TL;DR
- What it is: Getting cited and named in AI answers. The term comes from a 2023 paper.
- Google's view: It's still SEO. Its 2026 guide says so.
- What worked in the study: Quotations +40.9%. Statistics +30.6%. Keyword stuffing -8.3%.
- What to skip: llms.txt, chunking, schema as an AI trick. Google says ignore them.
- How to check: Ask each engine your buyer questions. Many times.
Does AI name your brand? Check your site.
Generative engine optimization means one thing.
Get your content cited and named in AI answers.
ChatGPT. Perplexity. Gemini. Copilot. Google AI Overviews.
The name comes from a 2023 research paper. Researchers at Princeton, IIT Delhi and elsewhere. Presented at KDD 2024.
They tested what moves a source up inside an AI answer.
- Citations, quotations and statistics: helped.
- Keyword stuffing: didn't.
Google's take? From its side, this is still SEO.
Here's what the study measured, what Google says to skip, and what to do first. Where the evidence is thin, I'll say so.
What is Generative Engine Optimization (GEO)?
The term comes from one paper: GEO: Generative Engine Optimization.
Pranjal Aggarwal and five co-authors. Published at the ACM SIGKDD conference, August 2024.
Their definition of a "generative engine": a system that searches for sources and writes an answer from them.
- Their 2023 examples: BingChat, Google's SGE, perplexity.ai.
- Today: ChatGPT, Perplexity, Gemini, Copilot, Google's AI Overviews and AI Mode.
It has other names. Answer engine optimization (AEO). Artificial intelligence optimization (AIO). Large language model optimization (LLMO). AI SEO.
Wikipedia notes there's no consensus definition.
I treat them as one job. I'll say GEO here.
The naming fight: AI SEO vs LLM SEO vs GEO and answer engine optimization.
Searched for "search generative experience optimization"? SGE is the old name for Google's AI Overviews. The paper names it as one example of a generative engine. See how to rank in Google SGE.
How a generative engine picks its sources
It doesn't read the whole web when you ask a question.
It grabs a few pages. Then writes from them.
The paper's example pipeline, close to BingChat's design at the time:
- A model rewrites the query.
- A search engine retrieves sources.
- A model summarizes each source.
- A model writes the answer.
Google's guide names two techniques behind its own AI features:
- Retrieval-augmented generation (RAG). Also called grounding. Google's core ranking systems pull relevant, up-to-date pages from the index. The model reviews them. Then writes a response with links to the pages that support it.
- Query fan-out. The model writes several related queries and runs them at the same time.
Google's fan-out example. You search "how to fix a lawn that's full of weeds". It might also run:
- "best herbicides for lawns"
- "remove weeds without chemicals"
- "how to prevent weeds in lawn"
So GEO is two jobs.
One hint from fan-out. Your page can be retrieved for a question the reader never typed.
My reading: answer the obvious follow-up questions on the same page.
Don't build one page per variation just to game rankings or AI answers. Google says that violates its spam policy.
GEO vs SEO
GEO doesn't replace SEO.
Google's own words, for its own AI features: optimizing for generative AI search is optimizing for the search experience. So it's still SEO.
Its guide says the AI features are rooted in its core ranking and quality systems. And that all existing technical SEO best practices stay worthwhile.
| SEO | GEO | |
|---|---|---|
| Goal | A page that ranks in a list of results | A page cited, or a brand named, inside an answer |
| You measure | Rankings, impressions, clicks | Mentions, citations and position in the answer |
| Where the work is | Your own site and links to it | Your site, plus the third-party pages the engines read about you |
| What stays the same | Crawlable, indexed pages with useful content | The same pages |
One catch. Google's guide covers Google only.
ChatGPT and Perplexity run their own crawlers. And make their own choices. The crawler checklist below covers both.
GEO or SEO? It's a naming fight. The work here doesn't depend on who wins.
The longer comparison: AI SEO vs traditional SEO.
What the original GEO study measured
People call it the Princeton study. It's the experiment behind the term.
The setup:
- GEO-bench: 10,000 queries from nine sources.
- Nine tactics. Nine ways to rewrite a page.
- The test engine: for each query, fetch the top five Google results. Have gpt-3.5-turbo write an answer from them. Five times, at temperature 0.7.
- One source per query got the rewrite. Chosen at random.
- The score: Position-Adjusted Word Count. Built from the word count and position of the sentences that cite your source.
Read it yourself: arXiv, the ACM library, or our Markdown copy of the paper.
The results:
| Tactic | Score | Change vs no optimization |
|---|---|---|
| Quotation Addition | 27.2 | +40.9% |
| Statistics Addition | 25.2 | +30.6% |
| Fluency Optimization | 24.7 | +28.0% |
| Cite Sources | 24.6 | +27.5% |
| Technical Terms | 22.7 | +17.6% |
| Easy-to-Understand | 22.0 | +14.0% |
| Authoritative | 21.3 | +10.4% |
| Unique Words | 20.5 | +6.2% |
| Keyword Stuffing | 17.7 | -8.3% |
(I computed the change from the printed scores in the paper's Table 1. Baseline: 19.3 for no optimization.)
The paper's text says Cite Sources, Quotation Addition and Statistics Addition gave 30 to 40 percent. The table gives Cite Sources a bit less: 27.5.
And keyword stuffing? The classic SEO trick. Below baseline. The authors say it gave little to no improvement.
Three limits. Read them.
- It was a test engine. The paper tested two: the authors' own setup, and Perplexity. Not ChatGPT, Gemini or Copilot. The authors say the methods may need to adapt as engines evolve.
- It scored visibility inside an answer. Word counts and GPT-3.5 grading. No real reader clicked anything. No test of search rankings.
- The live-engine check was small. Perplexity doesn't let you pick source URLs. So the authors uploaded the source text as files, on a 200-sample subset.
On Perplexity, the paper reports:
- Best tactics: 22% over baseline on word count. 37% on the subjective score.
- Keyword stuffing: about 10% worse on word count.
That tests how Perplexity writes from sources it's given. Not how it picks them.
The rank effect.
This is the part nobody quotes.
In a second test, every source in an answer got the same tactic at once. Who gains?
| Tactic | Source ranked 1st in Google | Source ranked 5th in Google |
|---|---|---|
| Cite Sources | -30.3% | +115.1% |
| Quotation Addition | -22.9% | +99.7% |
| Statistics Addition | -20.6% | +97.9% |
The underdog wins.
When everyone used the tactics, the source that already ranked first lost visibility. The source ranked fifth about doubled.
The table's caption says GEO is especially helpful for lower-ranked sites.
So the 30 to 40 percent gains are averages. These tactics help most if you're not already first.
It also varies by topic.
- Cite Sources: best on statement, fact, and law and government questions.
- Quotation Addition: best on People and Society, Explanation, and History.
- Statistics Addition: best on Law and Government, Debate, and Opinion.
The best pair: Fluency Optimization plus Statistics Addition. More than 5.5 percent above any single tactic. (On a 200-example subset.)
More numbers from many sources: AI SEO statistics.
What Google says to skip
Google's guide has a section called "Mythbusting generative AI search".
It lists what you can ignore. For Google Search.
| Advice you'll see | What the sources say | Verdict |
|---|---|---|
| Publish an llms.txt file | Google Search ignores it, so it neither helps nor harms. Ahrefs found 97 percent of valid llms.txt files got zero requests in May 2026 | Skip for Google |
| Chunk content into small pieces | Google: no requirement. Its systems handle several topics on one page, and there is no ideal page length | Skip |
| Rewrite pages just for AI | Google: AI systems understand synonyms and meaning, so you need no special style or every long-tail variant | Skip |
| Chase inauthentic mentions | Google: this is less helpful than it seems. Its ranking systems favor quality and other systems block spam | Skip |
| Add schema markup as an AI lever | Google: no special schema.org markup is needed, but keep it for rich results. Ahrefs: 1,885 pages that added JSON-LD showed no major citation gain against 4,000 control pages. All already had 100+ AI Overview citations, so the test says nothing about pages AI has not seen yet | Skip as a GEO tactic |
| Write a page for every fan-out query | Google: doing this mainly to manipulate responses violates its scaled content abuse policy | Don't |
Two of these have outside evidence. Both from Ahrefs.
llms.txt.
Ahrefs measured the 137,210 domains in its Web Analytics that had traffic in May 2026.
- 28% publish the file.
- About 38,000 of those files are valid.
- 97% of them got no requests in May 2026.
And no AI bot went looking for files that didn't exist.
One caveat. Google ignoring llms.txt is a statement about Google. Other agents may read Markdown or llms.txt. See Markdown for agents.
Schema.
Schema is common on cited pages. Ahrefs found that AI-cited pages were almost three times as likely to have JSON-LD.
But that may be correlation. Better-maintained sites add schema.
So Ahrefs ran a controlled comparison. Against matched control pages:
- Google AI Overviews: citations fell 4.6 percent.
- AI Mode: up 2.4 percent.
- ChatGPT: up 2.2 percent.
Ahrefs calls the last two indistinguishable from zero.
Does AI name you?
See if AI mentions your brand. Then get the pages to fix.
What to do first
Start with what an engine needs to find you. Then what it needs to use.
1. Let the engines fetch you
Google: the page must be indexed and able to show a snippet. Google adds that indexing and serving are not guaranteed.
ChatGPT: OpenAI's crawler page says OAI-SearchBot surfaces sites in ChatGPT search. Opt out, and you won't be shown in ChatGPT search answers. (You can still appear as a navigational link.) OpenAI recommends allowing it.
Perplexity: its crawler page says PerplexityBot surfaces and links sites in search results. It's not used for model training. Perplexity recommends allowing it.
User-agent: OAI-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
A few details:
- GPTBot is a separate choice. OpenAI uses it to crawl content that may train its models. Each setting is independent.
- ChatGPT-User and Perplexity-User act on a person's request. OpenAI says robots.txt rules may not apply to ChatGPT-User. Perplexity says Perplexity-User generally ignores robots.txt.
- IP ranges. Both vendors publish them. Both recommend allowing the search bot's ranges.
The free Page Inspector shows what AI can read on a public page. ChatGPT in depth: how to rank in ChatGPT.
2. Say something only you can say
Google calls this non-commodity content.
It says this will likely influence your presence in generative AI search more, in the long run, than any other suggestion in its guide.
Google's examples:
- Commodity: "7 Tips for First-Time Homebuyers".
- Non-commodity: a post about why the authors waived a home inspection.
Your own data. Your own tests. Your own opinion.
3. Add sourced numbers, quotes and citations. And write clearly
These were among the best tactics in the study. Fluency alone added 28.0 percent.
Add only real ones.
And remember the rank effect. Already first? Expect less.
4. Get onto the pages the engines cite
Two pages on our leaderboard show this. Read on 2 October 2026.
- "best SEO tool": the five engines cited 24 pages. 14 were listicles. 3 were comparisons.
- "best SEO company": 23 pages cited. 20 were listicles.
Two questions are a small sample. (The labels are the leaderboard's own.)
My reading: for "best X" prompts, the engines cite other people's lists. So GEO there means getting onto those lists.
How: getting into AI visibility listicles and how to get cited in AI answers.
5. Keep facts current
Google describes retrieval as a way to bring up-to-date pages into answers.
So the dates and numbers on your page should match today.
How to check whether GEO works
Google gives you a report. For the other engines, you ask them.
Google: the Generative AI performance report
Google announced it in Search Console on 3 June 2026. Since 31 August 2026 it's available to all websites.
It shows impressions, pages, countries, devices (Search only) and dates. For AI features in Search and Discover.
More: tracking AI in Search Console.
Everyone else: ask the engines
- Write the questions a buyer types. "Best X for Y". "X alternatives". "Is X worth it".
- Ask each engine.
- Write down five things: were you named? Cited with a link? Where in the answer? Which competitors appeared? Which sources were cited?
- Repeat.
Why repeat? Answers vary. The paper's authors sampled five answers per query for that reason.
So run each prompt several times, over several days. Report a rate. Not a single yes or no.
A GEO tracker or checker automates this. Methods: how to measure AI brand visibility.
Analytics: a weak signal
Analytics counts visits. Not mentions.
Ahrefs reports for March 2026, across the 74,752 websites on chatgpt-vs-google.com:
- All AI chatbots combined: 3.5 million visitors. 0.28 percent of web traffic.
- Google: 28.12 percent.
An answer can name your brand and send no click. So referral traffic counts only part of your visibility.
A tool for the manual check
The AI SEO Tracker report. $49, once.
Up to 50 prompts, written from your site. You confirm them. It scans ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews. Daily for 7 days. One brand.
You get mentions, citations, position in the answer, competitor gaps, the sources each engine cited and a prioritized action plan. You keep read, print and export access.
It doesn't scan Google AI Mode, Claude, Grok, DeepSeek or Meta AI. No free scan. No free trial.
Other options: best AI visibility tools and the AI visibility tracker comparison.
One warning from Google: no third-party tool has access to its internal ranking or AI systems.
That applies to every tracker. This one too. A tracker shows what the engines answer. Not why.
Sources
- Optimizing your website for generative AI features on Google Search, Google Search Central, last updated 10 July 2026
- Introducing Search Generative AI performance reports in Search Console, Google Search Central Blog, 3 June 2026
- GEO: Generative Engine Optimization, Aggarwal et al., KDD 2024 (ACM, Markdown copy)
- We tracked 1,885 pages adding schema, Ahrefs, 11 May 2026
- 97% of llms.txt files never get read, Ahrefs, 15 June 2026
- AI chatbot traffic, Ahrefs, data for March 2026
- OpenAI crawlers and Perplexity crawlers
- Generative engine optimization, Wikipedia, for terminology
- AI SEO Tracker leaderboard, read 2 October 2026
