Open Google today and the first thing on the page usually isn't a blue link — it's a paragraph the AI wrote, with a handful of small citations tucked underneath. Ask ChatGPT, Claude, or Perplexity the same question and there's no results page at all. You just get the answer.
That shift has created a discipline sitting next to SEO, not replacing it: Generative Engine Optimization, or GEO.
TL;DR
- SEO competes for a spot among ~10 blue links. GEO competes for one of the ~2–7 sources an AI actually cites.
- AI systems usually lift a single paragraph or sentence from a page, not the whole thing — structure matters more than ever.
- Ranking #1 on Google no longer guarantees an AI will mention you.
- The fixes aren't exotic: answer-first writing, real numbers, less hype, extractable formatting, fresh dates, off-site reputation — plus a couple of developer-specific moves like fixing your
robots.txt.
What Is GEO, Exactly? 🎯
Generative Engine Optimization is the practice of structuring content so AI systems — ChatGPT, Claude, Gemini, Perplexity, Google's AI Overviews — can find it, trust it, and cite it while generating an answer. Princeton researchers coined the term in 2023, and the industry has spent the years since arguing about what to call it. You'll also see AEO (Answer/AI Engine Optimization), AI SEO, and LLMO (large language model optimization) used more or less interchangeably. Nobody's fully settled on one name. Everyone's chasing the same goal: get cited by the system now standing between you and your reader.
The mental model that matters: SEO optimizes for a click. GEO optimizes for a mention.
SEO vs. GEO: What Actually Changed
| SEO | GEO | |
|---|---|---|
| Competing for | A spot among ~10 blue links | One of ~2–7 sources an AI cites |
| Unit that gets judged | The whole page | A single paragraph or sentence |
| Success looks like | A click to your site | A citation — the click is optional |
| What earns it | Keywords, backlinks, technical SEO, page experience | Clear answers, real data, structure, off-site trust |
| Being #1 guarantees | Strong visibility | Nothing. The model decides independently |
Why This Is Suddenly Everyone's Problem 📉
The numbers back this up. Depending on whose research you trust, somewhere between 58% and 68% of Google searches in the US now end without a click to an outside site — and when an AI Overview shows up on the page, that number climbs past 80%. Pew Research found that people click through on a source cited inside an AI Overview only about 1% of the time.
That's not the whole story, though. Seer Interactive's research found that brands who do get cited inside AI Overviews still pull in roughly 35% more organic clicks than brands who don't appear at all. The traffic didn't vanish — it consolidated around whoever gets mentioned.
This is spreading past consumer search, too. By some industry estimates, somewhere between a third and half of B2B research queries now start inside an LLM chat window instead of a search bar.
How AI Engines Actually Decide What to Cite 🔍
Here's the part that trips up a lot of developers: an AI engine doesn't just paste your question into a search box. Ask "what's the best VPN for streaming in Europe" and the model quietly breaks that into several smaller searches — something like "best VPN 2026," "VPN streaming speed comparison," and "VPN server locations Europe" — runs each one separately, then synthesizes an answer from whatever comes back across all of them. This is often called query fan-out, and if you've built a RAG (retrieval-augmented generation) pipeline yourself, it'll look familiar. It's the same pattern, just running one layer up, outside your control.
The practical consequence: your page doesn't need to rank #1 for one target keyword anymore. It needs to be the best-structured answer to any of the sub-questions the model might generate along the way. Trackers that measure the overlap between classic top-10 rankings and the sources AI engines actually cite report that overlap shrinking fast — some analyses put it dropping from roughly three-quarters down to under 40% over the past couple of years. Ranking well still helps. It's just no longer sufficient on its own.
The GEO Playbook: 8 Things You Can Do This Week 🛠️
1. Answer first, explain second
Put the actual answer in the first sentence or two, then back it up — classic inverted-pyramid structure, and exactly what a generative engine is trying to extract.
Buried: "Choosing the right database is one of the most consequential decisions a team makes, with many trade-offs to weigh — scalability, consistency, ecosystem maturity, long-term cost. In this piece we'll walk through several options before landing on a recommendation."
Answer-first: "For most new web apps, Postgres is the safest default — it covers relational and JSON workloads, has mature tooling, and scales further than most teams expect to need. Here's why, and when you'd reach for something else."
Both say roughly the same thing. Only one of them is quotable as a standalone answer.
2. Trade adjectives for numbers
"Extremely fast" gets filtered out. "40ms p95 latency across three regions" gets cited, because it's specific and checkable. Princeton's original GEO research found that adding concrete statistics, citing sources, and including direct quotes from credible people improved a page's odds of being cited by roughly 30–40% over unoptimized content. Specificity isn't just more persuasive to people — it's more retrievable by machines.
3. Kill the marketing voice
"World-class," "industry-leading," "revolutionary" — language that reads as a claim rather than a fact tends to get quietly discounted. Write the way you'd explain something to a smart colleague who has no reason to trust your adjectives: plainly, specifically, trade-offs included. The more neutral the tone, the more it reads as information instead of an ad — which, counterintuitively, makes it more persuasive.
4. Format for extraction, not for scrolling
AI systems typically lift one paragraph or one table row, not your whole page. Give them clean units to grab: short paragraphs, one idea per section, descriptive headers, comparison tables, and an FAQ block phrased the way people actually ask questions. If you wouldn't be comfortable pasting a single paragraph into a chat window as a standalone answer, it probably isn't extractable yet.
5. Timestamp everything
Generative engines weight freshness heavily, especially for anything volatile — pricing, statistics, version numbers, "best of" lists. Add a visible "Last updated" date, and actually update the numbers when you touch the page. An honest range from last month usually beats a precise-looking stat from two years ago.
6. Build a reputation the AI can find off your site
These systems don't only read your own copy — they weigh what Reddit threads, forums, review sites, and independent comparison articles say about you too. A page that claims "we're the best" carries little weight. A Reddit thread full of people independently agreeing carries a lot. Community reputation is now part of technical SEO, not a separate marketing workstream.
7. Check whether the bots can even reach you
This is the one developers skip and shouldn't. Most AI providers now run at least two kinds of crawler: one that harvests pages for model training, and one (sometimes two) that fetch pages live to answer a specific question and cite it. Blocking the wrong one quietly removes you from AI answers without you ever noticing.
Anthropic's current documentation, for example, splits Claude's crawling into three named bots: ClaudeBot (training data collection), Claude-SearchBot (indexing for Claude's search features), and Claude-User (fetching a page in real time because a person asked Claude a question that needed it). Anthropic states all three honor robots.txt — including Claude-User, which is a stronger commitment than some other providers make for their user-triggered fetchers. OpenAI runs a comparable split with GPTBot, OAI-SearchBot, and ChatGPT-User; Perplexity runs PerplexityBot for indexing and Perplexity-User for retrieval.
A reasonable starting point — allow retrieval and citation, opt out of training only if that's a deliberate policy choice:
# Traditional search — leave these open
User-agent: Googlebot
Allow: /
User-agent: Bingbot
Allow: /
# Live retrieval + citation bots — these fetch a page to
# answer a specific question and cite it. Block these and
# you disappear from AI-generated answers.
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: Claude-User
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
# Model-training bots — harvest content for future training
# runs. A data-policy decision, not a GEO one: blocking
# these doesn't affect today's citations.
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
User-agent: Google-Extended
Disallow: /
Sitemap: https://yoursite.com/sitemap.xml
Bot names change — Anthropic, OpenAI, and Perplexity have all deprecated or added user-agents in the past two years (Claude-Web and anthropic-ai are both retired). Verify the current strings in each provider's official docs before you ship this, and check your server logs afterward to confirm you got it right.
8. Mark up the facts a machine can quote directly
Structured data (schema.org via JSON-LD) turns your key facts into something a machine can parse without guessing at meaning. It won't guarantee a citation, but it removes ambiguity about what your page is actually claiming — especially useful for FAQs, specs, and how-tos:
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What is Generative Engine Optimization?",
"acceptedAnswer": {
"@type": "Answer",
"text": "GEO is the practice of structuring content so AI systems like ChatGPT, Claude, and Perplexity can find, trust, and cite it when generating an answer."
}
}
]
}
Where llms.txt Actually Fits 📄
You've probably seen /llms.txt described as "robots.txt, but for AI." It's a proposed convention — a markdown file at your site's root summarizing your most important pages in a format built for LLMs to skim instead of parsing full HTML.
The honest state of it: it's a community convention, not a ratified standard, and no major AI lab has publicly committed to reading it for search citations. Google's webmaster team has said outright that it doesn't use llms.txt, comparing it to the old meta-keywords tag. Large studies of live sites put adoption around one in ten, and traffic analysis shows major AI crawlers overwhelmingly skip the file and crawl HTML directly anyway.
So is it worth doing? Probably — for a different reason than "citations." It costs about an hour, and it's a genuinely useful entry point for AI coding agents and assistants that need to navigate documentation efficiently: a curated "here's what actually matters" map rather than a ranking lever. Treat it as cheap hygiene, not a growth strategy.
Measuring Whether Any of This Is Working 📊
Traditional analytics mostly can't see this activity — a citation inside an AI answer often generates zero referrer traffic. A newer category of "AI visibility" tools has emerged to track how often a brand gets mentioned across ChatGPT, Claude, Perplexity, and AI Overviews for a set of target queries, similar to how rank trackers work for classic SEO.
You don't need to buy one to start, though. Pick the 10–20 questions you most want to be the answer to, ask them yourself across a few AI assistants once a month, and note who gets cited. It's manual, but it's free — and for the first few months, it'll tell you more than a dashboard will.
The Bottom Line 🧭
None of this replaces SEO. You still need a technically sound, well-linked, genuinely useful site, because AI engines still lean heavily on the same trust signals search engines always have. What's changed is the target. You're no longer optimizing a page to outrank nine competitors for a spot on a results page — you're optimizing an answer, competing to be the source an AI trusts enough to quote when nobody's looking at a results page at all.
Have you checked whether AI crawlers can actually reach your site? Point 7 is the highest-leverage five minutes on this whole list — I'd genuinely be curious what you find in your logs.
Top comments (1)
I like the distinction between optimizing for a mention and measuring it separately from clicks. In a 284-brand Korean DTC scan we ran, 65.5% had zero appearances across 50 AI shopping-question checks, and the mean was just 0.648 appearances out of 50. That made “mentioned, cited, absent” a more useful baseline than treating schema or llms.txt completion as a success metric. I’d be curious whether you’d split your suggested monthly tracking into those three states.