2026 GEO Research Explained (I): Why Brand AI Visibility Needs Layered Measurement

"The brand has been mentioned by AI" is a starting point, not the conclusion of a GEO report.

The preprint *Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines*, made public on June 18, 2026, analyzed more than 100,000 AI search responses for over 100 brands from March through May 2026. Its most useful lesson for companies is not any single brand's appearance rate, but a more practical research framework: mentions, ranking, citation sources, and sentiment must be assessed separately.

This is the first article in the "2026 GEO Research Explained" series. Rather than discussing how to revise one article, it addresses the visibility baseline a company should establish before it begins optimization.

What brand visibility tiers did the paper find?

The study divided brands into three maturity tiers and reported mention rates in relevant AI responses from its first test: about 73% for globally recognized brands, 44% for established mid-sized or regional brands, and 11% for niche and small brands.

This does not mean every large brand will necessarily be recommended, nor that small brands have no opportunity. It shows only that, in the study's sample, question set, and observation window, existing brand awareness had a significant effect on initial visibility in AI search.

The real takeaway for companies is that they should not attribute a brand's initial absence directly to one weak piece of official-site content. Brand visibility in AI answers is jointly shaped by brand awareness, the source ecosystem, question types, platform retrieval mechanisms, and changes over time.

GEO projects should therefore distinguish between two goals. One is improving how a known brand is presented in competitor-comparison and purchase-decision questions. The other is getting a brand that AI has not yet understood into the candidate set for discovery, use-case, and category questions. The question sets, content, and evaluation criteria for these goals are different.

Why 78% of company-site citations does not mean official sites hold 78%

The paper reports that, when AI search responses provided citations, about 78% of citations fell into the company-website category. But the authors further note that most of these were third-party company sites, not the official sites of the brands analyzed. In their summary table, brand-owned websites accounted for only a small slice.

This finding is easy to misread as "expand the official site and it will receive more AI citations." A more accurate interpretation is that AI often uses information published by companies, product providers, retailers, reviewers, and industry service providers, but the information ecosystem is not the same as one brand's official website.

Companies need to manage three layers of sources at once.

The first layer is owned factual sources, such as product pages, pricing pages, help centers, case-study pages, specifications, and service-scope pages. They provide accurate, verifiable, and updateable primary information.

The second layer is third-party validation sources, such as media reviews, industry directories, specialist organizations, partners, and retail channels. They add independence and comparative context.

The third layer is discoverable content formats. The study found that ranked "best-of" lists were among the most frequently cited page formats in its sample. This suggests that companies should check whether they lack content entry points that can be externally evaluated, compared, and aggregated; it is not a reason to mass-produce ranking lists.

Why sentiment should not be merged with mentions into one score

The paper observed a 45.5% flip rate for sentiment labels and a 6.8% flip rate for mention signals, making the former about 6.7 times higher. In other words, whether AI mentioned a brand was relatively stable in this study, while how positively or negatively it described the brand was much less stable.

This matters for monthly GEO reports. If appearance count, ranking, and sentiment are combined into one overall score, high sentiment volatility can hide real visibility changes or create anomalies that do not actually exist.

A more useful approach is to divide metrics into three groups:

1. Coverage metrics: whether the brand appears, which questions it appears in, and which competitors co-occur with it.
2. Evidence metrics: which pages AI relies on, whether sources are outdated, and whether official and third-party evidence agree.
3. Presentation metrics: what recommendations, reservations, risk alerts, and sentiment tendencies are present.

With this approach, when a brand suddenly appears more negatively in AI answers, the team can first determine whether the cause is fewer mentions, replaced sources, changed facts, or variation in model wording, instead of rushing to change the entire site.

Research boundary: an observational baseline, not a promise of optimization results

The paper's authors explicitly state that the research is an observational baseline and does not support the conclusion that a given recommendation has causally increased brand visibility. Randomized controlled trials and closed-loop protocols remain part of the research agenda under development. Brand tiers also included manual coding by the researchers, and sentiment volatility may come from both model output and classifier noise.

For that reason, 73%, 44%, and 11% should not be used as industry promises or corporate KPIs. They are better used to remind teams that AI search does not begin on equal footing, and that different signals have different levels of reliability.

How companies can turn this paper into a monitoring practice

Do not start by pursuing a single "AI visibility score." First, establish a fixed question set for the brand and record brand mentions, recommendation position, citation sources, competitor co-occurrence, and answer presentation separately. For priority pages, maintain a source ledger that records which questions lead AI to use official sites, third-party media, product channels, and outdated pages.

GEO Radar at https://www.georadar.top can support multi-platform AI search visibility analysis, repeat testing of fixed question sets, competitor comparison, and structured reporting. It helps teams see how answers differ across platforms; whether a content change has produced improvement should still be determined through controls, records, and repeated tests.

The first GEO principle from this study is simple: establish a layered baseline before discussing the scale of optimization.

Sources for this article

• arXiv, June 18, 2026, *Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines*: https://arxiv.org/abs/2606.20065
• arXiv HTML full text, data scope, metrics, findings, and limitations: https://arxiv.org/html/2606.20065v1
• arXiv PDF, paper download: https://arxiv.org/pdf/2606.20065