Generative engine optimisation (GEO) is the practice of changing web content so that AI answer engines are more likely to use it, quote it and cite it in their generated responses. The term comes from a 2023 research paper by Aggarwal and colleagues, which defined GEO as a black-box optimisation framework for content creators and introduced a benchmark and visibility metrics to test it. This page explains what that paper actually measured, what the search and AI providers themselves say about appearing in AI answers, where GEO overlaps with ordinary search engine optimisation, and where it shades into manipulation that providers treat as spam or security researchers treat as an attack. It is written for publishers, marketers, communications teams and security staff who need a factual baseline rather than agency claims.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation and published research; it is not sponsored or endorsed by any provider.
Key takeaways
- GEO is a research term first. It was introduced in the paper "GEO: Generative Engine Optimization" (arXiv 2311.09735), listed as accepted to KDD 2024, which tested nine content changes against a simulated answer engine and a live one.
- The strongest lab results came from evidence, not keywords. Adding quotations and statistics raised measured visibility most, while keyword stuffing scored below the unoptimised baseline.
- Providers say no special optimisation is needed. Google states there are no additional requirements, special optimisations or special schema.org markup needed to appear in AI Overviews or AI Mode; eligibility rests on being indexed and snippet-eligible.
- Access controls decide eligibility before content does. Blocking search crawlers such as OAI-SearchBot or PerplexityBot, or applying nosnippet, removes content from those engines' answers regardless of how it is written.
- Manipulation is a known attack class. Research on "preference manipulation attacks" showed crafted content could steer production LLM search engines, and the same techniques fall under provider spam policies.
What is generative engine optimisation?
A generative engine, in the paper's terms, is a search system that retrieves several web sources and uses a large language model to write one synthesised answer with citations. Because the user may read the answer and never click, the paper argues that content creators lose visibility and control over how their material appears. GEO is the authors' name for methods that help a creator improve the visibility of their content inside those generated responses, together with the metrics used to measure that visibility.
Three features of the original definition matter for anyone using the term today:
- Black box: GEO assumes the creator cannot see or change the engine's ranking or model. Only the source content is changed.
- Visibility, not ranking: success is measured by how much of the generated answer is attributed to a source and how prominently, not by position in a list of blue links.
- Domain dependent: the authors report that which method works varies by subject domain, so they call for domain-specific approaches rather than a single recipe.
Commercial usage has since widened the term to cover almost any activity aimed at AI search visibility. On this page "GEO" means the research definition unless stated otherwise. The related term answer engine optimisation is covered in What Is Answer Engine Optimisation (AEO)? Definition, Signals and Measurement, and the engines themselves in the AI Answer Engines Hub.
What did the original GEO paper test?
The paper is by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. Version 1 was submitted to arXiv on 16 November 2023, version 2 on 28 May 2024 and version 3, the current version as of 15 November 2026, on 28 June 2024.
| Element | What the paper used |
|---|---|
| Benchmark | GEO-bench: 10,000 queries drawn from nine sources across 25 domains, split 8,000 train, 1,000 validation and 1,000 test, keeping a real-world mix of roughly 80% informational and 10% each transactional and navigational queries |
| Simulated engine | Two steps: Google search returns the top five sources, then gpt-3.5-turbo writes the answer at temperature 0.7, with five samples per query |
| Live engine | Perplexity.ai, as a deployed generative engine |
| Metric 1 | Position-Adjusted Word Count: words in sentences citing a source, weighted down exponentially by citation position, divided by the answer's total words |
| Metric 2 | Subjective Impression: seven sub-scores (relevance, influence, uniqueness, subjective position, subjective count, click probability, diversity) rated by GPT-3.5 in a G-Eval style template and normalised to the first metric |
| Methods tested | Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words, Technical Terms |
Each method rewrote one source page using a language model, then the engine regenerated its answer so the change in that source's visibility could be measured against the unmodified version.
What results did the GEO paper report?
The headline figure in the abstract is that GEO can improve visibility by up to 40%. The detailed tables are more specific, and they are the figures worth quoting.
| Method | Simulated engine, Position-Adjusted Word Count (overall) | Perplexity.ai, Position-Adjusted Word Count | Perplexity.ai, Subjective Impression |
|---|---|---|---|
| No optimisation (baseline) | 19.3 | 24.1 | 24.7 |
| Keyword Stuffing | 17.7 | 21.9 | 28.1 |
| Quotation Addition | 27.2 | 29.1 | 32.1 |
| Statistics Addition | 25.2 | 26.2 | 33.9 |
As of the version 3 paper (June 2024). Values are the paper's own scores, not percentages of traffic.
- Best methods: Quotation Addition and Statistics Addition performed best. The paper's Table 1 caption states the best methods improved on the baseline by 41% on Position-Adjusted Word Count and 28% on Subjective Impression in the simulated engine.
- Live engine: on Perplexity.ai, Quotation Addition gave about a 22% gain in word count and Statistics Addition about 37% on subjective impression, while Keyword Stuffing scored about 10% worse than the baseline on word count.
- Lower-ranked sources gained most: for a source ranked fifth in the retrieved set, Cite Sources raised visibility by 115.1%, Quotation Addition by 99.7% and Statistics Addition by 97.9%. For the top-ranked source the same methods reduced visibility by 30.3%, 22.9% and 20.6%. The authors conclude that GEO is especially helpful for lower-ranked websites.
- Domain fit: the paper lists the categories where each method did best, for example Statistics Addition in Law and Government, Debate and Opinion; Quotation Addition in People and Society, Explanation and History; Cite Sources in Statement, Facts and Law and Government.
The practical reading is narrow but useful: in this setup, content that carried verifiable evidence (quotes, figures, citations) was used more by the model, and the gain came largely at the expense of sources that were already ranked higher. Visibility is a share of a fixed answer, so one source's gain is another's loss.
What are the limits of the GEO research?
The authors list their own limitations, and several more follow from the design. Anyone citing GEO figures should carry these caveats with them.
| Limit | Source | Why it matters |
|---|---|---|
| Only two generative engines tested, and methods may need to change as engines evolve | Paper, Section 9 | Results from a 2023 to 2024 setup may not transfer to current engines, which use different models and retrieval |
| Queries drift over time, so the benchmark needs continual updating | Paper, Section 9 | A fixed benchmark ages quickly |
| Effect on search rankings not measured | Paper, Section 9 | A rewrite that helps in the answer could still change classic retrieval, which decides whether the page is fetched at all |
| Larger context windows may reduce the influence of ranking in future models | Paper, Section 9 | The rank-five advantage may shrink as engines read more sources |
| Automated query tags may be noisy | Paper, Section 9 | Per-domain findings are less certain than the overall figures |
| Subjective Impression is scored by a language model | Paper, metric definition | One model judging another's output is a proxy for human attention, not a measurement of clicks or conversions |
| Simulated engine used five sources per query | Paper, setup | Production engines use query fan-out across many searches (Google describes this for AI Overviews and AI Mode), so the competitive set is different |
None of the paper's metrics measure traffic, revenue or brand outcomes. Citation reporting from the engines themselves is covered in the AI Search Citation Reporting Tools Tracker.
What do providers say about appearing in AI answers?
Provider guidance is the other primary source. Providers are listed alphabetically; order is never a ranking.
- Google: its AI features page for site owners (last updated 10 December 2025) states that there are no additional requirements to appear in AI Overviews or AI Mode and no other special optimisations necessary, and that no special schema.org structured data is needed. A page must be indexed and eligible to be shown with a snippet. Google lists ordinary practices as worthwhile: allowing crawling in robots.txt and at the CDN or host, internal linking, page experience, textual content, structured data that matches visible text, and up-to-date Merchant Center and Business Profile information. It also says both features may use query fan-out, running several related searches to build one response. See Google AI Overviews Explained and Google AI Mode Explained.
- Microsoft: Bing's October 2026 announcement on training and grounding preferences says allowing grounding makes content eligible for AI responses, with links and citations pointing to the original content, and that no choice guarantees inclusion, traffic or payment. See Microsoft Copilot Web Answers Explained.
- OpenAI: its publishers and developers FAQ says any public website can appear in ChatGPT search, and that to be included in summaries and snippets a site should make sure it is not blocking OAI-SearchBot. ChatGPT adds utm_source=chatgpt.com to referral URLs so the traffic can be measured. See ChatGPT Search Explained.
- Perplexity: its crawler documentation says PerplexityBot is used to surface and link websites in Perplexity search results and recommends allowing it so a site can appear. See Perplexity AI Explained.
None of these providers publishes a GEO technique list or offers a way to buy organic placement in AI answers in the pages checked. Their shared message is about eligibility: be crawlable by the right agent, be indexed, and do not block the content from being used.
Which controls decide whether content can be used at all?
Before wording matters, a page has to be eligible. These are the controls documented by the providers, as of 15 November 2026. The full per-engine reference is in the AI Crawler and Grounding Controls Tracker, and policy design in the AI Crawler and Scraper Controls guide.
| Provider | Control | Documented effect on AI answers |
|---|---|---|
| nosnippet, max-snippet robots meta rules | nosnippet prevents content being used as a direct input for AI Overviews and AI Mode; max-snippet limits how much may be used | |
| Microsoft | ai-grounding page-level preference (from 1 February 2027) | disallow stops content informing generated responses; disallowextended limits it to Microsoft-operated experiences |
| OpenAI | OAI-SearchBot in robots.txt; noindex | Opted-out sites do not appear in ChatGPT search answers, though they can still appear as navigational links |
| Perplexity | PerplexityBot in robots.txt | Controls surfacing in Perplexity search results; Perplexity-User fetches for a user and generally ignores robots.txt |
A common and costly mistake is blocking a search agent while trying to block training. Training and search are separate agents or tokens at most operators, so a "GEO" programme that ignores robots.txt can be undone by a single line written by another team.
Where does GEO become manipulation?
The same research that shows content changes can raise visibility also shows they can be abused. Nestaas, Debenedetti and Tramèr (arXiv 2406.18382, July 2024) describe "preference manipulation attacks", in which crafted website content or plugin documentation steers an LLM to promote the attacker's products and disparage competitors. They demonstrated this on the Bing and Perplexity LLM search engines and on plugin APIs for GPT-4 and Claude, and describe a prisoner's dilemma: each party has an incentive to attack, but together they degrade answers for everyone.
Provider policies draw the line in familiar terms. Google's spam policies (last updated 28 August 2026) cover:
- Keyword stuffing: filling a page with keywords or numbers to manipulate rankings. The GEO paper's own results show it underperformed the baseline.
- Scaled content abuse: generating many pages mainly to manipulate rankings, explicitly including using generative AI tools to generate many pages without adding value for users.
- Hidden text: placing content solely to manipulate search engines and not easily viewable by people, which is also the usual carrier for instructions aimed at a model.
For security teams, instructions hidden in web pages and aimed at a model reading them are indirect prompt injection, described in AML.T0051 LLM Prompt Injection. Coordinated publishing to bias AI answers at scale is covered in LLM Grooming and AI Search Poisoning. The boundary is practical: adding a sourced statistic a reader can check is the evidence-led change the paper measured; hiding text or mass-producing pages to influence a model is spam under provider policy and, where it carries instructions, an attack.
How does GEO compare with SEO?
| Aspect | Search engine optimisation | Generative engine optimisation (research definition) |
|---|---|---|
| Output being influenced | A ranked list of links | A written answer that cites several sources |
| Success measure | Position, impressions, clicks | Share and prominence of the answer attributed to a source |
| Main lever | Crawlability, relevance, links, page experience | Content wording and evidence, given the page is already retrieved |
| Dependency | Stands alone | Depends on retrieval first: a page not fetched cannot be cited |
| Provider position | Long-standing published guidelines | Google: no special optimisation needed for AI features beyond Search fundamentals |
Because answer engines retrieve from a search index (their own or a partner's), GEO sits on top of search eligibility rather than replacing it.
How can GEO effects be measured honestly?
- Fix eligibility first. Confirm with each provider's tools and your server logs that the relevant search agents can fetch the page and that no nosnippet or noindex rule applies.
- Use provider-reported data. Google includes AI Overviews and AI Mode traffic in Search Console's Performance report under the Web search type; ChatGPT referrals carry utm_source=chatgpt.com.
- Change one thing at a time. The paper isolated one method per test; a site rewrite that changes structure, facts and links at once cannot be attributed.
- Sample answers repeatedly. The paper used five samples per query because answers vary between runs; a single screenshot is not a measurement.
- Track accuracy, not just presence. Record whether the answer represents the page correctly, since a citation attached to a wrong summary is a reputational risk.
Practitioner checklist
- Use "GEO" in its research sense in internal papers, and cite arXiv 2311.09735 version 3 with its limits when quoting figures.
- Audit robots.txt and robots meta rules so search agents (Googlebot, OAI-SearchBot, PerplexityBot, Bingbot) are not blocked by a training-only policy.
- Decide deliberately on nosnippet, max-snippet and, from February 2027, Bing ai-grounding values, and record the owner of each decision.
- Prefer evidence a reader can verify: attributed quotations, dated statistics with sources, and citations to primary documents.
- Reject keyword stuffing, hidden text and mass-generated pages; they breach Google's spam policies and performed worst in the GEO paper.
- Scan your own pages and user-generated content for hidden instructions aimed at AI readers, and treat any found as a security incident.
- Measure with provider reports and repeated answer sampling, and track accuracy of how your content is summarised.
- Recheck provider guidance quarterly; controls changed at Bing, Google and Anthropic during October 2026 alone.
Standards and references
- arXiv: GEO: Generative Engine Optimization (Aggarwal et al.)
- arXiv: Adversarial Search Engine Optimization for Large Language Models (Nestaas, Debenedetti, Tramèr)
- Google Search Central: AI features and your website
- Google Search Central: Robots meta tag, data-nosnippet and X-Robots-Tag specifications
- Google Search Central: Spam policies for Google web search
- Microsoft Bing Blog: Your content, your choice: clearer preferences for training and grounding
- OpenAI Help Center: Publishers and developers FAQ
- Perplexity: Perplexity crawlers
Related NHI Mgmt Group resources: AI Answer Engines Hub, What Is Answer Engine Optimisation (AEO)? Definition, Signals and Measurement, AI Crawler and Grounding Controls Tracker, LLM Grooming and AI Search Poisoning.