What is generative engine optimization (GEO)? A practical guide

Generative engine optimization (GEO) is the work of getting your pages read, quoted and cited when an AI system writes an answer in ChatGPT search, Perplexity, Gemini, Claude, or Google's AI Overviews and AI Mode. SEO earns you a position in a list of links, while generative engine optimization earns you a sentence inside the answer and, with luck, a source link beside it. In practice that means letting AI crawlers read your pages, putting the content in the HTML, writing passages with specific facts worth quoting, and making your brand easy to recognize across the web.

The term is nearly three years old, and much of what gets sold as GEO is SEO hygiene with a new label.

What is GEO? The meaning behind the acronym

GEO stands for generative engine optimization. A generative engine is any search product that answers with text written by a language model instead of a list of links.

A GEO win takes one of three forms. The answer names your brand, cites your page as a source, or repeats your facts without a link. Only the citation sends a click you can measure, but all three shape what a buyer thinks before they visit.

"GEO SEO" sometimes means geo-targeted local SEO, which is a different job. Here GEO always means the AI kind. It sits on top of SEO rather than replacing it, and our AEO vs GEO vs SEO comparison covers where the three overlap.

How generative engines find and pick sources

Most generative engines run a search before they write. They turn your question into one or more queries, fetch the results, pick the passages that answer it and write a response that cites some of them.

OpenAI's ChatGPT search help page says it rewrites a question into targeted queries for its search partners. Google's guidance on AI features calls its version query fan-out, several related searches across subtopics for one response. The cited page is often the one that answers a sub-question cleanly, not the one ranking first for the head term.

Training crawlers are a separate decision. OpenAI and Anthropic document GPTBot and ClaudeBot apart from their search bots, so you can block training and still appear in answers. Google says Google-Extended has no effect on Google Search, but blocking it opts you out of grounding in the Gemini app.

What the original GEO research paper found

The paper that named the field found that adding quotations, statistics and cited sources raised a page's share of generated answers, and keyword stuffing lowered it. GEO: Generative Engine Optimization by Pranjal Aggarwal and co-authors from Princeton, IIT Delhi and elsewhere was posted in November 2023 and accepted to KDD 2024.

The team built GEO-bench, 10,000 queries from nine datasets, and a test engine in which GPT-3.5-turbo answered each query from the top five Google results. A model rewrote one of the five sources with one of nine methods, and the team measured how many words of the answer came from that source, weighted by position.

Lower-ranked pages gained the most. When the rewritten page was the fifth result, citing sources raised its visibility by 115.1%, while the same edit cut the first result's visibility by 30.3%. The effect held on a live engine too. On 200 queries run through Perplexity, adding quotations raised the word-count measure by 22%.

Read it with its limits in mind. GPT-3.5 reading five pages in 2023 is not today's ChatGPT, and the metric counts words in the answer, not clicks or mentions. I take the paper as evidence for checkable specifics over vague claims, not as a promise of the 40% lift in its abstract.

A generative engine optimization checklist

Work through these in order. Each step is wasted if the one before it fails.

Let AI search crawlers into robots.txt

Open your robots.txt and look for groups that block AI search bots, by name or through User-agent: *. This setup allows the search bots and opts out of training:

# AI search and answer bots
User-agent: OAI-SearchBot
User-agent: PerplexityBot
User-agent: Claude-SearchBot
Allow: /

# Model training. Delete this group to allow training too.
User-agent: GPTBot
User-agent: ClaudeBot
Disallow: /

A crawler follows only the most specific group that names it, and OpenAI says changes take about 24 hours to reach its search systems. Then check your CDN. A bot rule there can return a 403 to a crawler that robots.txt invites in, and Cloudflare announced in July 2025 that new domains would block AI crawlers by default.

Put the content in the HTML

Most AI crawlers read the HTML your server returns and never run JavaScript. A December 2024 study by Vercel and MERJ found that none of the major AI crawlers rendered it, except Gemini, which uses Googlebot's infrastructure. Search the raw HTML for a phrase you can see in your browser:

curl -s -A "OAI-SearchBot" https://example.com/pricing | grep -c "Team plan"

A count of 0 means the crawler gets an empty shell, so render that content on the server. AI Crawler View compares the raw and rendered HTML and lists the headings, prices, links and structured data that only appear after scripts run.

Write passages a model can quote

Open each section with a direct answer that makes sense out of context, then back it with something checkable.

Weak:     Our platform helps agencies save time on reporting.
Citable:  Acme builds a client report from Google Analytics and Search
          Console in about four minutes. Agencies in our 2025 customer
          survey said the same report took them 45 minutes by hand.

The second version names the subject, gives a number and says where the number came from. That is the paper's statistics and cite-sources methods done by hand. Write your product's name instead of "it", because an engine lifts a passage without the paragraph above it. Date time-sensitive facts, and answer the obvious follow-up questions on the same page, because fan-out searches look for them. Skip keyword repetition.

Make your brand one entity

Models describe a company from everything they read about it, so give them one name, one URL and one logo, and link your profiles with Organization markup:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Acme Reports",
  "url": "https://acmereports.com",
  "logo": "https://acmereports.com/logo.png",
  "sameAs": [
    "https://www.linkedin.com/company/acme-reports",
    "https://github.com/acme-reports",
    "https://www.crunchbase.com/organization/acme-reports"
  ]
}
</script>

Use the same name in og:site_name, your WebSite markup and your page titles. "Acme", "Acme Reports" and "AcmeReports Inc." can read as three companies. A Wikidata item that lists your official site helps if you qualify for one. Brand Entity Checker tests these signals, including whether your sameAs profiles exist.

Get mentioned on the pages engines already cite

Run five recommendation prompts your buyers would type and look at what the answers cite. On many commercial topics the cited pages are comparison articles, review sites and forum threads, not vendor pages. Get listed where your category is compared, fix outdated facts in articles that mention you, and answer questions where those threads live. You need to be among the pages the engine reads, not first in Google.

Add llms.txt if you like, but expect little

llms.txt is a Markdown file at /llms.txt that summarizes your site and links to the pages a model should read. Jeremy Howard proposed it in September 2024. It is cheap to add and unproven for search. Google's guidance says you don't need AI text files to appear in AI Overviews or AI Mode, and I have not seen OpenAI, Anthropic or Perplexity say their search features read it. It helps most with developer docs that coding agents load on purpose. More in what llms.txt is and when it is worth adding.

How to measure GEO

Measure GEO by sampling the answers people get and by counting the visits AI products send you.

Write 20 to 30 prompts in your buyers' words, mixing definitions, "best X for Y" recommendations and comparisons with your main competitor. Run each in ChatGPT, Perplexity and Gemini with search on, and record whether you're named, whether your site is cited and who gets named instead. The same prompt gives different answers from run to run, so track a rate across several runs, once a month. AI Answer Visibility runs one prompt across those three engines and reports this for each answer.

For traffic, ChatGPT adds utm_source=chatgpt.com to the links it sends. Other assistants arrive as referrers such as perplexity.ai, gemini.google.com and claude.ai. Group them into one AI channel and accept an undercount, since some apps strip the referrer.

Google is the blind spot. It counts clicks from AI Overviews and AI Mode in Search Console's Web search type, mixed in with ordinary results. Much of GEO's value is a mention that shapes a shortlist, which analytics never records, so the prompt sample matters more than the traffic chart.

Check your site's GEO readiness

Our GEO tools each cover one step above and run on their own, with the credit price shown before a run starts. AI Crawler View costs 8 credits and finds content that needs JavaScript. Brand Entity Checker costs 8 credits and checks your name, logo, Organization markup, sameAs profiles and Wikidata item. Citability Analyzer costs 5 credits and flags sections that need a clearer answer, more context or supporting evidence. Its score is an editorial aid, not a prediction that an engine will cite you.

To measure, AI Answer Visibility costs 120 credits and reports whether ChatGPT, Perplexity and Gemini name or cite you for one question, and who they name instead. AI Citation Finder costs 160 credits and shows which pages Google AI Overviews and ChatGPT cite for a topic or a website. The full set, including a log analyzer for AI crawler visits, is on the GEO tools page.