Skip to content

Generative Engine Optimization

AI search is already deciding who gets cited.

AI search is rewriting who gets cited and who disappears. Most GEO advice is recycled SEO intuition. Here's what the actual research shows, and where it goes completely silent.

PublishedApril 11, 202620 min read

The short version

The one controlled experiment on the question is Aggarwal et al. (Princeton and IIT Delhi, KDD 2024), which found content changes can lift visibility in generative answers by up to 40%, with quotations from authoritative sources and statistics carrying explicit citations as the strongest levers. Everything else is platform documentation and industry measurement: Google says no special markup or AI-specific file is required, Pew found users click a traditional link 8% of the time when an AI summary appears against 15% when it does not, and Ahrefs found schema produces no citation lift. Follow-up papers extend the question to structure and to e-commerce rather than re-testing the original levers. The practical answer is unglamorous: cite your sources inline, give the figures, and write the answer where a reader will find it.

Opening finding
43.5%
of AI Overview citations come from pages outside the organic top 100
SE Ranking · 100,013 queries · July 2024
40%
ceiling on visibility gain from the GEO study's content methods (quotations were the single strongest)
Princeton / IIT Delhi · KDD 2024
−65%
organic CTR when an AI Overview is present and you're not cited
Seer Interactive · 42 organizations · 2025
8%
click rate on organic links when an AI Overview appears (vs 15% without)
Pew Research · real browsing data · March 2025, published July 2025

Chapter 01

What GEO is, and where the term came from

"Generative Engine Optimization" entered the academic literature in November 2023, when researchers from Princeton University and IIT Delhi published a paper by that exact title. It was presented at ACM KDD 2024, one of the top data mining conferences in the world. It's still the only peer-reviewed paper that systematically tested optimization strategies for AI search visibility.

The paper's authors, Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande, defined the problem precisely: search is shifting from returning a list of links to generating a synthesized answer. When that happens, visibility stops being about ranking position and starts being about whether the AI's response includes your content at all.

They built GEO-bench: 10,000 real queries paired with the top-5 Google results, tested 9 content optimization strategies, and measured which changes made the AI include more of a given page in its answer.

Two strategies stood out clearly: adding quotations from authoritative sources and adding statistics with explicit citations. Across 25 topic domains the framework's content methods lifted visibility by up to 40%, validated on real Perplexity.ai responses, not just simulations. (The paper reports several per-method figures depending on the metric; the headline ceiling is what holds up — treat the individual percentages as directional.)

StrategyVisibility improvement
Add quotations from sourcesTop lever in the studyStrongest method
Add statistics with citationsConsistent across 25 domainsEffective
Cite sources explicitlyEffective
Easy-to-understand languageModerate positive
Technical terminologyModerate positive
Keyword stuffingNo benefitZero / negative
40%
the ceiling on visibility gains from the GEO study's content methods — quotations were the single strongest lever
arXiv:2311.09735 / ACM KDD 2024
Follow-up work extends the question rather than re-running it. GEO-SFE (Yu et al., March 2026) holds content constant and varies structure instead — document architecture, chunking, visual emphasis — and reports a 17.3% citation lift across six engines. E-GEO (November 2025) builds an e-commerce testbed of 13,747 product queries and finds one rewriting pattern that holds across five engines, which is evidence that some GEO effect generalises. Neither re-tests quotations and statistics specifically, so the 40% figure above still rests on the original paper alone.

Chapter 02

What the platforms officially say (and what they won't)

Google
  • Standard indexability required
  • Snippet eligibility required
  • Ranking mechanism not disclosed
  • No special markup documented
  • llms.txt not supported (explicitly)
"No additional requirements to appear in AI Overviews... You don't need to create new machine readable files, AI text files, or markup." — Google Search Central, Dec 2025
Microsoft Bing — Most specific guidance
  • H1/H2/H3 headings reflecting specific questions
  • Short, focused sections (one idea each)
  • FAQ / HowTo / Product schema
  • Factual language with measurable facts
  • Clean, accessible HTML
  • Content consistent with authoritative sources
Guidance from Krishna Madhavan, Principal PM, Microsoft AI and Bing, Oct 2025. The often-quoted "+357% YoY to 1.13B visits (June 2025)" is not Bing's own figure: it is Similarweb counting referrals from all AI platforms to the top 1,000 sites.
OpenAI ChatGPT Search
  • Source selection mechanism not disclosed
  • Citations required in API responses
  • Query rewritten into targeted sub-queries
  • No publisher guidance published
Launched October 31, 2024 for all users. Uses "third-party search providers and content provided directly by partners."

Chapter 03

What an AI Overview does to your click rate

Pew Research Center published the most methodologically rigorous study on AI Overview CTR impact. Through March 2025 it tracked actual browsing behavior from over 900 U.S. adults who consented to a browser tracker (real clicks, not SERP scraping), and published the results that July. 68,879 unique Google searches analyzed.

Seer Interactive tracked 3,119 informational queries across 42 client organizations over 15 months (June 2024 – September 2025). Being cited in an AIO was associated with 35% higher organic CTR and 91% higher paid CTR. The authors explicitly note they can't prove causation. The figures below are that cut of the study; Seer has since published a third revision covering January 2024 to February 2026 over roughly 10,000 keywords, which reports a 55% organic CTR decline (1.41% to 0.64%) on a differently drawn sample. If you compare our table against Seer's live page and the headline numbers differ, that is why.

SparkToro's 2024 clickstream study, run on Datos panel data, found only 360 clicks per 1,000 US Google searches reach non-Google properties. 58.5% of US searches end with no click to the open web. Later readings put the figure higher still.

ScenarioOrganic CTRYoY change
AIO present, not cited0.52%−65%
AIO present, cited0.70%−49%
No AIO1.45%−46%
−65%
organic CTR drop when AIO present and you're not cited vs. the same queries before AIOs
Seer Interactive · Q3 2025 cut, superseded by their 2026 revision
8% vs 15%
CTR on organic links: with AI Overview vs. without
Pew Research · real navigation data
31.5%
The zero-click paradox:
Semrush found that keywords where AI Overviews appeared actually showed a slight decrease in zero-click rate, from 33.75% to 31.53%. AI Overviews tend to trigger on complex, research-type queries where users click more regardless of whether an AI summary appears. The Pew data controls for this better.
Semrush · 200,000+ keywords · Jan–Oct 2025

Chapter 04

What doesn't work

Schema markup as GEO driver
Microsoft mentions it. Google says you don't need it. Search Atlas found no correlation between schema coverage and citation rates (December 2024). No controlled experiment shows FAQ schema causes AI citation.
llms.txt
Proposed by Jeremy Howard (Answer.AI, September 2024). None of the major platforms — Google, OpenAI, Anthropic, Perplexity — have confirmed they read it. Adoption stats vary from 0.3% to 10% depending on measurement method.
Keyword stuffing
Princeton paper tested it directly. Zero improvement. AI search uses semantic retrieval, not keyword matching. A page with your target phrase 15 times is less likely to be cited than one with 2 uses and actual context.

Chapter 05

What the evidence actually supports

01
Cite your sources inline
Inline citations in body text, not a bibliography at the bottom. "According to a 2025 Pew study of 68,879 searches..." gives the AI something to anchor.
Up to 40% (Princeton GEO study)
02
Include verifiable statistics
"Most users prefer..." is ignored. "72% of users reported preferring X in a 2025 survey of 4,000 respondents" gets included. A number with a source gives the AI something to anchor.
Up to 40% (Princeton GEO study)
03
One section, one question
AI retrieval looks for the chunk of content that answers a specific sub-question. Focused sections get cited for more queries than blended pages.
04
No JavaScript gates
Bing's crawler doesn't click, expand accordions, or scroll carousels. Content behind any JavaScript interaction is invisible to crawler-based AI systems.
05
Write like a reference, not a product page
AI systems favor third-party authoritative sources over brand-owned content. A comparative page explaining how multiple solutions approach a problem gets cited more than a page that only describes your solution.
06
Keep content current
AI Overviews appeared on roughly 1 in 10 queries in 2025 (BrightEdge), and far more in some verticals than others. Pages with stale data get deprioritized in systems that actively compare claims against current web information.

Chapter 06

The honest uncertainty

A few things are genuinely unknown, and anyone claiming certainty about them is selling something.

Google's selection mechanism
Never published. The query fan-out description tells us how they search, not how they rank what they find. No controlled experiment on what changes cause Google AI to start citing a page.
GEO vs SEO causality
The two SE Ranking figures from mid-2024 look contradictory and are not: 93.67% of AI Overviews linked to at least one domain from the organic top-10, while 43.5% of the individual citations came from outside the top-100. An AIO typically anchors on a familiar top-10 domain and then reaches well past it for the rest. By early 2026, Ahrefs found the top-10 overlap had fallen from 76% to 38% after Gemini 3. Strong SEO correlates with AI citation but predicts it less and less.
AI slop contamination
"Retrieval Collapses When AI Pollutes the Web" (Yu, Kim and Kim, NAVER Corp., WWW '26) found that in the SEO scenario a 67% contamination of the pool led to over 80% exposure contamination, while answer accuracy stayed stable. The system looks fine while the web underneath it isn't.
Sources — primary only
Academic papers
Aggarwal et al. — GEO: Generative Engine Optimization (Princeton / IIT Delhi, KDD 2024)arxiv.org/abs/2311.09735
Yu et al. — GEO-SFE: Structural Feature Engineering (March 2026)arxiv.org/abs/2603.29979
E-GEO: A Testbed for Generative Engine Optimization in E-Commerce (November 2025)arxiv.org/abs/2511.20867
Chen et al. — GEO empirical study (September 2025)arxiv.org/abs/2509.08919
Yu, Kim, Kim — Retrieval Collapses When AI Pollutes the Web (WWW '26)arxiv.org/abs/2602.16136
Qian et al. — Citation Generation in LLMs (CCIR 2024)arxiv.org/abs/2410.11217
Venkit et al. — Answer Engine Evaluation (October 2024)arxiv.org/abs/2410.22349
Platform documentation
Google Search Central — AI features guidance (December 10, 2025)developers.google.com
Microsoft — Optimizing for AI search answers (October 8, 2025)ads.microsoft.com
Bing Webmaster — AI Performance tools (February 10, 2026)blogs.bing.com
OpenAI — Web search tool documentationdevelopers.openai.com
Jeremy Howard / Answer.AI — llms.txt proposal (September 3, 2024)answer.ai
Industry research
Pew Research Center — AI Overview CTR study (July 22, 2025)pewresearch.org
Seer Interactive — AIO CTR impact (September 2025)seerinteractive.com
Semrush — AI Overviews 10M+ keywords (2025)semrush.com
Semrush — Most-cited domains in AI (November 2025)semrush.com
SE Ranking — AI Overviews sources (July 2024)seranking.com
BrightEdge — One year of AI Overviews (May 14, 2025)brightedge.com
SparkToro — Zero-click search study (2024)sparktoro.com

Put the research to work

Find out if your pages are built to get cited

That SEO Agent analyzes your actual pages against the GEO signals this research identifies. Your real data, measured against what works.

See pricing

Early access — limited spots