What Is LLMO? Large Language Model Optimization — How It Works, Technical Requirements, and How It Differs From SEO [2026]

Who this is for

Marketing, PR, and SEO practitioners who heard the term "LLMO" and came to look it up, plus anyone working with AI search or LLMs. It assumes a beginner-to-intermediate reader and also serves somewhat more technical readers. If you already know AIO and GEO, it's a good way to reorganize LLMO in your mind as a story about the model layer.

Key figures

• ChatGPT has 900 million weekly active users (February 2026, per OpenAI's announcement / TechCrunch reporting). "Where people get their answers" is shifting toward the model side.
• In a study that asked eight AI search engines to identify sources, more than 60% of the answers were wrong (March 2025, Tow Center, Columbia University). An LLM's citations tend to be inaccurate, and misattribution does happen in practice.

To restate the definition:

LLMO (Large Language Model Optimization) refers to optimizing your information so that, in the process by which a large language model (LLM) generates its answers, that information is easy for the model to correctly retrieve, understand, and cite.

That single sentence is the core of this article, in a form an AI can easily extract. The key point is that it focuses on the model layer.

Rather than a standalone, independent field, LLMO is more accurately understood as one of the several names for optimizing AI search—the one that keeps its footing on the model itself. If AIO is "the umbrella concept that bundles the entire AI-search surface (Google's AI Overviews / AI Mode, ChatGPT Search, Perplexity, Gemini, and so on)," then LLMO sits inside that umbrella, with its viewpoint pulled toward the layer of "how the LLM handles information."

It overlaps with the neighboring concept of GEO, but the focus differs. GEO mainly looks at the output-side visibility of "how your brand gets cited or displayed inside a generative engine (AI search)." LLMO, by contrast, keeps its footing on the mechanism inside the model—how an LLM retrieves, generates, and cites information—and the technical conditions for delivering information to it. Even though both deal with "being cited by a model," GEO looks at the result (visibility) while LLMO looks at the mechanism just upstream of it—that's the sense of distance between them. We touch on the detailed differences lightly later and leave the deep dive to separate articles.

Organizing the terms with LLMO as the axis looks like this:

This is an overview showing the distance from LLMO to the other concepts. The detailed differences per term are left to What Is AIO, What Is GEO, and the individual "difference" articles (coming soon).

LLMO (Large Language Model Optimization) is a name that came into use on the ground in practice and the media. One relatively early documented use is a reference to the term by the trade publication Search Engine Land in October 2023. What matters, though, is this: LLMO has no single academic origin the way GEO does.

GEO was clearly formalized as "Generative Engine Optimization" in a 2023 academic paper (Aggarwal et al., "GEO: Generative Engine Optimization," arXiv:2311.09735, KDD 2024), so you can trace it back to a starting point. LLMO, on the other hand, began to be used in parallel by multiple writers as a practitioner-side label for the idea of "optimizing for LLMs." The way the term's contours wobble within the industry stems from this difference in provenance.

With that in place, there's something you must always keep in mind as a working premise. LLMO, GEO, AIO, and AEO are none of them official standards defined by Google, OpenAI, or anyone else. They are practitioner terms, and abroad you'll also see framings like "AEO and GEO are functionally near-synonyms" and "AI visibility is the most neutral umbrella term." This article, too, treats LLMO not as any platform's official specification but as "the practical domain of raising how easily your information gets retrieved, understood, and cited, in light of how LLMs handle information."

When it comes to Google in particular, you need to draw a line first before discussing the model layer. Google's AI-generated answers (AI Overviews / AI Mode) also pull from the same index as ordinary search, and there is no separate, AI-only ranking. Google itself explains in its official documentation that for AI features, too, the fundamentals of traditional SEO—useful, original content, crawlability, and quality—remain the foundation as they are. In other words, LLMO does not begin as "a brand-new magic tactic"; it begins from the understanding that you build on the existing foundation once you understand the mechanism by which models handle information.

To grasp LLMO as a practice, you first need to understand the flow by which an LLM produces an answer. Here we focus not on the outward shape of the generative engine but on the surrounding process by which the model retrieves, generates, and cites external information—from retrieval through generation to citation. (Outside practitioners can't control the model's internals themselves, so we focus on the process by which information reaches the model and gets selected.)

There are broadly two sources from which an LLM assembles its answer.

One is trained knowledge (parametric knowledge): knowledge the model took inside itself from training data and holds as parameters—the part where it answers without referring to anything external. The other is retrieval-augmented generation (RAG): a mechanism where, at generation time, the model searches for and retrieves external documents and answers in light of their contents. Answers that come with source links—like ChatGPT Search, Perplexity, or Google's AI Overviews—involve this RAG-style retrieval.

What often gets overlooked here is that retrieval and citation are separate stages. Even when the model retrieves many candidate pages, it only cites some of them in the final answer. In fact, according to observations by Profound, which works on AI-visibility analysis, about 85% of the pages ChatGPT retrieves never appear in the final answer. "Retrieved" does not equal "cited"; there is a large gap between retrieval and citation.

This two-layer structure is exactly why LLMO has both a "getting into training data" side and a "getting retrieved by RAG/search" side. The former works over the long term and indirectly (how your information is reflected in the model's next training). The latter works over a relatively short term and directly (whether your pages get picked up by the retrieval running right now). When thinking about LLMO, capturing these two separately is the starting point. The next chapter deals with a more fundamental problem lurking in the "citation" stage of that flow.

The most essential fact for understanding LLMO is that an LLM's citations are inherently incomplete. This is a point that rarely takes center stage in explanations of AIO or GEO, but in LLMO it is the very core.

A representative benchmark for evaluating this incompleteness is ALCE, from the Princeton NLP group (Gao, Yen, Yu, Chen. EMNLP 2023, arXiv:2305.14627). ALCE is a framework that automatically evaluates an end-to-end system—one that retrieves evidence supporting its claims and generates answers with citations—along three axes: fluency, correctness, and citation quality. Under this framework, it has been shown that even state-of-the-art LLMs leave many cases in which the citations they present do not adequately support their claims.

The incompleteness of citations has also been quantified in studies that manually evaluated generative search engines actually in operation. Liu et al., "Evaluating Verifiability in Generative Search Engines" (Nelson F. Liu, Tianyi Zhang, Percy Liang. EMNLP Findings 2023, arXiv:2304.09848) manually audited four then-representative generative search engines (Bing Chat, NeevaAI, perplexity.ai, and YouChat) and reported that, on average, only 51.5% of generated sentences were fully supported by their citations, and the share of presented citations that actually supported the sentence they were attached to was 74.5%. Those are primary figures: nearly half of the sentences are not adequately supported by their citations.

This "citations that can't fully support the claim" problem has been confirmed repeatedly in later measurements as well.

• The Tow Center's study at Columbia University (March 2025) gave eight AI search engines (ChatGPT Search, Perplexity, Gemini, and others) excerpts of articles and asked them to identify the headline, outlet, date, and URL; more than 60% of the answers were wrong, and even Perplexity, the most accurate, got about 37% wrong. Many chatbots tended not to say "I don't know," instead returning wrong or speculative answers.
• Amazon's study CiteFix (2025, arXiv:2504.15629) likewise references existing findings that the citation accuracy of major generative search engines sits at roughly 74%, which is consistent with Liu et al.'s 74.5% above. (CiteFix itself is a study that proposes a method to improve this accuracy through post-processing.)

What these mean is that "an LLM picks up information and cites it, but that citation is often inaccurate, and misattribution occurs." That is exactly why you design your information so the model retrieves, understands, and cites it correctly—and that is LLMO.

Two practical risks come with this. One is the risk that your brand isn't cited at all (its existence isn't even recognized). The other is the risk that your brand is cited incorrectly—depicted in a way that differs from the facts (misattribution). The latter is also why LLMO should be understood not as a mere "increase your exposure" activity but as a "manage the accuracy of your perception" activity. This point connects directly to the measurement discussion in Chapter 8.

Note that each of the figures above states its source and point in time. All are reports based on studies that actually evaluated the systems in question, but because the target (which engines were covered), the evaluation method, and whether causation was estimated differ from study to study, you should be careful not to simply generalize the figures by lining them up side by side.

For a model to cite you correctly, a technical condition has to be met first: "can the model even reach your information in the first place?" This technical side, which tends to be a bit player in explanations of AIO or GEO, occupies a corner of the main argument in LLMO.

(1) Are you being crawled?

Each AI company's crawlers are split by purpose and can be controlled independently. OpenAI's official documentation distinguishes GPTBot for training, OAI-SearchBot for search display, and ChatGPT-User, which fetches on the spot when a user references a URL, and explains that you can, for example, configure "I don't want it used for training, but I do want to appear in search results." Anthropic similarly distinguishes ClaudeBot for training, Claude-SearchBot for search, and Claude-User for user-initiated fetching. Note that the handling of legacy UA names (Claude-Web, anthropic-ai, and the like) may change, so at implementation you need to confirm against each company's latest official documentation.

A common mistake here is blanket-blocking all AI crawlers for the reason that "I don't want to be used by AI." This stops not just training but retrieval as well, and can amount to throwing away your own visibility in AI search.

(2) Can you be rendered?

The next gate after retrieval is rendering. An analysis by Vercel and MERJ (tracking over 500 million GPTBot fetches) showed that major AI crawlers (GPTBot, ClaudeBot, PerplexityBot, and others) do not execute JavaScript and read only the initial HTML (raw HTML). While GPTBot does fetch JavaScript files themselves (about 10% of its fetches), it did not execute them as code. The one major exception is Google's Gemini, which inherits Googlebot's rendering infrastructure and can therefore execute JavaScript. In other words, a site whose content display depends heavily on client-side JavaScript rendering may look like a nearly blank page to the major AIs other than Gemini, no matter how it grants access in robots.txt. On top of that, in B2B SaaS and e-commerce, a substantial number of sites are reported to be unintentionally blocking major LLM crawlers at the CDN layer—another point to confirm in a technical audit.

(3) Blocking can be a trade-off for visibility and traffic

There is also a report that the decision to "block AI just in case" leads to a loss of traffic. A study by Rutgers Business School and The Wharton School (Hangcheng Zhao and Ron Berman, arXiv:2512.24968, December 2025) analyzed the change after major news publishers (roughly the top 30) blocked GenAI crawlers in robots.txt, and reported that monthly visits on a SimilarWeb basis fell by about 23.1% and human browsing on a Comscore basis fell by about 13.9%. The fact that a decline shows up even in the human panel suggests it's hard to explain as mere bot removal. That said, the subjects were primarily news publishers, and the effect was heterogeneous by publisher size (some mid-sized ones actually increased); it's also an observation from an early phase of the technology, so it can't be applied uniformly across all industries. Blocking should be captured as a trade-off between the benefit of "not being used for training data" and the drawback of "losing visibility and traffic in AI search."

(4) llms.txt is not the solution to the technical conditions

What often gets brought up on the technical side is llms.txt, but Google has officially stated that it does not support llms.txt in its own search (per remarks from Search Relations' Illyes and Mueller). Mueller likened it to the "keywords meta tag," which has long been ignored because it's manipulable by self-declaration.

What technically works in LLMO is not installing a new AI-only file but ensuring "is it in a crawlable state?" and "is it in a state where it can be rendered and its contents read?" Even if you install a special machine-readable file or schema as "a tactic that works in Google," it's meaningless if that premise has collapsed. What works technically is, in the end, crawlability and rendering—not installing an AI-only file—and that's a line you can't leave out when discussing LLMO.

The background against which LLMO came to be discussed boils down, in the end, to a single point: "LLMs came to be used at scale in daily life." Here, rather than lining up market statistics, we'll keep to a narrow sense of scale.

ChatGPT's weekly active users reached 900 million (February 2026, per OpenAI's announcement / TechCrunch reporting). And usage at this scale is happening not just on Google but spread across multiple models—ChatGPT, Gemini, Claude (the Gemini app, too, was announced at Google I/O 2026 in May 2026 to have surpassed 900 million monthly active users—more than double the 400 million a year earlier).

In other words, "where people get their answers" is shifting toward the model side, and that place is not concentrated in one spot—this structure is what surfaces model-layer optimization (LLMO) as a practical challenge. More detailed statistics on the market background are covered in What Is AIO and What Is GEO, so here we keep to the minimum that shows LLMO's necessity.

Around LLMO there are misconceptions seen repeatedly in practice. We organize them alongside the correct understanding.

• "LLMO is Google's official optimization method." → Misconception. LLMO is a practitioner term, not an official standard. As for Google, AI Overviews / AI Mode also pull from the same index as ordinary search, and there is no separate ranking.
• "LLMO = installing llms.txt." → Misconception. Google ignores llms.txt. What works technically is crawlability and rendering, not installing an AI-only file.
• "Get into the training data and you're set." → Misconception. There's also the side of being retrieved by RAG/search, and that citation is incomplete, with misattribution occurring. Getting in is not itself the goal.
• "You should block all AI crawlers." → Misconception. Training and search are controlled separately, and stopping the search crawlers loses your visibility in AI search. There are also reports that blocking leads to a traffic decline.
• "LLMO is something entirely separate from GEO." → Misconception. The two overlap in "being cited by a model." What differs is the focus: LLMO keeps its footing on the model layer and technical conditions, GEO on visibility inside generative engines and an academic framework. The details are covered in a separate article, "LLMO vs GEO" (coming soon).

You can't measure LLMO's outcomes with a single number the way you can with search rank. What you should measure is "whether, inside an LLM's answer, your brand is cited or mentioned correctly and in a desirable form."

What to measure

Concretely, you look at items like these.

• AI citation rate: in what share of answers your brand is cited or mentioned.
• Share of voice (within answers): how much your brand appears compared with competitors.
• Source URL: which page is being cited (the one you intended, or an unintended one).
• Citation accuracy / presence of misattribution: an especially important item in LLMO. As ALCE and the Tow Center showed in the previous chapter, citations tend to be inaccurate, and your brand being cited in a way that differs from the facts (misattribution) does happen in practice.
• Sentiment (positive / negative / neutral) and how you're treated within the answer: are you being recommended, or is it just a passing note? This needs to be seen separately from mere volume of exposure.

Measurement conditions

An LLM's answers wobble every time you run them. Even for the same question, the content cited changes with the phrasing of the query, the region, the language, the login state, and—since LLMO is a story about the model layer, especially—the difference between models. A single measurement is therefore insufficient; you need to run it repeatedly and capture it in confidence intervals. "I asked once and it did / didn't show up" does not reflect reality.

Volatility (citation drift)

This wobble is on a scale you can't ignore. By Profound's observation, roughly 40–60% of the source domains cited by AI turn over each month. In other words, the state of "having been cited once" is not stable, and unless you keep measuring continuously, you'll lose sight of what's actually happening.

How far Google Search Console can measure

As a measurement tool, first pin down the official Google Search Console (GSC). On June 3, 2026, Google added a generative-AI performance report to GSC (Google Search Central Blog). This made it possible to check, in a dedicated view by page, country, device, and date, the impressions where your site's URLs appeared inside AI Overviews / AI Mode (and Discover's generative-AI features). Previously, AI-derived appearances were buried in the aggregate "Web" search figures and couldn't be separated, so this is a step forward.

But the limits are just as clear. First, it's a staged rollout, initially limited to a subset of sites (starting from the UK). Second, what you can grasp is mainly appearances (impressions) and pages; clicks, CTR, and per-query data are not included. And above all, what GSC shows is limited to the appearance of your site's URLs inside Google's generative-AI features. Brand-name-only mentions, the impression of the answer text itself (favorable or not), misattribution, and visibility across engines other than Google such as ChatGPT and Perplexity cannot be measured by GSC alone. This is where you need to combine it with your own prompt-based measurement.

The development into AI Perception Management

As we've seen, an LLM's citations are incomplete, and they shift month by month. In this situation, rather than "run a tactic once and you're done," you need an operating practice of continuously measuring how the model cites and depicts your brand, and correcting course as needed. This is already a domain beyond the frame of an individual "LLMO tactic"—it's about managing AI-space perception itself.

Supporting this continuous measurement and management of perception is Vaipm's area of focus. Vaipm standardizes AI-space perception through stateless, repeated measurement across four AIs (25 runs each), and works to continuously surface citation rate, share of voice, misattribution, and sentiment. Capturing LLMO not as a one-off tactic but as the entry point to the continuous management of perception—that is the essence of this domain. The detailed procedures for measuring results are covered in a separate article, "Measuring AI-search visibility" (coming soon).

Finally, let's restate the principle confirmed in this article. What works technically in LLMO is not installing an AI-only file but ensuring your information is in a state where it "arrives and can be read." Concretely, the three foundations are: not mistakenly blanket-blocking the major AI crawlers (for training and for search), your content not looking blank because it depends heavily on client-side JavaScript, and not unknowingly turning away LLM crawlers at the CDN layer. When these collapse, the information on your own site becomes hard for models and search-retrieval systems to reach. (That said, the possibility of being recognized via third-party mentions or via existing training data still remains.)

Concrete implementation steps and code examples are beyond this article's remit, so we don't cover them. If you're moving on to the hands-on stage, see the separate article "LLMO tactics" (coming soon). This article's job is limited to "what works, and why (the principle)."

Q. What is LLMO?

LLMO (Large Language Model Optimization) refers to optimizing your information so that, in the process by which a large language model generates its answers, that information is easy for the model to correctly retrieve, understand, and cite. It's one sub-concept under AIO, which covers the whole AI-search surface, and it's a practitioner term focused in particular on the "model layer"—how an LLM retrieves, generates, and cites information. It is not an official standard.

Q. How is LLMO different from GEO?

The two overlap in "being cited by a model," but their focus differs. GEO mainly looks at the output-side visibility of "how your brand gets cited or displayed inside a generative engine (AI search)." LLMO keeps its footing on the mechanism inside the model—how an LLM retrieves, generates, and cites information—and on the technical conditions for delivering information to it. Also, GEO has an origin formalized in an academic paper, whereas LLMO is a practitioner term. See the individual "difference" articles for details.

Q. How is LLMO different from AIO?

AIO is the umbrella concept that bundles the "whole AI-search surface"—Google's AI Overviews / AI Mode, ChatGPT Search, Perplexity, Gemini, and so on. LLMO can be organized as a name that sits inside that umbrella, with its viewpoint pulled toward the model layer. Note, though, that this hierarchy itself—"AIO is the broad umbrella, LLMO is a part of it"—is not an official standard but merely a practitioner-side way of organizing things. The detailed line-drawing between the terms is covered in the What Is AIO article.

Q. What works technically in LLMO?

The most foundational things are "crawlability" and "renderability." The premise is that you're not blocked by the major AI crawlers, and that your content doesn't look blank because it depends on client-side JavaScript. An analysis by Vercel/MERJ showed that the major AI crawlers (GPTBot, ClaudeBot, PerplexityBot, etc.) do not execute JavaScript and read only the initial HTML (Gemini being the sole exception), so information that depends on client-side rendering won't arrive. The starting point is ensuring these two things, not installing a special AI-only file.

Q. Will placing an llms.txt count as LLMO?

At least in Google search, it doesn't work. Google has officially stated that it does not support llms.txt (per remarks from Illyes and Mueller). What works technically is crawlability and renderability, not installing an AI-only file. llms.txt can have meaning in limited use cases such as developer documentation, but it's not something you can recommend as "a tactic that works in Google."

Q. Is LLMO about getting into the training data, or about search retrieval?

It has both sides. An LLM's information sources are the knowledge it took inside from training data (parametric knowledge) and the RAG that searches for and retrieves external information at answer time. The former works over the long term and indirectly; the latter over a relatively short term and directly. It's not "get into the training data and you're safe"; whether you're picked up by the retrieval running right now also matters, and that citation is incomplete. You need to think about both sides separately.

Q. Is it safer to block all AI crawlers?

No. AI crawlers are split into training and search, and can be controlled independently (e.g., OpenAI's GPTBot is for training, OAI-SearchBot is for search). Blanket-blocking everything stops not just training but retrieval, losing your visibility in AI search. There's also a report that, for major news publishers, monthly visits fell about 23.1% after blocking (Zhao & Berman, 2025, arXiv:2512.24968). Blocking should be captured as a trade-off between the benefit of "not being used for training" and the drawback of "losing visibility in AI search."

Q. How do you measure citation accuracy and misattribution?

You put questions about your brand to an LLM and check whether what's cited or mentioned in the answer matches the facts and whether the source URL is correct. But because an LLM's answers wobble run to run and model to model, you need to run them repeatedly, not once, and capture them in confidence intervals. There's also an observation that source domains turn over 40–60% each month (Profound), so continuous measurement is a premise.

Q. Is LLMO an official Google optimization method?

No. LLMO, GEO, AIO, and AEO are none of them official standards defined by Google or OpenAI; they are practitioner terms. In Google's case especially, AI Overviews / AI Mode pull from the same index as ordinary search, and there is no separate, AI-only ranking. The fundamentals of traditional SEO (useful, original content, crawlability, quality) are the foundation, as Google itself explains.

Q. How long until LLMO shows results?

It can't be said in a single sweep. On the RAG/search-retrieval side, results may be reflected relatively quickly; the side reflected in training data is long-term. On top of that, citation is unstable, with an observation that source domains turn over 40–60% each month. So rather than judging "when it works" from a single shot, the realistic approach is to measure continuously and read it as a trend.

Q. Are LLMO and SEO separate things?

Not entirely separate. The foundation—crawlability and content quality—is shared with SEO. Google's AI features in particular sit on an extension of traditional SEO, as Google itself explains. What LLMO adds is the model-layer viewpoint of "how the model retrieves, generates, and cites information," and a way of thinking about information design and measurement that presupposes the incompleteness of citations.

(Note: On May 7, 2026, Google retired the display of FAQ rich results. FAQ content has meaning because it can be cited in AI search, but you can't treat FAQ schema as "a quick-win tactic that stands out in Google.")

To organize the key points.

• LLMO is optimizing your information so that, in the process by which an LLM generates its answers, it's easy for the model to correctly retrieve, understand, and cite. It's a sub-concept under AIO (the umbrella)—a practitioner term focused on the model layer.
• It is not an official standard. In Google's case especially, there is no separate ranking, and traditional SEO is the foundation.
• An LLM's information sources are two-layered (training data and RAG), and retrieval and citation are separate stages. Even when retrieved, much of it isn't cited.
• An LLM's citations are inherently incomplete (ALCE, Tow Center, CiteFix). That's exactly why you need "information design so it's cited correctly" and attention to misattribution.
• What works technically is crawlability and renderability. llms.txt doesn't work in Google search.
• Outcomes can't be measured by a single rank; you need to look at citation rate, share of voice, misattribution, and sentiment continuously, through repeated measurement.

As long as an LLM's citations are incomplete and shift month by month, this isn't a "run a tactic once and you're done" affair—it naturally connects to the ongoing practice of measuring and managing AI-space perception. Vaipm is working to standardize that AI Perception Management through stateless, repeated measurement across four AIs (25 runs each).

Articles to read next

• What Is AIO (the umbrella concept for AI-search optimization)
• What Is GEO (visibility inside generative engines)
• LLMO tactics (concrete practice for meeting the technical conditions; coming soon)
• Measuring AI-search visibility (measurement in detail; coming soon)

For the main claims and figures in this article, we show the underlying studies and surveys along with their source and point in time. Because the AI-search domain changes quickly, confirm the latest version of each primary source at implementation.

• A framework for automatically evaluating LLM citation quality (ALCE) — Gao, Yen, Yu, Chen. "Enabling Large Language Models to Generate Text with Citations." EMNLP 2023 (arXiv:2305.14627). Evaluates along three axes: fluency, correctness, and citation quality.
• Verification of generative-search-engine citations (51.5% / 74.5%) — Nelson F. Liu, Tianyi Zhang, Percy Liang. "Evaluating Verifiability in Generative Search Engines." EMNLP Findings 2023 (arXiv:2304.09848). Of generated sentences, an average of 51.5% were fully supported by citations, and the share of presented citations that supported their corresponding sentence was 74.5% (a human audit of Bing Chat, NeevaAI, perplexity.ai, and YouChat).
• Source accuracy of AI search engines (over 60% wrong / Perplexity about 37%) — Klaudia Jaźwińska and Aisvarya Chandrasekar. "AI Search Has a Citation Problem." Tow Center for Digital Journalism, Columbia University (March 2025).
• Reference to citation accuracy of about 74% (CiteFix) — Amazon, "CiteFix," 2025 (arXiv:2504.15629). References the citation accuracy of existing research and proposes an improvement method through post-processing (it is not a study that measured this figure itself).
• Purpose-based control of AI crawlers (GPTBot / OAI-SearchBot / ChatGPT-User) — OpenAI's official crawler documentation.
• Purpose-based control of AI crawlers (ClaudeBot / Claude-SearchBot / Claude-User) — Anthropic's official crawler documentation. The handling of legacy UA names may change, so confirm against the latest version at implementation.
• Google's AI features and the fundamentals of optimization / there is no separate, AI-only ranking in Google search — Google Search Central official documentation (optimization guidance for generative-AI search).
• Google Search Console generative-AI performance report (June 3, 2026) — Google Search Central Blog (2026-06-03). Displays impressions within generative-AI features across AI Overviews / AI Mode / Discover. Staged rollout, initially from the UK. Does not include clicks, CTR, or per-query data.
• AI crawlers not executing JavaScript — Vercel/MERJ, "The rise of the AI crawler." Zero traces of JavaScript execution across over 500 million GPTBot fetches. Major AI crawlers (GPTBot, ClaudeBot, PerplexityBot, etc.) don't execute it; Google Gemini is the sole exception.
• Crawler blocking and traffic (about 23.1% / about 13.9%) — Hangcheng Zhao (Rutgers Business School) and Ron Berman (The Wharton School), December 2025 (arXiv:2512.24968). For major news publishers (roughly the top 30), after blocking GenAI crawlers in robots.txt, SimilarWeb monthly visits fell about 23.1% and Comscore human browsing fell about 13.9%. The effect was heterogeneous by publisher size. Note: this paper also has a later revised version (2026; about a 7% decline over six weeks on a weekly basis), and the figures differ depending on how the target and period are taken.
• The gap between retrieval and citation (about 85%) / the churn of cited source domains (40–60% monthly) — Observations by Profound (AI-visibility analysis).
• Google search not supporting llms.txt — Google Search Relations (remarks by Gary Illyes and John Mueller). Mueller likened llms.txt to the "keywords meta tag," which has been ignored because it's manipulable by self-declaration.
• The academic origin of GEO — Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande. "GEO: Generative Engine Optimization." KDD 2024 (arXiv:2311.09735, Princeton University et al.). The basis for the fact that, among the practitioner terms, only GEO has an academic origin.