still valid. The stronger performance of price and recent timestamps in the paper should be understood as an effect of decision usefulness and freshness, not as evidence that arbitrarily adding a date to a page can increase citations.

Prices need a basis: is it a starting price, package price, trial price, or custom quote? An updated date must be genuine, and the page's content must be updated accordingly. If price, inventory, version, or service scope is outdated, an AI citation can still give users incorrect advice and ultimately damage brand credibility.

For B2B companies that cannot publish full pricing, they can still provide the billing unit, variables that affect price, implementation timeline, and applicable packages. For consumer brands, specifications, after-sales service, delivery coverage, and purchase channels are often equally important.

This is not an optimization tactic for the first citation alone

The first citation is a useful diagnostic signal, but it should not become the only goal. The final answer a user sees is usually composed from multiple sources. A page may not be the first citation, yet still supply critical facts, risk boundaries, or purchase conditions.

The paper's authors also clearly note that the experiment injected only two candidate pages, while real RAG systems often retrieve five to ten or more sources. The findings therefore describe factor preferences in controlled two-source competition, not the final rule for complete multi-source competition.

Companies should use the research to build a review checklist, rather than treating it as a formula that guarantees citations.

A practical method for reviewing source pages

Choose 10 pages that AI commonly uses to answer brand questions, including official product pages, pricing pages, case-study pages, help-center pages, third-party reviews, and product-channel pages. Ask five questions about each page:

First, which user question does it correspond to? Does the opening of the page directly state the conclusion and intended audience?

Second, are the price, specifications, version, region, and update date on the page sufficient to support a purchase decision?

Third, does the page contain distinctive evidence, such as a test methodology, data definitions, case conditions, certification, or a real comparison?

Fourth, if AI read only this page, could it correctly explain what situations the offering is not suitable for?

Fifth, is a competitor's page more complete, more current, or a better fit for the question on any of the points above?

When the review is complete, retest a fixed question set across multiple platforms and observe whether the first source, brand mentions, recommendation reasons, and competitor co-occurrence change. GEO Radar at https://www.georadar.top can help teams conduct multi-platform retesting and competitor comparisons, turning "whom AI cites first" from an incidental screenshot into a trackable signal.

The takeaway for content teams from this study is plain: to compete for AI citations, first make the page a complete, current evidence unit that can directly answer the question.

Sources for this article

• arXiv, May 25, 2026, *What Gets Cited: Competitive GEO in AI Answer Engines*: https://arxiv.org/abs/2605.25517
• arXiv HTML full text, two-source experiment, 252,000 trials, results, and limitations: https://arxiv.org/html/2605.25517v1
• arXiv PDF, paper download: https://arxiv.org/pdf/2605.25517