GEO, short for generative engine optimization, is the practice of shaping web content, site access and a brand's wider presence online so that generative artificial intelligence (AI) systems retrieve, cite or name it when they compose an answer. The targets include ChatGPT, Perplexity, Microsoft Copilot and Google's AI Overviews and AI Mode, which respond in prose rather than with a ranked list of links. Success means inclusion in the answer, not position on a results page.

It exists because the answer has absorbed much of the click. A results page sends users to websites; a synthesised response often settles the query itself, citing a handful of sources. GEO is the label for competing over those citations and mentions.

How answer engines pick sources

Most answer engines follow the same broad sequence, known in AI research as retrieval-augmented generation (RAG). GEO work targets each stage.

Access comes first. Google's guide to its generative features, last updated on July 10, 2026, requires a page to be indexed and eligible to appear in Search with a snippet, so a nosnippet directive removes it from consideration. OpenAI's publisher documentation makes inclusion in ChatGPT search summaries conditional on not blocking its OAI-SearchBot crawler.

Query rewriting follows. Models rarely search for the words a user typed. Google calls the technique query fan-out: concurrent, related queries the model generates to fetch more results. Across 5 million such queries collected in April 2026, Tomek Rudzki of Peec AI found ChatGPT issued 2.1 per prompt, against 1.4 for Perplexity and 6.8 for Grok, adding the word "best" to 15.33% of them. In 21 of 27 ChatGPT conversations examined by the consultant Suganthan Mohanadasan in July 2026, the first search named a brand the user never typed; brands named in those queries reached the answer 68.9% of the time, against 2.1% for brands retrieved but unnamed. Only 3.1% of 3,554 retrieved pages were cited.

Retrieval and synthesis close the loop. Google says its generative features use its core ranking systems to fetch pages from the Search index, so organic rankings and AI citations tend to overlap. Citation density varies: across 126 million US prompts, Semrush counted 15.4 sources per ChatGPT response against 3.3 for Gemini, and found that engines often mention brands they do not cite.

Measurement mostly sits outside the engines. Vendors such as Profound, Peec AI and Semrush run fixed prompt sets and log mention rate, citation rate, sentiment and share of voice. HubSpot's AEO tool, launched on April 14, 2026, costs $50 a month and scores sentiment from -100 to +100. Bing Webmaster Tools added an AI Performance dashboard in preview on February 10, 2026, reporting citation counts and the grounding queries used to retrieve pages.

The tactics fall into three groups. Content changes add statistics, quotations and attributed sources. Off-site work pursues coverage, reviews and listings on the third-party pages engines retrieve. Technical work covers crawler permissions, rendering and, more contentiously, machine-readable files and structured data.

On the buy side, GEO is run by brands, search and content agencies and public relations (PR) teams. The engine operators set the rules unilaterally. Publishers sit in between, as the sources being cited.

Origin and evolution

The term comes from a paper posted to arXiv on November 16, 2023 by Pranjal Aggarwal of the Indian Institute of Technology (IIT) Delhi; Vishvak Murahari, Karthik Narasimhan and Ameet Deshpande of Princeton University; and two independent researchers, Tanmay Rajpurohit and Ashwin Kalyan. It followed Microsoft's Bing Chat, launched in February 2023, and Google's Search Generative Experience, unveiled that May, which the authors called "generative engines".

The team built GEO-bench, 10,000 queries from nine datasets, 80% of them informational. For each, the top five Google results were fed to OpenAI's GPT-3.5-turbo, which wrote a cited answer, and nine rewriting methods were applied to the sources. Adding citations, quotations and statistics raised a source's position-adjusted word count, its share of the answer weighted by position, by 30% to 40%; the best method reached 41%. Keyword stuffing, a classic search tactic, scored 17.7 against an unoptimised baseline of 19.3. A follow-up on 200 queries using Perplexity showed gains of up to 37% on a second, model-scored metric.

The most quoted result concerned rank. When all five sources were rewritten at once, adding citations raised the fifth-ranked page's visibility by 115.1% and cut the top-ranked page's by 30.3%. The paper was presented at KDD 2024, a data mining conference held in Barcelona from August 25 to 29, 2024.

Commercial adoption followed. On February 19, 2024, Gartner forecast that traditional search engine volume would fall 25% by 2026. Google launched AI Overviews in the US in May 2024, and OpenAI launched ChatGPT search on October 31, 2024. On May 28, 2025, Zach Cohen and Seema Amble of Andreessen Horowitz argued that GEO was rewriting the rules of search, citing a search engine optimization (SEO) market worth more than $80 billion and ChatGPT prompts averaging 23 words, against about four for a typical search.

Why it matters for marketers

In the first randomised field experiment on the subject, Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University hid AI Overviews from a random group of Chrome users, and outbound organic clicks rose from 0.37 to 0.62 per search.

For publishers, fewer clicks mean fewer pageviews to monetise programmatically, even as their pages supply the citations brands chase. For advertisers, visibility is concentrated: only 36 of more than 1,200 brands tracked by Semrush held top-100 visibility on all four platforms it measured, every month. In a Fractl survey of 343 US marketing decision-makers published on July 21, 2026, 19% had adopted the GEO label internally, yet 81% still called the work SEO.

Limitations and disputes

The first dispute is whether GEO is a discipline at all. Google's guide, published on May 15, 2026, says optimising for generative AI search is still SEO, and that Google Search does not use llms.txt files, needs no special schema and does not require content to be chunked. Brendon Kraham, a vice president in Google's ads business, made the case to chief marketing officers under the title "Good SEO is good GEO". John Mueller of Google wrote on August 14, 2025 that the stronger the push of new acronyms, "the more likely they're just making spam and scamming".

Others disagree. Mike King of iPullRank argued that Google confuses what it does not use with what sites do not need, when other engines retrieve differently. Microsoft Advertising's January 2026 retail playbook treats AEO and GEO as separate practices. Google, which sells search advertising, has an interest in describing the shift as continuity; vendors of GEO tools have the opposite interest.

The evidence is thin. The founding paper used a five-source pipeline, GPT-3.5 and AI-written rewrites; its Perplexity test uploaded source text as files rather than using live retrieval, and the authors did not measure effects on rankings. C-SEO Bench, a 2025 benchmark by Haritz Puerto and colleagues accepted at the NeurIPS conference, found most such methods "largely ineffective" and often harmful to ranking, with gains shrinking as more competitors adopted them. Kraham also cited record Search query volumes in early 2026, against Gartner's forecast of decline.

Measurement is unstable. Answers vary with location, conversation history and custom instructions; Andrea Volpini, founder of WordLift, said there is "no reliable way to track LLM users".

Manipulation is the final problem. Self-promotional "best of" listicles became a standard tactic. The SEO consultant Lily Ray saw dozens of companies hit by what appeared to be a targeted action on January 20, 2026; one had published 2,000 articles each claiming first place in its category. On May 15, 2026, Google stated that its spam policies apply to generative AI responses, with pages served only to AI crawlers treated as cloaking. Technical shortcuts have made little headway: llms.txt, proposed by Jeremy Howard on September 3, 2024, featured on only 7.4% of Fortune 500 websites in March 2026.

GEO and adjacent terms

SEO optimises for ranked positions in a list of links. GEO optimises for inclusion in a synthesised answer, which in Google's case is drawn largely from those same rankings.

AEO, or answer engine optimization, is usually a synonym. Microsoft separates the two, assigning AEO to agents and assistants that need precise, structured data and GEO to authority in generative search. LLM optimization (LLMO) and AI search optimization are other variants, used by 15% and 48% of Fractl's respondents respectively.

Geo-targeting is unrelated. In advertising, "geo" usually means location, and campaign settings labelled GEO refer to countries and regions.

Prompt injection hides instructions for a model inside content. GEO persuades through what a page says; injection tries to command the model, and Google and Microsoft have classified such prompts as spam.

Recent developments

Measurement is moving inside the platforms. Google began rolling out Search Generative AI performance reports in Search Console in June 2026, covering AI Overviews and AI Mode for a subset of sites, with impressions but no clicks. Regulation is pushing further: an order by the UK Competition and Markets Authority (CMA) dated June 3, 2026 requires Google to give publishers an opt-out from its AI features without a ranking penalty, to supply engagement metrics by December 3, 2026 and to offer page-level controls by March 3, 2027.

A Scrunch study published on September 24, 2026, covering seven platforms and about 10,000 cited URLs, found that a cited self-promoting listicle lifted its author's recommendation rate from about 4% to 7%. A competitor was recommended instead in 24.3% of those responses, and no brand at all in 63.9%.

Nearly three years on, whether GEO is a discipline, a relabelling of SEO or a sales pitch depends, as of October 2026, on who is asked.

Timeline

  • February 7, 2023: Microsoft launches Bing Chat, combining its search engine with a large language model
  • May 10, 2023: Google unveils the Search Generative Experience at its I/O conference
  • November 16, 2023: Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande post "GEO: Generative Engine Optimization" to arXiv
  • February 19, 2024: Gartner forecasts a 25% fall in traditional search engine volume by 2026
  • May 14, 2024: Google launches AI Overviews in the US
  • June 28, 2024: The final revision of the GEO paper is posted
  • August 25-29, 2024: The GEO paper is presented at KDD 2024 in Barcelona
  • September 3, 2024: Jeremy Howard proposes the llms.txt file format
  • October 31, 2024: OpenAI launches ChatGPT search
  • May 28, 2025: Andreessen Horowitz publishes its case for GEO as the successor to SEO
  • June 6, 2025: C-SEO Bench is posted to arXiv, later accepted at NeurIPS 2025
  • August 14, 2025: John Mueller links aggressive promotion of AI search acronyms to spam
  • November 2025: Peec AI raises $21 million
  • January 2026: Microsoft Advertising publishes a retail playbook separating AEO and GEO
  • January 20, 2026: Lily Ray observes dozens of companies hit by an apparent targeted Google action against self-promotional content
  • February 10, 2026: Bing Webmaster Tools adds an AI Performance dashboard
  • April 14, 2026: HubSpot launches its AEO tool
  • May 15, 2026: Google publishes its guide to generative AI features and extends its spam policies to AI responses
  • June 3, 2026: The CMA orders Google to offer publishers AI feature controls
  • June 2026: Google begins rolling out generative AI performance reports in Search Console
  • July 21, 2026: Fractl publishes its survey on AI search terminology
  • September 24, 2026: Scrunch publishes its study on self-ranking listicles in AI answers
  • December 3, 2026: Deadline for the CMA's engagement metric requirements
  • March 3, 2027: Deadline for the CMA's page-level control requirements

Summary

Who. Brands, search and content agencies and PR teams practise GEO, supported by monitoring vendors such as Profound, Peec AI, Semrush and HubSpot. The rules are set by the operators of answer engines, chiefly Google, OpenAI, Microsoft and Perplexity, while publishers supply the pages that get cited. The term was coined by researchers at IIT Delhi and Princeton University.

What. The practice of adapting content, site access and off-site presence so that generative AI systems retrieve, cite, name or recommend a source in their answers. Visibility is measured by mentions, citations and share of answers rather than by ranking position.

When. The term dates from a paper posted on November 16, 2023 and presented at KDD 2024. Commercial use spread after AI Overviews launched in May 2024 and ChatGPT search in October 2024, and the label's status was contested by Google in guidance published on May 15, 2026.

Where. In ChatGPT, Google AI Overviews and AI Mode, Perplexity, Microsoft Copilot, Gemini and similar assistants, and on the third-party websites from which those systems draw their sources.

Why. Answer engines settle many queries without a click, shifting value from ranking to inclusion. Whether that shift demands a new discipline or is ordinary SEO under a new name remains disputed, and the evidence that specific GEO techniques work outside laboratory conditions is limited.