Generative engine optimization (GEO) is the work of getting your brand named, and your pages cited, in answers written by AI systems such as ChatGPT, Gemini, Claude, Perplexity and Google's AI Overviews and AI Mode. These engines run a few web searches, read a handful of pages and write one answer. GEO makes your pages the ones they read and your business one of the names they write.
It matters because buyers now ask these engines for recommendations directly. 44% of AI search users call it their primary source of insight, ahead of traditional search at 31% (McKinsey, 2025), yet only 16% of brands systematically track how they appear there (same study).
This guide covers where GEO comes from and what the original research found, how each engine chooses its sources, what Google itself says to do and not do, the tactics with evidence behind them, how to measure results when answers change from run to run, and the tools. It includes Rank.ai's own data from 11,726 AI answers to real buyer questions.
GEO, AEO, LLM SEO and AI SEO: one job, several names
| Term | Where it comes from | What it emphasises |
|---|---|---|
| Generative engine optimization (GEO) | The 2023 research paper that coined it | Engines that generate an answer from sources |
| Answer engine optimization (AEO) | Older SEO usage for featured snippets and voice | Being the direct answer |
| LLM SEO, LLM optimization (LLMO) | Practitioner shorthand | The language models themselves |
| AI SEO, AI search optimization | Broad marketing term | Everything AI-related in search, sometimes including using AI tools for SEO |
The tactics are the same whichever name you use. Our AEO vs SEO entry and GEO vs SEO comparison go deeper on the overlap with classic SEO.
Where the term came from: the GEO paper
The name was coined in GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, posted to arXiv in November 2023 and published at KDD 2024. Authors were affiliated with Princeton University and IIT Delhi, among others.
The team built a benchmark of 10,000 real queries (GEO-bench), took the pages a generative engine used to answer them, rewrote one page at a time with nine methods, and measured how much of the final answer came from that page. Their main metric, position-adjusted word count, rewards a source that supplies more of the answer and appears earlier in it.
Their headline results (Table 1); the percentage column is our arithmetic against the baseline:
| Method tested | Position-adjusted word count | Change vs. no optimization |
|---|---|---|
| No optimization (baseline) | 19.5 | 0% |
| Keyword stuffing | 17.8 | 9% worse |
| Unique words | 20.7 | +6% |
| Authoritative tone | 21.8 | +12% |
| Easy-to-understand wording | 22.2 | +14% |
| Technical terms | 23.1 | +18% |
| Cite sources | 24.9 | +28% |
| Fluency optimization | 25.1 | +29% |
| Statistics addition | 25.9 | +33% |
| Quotation addition | 27.8 | +43% |
What practitioners should take from it:
- Evidence beats vocabulary. Quotations, statistics and citations did the most. Keyword stuffing scored below doing nothing.
- Lower-ranked pages gained most. For a source ranked fifth in search, citing sources raised visibility by 115.1%, while the top-ranked source lost 30.3% with the same edit. Better evidence lets a smaller site take a bigger share of the answer.
- It varies by topic. Different methods won in different domains, so test on your own questions before rewriting a whole site.
The limits matter too. The experiments ran on 2023-era models in a controlled setup, with a follow-up check on Perplexity. Production engines have changed since. The principle has held up in later studies: engines prefer pages that hand them specific, quotable, sourced facts.
How AI engines choose which brands and pages to use
Every major engine follows the same broad pipeline. Each step is a place you can win or lose.
1. The engine decides whether to search
A model can answer from training data or search the web first. ChatGPT "may search the web automatically when your question would benefit from current information". Across all prompts, searching is the minority: about 1 in 7 free, logged-out ChatGPT responses (14.4%) ran a web search, across 66.6 million responses (Ahrefs, 2026). Buying questions are different. With search available, Rank.ai's sample of buyer questions saw ChatGPT search for 100% of answers, Claude 98.4% and Gemini 95.6%.
Two layers decide the answer, and they need different work:
| Layer | What it decides | How it changes |
|---|---|---|
| Training data | Whether the model knows your brand and links it to your category | Slowly, through wide coverage, and only when a model is retrained |
| Live search | Which pages get read and cited, and which brands get named, for this question | As soon as a better page is indexed and ranking |
2. It rewrites the question into searches
OpenAI says ChatGPT "typically rewrites your query into one or more targeted queries". Google calls its version query fan-out: AI Overviews and AI Mode issue several related searches across subtopics.
The rewrite shortens. In Rank.ai's data, buyer questions averaged 11.7 words; the searches ChatGPT and Claude ran averaged 6.5 words, and Gemini's 5.1. A question like "what's the best accounting firm in Denver for a startup that just raised a seed round" becomes something like "best startup accounting firms Denver". That short phrase is what you need to rank for.
3. It pulls results from an index
| Engine | Where its web results come from | Crawler to allow |
|---|---|---|
| ChatGPT | Third-party search partners plus OpenAI's own crawler (OpenAI) | OAI-SearchBot |
| Google AI Overviews and AI Mode | Google's own index; a page must be indexed and eligible for a snippet (Google) | Googlebot |
| Gemini | Grounding with Google Search | Googlebot |
| Perplexity | Its own index, built by PerplexityBot, plus live fetches (Perplexity) | PerplexityBot |
| Claude | Claude's web search; Anthropic runs Claude-SearchBot to improve its search results (Anthropic) | Claude-SearchBot |
If you are missing from the index an engine uses, nothing else in this guide helps on that engine.
4. It reads a few pages and picks passages
Engines find more than they read. 58% of pages ChatGPT retrieves are never cited, and the citation rate falls from 58.4% for the first result slot to 14.2% for the tenth (AirOps and Kevin Indig, 2026). Pages that state the answer plainly, near the top, under a heading that matches the question, are the easiest to use: close heading matches were cited 41% of the time against 30.2% for weak ones (same study).
5. It writes the answer, names brands and cites sources
The answer names the businesses that the pages it read agree on, and cites the pages that supplied clear answers. A brand that appears on three of the five pages the engine read is far more likely to be named than a brand with one excellent page of its own. For "best X" questions, 43.8% of the pages ChatGPT cited were "best X" list posts (Ahrefs, 2025).
What the answers actually look like: Rank.ai data
Rank.ai asked ChatGPT (gpt-4o-mini), Claude (claude-haiku-4-5) and Gemini (gemini-2.5-flash) 102 buyer questions tracked by 15 Rank.ai customer accounts, through each model's API with Rank.ai's web search tool, from May 7 to October 16, 2026. That produced 11,726 fresh answers (cached repeats excluded) and 74,761 citations. Source: Rank.ai analysis of 11,726 AI answers, May to October 2026.
| Measure | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Answers naming no business | 61.0% | 63.4% | 61.9% |
| Answers naming 5 or more | 21.7% | 22.2% | 20.8% |
| Businesses named per answer, average | 2.03 | 2.47 | 2.53 |
| Answers with at least one citation | 96.3% | 80.1% | 94.4% |
| Citations per answer, average | 6.34 | 5.40 | 7.38 |
| Distinct domains cited | 2,448 | 2,577 | 1,230 |
| Answers naming the brand that tracks the question | 6.4% | 8.1% | 11.0% |
| Answers citing that brand's own site | 3.7% | 4.5% | 7.5% |
| Average answer length (characters) | 2,139 | 1,706 | 1,666 |
What it shows:
- Answers are all or nothing. The median number of businesses per answer was zero on every engine. An answer either names none or names a list, so GEO is about getting onto lists.
- Engines disagree. Gemini named the tracking brand almost twice as often as ChatGPT. Measuring one engine misses most of the picture.
- Gemini cites more, from fewer sites. It produced the most citations from the fewest distinct domains, so the few sites it trusts in your category matter more there.
- No single site dominates. Reddit was the most-cited site, at 4.01% of all citations (Gemini 4.53%, ChatGPT 3.74%, Claude 3.62%). YouTube took 2.09%, LinkedIn 1.73%, Facebook 0.87%, Quora 0.59%, and Wikipedia under 0.3% on every engine. Everything else was spread across thousands of sites, so you find the ones that matter for your category by reading your own answers.
More figures, each sourced, are on our GEO and AEO statistics and AI search statistics pages.
What Google says to do, and what it says you don't need
In July 2026 Google published a guide to optimizing for generative AI in Search. Its core position: AI features run on Google's core ranking and quality systems, and "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." Much popular GEO advice conflicts with it, so here it is side by side.
| Google says do | Google says you don't need to |
|---|---|
| Keep doing foundational SEO: crawlable pages, sound JavaScript, good page experience | Create "new machine readable files, AI text files, markup, or Markdown", llms.txt included |
| Write original, first-hand, people-first content that isn't a commodity | Break content "into tiny pieces" for AI |
| Use clear sections and headings, with good images and video | Rewrite content specially for AI or chase long-tail variants |
| Keep Google Business Profile (local) and Merchant Center (products) complete | Add structured data for AI: it "isn't required for generative AI search" |
| Measure with Search Console's generative AI reporting | Trust tools claiming Google data: "No third-party tool has access to our internal ranking or AI systems" |
Google also warns that mass-producing pages aimed at fan-out searches can count as scaled content abuse, and against chasing inauthentic mentions.
Two caveats. This is Google's guidance for Google's AI features; ChatGPT, Claude and Perplexity have their own crawlers and partners, covered above. And Google says creating llms.txt is "completely fine" but won't help or hurt in Search. Our llms.txt entry covers the trade-offs.
What the evidence says works
The factors with published evidence behind them, and what to do about each:
| Factor | Evidence | What to do |
|---|---|---|
| Ranking in search | 38% of URLs cited in AI Overviews rank in Google's top 10 (Ahrefs, 2026); 54.5% of AI Overview citations came from pages that rank organically (BrightEdge, 2025) | Rank for the short searches engines run |
| Brand mentions across the web | 0.664 correlation with AI Overview visibility, against 0.326 for Domain Rating and 0.218 for backlinks (Ahrefs, 2025) | Get named on review sites, lists and publications |
| Being on "best X" lists | 43.8% of pages ChatGPT cited for "best X" prompts were list posts (Ahrefs, 2025) | Earn a place on the lists that rank for your category |
| Evidence on the page | +28% to +43% share of answer from citations, statistics and quotations (GEO paper) | Add numbers, named sources and short quotes |
| Headings that match the question | 41% citation rate for close matches vs 30.2% for weak (AirOps, 2026) | Use the buyer's question as an H2 |
| Focused pages | 500 to 999 words had the best citation rate (34.3%); over 5,000 words fell to 28.6% (same study) | One page per question, answered early |
| Freshness | AI-cited pages are 25.7% fresher than Google's organic results (Ahrefs, 2025) | Update money pages when facts change and show the date |
| Cited and mentioned together | Brands both cited and mentioned were 40% more likely to resurface across runs (AirOps, 2025) | Aim for your page to be the source and your brand to be the name |
| Schema markup | +2.2% change in ChatGPT citations after adding JSON-LD, no different from zero (Ahrefs, 2026) | Keep it accurate for rich results; don't expect it to move AI answers |
GEO tactics: a checklist
Work through these in order. The early items are cheap and blocking; the later ones take longer and compound.
| # | Tactic | What to do | How to check |
|---|---|---|---|
| 1 | Crawl access | Allow Googlebot, Bingbot, OAI-SearchBot, Claude-SearchBot and PerplexityBot in robots.txt and at your CDN or firewall | AI crawler checker |
| 2 | Indexing | Make sure key pages are indexed in Google and Bing | Search Console and Bing Webmaster Tools |
| 3 | Question list | Write 20 to 50 questions buyers ask, in their words | Every page you plan maps to one |
| 4 | Short searches | Turn each question into the 4 to 7 word search an engine would run, and check who ranks | Page 1 of Google and Bing for that phrase |
| 5 | One page per question | Answer it in the first two sentences, with the question as a heading | Read the first 100 words aloud: do they answer it? |
| 6 | Quotable specifics | Prices, numbers, named methods, dates, credentials, short quotes from named people | Count the facts a model could lift word for word |
| 7 | Cite sources | Link the study or data behind each claim | Every statistic has a link |
| 8 | Third-party presence | Get onto the lists, review sites, directories and threads that engines cite for your questions | Look at the citations in real answers |
| 9 | Consistent facts | Same name, prices, services, locations and hours on your site, Business Profile and listings | Search your brand and compare |
| 10 | Freshness | Update the pages that answer money questions, with a visible date | "Updated" date on the page |
| 11 | Original data | Publish numbers only you have: surveys, benchmarks, pricing data | Would another site cite this page? |
| 12 | Measure | Ask the same questions on a schedule across engines | Mention and citation rates over weeks |
Crawlers: which to allow
Search crawlers decide whether you can be cited. Training crawlers only affect what future models learn. You can block one kind and allow the other.
| Crawler | Company | Purpose | If you block it |
|---|---|---|---|
| OAI-SearchBot | OpenAI | Surfaces sites in ChatGPT search | Your pages aren't shown in ChatGPT search answers (OpenAI) |
| GPTBot | OpenAI | Training data | Content isn't used to train OpenAI models; search is a separate setting |
| ChatGPT-User | OpenAI | Fetches a page a user's request needs | Pages can't be read on demand; OpenAI says it doesn't decide search inclusion |
| Google-Extended | Gemini training and grounding | Google says it "does not impact a site's inclusion in Google Search" (Google) | |
| PerplexityBot | Perplexity | Surfaces and links sites in Perplexity search; not used for training (Perplexity) | You drop out of Perplexity's index |
| Claude-SearchBot | Anthropic | Indexes content to improve Claude's search results | Lower visibility in Claude's search (Anthropic) |
| ClaudeBot | Anthropic | Training data | Future content excluded from training |
Check your CDN and bot protection too. Firewalls can block crawlers that robots.txt allows; OpenAI publishes its crawler IP ranges so you can allow them.
A worked example: one question, start to finish
Say you run a dental practice in Denver and the question you care about is "best emergency dentist in Denver".
- Ask it. Put the question to ChatGPT with search on, Gemini, Claude and Google, three times each on different days. Write down every practice named and every page cited.
- Sort the citations. You will usually see a mix: a few practice websites, one or two "best dentists in Denver" lists, a review platform, maybe a Reddit thread. Note which mention you.
- Find the gap. Suppose four competitors are named, each has a page about emergency appointments with hours, same-day availability and prices, and two appear on a local "best emergency dentists" list. You have a generic services page.
- Fix your page. Publish an "Emergency dentist in Denver" page that states, in the first paragraph, your same-day hours, what counts as an emergency, typical costs, insurance accepted and your address, with a short FAQ.
- Fix the third-party gap. Ask the list publishers to consider you, check your Business Profile hours and categories, and ask recent emergency patients for reviews.
- Ask again. Track whether you start being named, and which page gets cited.
That loop (ask, read the sources, close the gap, ask again) is GEO in practice. For one engine in more depth, see how to rank in ChatGPT.
How to measure generative engine optimization
AI answers change from run to run, so measurement needs a method. SparkToro and Gumshoe found ChatGPT and Google's AI returned the same brand list twice in fewer than 1 in 100 runs, and the same order in about 1 in 1,000 (SparkToro, 2026).
Build a fixed question set. 20 to 50 questions split by intent: "best X", "X vs Y", "how much does X cost", "is X worth it". Keep the wording fixed so any change comes from the engines.
Ask on a schedule, across engines. Daily or weekly, in at least ChatGPT, Gemini and Claude, with a dated record of each answer.
Track four numbers per question:
| Metric | What it tells you |
|---|---|
| Mention rate | Share of answers that name your brand |
| Citation rate | Share of answers that link one of your pages |
| Share of voice | Your mentions as a share of all brand mentions across the question set |
| Who's named instead | The competitors and pages you need to beat |
Watch traffic, with caveats. ChatGPT adds utm_source=chatgpt.com to referral links, so you can build a GA4 channel group for AI referrals. Google counts AI Overviews and AI Mode clicks inside web search in Search Console and points site owners to its generative AI reporting there. Clicks undercount visibility, because many people read the answer and never click: when an AI summary appeared, Google users clicked a traditional result in 8% of visits, against 15% without one (Pew Research Center, 2025).
The visitors who do click are valuable. Across 1,277 sites, 1.66% of visitors from AI assistants signed up, against 0.15% from search (Microsoft Clarity, 2025). The LLM visibility entry defines each metric in more detail.
GEO for local businesses and ecommerce
Local. AI engines lean on reviews and listings for local questions. 45% of US consumers used ChatGPT or other AI tools for local recommendations in BrightLocal's 2026 survey of 1,002 (BrightLocal). In a loc.ai prototype index of 460 AI picks of plumbers across five cities, only 12% of the picks were in the Google Maps top 10, and 1 in 18 had no matching Maps listing at all (loc.ai local AI index, October 2026). The Maps pack and AI answers are different races. Keep your Business Profile complete, build reviews, and publish city-specific service pages with prices and hours. Our local business AI search post covers the detail.
Ecommerce. Product questions pull from product data. Google says to keep Merchant Center feeds complete, and OpenAI lets merchants apply to share a product feed for ChatGPT's shopping results. Product pages need full specs, prices and reviews in the page text.
GEO tools
GEO tools fall into a few groups:
| Group | What it does | Examples |
|---|---|---|
| AI visibility trackers | Ask your questions in AI engines on a schedule and record who's named and cited | Rank.ai, Profound, Peec AI, Otterly, Semrush AI Visibility Toolkit, Ahrefs Brand Radar |
| Content tools | Research, brief, write and optimize the pages | Surfer, Clearscope, Frase, the content features in Semrush and Ahrefs |
| Technical tools | Crawl the site, check bots and structured data | Screaming Frog, the AI crawler checker, the schema validator |
| Free spot checks | One question, one moment | The LLM SEO tool, the ChatGPT SEO check, the AI grade |
Our AI SEO tools list compares them with prices checked on each vendor's site, and the best AI visibility tools post compares the trackers.
Rank.ai covers the full loop in one plan. Its AI visibility plans ask your customers' questions in ChatGPT, Claude and Gemini every day through each engine's API with web search on, record every brand named and every page cited, show the questions where competitors are named and you aren't, then write the articles that answer them and publish them to your site. Self-serve plans start at $49 a month. To see where you stand first, the free AI grade puts 12 of your customers' questions to all three engines and scores you out of 100.
Doing GEO yourself or hiring help
Most of the checklist above is work a marketing team can do: the questions, the pages, the crawler settings, the measurement. Hiring makes sense when you have more questions than writers, or several markets to cover. If you are comparing agencies, our page on choosing a generative engine optimization agency lists the questions to ask any of them, and generative engine optimization services describes what Rank.ai's done-for-you program includes: we run Rank.ai for you, from $2,000 a month.
FAQ
What is generative engine optimization in simple terms?
It is the work of getting AI assistants like ChatGPT, Gemini and Claude to name your business and link your pages when people ask them questions. It overlaps with SEO, because those assistants search the web and read pages that rank.
Is SEO dead now with AI?
No. AI engines search an index before they answer, and Google says optimizing for its AI features "is optimizing for the search experience, and thus still SEO." 38% of URLs cited in AI Overviews rank in Google's top 10. SEO is now an input to AI answers as well as a source of clicks.
What is the difference between SEO and AIO (or GEO)?
SEO aims for a position in a list of search results, measured with a rank tracker. GEO, sometimes called AIO (AI optimization), aims to be named and cited inside a written AI answer, measured by asking the same questions repeatedly and counting mentions and citations. The tactics overlap; GEO adds third-party mentions, quotable facts and a different way of measuring.
What is the best tool for generative engine optimization?
It depends on the job. An AI visibility tracker shows where you are named and who is named instead; a content tool helps write the pages; a crawler checks access. Rank.ai combines the tracking with articles written and published to your site. Our AI SEO tools list compares the main options.
Does llms.txt help with GEO?
There is no evidence it helps. Google says you don't need AI text files to appear in its AI features, and that creating one won't help or hurt in Search. It is cheap housekeeping at best.
How long does GEO take to work?
Crawler fixes register within about a day on ChatGPT, per OpenAI. A new page has to be indexed and ranking first, which takes days to weeks. Third-party mentions take months. Judge results over two to three months of repeated measurement.
Can anyone guarantee being recommended by ChatGPT?
No. OpenAI says placement in ChatGPT search is not guaranteed, and answers change between runs. Be wary of any agency or tool that promises guaranteed recommendations.
What is the difference between GEO and AEO?
They describe the same work. AEO (answer engine optimization) is the older term from featured snippets and voice search; GEO comes from the 2023 research paper on generative engines. See our answer engine optimization entry.
