GEO without the hype: what AI search visibility actually depends on

Every few months a new acronym arrives promising a shortcut into AI answers. GEO, AEO and LLMO all describe the same ambition: to be the source an AI system draws on when it answers a question about your field. The ambition is reasonable. Much of the advice attached to it is not.

This article separates what has public evidence behind it from what is still guesswork. It is written for companies deciding how much attention, and budget, AI search visibility deserves.

What Google actually says

Google's own documentation on AI features in Search, covering AI Overviews and AI Mode, is unusually direct. It states that there are no additional requirements to appear in these features and no special optimisation needed beyond the existing SEO fundamentals. To be eligible as a supporting link, a page must be indexed and eligible to be shown in Search with a snippet.

That does not make AI search irrelevant. It means the foundation is the same one that has always mattered: pages that can be crawled, are technically sound, and answer questions clearly and credibly. Anyone selling a separate “AI ranking” layer for Google should be able to explain why Google's own guidance is wrong.

What the research shows

The term GEO comes from an academic paper by Aggarwal and colleagues, published at KDD 2024. The authors built a benchmark of queries, applied different rewriting strategies to source content, and measured how visible that content became in generated answers. Adding citations to sources, quotations from relevant authorities and concrete statistics improved visibility in their benchmark. Traditional keyword stuffing did not help much.

Two caveats matter. First, the effect varied by domain: a strategy that worked well for one type of query did little for another. Second, a benchmark is a controlled environment, not a live commercial search engine whose systems change continuously. The paper is a useful signal about what generative systems tend to favour, namely content that is specific, sourced and quotable, not a recipe with predictable returns.

Brand mentions: a correlation worth understanding

Several SEO software vendors have published studies comparing which brands appear in AI answers with signals such as backlinks, branded search volume and mentions of the brand on other websites. A recurring finding is that brands mentioned widely across the web tend to be more visible in AI answers, in some studies more strongly than backlinks alone would predict.

These are correlations from vendors with an interest in the topic, using their own datasets and methods. They do not show that mentions cause visibility. Well-known companies are mentioned more and are also more likely to have strong content, links and demand. The practical reading is modest: being discussed accurately and often in credible places is good for a company, and AI systems appear to reflect that.

llms.txt: harmless, optional, unproven

llms.txt is a proposed convention for a plain-text file that summarises a site for language models. It is cheap to create, and we publish one on our own site as a courtesy. We have not seen public evidence that it affects whether or how often a site is cited, and major AI search providers have not documented that they rely on it. Treat it as a tidy extra, never as a lever.

Evidenced versus speculative

Evidenced

Pages must be crawlable, indexable and snippet-eligible to be used by Google's AI features (Google documentation)

Evidenced

Citations, quotations and statistics improved visibility in a controlled GEO benchmark, with effects varying by domain (Aggarwal et al., 2024)

Correlational

Widely mentioned brands tend to appear more in AI answers (vendor studies, causation not shown)

Speculative

llms.txt, special “AI-friendly” markup and prompt-shaped content as ranking levers

Unsupported

Guaranteed citations or placement in any AI system

Crawlability still decides who gets considered

AI systems can only cite what their crawlers can read. That sounds obvious until you look at modern immersive websites, where copy is sometimes rendered inside a canvas, loaded only after interaction, or hidden behind client-side scripts. To a crawler, those pages can be nearly empty.

It is also worth checking robots.txt. Several AI providers now distinguish between crawlers used for model training and those used for search and retrieval. OpenAI, for example, documents separate user agents for each. Blocking training while allowing search crawlers is a legitimate choice, but it should be a choice, not an accident left over from a blanket rule.

What we actually do

• Entity clarity: every important page states plainly who the company is, what it does, for whom and where, in sentences that can be quoted without context.
• Factual, citable content: specific claims, named sources and real numbers where they exist; qualitative statements where they do not.
• Crawlable experiences: all copy lives in server-rendered HTML, even when the page is built around 3D or motion. The visual layer enhances the content; it never contains it.
• Consistent brand facts: the same name, description and contact details in the visible page, the structured data and external profiles.
• Honest measurement: Search Console data, referral traffic from AI assistants where analytics can identify it, and periodic manual checks of how the company is described. We report what we can see and say what we cannot.

Sources

• Google Search Central — AI features and your website: developers.google.com/search/docs/appearance/ai-features
• Aggarwal et al. — GEO: Generative Engine Optimization (KDD 2024): arxiv.org/abs/2311.09735
• llms.txt proposal: llmstxt.org
• OpenAI — Overview of OpenAI crawlers: platform.openai.com/docs/bots