The Library of Babel

In Borges' The Library of Babel, the universe is an endless library. Every possible book exists somewhere inside it: every true history, every false one, every biography, every prophecy, every index that might explain where the useful volumes are hidden.

The horror of the story is not scarcity. It is abundance without orientation. The librarians are surrounded by total information and still cannot know what matters. A perfect catalog would save them, except that the catalog would also need to be found, interpreted, trusted, and probably cataloged again.

The internet is beginning to feel less like a public square and more like that library. The difference is that the new catalogers are not human. They are AI systems, and their cataloging decisions increasingly determine what people find before they ever reach a source.

Visibility is no longer ranking

Search used to expose its machinery more plainly. You typed a query, received a list of links, and did at least some of the work of choosing. That system had many problems, but it preserved a visible distance between the question, the sources, and the answer.

AI search collapses that distance. It reads, chooses, compresses, and speaks. The user often sees a finished answer before seeing the evidentiary trail. That makes visibility less like shelf placement and more like memory. A brand, publisher, institution, or idea has to become something the model can retrieve, cite, and synthesize.

Semrush's AI Visibility Index makes this shift concrete. Its initial study analyzed 2,500 prompts across major verticals, comparing how brands and sources appeared in systems like ChatGPT and Google AI Mode. The later 2026 expansion scaled the work to 126 million U.S. prompts. The point is not that one leaderboard is final. It is that discovery has become measurable in a new way: not simply by where a page ranks, but by whether it enters the answer at all.

Traditional SEO was a contest over the list. AI visibility is a contest over the summary.

Mentions and citations are different kinds of power

The useful distinction is between being mentioned and being cited.

A mention means the system has decided you belong in the answer. You exist in the cultural map. People talk about you, compare you, recommend you, complain about you, and the model has absorbed enough of that surrounding language to place you in the scene.

A citation means the system has decided your page can support a claim. You exist in the machinery of fact. Your content is structured, reachable, parseable, and authoritative enough to serve as evidence.

Those are not the same achievement. A company can be discussed constantly and still fail to become a source. Another can be cited repeatedly while remaining invisible as a named actor. One is folklore. The other is plumbing.

This matters because AI systems are assembling answers from both layers at once. The social layer tells them what belongs in the conversation. The source layer tells them what can be used to justify the conversation. Brands that live in only one layer become lopsided. They are famous but unsupported, or useful but anonymous.

The maps do not agree

The deeper problem is that different AI systems build different maps.

Ask two systems the same ordinary commercial question and they may mention similar names while grounding those answers in very different archives. One may lean on community discussion. Another may prefer professional reviews, official documentation, comparison sites, or structured databases. The answer can sound equally confident in both cases while resting on different assumptions about trust.

This is not just an SEO problem. It is an epistemic one.

We used to worry about filter bubbles because people with different preferences were shown different information. AI search introduces something stranger: people can ask neutral questions and receive definitive answers assembled from different source-worlds. The divergence happens before ideology, inside the retrieval and citation layer.

In Borges' library, every book exists, but most are useless noise. In AI search, the library is finite enough to be crawled but large enough that selection becomes destiny. The argument is no longer only about what is true. It is about which evidence gets admitted into the room where truth is being summarized.

Machine-legibility is literacy

There is a mundane technical lesson hiding underneath the philosophy: if machines cannot read you, they cannot remember you.

Static HTML, clean metadata, accessible pages, canonical URLs, structured markup, and crawlable documentation sound like chores. They are now part of public legibility. A beautiful JavaScript interface that hides the underlying content from crawlers is a locked cabinet in the library. Humans may open it. The new librarians may not.

This creates a strange inversion. The best human experience can still be a poor machine experience. A site can be visually polished, fast enough for visitors, and persuasive to a buyer while remaining opaque to the systems that increasingly mediate discovery. If the model cannot parse your documentation, terms, claims, authorship, or product details, it has little reason to cite you.

That is why the AI visibility conversation should not be reduced to prompt tricks. The durable work is infrastructural: publish useful pages, expose them plainly, make claims easy to verify, and keep the source of truth close to the surface.

The commons will be gamed

There is also a timing problem. AI systems currently draw signal from a web that was not entirely built for them. Reddit threads, Wikipedia pages, reviews, docs, forums, and old blog posts still contain a great deal of unoptimized human residue. That residue is valuable because it was not all produced for the metric now being measured.

Once AI visibility becomes a budget line, the signals will change.

Reddit will attract more synthetic enthusiasm. Wikipedia will see more pressure from reputation teams. Review sites will be shaped by the knowledge that they are not merely influencing readers but feeding answer engines. Documentation will be written for machines as much as developers. The commons will not disappear, but it will become more contested.

This is the old SEO story in a new medium. Measurement creates optimization. Optimization degrades the signal. The degraded signal requires new measurement. The catalog expands, then needs another catalog.

The test

The practical test is simple.

If people mention you but AI systems do not cite you, you have cultural presence without evidentiary authority. You are part of the story, but not part of the proof.

If AI systems cite you but do not mention you, you have evidentiary utility without identity. You help construct the answer, but someone else receives the memory.

If neither happens, you are outside the map.

If both happen, you have become part of the new catalog.

It is tempting to treat that as a marketing problem. It is larger than that. The same mechanics will shape how people learn about medicine, money, law, software, politics, and history. The winners will not simply be those with the best pages, or the loudest audiences, or the most backlinks. They will be those who become legible to both humans and machines without collapsing into noise.

The Library of Babel was terrifying because everything was already written and almost nothing could be found. Our problem is different. The books can be found. The question is which librarian you asked, what it could read, and what it decided was worth remembering.

Sources

[1] Semrush: AI Visibility Index

[2] Semrush: 2026 AI Visibility Index expansion

[3] Semrush: Google AI Mode comparison study

[4] Xu, Iqbal, and Montgomery: Measuring Google AI Overviews