What is GEO (generative engine optimization) and how does it work?
GEO (generative engine optimization) is the practice of preparing a website so that AI answer engines – ChatGPT, Perplexity, Gemini, Google AI Overviews, Copilot – can find it, understand it and cite it inside their generated answers. Unlike classic SEO, where the goal is a ranking position and a click, the goal of GEO is to be the source the model quotes and names. It works through a mix of technical access for AI crawlers, content written in short, self-contained, factual blocks, structured data, and a recognizable brand presence across sites the models already trust.
What does GEO actually optimize?
GEO optimizes three things at once: whether an AI system can reach your pages, whether it can extract a clean answer from them, and whether it has a reason to prefer your answer over a competitor's.
The unit of optimization changes. In SEO the unit is a page competing for a keyword. In GEO the unit is a passage – a paragraph, a list, a table row – that can be lifted out of context and still make sense. A model does not quote your whole article; it quotes the two sentences that answer the question.
The second change is the outcome. A GEO "win" is a citation, a brand mention, or a recommendation inside an answer, often with no click at all. That makes brand visibility and measurement harder, but it does not make it optional: if a user asks "best CRM for small law firms" and three vendors are named, the ones that are not named lose the consideration stage entirely.
How is GEO different from SEO?
They overlap heavily. Most AI answer engines lean on a search index (their own, Bing's, or Google's) to retrieve candidate pages, so a site invisible in classic search is usually invisible in AI answers too. The differences are in emphasis.
| Aspect | Classic SEO | GEO |
|---|---|---|
| Goal | Ranking position, click | Citation, mention, recommendation |
| Unit | Page vs. keyword | Passage vs. question |
| Query type | Short keywords | Long conversational questions |
| Content shape | Long pages, internal links | Short self-contained blocks, direct answers |
| Technical core | Crawlability, Core Web Vitals, sitemaps | AI crawler access, structured data, clean HTML, llms.txt |
| Trust signals | Backlinks, domain authority | Mentions and consistency across third-party sources |
| Measurement | Rankings, impressions, clicks | Share of citations, AI referral traffic, answer wording |
Practical conclusion: GEO is not a replacement for SEO. It is a layer on top of it, plus a rewrite of how your pages present answers.
How do AI answer engines choose which sources to cite?
The pipeline differs between products and changes often, but the well-established general shape looks like this:
- Query interpretation. The user's question is rewritten and often fanned out into several search queries ("best CRM for small law firms" becomes several narrower searches).
- Retrieval. The system pulls candidate documents from a search index, a live crawl, or both.
- Chunking and ranking. Documents are split into passages; passages are scored for relevance to the rewritten query.
- Grounding and generation. The model writes an answer using the top passages and attaches citations to the sources it used.
Two implications matter. First, your page competes at passage level, so a buried answer loses to a competitor's clearly labelled one. Second, if your content cannot be parsed – it is rendered only in JavaScript, locked behind a cookie wall, or sits inside an image – it never enters step 3.
Beyond relevance, systems tend to favour sources that look verifiable: pages with clear authorship, dates, specific figures with attribution, and agreement with other sources on the same facts. Where a claim is contested, the model usually prefers the version repeated across several independent sites.
Which AI crawlers need access to your site?
Different bots do different jobs. Some collect training data, some fetch pages live when a user asks a question, some index for AI search. Blocking all of them removes you from AI answers; blocking none means your content may also be used for training. That is a business decision, and the two can be separated.
| User agent | Operator | Typical role |
|---|---|---|
GPTBot | OpenAI | Crawling, training data |
OAI-SearchBot | OpenAI | Indexing for ChatGPT search |
ChatGPT-User | OpenAI | Live fetch on user request |
PerplexityBot | Perplexity | Indexing for answers |
Perplexity-User | Perplexity | Live fetch on user request |
ClaudeBot | Anthropic | Crawling for Claude |
Google-Extended | Gemini / AI grounding permissions (not a separate crawler) | |
Bingbot | Microsoft | Index behind Bing and Copilot |
Applebot-Extended | Apple | AI usage permissions |
CCBot | Common Crawl | Open dataset used by many models |
A robots.txt that allows AI search but blocks bulk training collection:
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Perplexity-User
Allow: /
User-agent: GPTBot
Disallow: /
User-agent: CCBot
Disallow: /
User-agent: *
Allow: /
Sitemap: https://example.com/sitemap.xml
Check your CDN and WAF too. Bot-protection rules often block AI crawlers by default even when robots.txt allows them, and the only way to notice is in server logs.
What should content look like so an AI can quote it?
Write so that any 2–4 sentence block can be pulled out and still be correct and complete.
- Answer first. Put the direct answer in the first paragraph under the heading, then explain. Models tend to pull the clearest early statement.
- Question-shaped headings. "How much does a GEO audit cost?" matches how people ask; "Pricing" does not.
- Self-contained paragraphs. Avoid "as mentioned above" and pronouns that depend on earlier text.
- Facts with handles. Numbers, dates, versions, named steps, units and prices are more quotable than adjectives.
- Tables and lists for comparisons. Structured text survives chunking better than prose.
- Entity clarity. Name your product, category and location explicitly on the page instead of relying on the reader knowing the context.
- Visible freshness. Show a published and updated date, and keep facts current; stale figures lower the chance of being used.
Avoid the opposite pattern: a 3000-word page where the answer appears in paragraph nine, wrapped in marketing language.
Do llms.txt and schema.org help?
Schema.org markup is well established. It gives machines an unambiguous version of what the page says – product, price, author, FAQ, organization – and it feeds knowledge graphs that AI systems draw on.
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [{
"@type": "Question",
"name": "What is GEO?",
"acceptedAnswer": {
"@type": "Answer",
"text": "GEO (generative engine optimization) is the practice of preparing a site so AI answer engines can find, understand and cite it."
}
}]
}
Also keep Organization markup consistent across the site, with sameAs links to your profiles, so the model can connect mentions of your brand to one entity.
llms.txt is a proposed convention: a Markdown file at the root that lists your most important pages with short descriptions, so a model can orient itself quickly. It is not officially supported by the major engines, and there is no public confirmation that it affects citations today. It costs almost nothing to publish, so treat it as a low-risk bet, not a ranking factor.
# Example Company
> B2B scheduling software for clinics.
## Docs
- [Getting started](https://example.com/docs/start): installation and first booking.
- [API reference](https://example.com/docs/api): endpoints, auth, limits.
## Company
- [Pricing](https://example.com/pricing): plans and current prices.
- [About](https://example.com/about): team, location, contact.
How do you measure GEO results?
There is no equivalent of Search Console for AI answers yet, so measurement is assembled from several sources.
- Prompt testing. Build a list of 30–100 questions your buyers would ask, run them across ChatGPT, Perplexity, Gemini and Copilot on a schedule, and record whether you are cited, mentioned, or absent. Answers vary between runs, so repeat and track a share, not a single result.
- Referral traffic. Filter analytics for referrers such as
chatgpt.com,perplexity.ai,copilot.microsoft.comandgemini.google.com. Volumes are usually small relative to search but often convert differently [check: your own analytics]. - Server logs. Confirm that AI crawlers actually reach your pages and are not blocked by the CDN.
- Brand queries. Rising branded search and direct traffic can indicate answer-level exposure that never produced a click.
- Answer wording. Note which phrases the models repeat about you. If the description is wrong, fix the page that produced it.
What is a realistic first GEO sprint?
- Audit crawler access: robots.txt, CDN rules, server-rendered HTML for key pages.
- Pick the 10 pages that matter commercially and add a direct answer in the first paragraph of each.
- Convert headings into the questions customers ask, and add a short FAQ block to each page.
- Add or fix
Organization,Product,ArticleandFAQPageJSON-LD. - Publish
llms.txtand keep it in sync with the sitemap. - Replace vague claims with specific, checkable facts and dates.
- Build or update third-party presence: directories, review sites, comparison articles and expert roundups in your category, so the model sees you mentioned outside your own domain.
- Set up the prompt list and run the first baseline measurement before the changes land, so you can compare later.
FAQ
Is GEO replacing SEO?
No. Retrieval in most AI answer engines still runs through a search index, so technical SEO, crawlability and authority remain prerequisites. GEO changes how you structure and phrase content, and adds citation tracking to your reporting.
How long does it take to appear in AI answers?
It depends on how often the engine recrawls you and whether the answer is generated live or from a cached index. Live-fetching tools can reflect changes within days; index-based ones take longer. Treat weeks, not hours, as the normal horizon.
Does blocking GPTBot remove me from ChatGPT answers?
Not necessarily. GPTBot is primarily associated with crawling for training, while search and live fetching use different agents. Blocking the search-related agents is what removes you from answers, which is why the two should be configured separately.
Do I need llms.txt?
It is optional. No major engine has publicly confirmed it uses the file, but it is cheap to publish and harmless if ignored. Prioritize crawler access, structured data and answer-first content first.
Can small sites be cited at all?
Yes. Passage-level retrieval rewards precise, specific answers, so a narrow site with exact, well-marked facts can be quoted on a question where a large generic page is vague. Topical focus beats raw domain size for specific questions.
If you want to know where your own site stands, run a free check in Zgadka: it gives GEO and SEO scores, a full report with fixes and ready-made files (robots.txt, llms.txt, schema.org, FAQ), an action plan you can hand to a developer, and tracking of how often you are cited in AI answers.