What is Generative Engine Optimization?
GEO is the work of making a brand usable as a source by generative search engines and AI assistants. What decides it is not backlink authority but crawler access, machine readability and entity clarity. This guide explains the mechanism and the measurement method, with sources.
GEO is the work of making a site something generative engines — ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews — use and cite when they write an answer. The deciding signals are crawler access, content readable without JavaScript, a clear entity definition and short passages that stay accurate out of context. Results are probabilistic; measurement is a sample.
DEFINITION
Generative Engine Optimization (GEO)
GEO is the work of turning a website into a source that generative AI engines use and cite while composing an answer. It targets attribution inside the generated text rather than a rank position.
Why it matters: part of a corporate buyer's supplier research now finishes inside an assistant window. If your name never appears in that window, you are not among the companies being weighed for the shortlist.
The short definition in the glossary Scope and deliverables of the GEO service
What is GEO?
GEO (Generative Engine Optimization) is the work of turning a website into a source that generative AI engines use and cite while writing an answer. The target is not a rank position but the appearance of the brand name and a page link inside the generated text. The term was defined by an academic paper in 2023.
In classic search visibility the user sees ten links and picks one. In a generative engine the user sees a single block of answer text, and the sources appear as a few links inside or beside it. The difference is not one of degree but of kind: in the first case visibility is a rank position, in the second it is being the raw material of the answer.
That is why GEO is not about producing more content but about making content machine-retrievable. If an engine cannot read your page, if you have not defined what your company does in a machine-readable form, or if the answer to the question does not sit on your page as a short self-contained passage, no amount of authority puts you in that answer.
GEO is a separate channel, but not a discipline built from nothing. Crawlability, indexability, factual accuracy and structured data are the shared foundation. What is new sits on top of it: crawler access policy, entity clarity and passage writing.
How does GEO differ from AEO and SEO?
All three share the same technical foundation and differ in target surface. SEO targets ranking on the results page, AEO targets being quoted directly in the answer box, and GEO targets being cited inside the text a generative engine writes. None replaces another, and their units of measurement differ, so they cannot be added up in one report.
| Dimension | SEO | AEO | GEO |
|---|---|---|---|
| Goal | Rank on the results page | Be quoted in the answer box | Be cited in the generated answer |
| Surface | Search results page | Featured snippet, knowledge panel | Assistant answer, AI Overviews |
| Primary signal | Relevance and link authority | Question-answer structure, passage clarity | Access, entity clarity, citability |
| Unit of measurement | Position, impression, click | Answer box visibility | Sampled mentions and citations |
| Nature of the result | Relatively stable | Relatively stable | Probabilistic; the same question can return a different answer |
| Click expectation | Direct click | Fewer clicks, more visibility | Brand exposure can happen without any click |
The last row is the one most argued about in corporate reporting: being named in a generated answer does not always convert into a measurable session. A GEO report is therefore not a traffic report; it is a record of mentions and citations.
How does a generative engine choose a source?
While writing an answer, a generative engine first splits the question into sub-queries, gathers candidate pages from live search and its own index, extracts the passages that answer the question, and grounds the answer in those passages. The source list is simply the set of pages those passages came from.
The chain has four links, and each one depends on the previous one:
- Access. Can the crawler reach the page at all? A robots.txt rule, a login wall, a bot filter or a geographic block breaks the chain at this step.
- Parsing. Is the content readable in the first HTML response? If the main text is produced only on the client, the page is empty for many crawlers.
- Entity resolution. Whose page is this, what does the company do, in which country? Without structured data stating it, the engine treats the page as anonymous text rather than as a source.
- Citation. Does the answer sit there as a passage that can be lifted out without the surrounding sentences? If the answer arrives three paragraphs later, it cannot be extracted.
These four steps explain why GEO work starts on the server and schema side, before any writing. On a site where the first two links are broken, no passage reaches the answer no matter how carefully it is written.
In practice the link that breaks most often is the second. Corporate product tables, technical document lists and contact details are frequently rendered inside a component that only runs in the browser. Look at the raw HTML response and that content is not there.
Should you allow AI crawlers?
There is no single correct answer; this is a strategy decision. Allowing them lets your content be a source in generative answers, but it is also read and summarised. Blocking them keeps your content off that channel, and third-party sources fill the gap about your brand. The decision can be taken crawler by crawler.
Taking the decision properly requires one distinction first: different crawlers from the same company do different jobs. The crawler that gathers data for training is not the same as the one that fetches your page at the moment a user asks a question. The table below summarises what the decision means for each crawler.
| Crawler name | Operator | What it does | If blocked |
|---|---|---|---|
GPTBot |
OpenAI | Collects content for model training | Content stays out of training data; the route to being a live source is not closed |
OAI-SearchBot |
OpenAI | Feeds the index behind the search surface | The page does not appear in ChatGPT search results |
ChatGPT-User |
OpenAI | Fetches the page on a user's request, in the moment | The page cannot be read even when the user pastes the link |
ClaudeBot |
Anthropic | Collects content | Content cannot be used on that channel |
PerplexityBot |
Perplexity | Feeds the search index | The page does not appear in Perplexity source lists |
Google-Extended |
Gemini apps and Vertex AI grounding | Does not affect Google Search ranking or AI Overviews eligibility | |
CCBot |
Common Crawl | Builds an open web archive | Content closes for every third party that uses the archive |
Applebot-Extended |
Apple | Permission for generative model training | Applebot's search crawl continues; training use stops |
Sources: OpenAI bot documentation ↗ (opens in a new tab) · Google Search Central, crawler list ↗ (opens in a new tab) · Perplexity bot documentation ↗ (opens in a new tab) · Common Crawl CCBot ↗ (opens in a new tab) · Accessed: 2026-07-29.
Allowing is consistent when
- Your buyer researches at length before purchase and asks comparison questions
- What you publish is already public technical information and corporate description
- Outdated or wrong information about the brand is in circulation and you want to publish the correct version
- Your revenue model from content is starting commercial conversations, not page impressions
Blocking is consistent when
- The content itself is the product: subscription data, research reports, courses, archives
- Publication rights are contracted with third parties
- Customer-specific pricing or contract details are visible on the pages
- Your legal team has set a written restriction on this kind of use
There is a middle path as well: block the training crawlers while allowing the ones that fetch a page on a user's request and the ones that feed a search index. What matters is that the decision is written down with its reasoning and documented as comment lines in robots.txt.
What is llms.txt and how is it written?
llms.txt is a proposed file at the root of a site that summarises in plain text what the site is and which pages matter most. It was published at llmstxt.org in 2024. No major engine provider has confirmed in public documentation that it uses this file for source selection.
For that reason we do not present the file as a visibility promise. It costs little, carries no risk, and when written well it also gives human readers a clean index of the site. Oversold, it produces false expectations.
The correct format is a short index: a heading, a one-paragraph summary, a sectioned list of links and contact details. The wrong format is dumping the entire text of the site into one file. Such a file goes unmaintained, contradicts the real content, and the moment it contradicts it stops being a trustworthy signal.
This site's own llms.txt file and the robots.txt file documenting its AI crawler policy are open as live examples; both can be read directly in the browser.
# Company Name
> One-paragraph summary: what the company does, who it serves,
> in which region and in which languages it works.
Citation policy: content may be quoted with attribution.
## Services
- [Service name](https://example.com/service/): one-sentence scope
- [Service name](https://example.com/service-2/): one-sentence scope
## Reference content
- [Guide title](https://example.com/guide/): what it answers
## Contact
- Address, phone, email (identical to the site)24.5 MB
The size of a single llms-full.txt file containing the entire text of a Turkish digital agency site, measured during a competitor review. At that scale a file is neither parseable nor maintainable in practice. For comparison: this site's llms.txt is 6.0 KB.
Source: Baki Bilisim competitor analysis, company not named · Measured: 2026-07-29
What are entity clarity and the sameAs graph?
Entity clarity means the company is recognised by machines as one consistent object: the same name, address, phone number and description on every surface. The sameAs graph is the list of links to the other profiles that entity owns, letting engines connect separate records to the same company.
While writing an answer, a generative engine has to resolve one question: is the company on this page the same company as the record in that directory? Name similarity is weak evidence for that. Strong evidence is structured data on the page itself that identifies the company.
Three rules are enough. First: the address, phone number and company name are written identically on the site, in business profiles and in directory records. Second: only profiles you genuinely own go into the sameAs list. Third: no unverified field is ever written. A field left empty does no harm; an invented field lowers the credibility of the whole graph.
{
"@type": "Organization",
"@id": "https://example.com/#organization",
"name": "Company Name",
"url": "https://example.com/",
"telephone": "+90XXXXXXXXXX",
"address": {
"@type": "PostalAddress",
"streetAddress": "Street, No",
"addressLocality": "Izmit",
"addressRegion": "Kocaeli",
"addressCountry": "TR"
},
"sameAs": [
"https://www.linkedin.com/company/...",
"https://www.instagram.com/..."
]
}The full graph, type selection and @id architecture are a separate subject; the Schema.org and JSON-LD guide covers them with copyable examples.
How is GEO measured?
A fixed question set is run periodically against the same engines with a fixed method; whether the brand was mentioned, which page was cited and whether the information was accurate are recorded with a timestamp. This is a sample rather than a census, and the report says so plainly.
| Method | What it tells you | Its limit |
|---|---|---|
| Fixed question set sampling | Whether the brand is mentioned for defined questions, and which page is cited | It is a sample; output can change on the same day |
| Answer accuracy record | Whether what the engine says about the company is correct | Correction is indirect, through source signals |
| Server log analysis | Which AI crawler fetches which pages, and how often | Shows access, not citation |
| Assistant referral traffic | Sessions arriving from assistant surfaces | Never sees click-free mentions |
| Structured data validation | That the entity definition is read without errors by machines | Measures the precondition, not the visibility |
The one requirement of measurement is that the method stays fixed. If the question set, the engine list or the record format changes between periods, the comparison loses its meaning. That is why the first job is the baseline: if the state before the change was not recorded with a timestamp, no improvement can be evidenced afterwards.
What should never be written into a measurement is a claim resting on a single screenshot. Generative model output is probabilistic; one successful answer is not evidence, a repeated record is.
Where do the numbers in this guide come from?
Every number on this page carries its source and access date. We use only publicly verifiable documents and our own measurements. A ratio or percentage whose source cannot be shown is not published here; we do not repeat the unsourced impact claims circulating in the industry. The four figures below are this guide's base findings.
2023
The year the term Generative Engine Optimization was defined by an academic paper. The paper proposes a framework for measuring source visibility in generative engine answers.
Source: arXiv — “GEO: Generative Engine Optimization” (arXiv:2311.09735) · Accessed: 2026-07-29 · arxiv.org ↗ (opens in a new tab)
0
The number of dedicated markup formats, application forms or paid routes defined for appearing in Google AI Overviews. Eligibility follows the same rules as the page's eligibility for ordinary search results.
Source: Google Search Central — “AI features and your website” · Accessed: 2026-07-29 · developers.google.com ↗ (opens in a new tab)
2024
The year the llms.txt proposal was first published. It offers a site summary and priority links to language models as plain text; adoption has not been confirmed by engine providers.
Source: llmstxt.org — Jeremy Howard (Answer.AI) · Accessed: 2026-07-29 · llmstxt.org ↗ (opens in a new tab)
4.3 KB
The size of this site's published llms.txt file. It contains a corporate summary, links to services and guides, and contact details; it does not contain a text dump of the site.
Source: Baki Bilisim, own measurement · Measured: 2026-07-29 · view the file
How do you set up an AI policy and llms.txt?
Six steps: measure the current state, decide the access policy in writing, apply it in robots.txt, verify the content is readable without JavaScript, publish a compact llms.txt, then validate and record the result with a timestamp. Order matters: work that does not start with measurement cannot be evidenced afterwards.
-
Measure and record the current state
Record which AI crawlers reached the site over the last 90 days, which pages they fetched and what the current robots.txt says. This record is the comparison baseline for the next period.
-
Decide the access policy in writing
Write the allow or block decision for each crawler name, with its reasoning, in a one-page note. Training, answer generation and search visibility run under separate names, so the decision is not a single switch.
-
Apply the decision in robots.txt
Write a separate
User-agentblock for each crawler name; a wildcard rule alone is not treated as sufficient. Add the reasoning as comment lines at the top so whoever opens the file a year later can read the intent. -
Verify the content is readable without JavaScript
Fetch the raw HTML response and look for the main text, headings, tables and answer passages in it. Content produced only on the client is invisible to many crawlers, and the chain breaks here.
-
Write and publish a compact llms.txt
Place a plain-text file at the root that summarises what the company does and links to the priority pages. Target size is under 6 KB. The file is an index, not a dump of the site.
-
Validate, measure and record with a timestamp
Pass the structured data through a validator with zero errors, run the fixed question set against the generative engines and record the result with a timestamp. Without a record there is no improvement claim either.
What are the eight most common mistakes?
Most mistakes are about order and scale rather than content quality: expecting visibility from behind a closed door, inflating a file for no reason, giving the answer late, and reporting results without a measurement method. The table below lists eight, why they fail and what replaces them.
| Mistake | Why it fails | Instead |
|---|---|---|
| Expecting visibility while crawlers are blocked | The first link of the chain is broken; nothing after it is ever read | Turn the policy into a deliberate decision and document it in the file |
| Turning llms.txt into a dump of the site | It goes unmaintained, contradicts the real content and cannot be parsed | Keep it under 6 KB as an index and verify every link |
| Giving the answer in the third paragraph | No extractable passage forms; the engine moves to another source | Put a self-contained 40–60 word answer right under the heading |
| Publishing five pages that answer the same question | Which page to pick becomes unclear, and often none is picked | Give each question one canonical page and link the rest to it |
| Listing profiles you do not own in sameAs | Entity resolution attaches to the wrong record and trust falls | List only profiles you have administrative access to |
| Schema that contradicts the visible content | The markup asserts a fact that is not on the page | Give every marked-up field a visible counterpart on the page |
| Rendering the main content only with JavaScript | The raw HTML response looks empty and no passage can be extracted | Render text, tables and contact details on the server |
| Reporting “AI visibility” without a method | A single screenshot is not repeatable evidence | Fixed question set, fixed engine list, timestamped record |
Check these sixteen items before you publish
The list has three parts: access and readability, entity clarity, then citability and measurement. Work through them in order; until the access part is complete, the other two have nothing measurable to show. Copy the list into your own project tracker and give every item one owner.
# GEO checklist — access and readability
[ ] Allow/block decision for every AI crawler written explicitly in robots.txt
[ ] The reasoning behind the decision documented as comment lines in the file
[ ] Main content readable in the first HTML response with JavaScript disabled
[ ] nosnippet and narrow max-snippet limits reviewed deliberately
[ ] Sitemap and robots.txt agree with each other; no broken redirects
# Entity clarity
[ ] Organization and LocalBusiness nodes defined with single, consistent @id values
[ ] Address, phone number and company name identical across every surface
[ ] sameAs contains only profiles the company owns
[ ] No unverified field (postal code, opening hours, price) invented
[ ] Structured data returns 0 errors in the validator
# Citability and measurement
[ ] Every key page has a question-form heading and the answer in the first sentence
[ ] Answer passages are 40-60 words and stay accurate out of context
[ ] Every number carries a visible source and access date
[ ] The visible "last updated" date matches dateModified exactly
[ ] llms.txt is live, under 6 KB, and its links resolve
[ ] Fixed question set, method note and timestamped baseline all recordedWhat do the terms in this guide mean?
Each of the ten terms below is explained in the glossary with a short definition that stands on its own. Read them here in the context where they appear, then follow the link for the full entry. Every glossary term has its own anchor and is written to be quoted directly.
- GEO — the work of being cited as a source in a generative engine's answer.
- AEO — the work of being quoted directly in an answer box.
- AI Overviews — the generative summary surface in Google search results.
- Crawler — software that fetches pages automatically.
- llms.txt — a proposed file offering a site summary and priority links in plain text.
- Entity — a singular object recognised by machines: a company, person, product or place.
- sameAs — the list of links to the other profiles an entity owns.
- JSON-LD — the recommended format for embedding structured data in a page.
- Knowledge graph — a network of entities and the relationships between them.
- LCP — how long the main content takes to become visible; one of the preconditions of access.
Suggested reading order from here: how to write an extractable passage, then the LCP, INP and CLS thresholds. Every guide is listed together in the knowledge centre.
Frequently asked questions about GEO
What is generative engine optimization?
Generative engine optimization (GEO) is the work of turning a website into a source that generative AI engines use and cite while writing an answer. The target is not a rank position but the appearance of the brand name and a page link inside the generated text.
The term was defined by an academic paper published in 2023, which proposed a way of measuring source visibility in generative engine answers.
How does GEO differ from AEO and SEO?
All three share the same technical foundation and differ in target surface. SEO targets ranking on the search results page, AEO targets being quoted directly in the answer box, and GEO targets being cited as a source inside the text a generative engine writes.
GEO does not replace SEO: without crawlability, indexability and accurate content, generative engines cannot use the page either. The most practical difference is the unit of measurement; in GEO what is measured is not position but a record of mentions and citations.
Should I allow AI crawlers?
There is no single correct answer. If you allow them, your content can be used as a source in generative answers; in exchange it is read, summarised and may be consumed without a click. If you block them, your content cannot be used on that channel and third-party sources fill the information gap about your brand.
The decision is not a single switch: training, answer generation and search visibility are handled under different crawler names. For example, according to Google's documentation, Google-Extended affects only Gemini apps and Vertex AI grounding; it does not affect Google Search ranking or AI Overviews eligibility.
What is llms.txt and how is it written?
llms.txt is a proposed file, published at llmstxt.org in 2024, that sits at the root of a site and summarises in plain text what the site is and which pages matter most. None of the major generative engine providers has confirmed in public documentation that it uses this file for source selection.
The correct format is a short index: a heading, a one-paragraph summary, a sectioned list of links and contact details. Versions that dump the whole site into one file go unmaintained and end up contradicting the real content. This site's own file is 6.0 KB and can be read directly.
How is GEO measured?
A fixed question set is run periodically against the same engines with a fixed method. Each run records, with a timestamp, whether the brand was mentioned, which page was cited and whether the information given was accurate.
This is a sample, not a census. Two signals support the measurement: the frequency of AI crawler access in server logs, and referral traffic from assistant surfaces. A single screenshot does not count as evidence.
What if an AI engine describes my company incorrectly?
The engine's output cannot be edited directly; the only thing you can correct is the source signals behind it. Start by tracing where the wrong information comes from: in most cases it is an outdated directory record, an inconsistent address or phone number, a removed page, or a fact that is nowhere stated plainly on the site.
Publish the correct information on the site — visible, dated and marked up with structured data — and correct the inconsistent external records. Then run the same question set again and record the change with a timestamp.
Questions about process, pricing and contracts are collected on the general questions page.
See the measured state of your website within five working days.
The audit is free and creates no obligation to work with us. The report itemises findings on AEO answerability, AI crawler access, lab-measured Core Web Vitals, structured data validity and accessibility (automated scan).
We work with corporate-scale, multi-location or multilingual organisations. One-off small jobs fall outside our scope; in that case we point you to smaller studios.