What Is GEO? Generative Engine Optimization Explained — Definition, Origin, and the Research [2026 Definitive Guide]

To understand GEO correctly, three points matter. First, among the practitioner terms AIO, LLMO, and AEO, GEO is the only one whose concept has been defined, formalized, and empirically validated in an academic paper — giving it the distinctive position of "not an official standard, but academically grounded" (see Section 2). Second, what its founding paper proved to be effective was factual density — statistics, sources, and quotations — not keyword manipulation or special markup (see Section 3). Third, GEO is not a separate "ranking" that replaces SEO. At least on Google Search, generative AI answers are built on the same foundation as ordinary search (crawlability, useful content, quality), and GEO is a way of thinking that — on top of that foundation — narrows its focus to how you get cited inside the generated answer (see Section 6).

Who this article is for

Marketing, PR, and SEO professionals who came here after hearing the term "GEO," and practitioners interested in adapting to AI search. Written for beginners through intermediate readers, including those who already know AIO.

Numbers to grasp first

• The founding paper reports that GEO methods improved visibility inside generative-engine answers by up to 40% (Aggarwal et al., "GEO: Generative Engine Optimization," arXiv:2311.09735, KDD 2024). Adding statistics, sources, and quotations in particular delivered improvements exceeding 40%.
• Use of generative engines is expanding rapidly: ChatGPT's app reached 1 billion monthly active users, for example (June 2026, as reported by Reuters based on Sensor Tower data). Being cited inside the answer — rather than ranking — is increasingly the starting point for visibility.

GEO (Generative Engine Optimization) is the practice of making your information more likely to be cited, referenced, or mentioned inside the answers that generative engines synthesize. Here, a "generative engine" is a system that — instead of returning a list of relevant links like a traditional search engine — retrieves information from multiple sources and uses a large language model (LLM) to synthesize and summarize them into a single answer with citations. Concretely, this includes ChatGPT Search / answers that involve browsing, answer engines such as Perplexity, and Google's AI Overviews / AI Mode.

GEO is one of several terms used in the practice surrounding AI search. Closely related terms include AIO, LLMO, and AEO, which in practice are often used with overlapping meanings. Here, as an overview, we organize the sense of distance between each term when viewed with GEO as the axis (the detailed definitions and origins of each term, and the precise differences between each pair of terms, are handled in the "What Is AIO" article and the various "difference between ___" articles. The focus of this article is on understanding GEO on its own).

What this table shows is that GEO is often positioned within the broad umbrella of AIO, yet — unlike the other terms — it clearly has an academic origin. Of GEO, AIO, LLMO, and AEO, GEO is the only one whose concept was defined, formalized, and validated in a paper. That said, hierarchical relationships such as "umbrella and sub-concept" are not standardized in the industry; this is only a map that this article draws for the sake of organization. The next section confirms this property of GEO — its "unique academic origin" — through the concrete fact of the founding paper.

The origin of GEO can be pinpointed precisely. It is the paper "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik R. Narasimhan, and Ameet Deshpande. The preprint first appeared on November 16, 2023 (arXiv:2311.09735), and was subsequently accepted and published at KDD 2024 (the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, August 25–29, 2024, Barcelona), pp. 5–16. The authors are researchers based primarily at Princeton University, along with Georgia Tech, the Allen Institute for AI, and IIT Delhi.

Why is this single paper decisive for understanding GEO? Because whereas AIO, LLMO, and AEO spread organically as "labels" within practice and the media, GEO is the only term whose very concept was defined within a paper, whose measurement framework was designed, and whose effects were tested experimentally. In other words, GEO has an origin the other practitioner terms lack: it was "academically formalized."

That said, this fact must be received with care. That GEO was proposed and formalized in a paper is entirely separate from GEO being an "official standard." GEO is not a specification set by Google or OpenAI; it is an optimization paradigm proposed by researchers. In fact, regarding AI search, Google has officially stated the position that generative AI search too is an extension of traditional SEO, and that no separate "AI ranking" or special mechanism exists. Accordingly, this article does not treat GEO as an "official name" or an "officially proposed standard"; it treats it as "a practical way of thinking that was proposed and formalized in a paper and has empirical backing." Blurring this distinction would create the misconception that GEO is an authoritative, official specification.

Below, as the core of this article, we look in detail at what that founding paper specifically defined and proved.

What someone who searched for the term "GEO" most wants to know is probably the substance: "In the end, what did that paper reveal?" We make this the center of the article and lay out the paper's empirical content in order.

3-1. What Is a Generative Engine?

The first thing the founding paper did was define a new category: the "Generative Engine (GE)." The paper positioned this as a new kind of system that combines a traditional search engine with a generative model. Technically, it retrieves relevant documents from a database (such as the internet) and uses a large language model to generate an answer grounded in those sources. In the generative engine the paper envisions, citations are shown in the answer so users can verify the information (though in actual services, the format and explicitness of citations vary by engine). As representative examples at the time of writing, the authors list BingChat, Google's SGE, and perplexity.ai.

The processing the paper describes for a generative engine roughly follows this flow: it receives the user's query, first reformulates it into several more tractable queries, and retrieves relevant sources with a search engine. Next, a summarization model summarizes each source, and finally a response-generation model assembles a single answer with citations based on those summaries. The nature of the output differs fundamentally from traditional search, which lines up "ten blue links."

And what the paper emphasized is the "third party" problem this shift creates. Because a generative engine can complete an answer without the user navigating to a site, it reduces organic traffic to sites and can hurt content creators (the creator economy). Moreover, the generative engine is a black box, and creators can barely control or grasp when and how their content is ingested and how it is portrayed. GEO was proposed as a response to this "opacity for creators."

3-2. Why "Rank" Doesn't Carry Over As-Is

An important point the founding paper makes is that the traditional way of measuring visibility — "rank" — does not carry over to generative engines. In a traditional search engine, visibility is largely determined by "average position across a set of actual queries." It is a one-dimensional metric: which slot you appear in on a linearly ordered list. But a generative engine returns rich, structured answers, weaving multiple sources into one another and embedding them within a single block. Length, position, and style all vary. As a result, visibility becomes "multifaceted and nuanced," and cannot be captured by the single yardstick of rank.

So the paper defined a new visibility (impression) metric for generative engines. As objective metrics, it proposes a "Word Count" metric that normalizes, against the whole answer, the amount of text tied to a given source, as well as a position-adjusted version that factors in where the source appears in the answer using exponential decay. Further, to measure the effect on the reader's attention that objective metrics alone cannot capture, it defines a "subjective impression" metric across seven aspects — (1) relevance to the query, (2) influence of the citation, (3) uniqueness of the presented content, (4) subjective position, (5) subjective word count, (6) probability of being clicked, and (7) diversity of the presented content — measured with an LLM-based evaluation method (G-Eval). In short, GEO's notion of visibility measures not the binary of "whether you were cited," but "with how much text, in what position, and with what influence you were cited." This idea also forms the foundation of the measurement discussed later (Section 7).

3-3. GEO-bench and the Nine Methods That Were Tested

To evaluate systematically based on these metrics, the paper built a benchmark called GEO-bench. It consists of roughly 10,000 queries spanning diverse domains and sources, specially adapted for generative engines.

On this GEO-bench, the paper tested nine optimization methods: (1) authoritative phrasing (increasing persuasiveness and authority), (2) keyword stuffing (including more of the query's keywords — a keyword-heavy method that current SEO also discourages), (3) statistics addition (replacing qualitative descriptions with quantitative statistics), (4) citing sources, (5) quotation addition, (6) simplification (making the language easier to understand), (7) fluency optimization, (8) unique-words addition, and (9) technical-terms addition. Of these, (3), (4), and (5) involve adding new content, while the rest are classified as methods that improve how existing content is "presented." In the experiments, each method was implemented as a text transformation with GPT-3.5, and to reduce statistical variance in the results, five responses were generated per query.

3-4. What Worked and What Didn't (The Core Finding)

And here is the core conclusion of the GEO paper. Adding statistics, citing sources, and adding relevant quotations were especially effective, improving visibility inside generative-engine answers by up to 40% (over 40%) across various queries. On the other hand, keyword stuffing — a keyword-heavy method, over-packing the query's terms, that current SEO also discourages — was low in effectiveness and often worked against the content. Furthermore, effects varied greatly by domain (field), and because no single universal method exists, the paper concludes that domain-specific optimization is necessary.

The implications of this finding are large. What worked in GEO was the "factual density of the substance" — statistics, sources, and quotations — not keyword manipulation, nor the special markup discussed later. This result shows there are limits to trying to force AI-search visibility through keyword-heavy rank tactics alone. In GEO, on top of the traditional SEO foundation, designing concrete information that is easy to use as evidence inside an answer becomes important. Concrete statements backed by facts can be an important element of content that generative engines are likely to cite — this is the theoretical backbone of GEO, and the central insight that anyone who looked up "What is GEO" should grasp first.

Behind the practical need for the idea of GEO is the fact that generative engines spread to the mass market in a short time. Here, rather than covering the statistics for the whole market (that is handled in "What Is AIO"), we present a focused set of representative numbers sufficient to explain why GEO is necessary. Because the figures all change quickly, we note the source and point in time for each.

ChatGPT and Perplexity — the examples the founding paper cited as generative engines — have both seen rapid growth in use over the past few years. ChatGPT's app reached 1 billion monthly active users (June 2026, as reported by Reuters based on Sensor Tower data), becoming the fastest app ever to reach that scale. Perplexity, an answer engine that shows its sources explicitly, is smaller in scale but is raising its presence as one form of generative engine. In other words, the situation the paper warned about in 2023 — "generative engines are here to stay, and left unaddressed they will disadvantage the creator economy" — has since become even more real.

What is happening here is a shift in thinking: the starting point of visibility has moved from "winning a rank" to "being cited inside the generated answer." Traditionally, the premise was that ranking high in search results brought people to your site. But the more often a generative engine completes the answer, the more that being cited, mentioned, or referenced inside the AI's answer — before ever winning a link click — becomes the starting point of visibility. GEO is positioned as a way of thinking that responds to this shift.

GEO did not end with the 2023 founding paper; it is a developing field in which research has continued. Here we introduce several follow-up studies that build on the founding paper. Because all of them are, as of this writing, at the preprint (arXiv) stage, and their peer-review status must be checked individually, we treat them accurately as "proposed" or "reported," without asserting them as settled.

• E-GEO (from e-commerce specialization to general patterns): "E-GEO: A Testbed for Generative Engine Optimization in E-Commerce" by Puneet S. Bagga et al. (arXiv:2511.20867, 2025) is proposed as the first GEO benchmark specialized for e-commerce (EC). It reportedly includes over 7,000 realistic multi-sentence product queries, compares 15 rewriting heuristics, formalizes GEO as a tractable optimization problem, and proposes a lightweight iterative prompt-optimization algorithm. What draws attention is the report that this optimization surfaced a domain-agnostic pattern that held stably across fields. Whereas the founding paper stated that "no single universal method exists and effects depend on the domain," E-GEO — starting from validation in the specific field of EC — suggests the existence of a common pattern that works across fields. It can be positioned as research that searches for general patterns from field-specific validation.
• IF-GEO (extension to multiple queries): "IF-GEO: Conflict-Aware Instruction Fusion for Multi-Query Generative Engine Optimization" (arXiv:2601.13938, 2026) is proposed as a framework for handling the conflicts that arise when integrating GEO-style instructions (methods) across multiple queries. It is research showing that the problem setting is broadening from single-query optimization to optimization across multiple queries.
• The Rise of AI Search (impact on the market and society): "The Rise of AI Search: Implications for Information Markets and Human Judgement at Scale" by Sinan Aral, Haiwen Li, and Rui Zuo of MIT (arXiv:2602.13415, 2026) is reported as a large-scale empirical study that ran 24,000 queries across 243 countries and collected 2.8 million AI-search and traditional-search results. It shows, among other things, that from 2024 to 2025 the countries where Google AI Overviews is offered expanded from 7 to 229, and that search is shifting from "navigation (choosing from a list)" to "synthesis (reading a single answer delivered in one voice)." It can be called research that substantiates the very market structure in which GEO is needed.
• SAGEO Arena (an evaluation environment covering all stages from search to generation): "SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization" (arXiv:2602.12187, 2026; a Work-in-Progress preprint) is reported to propose a more realistic, reproducible evaluation environment covering all stages of generative AI search — retrieval (search), re-ranking, and generation. It raises as an issue that prior GEO benchmarks abstracted away everything up to just before generation and discarded the structured information (such as schema) of web pages, and reports that structured information can help at the retrieval and selection stages. However, this is a research finding aimed at general retrieval-augmented generation (RAG) pipelines, and it concerns a different layer from Google's official position that "no special schema is needed for Google Search" (see Section 6) — the two do not contradict each other; they are about different layers.

Note that, beyond academic research, some vendors have published correlation data on GEO's effects. For example, Ahrefs reports, for brand visibility in Google AI Overviews, a correlation with brand mentions on the web (a correlation coefficient of 0.664, among others) (2025). However, this is correlation, not causation. It cannot be reinterpreted as "increasing mentions will necessarily get you cited." Rather, Google itself explicitly treats acts of artificially manufacturing links or reputation for the purpose of manipulating search rankings (such as link spam and site-reputation abuse) as violations of its spam policies, taking the position that artificially manufactured authority is not rewarded. Therefore, Ahrefs's correlation should not be reinterpreted as "you just need to artificially inflate mentions." And, importantly, this correlation is a separate matter from the causal finding the founding paper validated experimentally — that "factual density (statistics, quotations) is effective." The two must not be conflated.

Precisely because GEO has an academic origin, it is, ironically, easily mistaken for an "authoritative official standard." Here we organize the misconceptions that most easily arise when discussing GEO, along with the correct framing.

Misconception 1: "GEO is an official Google standard."

This is wrong. GEO is a concept proposed and formalized in a paper; it has academic backing, but it is not an official specification set by Google or OpenAI. "Being academically formalized" and "being an official standard" are different. Talking about GEO in sales or articles as if it were "an official Google specification" is contrary to fact.

Misconception 2: "GEO is something entirely separate from SEO."

This is not accurate. At least on Google Search, AI Overviews / AI Mode also assemble answers from the same index as ordinary search, so there is no separate "AI ranking" just for GEO; the traditional SEO foundation — crawlability, useful and original content, quality — works as is. The accurate view is that GEO is a way of thinking that, on top of that foundation, narrows its focus to how you get cited inside the generated answer.

Misconception 3: "Adding schema or llms.txt will give you an advantage in GEO."

For Google Search, this is wrong. Thinking of schema or llms.txt as an "advantage switch you just flip" misidentifies where GEO takes effect. Google explicitly states that no new machine-readable files, AI-specific markup, or special schema.org structured data are needed to appear in generative AI search, and it has officially said that it ignores llms.txt (placing it neither helps nor harms Google Search visibility or rankings). What the GEO founding paper showed to be "effective" was factual density — statistics, sources, and quotations — not the placement of markup or files. (The details of the limits of FAQ rich results and llms.txt are covered in their respective articles.)

GEO's results are not determined solely by the binary of "whether you were cited or mentioned inside the generated answer." You need to look at how much, in what position, and with what context and influence you were treated. Measurement, therefore, targets not traditional search rankings but how you are treated inside the AI's answer.

The main items to measure are: AI citation rate, in-answer share (Share of Voice), how you are treated in the answer (not display rank, but amount, position, and influence — continuous with the visibility metric of the founding paper), the source URL cited, impression (positive / negative / neutral), and whether there is misattribution. The key to GEO measurement is that it does not stop at the binary of "whether you were cited."

At the same time, the conditions of measurement require attention. A generative engine's answers change with query variation, region, language, login state, and model differences. And they fluctuate on every run. So rather than measuring once, you need to run repeatedly and capture the results as confidence intervals. In fact, Profound's vendor measurement (large-scale domain tracking) reports that the cited source domains turn over by as much as 40–60% per month. Citations are startlingly unstable; you cannot measure once and be done.

This very fact is why the perspective of AI Perception Management (AIPM) becomes necessary beyond GEO. Measuring the results of GEO (citation optimization in generative engines) continuously on the premise of fluctuation and misattribution, and managing them — including incorrect information and unfavorable portrayals — is the domain Vaipm addresses. Vaipm is not merely a "GEO tool." Its distinguishing feature is that it makes citations, mentions, and impressions on AI-search surfaces visible through repeated measurement with confidence intervals that account for run-to-run fluctuation, and goes as far as diagnosing "why you are not being cited." What readers who understood GEO as an entry point come to face next is this perspective of "continuously managing perception," and Vaipm is a platform that supports this continuous measurement and perception management.

(The detailed definition of AIPM and its relationship to AIO and GEO are covered in the "What Is AI Perception Management" article, and the specific methodology of measuring results is covered in the "How to Measure AI Search Visibility" article.)

Finally, we confirm only the starting point for practicing GEO. As seen in Section 3, what works most is factual density — statistics, sources, and quotations — while keyword-heavy methods tend to backfire, and effects vary by domain. You need an attitude of not relying on a single universal solution, but validating what works in your own field.

That said, the concrete implementation steps, code, and checklists are deliberately not covered in this article. That is the province of the "GEO tactics" article. The role of this article is to present the "map" of what GEO is and what has been proved. For detailed procedures at the hands-on stage, please refer to the "GEO tactics" article.

Q1. What is GEO?

GEO (Generative Engine Optimization) is the practice of making your information more likely to be cited or referenced inside the answers that "generative engines" such as ChatGPT, Perplexity, and Google AI Overviews synthesize. A generative engine retrieves multiple sources and uses an LLM to synthesize and summarize them into a single answer with citations. What distinguishes GEO is that, rather than competing for search rankings, it works to increase visibility (amount, position, and influence) inside the answer the AI generates.

Q2. What is the difference between GEO and SEO?

SEO is the field that handles the optimization of search rankings, while GEO is a way of thinking focused on visibility inside generative-engine answers. However, the two are not separate: at least on Google Search, GEO too is built on the foundation of traditional SEO (there is no separate "AI ranking"; for details, see Misconception 2 in Section 6 of this article, and the "Difference Between GEO and SEO" article).

Q3. How is GEO different from AIO?

AIO (AI Optimization) is a broader, more elastic umbrella term for visibility across AI-search surfaces in general, and GEO is a concept often positioned within it. The decisive difference is origin: GEO is the only term defined and validated in a paper, whereas AIO is a term that spread organically in practice and has no single official definition. The detailed differences are handled in the "Difference Between AIO and GEO" article.

Q4. What actually works in GEO?

According to the founding paper's validation, what worked most was raising "factual density" — adding statistics, citing sources, and adding relevant quotations — which improved visibility by up to over 40%. Conversely, keyword-heavy methods (also discouraged in current SEO) were low in effectiveness and often counterproductive. Effects depend on the domain, and no single universal method exists. Concrete implementation steps are covered in the "GEO tactics" article.

Q5. Is GEO an official standard?

No. GEO is a concept proposed and formalized in a paper and has academic backing, but it is not an official specification set by Google or OpenAI. "Being academically formalized" and "being an official standard" are different (for details, see Misconception 1 in Section 6 of this article).

Q6. What are the key points of the GEO founding paper?

It is "GEO: Generative Engine Optimization" by Pranjal Aggarwal et al. (arXiv:2311.09735, accepted at KDD 2024). The paper defines the new category of "generative engine" and a visibility metric, builds GEO-bench with roughly 10,000 queries, and tests nine methods. It demonstrates that adding statistics, sources, and quotations improves visibility by up to over 40%, while keyword stuffing is low in effectiveness, and that effects depend on the domain.

Q7. Will adding schema or llms.txt give me an advantage in GEO?

For Google Search, no. Google explicitly states that no special machine-readable files, markup, or additional schema.org structured data are needed, and it says it ignores llms.txt. What works is factual density — statistics, sources, and quotations (for details, see Misconception 3 in Section 6 of this article).

Q8. How do you measure GEO results?

You measure "how you are treated inside the AI's answer," not search rankings. Concretely, this includes AI citation rate, in-answer share (Share of Voice), how you are treated (amount, position, influence), the source URL cited, impression, and whether there is misattribution. Because answers fluctuate greatly on every run and cited sources turn over 40–60% per month, you need to measure repeatedly rather than once and capture results as confidence intervals. The detailed methodology is covered in the "How to Measure AI Search Visibility" article.

Q9. Is GEO only about ChatGPT?

No, it is a way of thinking that concerns generative engines in general. It includes not only ChatGPT Search and answers that involve browsing, but also answer engines such as Perplexity and Google's AI Overviews / AI Mode. Each engine has different citation tendencies, and the rate at which they cite the same URL is not necessarily high, so a realistic approach is to look across the major surfaces together.

Q10. Will GEO remain a valid way of thinking going forward?

GEO is a developing field in which research has continued since the founding paper, with follow-up studies appearing such as E-GEO (specialized for e-commerce), IF-GEO (multiple queries), The Rise of AI Search (market impact), and SAGEO Arena (evaluation environment). Many are at the preprint stage and their peer-review status requires individual verification, and because the AI-search domain changes quickly, we recommend continuously checking figures and trends against the latest primary sources.

Q11. Are GEO and LLMO the same?

They largely overlap but differ in perspective. LLMO (Large Language Model Optimization) focuses on "the layer of the mechanism — the large language model," while GEO focuses on "visibility inside the answer a generative engine returns," and in practice they are often used with overlapping meanings. The detailed differences are handled in the "What Is LLMO" and "Difference Between LLMO and GEO" articles.

Let's organize the key points of this article.

• GEO is the practice of making your information more likely to be cited or referenced inside the answers that generative engines synthesize — a way of thinking that raises "how you get cited inside the answer" rather than search rankings.
• Among the practitioner terms AIO, LLMO, and AEO, GEO is the only one whose concept was defined, formalized, and validated in a paper (Aggarwal et al., KDD 2024). However, "academic backing" and "official standard" are different, and it is a mistake to treat it as an official specification.
• What the founding paper showed to be "effective" was factual density — statistics, sources, and quotations (improving visibility by up to over 40%), while keyword stuffing (a keyword-heavy method also discouraged in current SEO) is low in effectiveness and can backfire. Effects depend on the domain.
• GEO is a developing field, with follow-up research continuing: e-commerce–specific (E-GEO), multi-query (IF-GEO), and market impact (The Rise of AI Search). All are at the preprint stage, and their peer-review status requires verification.
• GEO's results are unstable (cited sources turn over 40–60% per month) and carry a risk of misattribution. That is why repeated measurement and continuous perception management (AIPM) — not a one-off — become necessary.

Articles to read next

• To learn the umbrella concept and the big picture: What Is AIO? The Definition, Etymology, and Big Picture of AI Search Optimization
• To implement the methods: The Complete Guide to GEO Tactics / Factual-Density Content / How to Structure Sentences That AI Will Cite
• To learn the differences in depth: The Difference Between AIO and GEO / The Difference Between LLMO and GEO / The Difference Between GEO and SEO
• To learn how it works: How AI Search Works (Indexing, RAG, Query Fan-Out)
• To learn measurement: How to Measure AI Search Visibility / What Is AI Share of Voice
• To deepen the concepts: What Is AI Perception Management / What Is LLMO

GEO is a way of thinking about "being cited inside the generated answer," but winning a citation is not itself the goal. Those citations turn over within weeks (cited source domains vary by as much as 40–60% per month), and their content is not always accurate. That is exactly why you need a perspective that does not end with a single effort — one that continuously measures and manages "how generative AI cites and portrays you." That is the AI Perception Management that lies beyond GEO, and the domain Vaipm addresses.

The figures and facts in this article have been verified as of July 2, 2026 against the following primary sources and research. Because the AI-search field changes quickly, please re-verify against the latest primary sources when citing in practice.

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, "GEO: Generative Engine Optimization" (arXiv:2311.09735, KDD 2024, Barcelona, pp. 5–16) — the definition of a generative engine, GEO-bench (roughly 10,000 queries), the nine methods, statistics/sources/quotations delivering up to over 40% visibility improvement, the low/negative effect of keyword stuffing, and domain dependence
• Bagga, Farias, Korkotashvili, Peng, Wu, "E-GEO: A Testbed for Generative Engine Optimization in E-Commerce" (arXiv:2511.20867, 2025; preprint; the first GEO benchmark specialized for e-commerce / over 7,000 product queries / 15 rewriting heuristics / reports the possibility of a domain-agnostic pattern from the optimization)
• "IF-GEO: Conflict-Aware Instruction Fusion for Multi-Query Generative Engine Optimization" (arXiv:2601.13938, 2026; preprint)
• Aral, Li, Zuo, "The Rise of AI Search: Implications for Information Markets and Human Judgement at Scale" (arXiv:2602.13415, 2026; preprint; 243 countries, 24,000 queries, 2.8 million results; AI Overviews availability expanding from 7 to 229 countries)
• "SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization" (arXiv:2602.12187, 2026; preprint / Work in Progress; an evaluation environment covering all stages of retrieval, re-ranking, and generation; reports that structured information can help at the retrieval and selection stages. Note: this is a research finding aimed at general RAG pipelines and concerns a different layer from Google's official position that "no special schema is needed")
• Google Search Central, "AI features and your website" and "Optimizing your website for generative AI features on Google Search" (the primary basis for: no separate "AI ranking" exists and traditional SEO is the foundation / no special files or schema are needed / llms.txt is ignored)
• Google Search Central, "Spam policies for Google web search" (last updated May 15, 2026; explicitly covering link spam, site-reputation abuse, keyword stuffing, etc.; the primary basis for: artificially manufactured authority created to manipulate search rankings is not rewarded)
• Reuters (ChatGPT app, 1 billion monthly users, June 2026, Sensor Tower data)
• Ahrefs (a correlation of 0.664 between brand visibility in Google AI Overviews and web brand mentions, 2025. Note: correlation, not causation; the subject is brand visibility in Google AI Overviews)
• Profound (40–60% monthly turnover of cited source domains. Note: vendor measurement)