Generative Engine Optimization (GEO)
Making content citable by AI answer engines
The umbrella term for content and technical work that raises the chance a page appears and is cited as a source in answers from generative engines such as ChatGPT search, Perplexity and Google AI Overviews.
A classic search engine returns ten blue links; SEO tries to climb that list. Generative engines first retrieve source pages, then an LLM summarizes them into a single answer and shows some of them as citations. At heart this is a web-scale RAG pipeline: query → retrieval → reranking → generation. GEO aims to get your content into the answer and the source list at the end of that pipeline.
The term comes from the paper "GEO: Generative Engine Optimization" by Aggarwal et al. (arXiv November 2023, KDD 2024). The authors built GEO-bench, a 10,000-query benchmark, and tested nine methods such as adding statistics or adding quotations and citations from credible sources to the source text. In their own setup (top 5 Google results + GPT-3.5 writing the answer) the best methods raised visibility by up to 40%, and up to 37% on Perplexity.ai. Classic SEO-style keyword stuffing did not help.
In practice GEO is not new magic: a page must first be findable (crawlable, indexed, with the relevant bots allowed in robots.txt), then worth retrieving (answers the question directly, defines clearly, is current), then worth citing (verifiable numbers, sources, original information). Google says AI Overviews and AI Mode need no extra optimization and that SEO fundamentals still apply. Most of the evidence is still academic experiments or vendors' own statements; the engines are black boxes and change often.
Picture a journalist looking for experts for a story. In the old world (SEO) the goal was to be at the top of the phone book. Now the journalist talks to five experts and quotes only two of them.
The quoted expert isn't the loudest one; it's the one who answers the question in one clear sentence, gives a number, and can point to a source when asked "how do you know?". GEO is writing your content like that second kind of expert.
A software company's "What is SSO?" page opens with a long marketing intro and buries the definition in the third paragraph. The team reworks it: a two-sentence direct definition up top; a table comparing SAML and OIDC; links to the official specs; a "last updated" date; and a check that robots.txt actually allows OAI-SearchBot, Claude-SearchBot and PerplexityBot so the page can be crawled.
To measure, they track chatgpt.com and perplexity.ai referrals as a
separate channel in analytics and check the page's Copilot citations
in Bing Webmaster Tools' AI Performance report. They also know they
can't prove causation: answer engines can pick different sources for
every query.
# Bots needed for search/citation: allowed
User-agent: OAI-SearchBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
# Bots crawling for model training: blocked (an optional choice)
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
# Gemini training/grounding; does NOT affect Google Search or AI Overviews
User-agent: Google-Extended
Disallow: /
Sitemap: https://example.com/sitemap.xmlimport re
from collections import Counter
BOTS = ["GPTBot", "OAI-SearchBot", "ChatGPT-User", "ClaudeBot", "Claude-User",
"Claude-SearchBot", "PerplexityBot", "Perplexity-User", "Googlebot", "bingbot"]
REFERRERS = ["chatgpt.com", "perplexity.ai", "copilot.microsoft.com", "gemini.google.com", "claude.ai"]
bots, refs = Counter(), Counter()
line_re = re.compile(r'"[A-Z]+ (?P<path>\S+) [^"]*" \d+ \S+ "(?P<ref>[^"]*)" "(?P<ua>[^"]*)"')
with open("/var/log/nginx/access.log") as f:
for line in f:
m = line_re.search(line)
if not m:
continue
bots.update(b for b in BOTS if b in m["ua"])
refs.update(r for r in REFERRERS if r in m["ref"] or f"utm_source={r}" in m["path"])
print("Bot requests:", bots.most_common())
print("AI referrals:", refs.most_common())- Content sites, documentation, glossaries and guides that get traffic from informational queries
- When you want your product or brand to come up when people ask AI assistants 'which tool for X?'
- As the next layer once classic SEO fundamentals are solid
- When your content can offer verifiable data, sources and original information
- While technical SEO basics (indexing, crawlability, speed) are broken; fix those first
- If you expect results just from adding llms.txt or special schema; their effect on citations is unproven
- If 'optimizing' means hidden text or prompt injection aimed at AI; that is spam and a risk
Generalizing one paper's number
The 'up to 40%' was measured on the GEO paper's own benchmark with its own GPT-3.5-based engine. The same paper shows the effect varies a lot by domain and rank, and that top-ranked pages lost visibility with some methods.
Confusing training bots with search bots
Blocking GPTBot does not drop you from ChatGPT search; blocking OAI-SearchBot does. Blocking Google-Extended does not affect AI Overviews; that is governed by Googlebot.
Misattributing incoming traffic
Some clicks from AI assistants carry no referrer and show up as 'direct'. OpenAI says it adds utm_source=chatgpt.com automatically, but third-party measurements report it is not on every link. Read the numbers as a lower bound.
Focusing only on your own site
A 2025 University of Toronto study reports that AI search leans strongly towards independent, third-party ('earned') sources over a brand's own site.
Promising certainty about a black box
No engine discloses its ranking or citation logic, and results can shift with the wording of a query. Be sceptical of services selling 'guaranteed citations'.