What Is Generative Engine Optimization (GEO)?

GEO in one sentence

Generative engine optimization (GEO) is the practice of making your pages more likely to be retrieved, quoted and linked by systems that write answers rather than list results: Google AI Overviews and AI Mode, ChatGPT search, Perplexity, Microsoft Copilot and Claude. It overlaps heavily with SEO because those systems draw on search indexes, but it is judged by citations rather than rankings.

That is the whole idea. The rest of this guide covers where the term came from and what the original research actually measured, how each engine picks its sources according to its own documentation, what the evidence supports, what is contested, and how to measure any of it. If you only want the short comparison, read GEO vs SEO. This is the long version.

Where does the term “generative engine optimization” come from?

The term was coined in a November 2023 paper by researchers at Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi, published at KDD 2024. It introduced a benchmark called GEO-bench and tested nine content edits for their effect on a source’s visibility in generated answers. The headline result was a lift of up to about 40 percent.

Research shows: the paper, GEO: Generative Engine Optimization by Pranjal Aggarwal and colleagues, is the origin of the acronym and still the only controlled experiment most GEO advice rests on. It is worth knowing exactly what it did.

What the paper tested

• Queries. GEO-bench holds 10,000 queries drawn from nine sources (MS MARCO, ORCAS-I, Natural Questions, AllSouls, LIMA, Davinci-Debate, Perplexity Discover, ELI5 and GPT-4-generated questions). Roughly 80 percent are informational.
• The engine. The authors built their own generative engine: fetch the top five Google results for a query, then have GPT-3.5-turbo write an answer with citations. They describe it as mimicking commercial engines such as Perplexity. A 200-query subset was also run through Perplexity itself.
• The edits. Each source page was rewritten with one of nine strategies, from adding statistics to stuffing keywords, and the answer was regenerated.
• The metrics. Position-Adjusted Word Count measures how many words of the answer cite your source, weighted toward the top of the response. Subjective Impression is a GPT-3.5 rating across seven sub-scores such as relevance, influence and follow-up likelihood.

What the numbers were

Edit to the source page	Position-adjusted word count	Change	Subjective impression	Change
No optimization (baseline)	19.5	baseline	19.3	baseline
Quotation addition (relevant quotes from credible sources)	27.8	+43%	24.7	+28%
Statistics addition (replace qualitative claims with numbers)	25.9	+33%	23.7	+23%
Fluency optimization (smoother prose)	25.1	+29%	21.9	+13%
Cite sources (add citations)	24.9	+28%	21.9	+13%
Technical terms	23.1	+18%	21.4	+11%
Easy-to-understand language	22.2	+14%	20.5	+6%
Authoritative tone	21.8	+12%	22.9	+19%
Unique words	20.7	+6%	20.4	+6%
Keyword stuffing	17.8	-9%	20.2	+5%

Percentage changes are computed from Table 1 of the paper against the baseline rows; the authors’ own headline figures are 41 percent and 28 percent for the best method. On the Perplexity subset, quotation addition lifted word count by about 22 percent and statistics addition lifted the subjective score by about 37 percent, while keyword stuffing lost about 10 percent.

What the paper does not say

Three limits matter when someone sells you “the GEO method”.

1. It measured a benchmark, not a product. The main engine was GPT-3.5 over Google’s top five results in 2023. The authors list “only two generative engines” and “methods may need to change as engines evolve” among their limitations. Nothing was measured on AI Overviews, AI Mode, ChatGPT search or Copilot, none of which existed in their current form.
2. It assumed you were already retrieved. Every source was in the top five before any edit. The paper did not test whether an edit gets a page into the candidate set at all, and it explicitly did not evaluate the effect on search rankings.
3. The gains were uneven by rank and domain. Lower-ranked sources gained most. For the cite-sources edit, the rank-one source lost about 30 percent of its share while the rank-five source more than doubled. Which edit worked best also varied by topic: statistics helped law and debate queries most, quotations helped history and explanation queries.

Our view: the paper is good evidence that quotable, specific, sourced writing gets more of an answer attributed to it once you are in the candidate set. It is not evidence for a percentage you can promise a client.

How do generative engines choose sources?

Most generative engines use retrieval-augmented generation: they run one or more searches against an index, pull the top documents or passages, and hand them to a model that writes the answer and cites what it used. Your page has to be in that index, retrievable for the engine’s sub-queries, and quotable once it is retrieved.

The retrieval step is what SEO people recognize. The vocabulary is new: retrieval-augmented generation is the pattern, grounding is what Google and Microsoft call feeding retrieved content to the model, and the search itself is often several searches.

Google says: both AI Overviews and AI Mode “may use a ‘query fan-out’ technique, issuing multiple related searches across subtopics and data sources”, per the AI features documentation. Robby Stein’s March 2025 AI Mode announcement describes the same thing: the system issues multiple related searches concurrently across subtopics and brings the results together. Google has never published how many sub-queries a fan-out runs. The practical consequence is that a page can be cited for a sub-question the user never typed. We cover this in query fan-out explained.

The engines differ in where the index comes from, which bot you must allow, and whether you can see any of it. This table is built from each company’s own documentation.

Engine	Where its sources come from	The bot that matters	robots.txt behavior	Publisher reporting
Google AI Overviews and AI Mode	Google’s Search index. A page “must be indexed and eligible to be shown in Google Search with a snippet”.	Googlebot. There is no separate AI crawler. Google-Extended is a training and grounding token for Gemini products and “does not impact a site’s inclusion in Google Search”.	Googlebot rules apply. nosnippet and max-snippet also limit what can be used as “a direct input for AI Overviews and AI Mode”.	Search Console Performance report (Web) and the generative AI performance report.
ChatGPT search	OpenAI’s own crawl. “OAI-SearchBot is for search.”	OAI-SearchBot (search), ChatGPT-User (fetches triggered by a user), GPTBot (training only). Each setting is independent.	Sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers”. For ChatGPT-User, “robots.txt rules may not apply”.	None. GA4 referrals only.
Perplexity	Perplexity’s own index, built by PerplexityBot, which “is not used to crawl content for AI foundation models”.	PerplexityBot (index), Perplexity-User (fetches triggered by a user).	Perplexity recommends allowing PerplexityBot. Perplexity-User “generally ignores robots.txt rules”.	None. GA4 referrals only.
Microsoft Copilot and Bing AI summaries	Bing’s index. Copilot Chat “generates a search query that it sends via secure connection to the Bing search service”.	bingbot.	bingbot rules apply. data-nosnippet content is “excluded from snippets and AI summaries”.	Bing Webmaster Tools AI Performance report.
Claude	Anthropic’s search index plus live fetches.	Claude-SearchBot (index), Claude-User (fetches triggered by a user), ClaudeBot (training).	All three “respect ‘do not crawl’ signals by honoring industry standard directives in robots.txt”.	None.

Sources: Google crawlers, OpenAI bots, Perplexity bots, Anthropic crawlers.

Two details trip people up. First, blocking GPTBot does not remove you from ChatGPT search; blocking OAI-SearchBot does. Second, the Perplexity and ChatGPT user-triggered fetchers can still read a page you have disallowed, because the request was made by a person. You can check how your robots.txt treats every one of these bots with the robots.txt tester.

GEO vs SEO vs AEO: what is the difference?

SEO earns a ranking in a list of links. AEO earns the single extracted answer, historically a featured snippet and now the passage an AI quotes. GEO earns citation inside a generated answer that draws on several sources. All three depend on being crawled, indexed and clearly written; they differ in the unit of success and in measurement.

	SEO	AEO	GEO
Unit of success	Ranking position for a query	Being the extracted answer	Being cited in a multi-source generated answer
Primary surfaces	Organic listings	Featured snippets, People Also Ask, voice assistants, the direct-answer passage inside an AI response	AI Overviews, AI Mode, ChatGPT, Perplexity, Copilot, Claude
What you work on	Crawlability, relevance, links, page experience	Question-led structure, concise answers, entity clarity	Everything in SEO, plus quotable facts, statistics, citations, and mentions on other sites
How you measure	Rankings, clicks, impressions	Snippet ownership, impression share	Citation counts, share of voice across prompts, AI referrals
Main risk	Losing rank	Answering so well nobody clicks	Being summarized without attribution or traffic
Read more	How to do an SEO audit	What is answer engine optimization	This guide

The acronyms are marketing categories, not separate systems. AI SEO and LLMO are used for the same thing. Pick one vocabulary for your team and stop there.

What does the evidence actually support?

The evidence supports five things: Google needs no special markup or files, only an indexed and snippet-eligible page; specific, sourced writing earns more of a generated answer in a controlled benchmark; brand mentions correlate with AI visibility more strongly than links; adding schema produced no measured citation lift; and Microsoft wants discrete, dated, verifiable facts. Each carries its evidence label.

1. Google says no special markup, file or optimization is needed. “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”, and “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.” The requirements are the ordinary ones: indexed, snippet-eligible, not blocked for Googlebot. John Mueller’s May 2025 post adds the content advice: create “non-commodity content”, keep structured data consistent with visible text, and make sure important content “is available in textual form”.
2. Research shows specific, sourced writing earns more of the answer. In the KDD 2024 benchmark, adding quotations, statistics and citations raised a source’s share of the generated response by roughly 28 to 41 percent on the main metric, while keyword stuffing lowered it. That is a benchmark result (see the limits above), but the direction is consistent with what Google and Microsoft recommend.
3. Research shows brand mentions correlate with AI visibility more than links do. Ahrefs’ analysis of 75,000 brands (May 2025) found a Spearman correlation of 0.664 between branded web mentions and AI Overview brand visibility, against 0.218 for backlinks and 0.326 for Domain Rating. The authors say plainly that correlation is not causation, and the sample was limited to domains above DR 40.
4. Research shows adding schema did not buy citations. Ahrefs’ May 2026 study tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against matched controls. Citations moved by minus 4.6 percent on AI Overviews, plus 2.4 percent on AI Mode and plus 2.2 percent on ChatGPT, none of it a meaningful uplift. Caveat: every page in the sample was already heavily cited, so the study says nothing about pages AI systems do not yet see.
5. Vendor claim, worth reading as a statement of intent rather than a measurement: Microsoft’s search team wrote in May 2026 that the unit of value shifts to “groundable information – discrete, supportable facts with clear provenance”, and that “abstention is a valid outcome when support is missing, stale, or conflicting”. Microsoft Advertising’s October 2025 guidance tells publishers to write sentences that “make sense even when pulled out of context” and not to “hide important answers in tabs or expandable menus”.

Our view: most of GEO is SEO done with more discipline about facts and sources, plus an off-site program that earns mentions. The remainder is measurement.

What is contested, and what is myth?

Four common claims have no documented support: that llms.txt affects inclusion, that content must be “chunked” to a size, that schema markup earns citations, and that buying mentions moves a model. Two more are genuinely contested: whether GEO is “just SEO”, and how much traffic AI Overviews remove, where published studies use different methods and reach different numbers.

llms.txt

The proposal is a Markdown file at /llms.txt that summarizes your site for language models. Google says Search does not use it; John Mueller has repeatedly said no major AI system currently reads it, and has compared it to the keywords meta tag. OpenAI, Anthropic and Perplexity do not document any use of it either. It is harmless and takes ten minutes, but it is not a ranking or citation input. Our llms.txt guide covers when it is still worth publishing.

“Chunking” your content for AI

The claim is that you should write in fixed-size blocks because retrieval systems split pages into chunks. Retrieval systems do split pages, but every engine does its own splitting and none publishes the rules. What the documentation does support is narrower: self-contained sections with a clear heading and a direct statement near the top. Microsoft’s guidance calls these concise answers of one to two sentences that directly address a question. That is good writing, not a chunk size.

Schema markup “for AI”

Structured data is not a citation signal that any engine documents. Google’s advice is to keep markup consistent with visible text; Microsoft’s is that schema helps systems “understand your content”. Ahrefs’ causal test found no uplift. Keep your structured data valid and honest because it still supports rich results, and validate it with the schema markup validator. Do not buy a “GEO schema package”.

Buying mentions

Because mentions correlate with AI visibility, some vendors sell placements on listicles and directories. Three problems: correlation is not causation, so there is no evidence the bought mention moves the model; mentions on low-quality pages are the ones least likely to be retrieved; and Google says creating links “primarily for the purpose of manipulating search rankings” is link spam, which still applies to the index every AI feature reads.

“GEO is just SEO”

This is Google’s position and it is largely right for Google’s own features, which use the same index, crawler and controls. It is less complete for ChatGPT and Perplexity, which run their own crawlers and have their own robots.txt entries, and for Copilot, where Bing’s index matters and most sites have never looked at Bing Webmaster Tools. “Just SEO” also skips the measurement problem: rankings are not citations, and no rank tracker shows you an AI Mode answer.

AI Overview click-through rates

Studies disagree, and the methods differ. Research shows (Ahrefs, February 2026) a 58 percent lower average click-through rate for the top result when an AI Overview is present, updating its earlier 34.5 percent figure; the comparison is desktop only and uses a forecast counterfactual. Research shows (Pew Research Center, July 2025) that panel users clicked a traditional result on 8 percent of visits with an AI summary versus 15 percent without, from 900 US adults in March 2025. Google says clicks from pages with AI Overviews “are higher quality”. None of these measures the same thing, so quote the method with the number.

How do you measure generative engine optimization?

Measure GEO in four places: Search Console’s generative AI performance report for AI Overviews and AI Mode impressions, Bing Webmaster Tools’ AI Performance report for Copilot citations, a fixed panel of prompts you re-run on a schedule, and AI referral sessions in GA4. None is complete on its own; together they show direction.

Google Search Console

Google says the Search generative AI performance reports, announced in June 2026 and rolled out to all sites by the end of August 2026, show how often your URLs appeared in AI Overviews and AI Mode, broken down by page, country, device and date. The report shows impressions; clicks and queries remain merged into the overall Performance report under the “Web” search type. So you can see which pages appear in AI features, but not what was asked or whether anyone clicked.

Bing Webmaster Tools

Microsoft’s AI Performance report (public preview since February 2026) shows total citations across Copilot, Bing AI summaries and partner integrations, average cited pages per day, page-level citation counts and a sample of the grounding queries used to retrieve your content. A June 2026 update added intent and topic groupings and Citation Share, defined as “the percentage of citations attributed to your site out of all citations shown across all sites” for the same grounding query. Citations are not traffic; Microsoft says so in the help text.

A prompt panel

Nothing reports ChatGPT, Perplexity or Claude citations to publishers. The repeatable substitute is a panel: 20 to 50 prompts that matter to your business, written the way customers ask, run on the same day each month in each engine, with the cited domains recorded in a sheet. From that you get AI share of voice over time, which is the closest thing to a rank tracker this channel has. The AI visibility checker runs the readiness side of this for a single URL: which bots can reach it, whether it is snippet-eligible, and whether the content is extractable without JavaScript.

GA4 referrals

Referral traffic from chatgpt.com, perplexity.ai, copilot.microsoft.com, claude.ai and gemini.google.com is visible in GA4 as ordinary referrals; a custom channel group that matches those hostnames gives you one line to watch. Two caveats. Clicks from AI Overviews and AI Mode arrive as google / organic and cannot be separated in GA4. And Google says to measure “the full value of your traffic”, meaning conversions rather than sessions, which is fair advice for a channel that sends fewer, better-informed visitors.

A GEO checklist

Run the free SEO audit first; it covers the first six items for any URL in about a minute. Then work down the list.

• Googlebot, bingbot, OAI-SearchBot, PerplexityBot and Claude-SearchBot are allowed in robots.txt, and you have decided separately about GPTBot, ClaudeBot and Google-Extended (training).
• The page is indexed in Google and Bing and carries no noindex or nosnippet that you did not intend.
• Main content is in the HTML as text, not injected by JavaScript or hidden behind tabs that load on click.
• Each section opens with a direct statement that stands alone when quoted.
• Claims are specific: numbers with their source, named studies, dated statements.
• You cite primary sources with links, and you are quotable in return (a definition, a figure, a comparison someone else would cite).
• Author and organization are identifiable on the page and consistent across the web.
• Structured data validates and matches the visible text; nothing in markup is missing from the page.
• Page updates are real and dated; stale facts are the thing Microsoft says produces wrong answers.
• The brand is mentioned on pages you do not control: reviews, comparisons, industry publications, communities.
• Search Console’s generative AI report and Bing’s AI Performance report are both set up and reviewed monthly.
• A prompt panel exists, with a fixed list, a fixed day and a sheet of cited domains.

What to do next

Start with eligibility, because it is binary and cheap: check robots.txt for every bot in the table above, confirm the page is indexed in both Google and Bing, and remove any snippet restrictions you did not mean to set. Then rewrite your three most important pages so that every section opens with a quotable, sourced statement, and remove any content that is only reachable by clicking. Set up the two publisher reports and a prompt panel before you change anything else, so you can tell in three months whether the work moved citations. The AI visibility guide turns this into a platform-by-platform tutorial, and GEO vs SEO covers when the extra effort is worth it.

Frequently asked questions

Is generative engine optimization different from SEO?

Mostly no. Every major generative engine draws on a search index, so crawlability, indexing and clear writing are shared. GEO adds a different unit of success (citations rather than rankings), a stronger emphasis on quotable facts and off-site mentions, and different measurement.

Does GEO really give a 40 percent visibility boost?

That figure comes from the KDD 2024 benchmark, where adding quotations lifted a source's share of the generated answer by about 41 percent on the paper's main metric. It was measured in a controlled setup with GPT-3.5 and Google's top five results, not on live products, so treat it as direction rather than a promise.

Do I need an llms.txt file for GEO?

No. Google has said it does not use llms.txt, and no major engine documents using it for retrieval. It is harmless to publish, but it is not what gets you cited. Indexing, snippet eligibility and clear content do that.

Which crawlers do I need to allow for GEO?

Googlebot for AI Overviews and AI Mode, OAI-SearchBot for ChatGPT search, PerplexityBot for Perplexity, bingbot for Copilot, and Claude-SearchBot for Claude. Blocking GPTBot, ClaudeBot or Google-Extended affects training, not search inclusion.

How do I measure generative engine optimization?

Use Search Console's generative AI performance report for AI Overviews and AI Mode impressions, Bing Webmaster Tools' AI Performance report for Copilot citations, a fixed panel of prompts you re-run monthly, and AI referral sessions in GA4. No single source is complete.

Does schema markup help with AI citations?

There is no measured lift. Ahrefs' May 2026 test of 1,885 pages that added JSON-LD found no uplift on AI Overviews, AI Mode or ChatGPT, although every page in that sample was already heavily cited. Keep schema accurate and matching the visible text; do not expect it to buy citations.