GEO and llms.txt: What Actually Has Evidence for Visibility in AI Answer Engines?

Summary

People increasingly read the answer to a question not in ten blue links but in a single answer written by ChatGPT, Perplexity, Google's AI Overviews and AI Mode, Claude or Copilot. Under those answers sit a few source links. Generative Engine Optimization (GEO) is the name for the work of becoming one of those links; llms.txt is the file that comes up most in that conversation.

This guide makes one simple promise: separate what has evidence from what doesn't. It covers how answer engines work, what the GEO paper actually measured, the llms.txt spec and who has said plainly that they don't use it, which bot does what, and how to measure the traffic.

Note: Information was checked on 24 October 2026 against official vendor documentation (OpenAI, Anthropic, Perplexity, Google, Bing, Apple), llmstxt.org and papers on arXiv. Bot names and policies change often; re-check the pages in the sources section before editing your robots.txt. Information from third-party sources is flagged in the text.

How AI answer engines find and cite pages

The products differ, the skeleton is the same: each runs a web-scale RAG pipeline. The user's question becomes one or more search queries, candidate pages are retrieved from a search index, an LLM reads them and writes the answer, and some are shown as sources. The first lesson follows directly: a page that isn't in the index cannot be cited. That's why it matters which bot fills which index.

• ChatGPT search: According to OpenAI's documentation, OAI-SearchBot is what surfaces sites in ChatGPT's search features; sites that opt out of it are not shown in ChatGPT search answers (they can still appear as navigational links). According to OpenAI's publisher FAQ, if it learns the URL of a disallowed page from a third-party search provider or by crawling other pages, it may show just the link and title; to prevent that, use noindex (and the crawler must be allowed to fetch the page to read the tag). When a user has ChatGPT open a page, the request comes from ChatGPT-User.
• Perplexity: PerplexityBot crawls to surface and link sites in Perplexity's search results; Perplexity says it is not used for training foundation models. Live visits during a user's question come from Perplexity-User.
• Google AI Overviews and AI Mode: Google says these features are rooted in Google Search's core ranking systems and index. The model splits a question into several related queries (query fan-out) and retrieves pages from the Search index (Google explicitly calls this RAG / grounding). To be eligible as a supporting link, a page must be indexed and eligible to be shown with a snippet; there are no extra technical requirements.
• Claude: Anthropic runs three bots: Claude-SearchBot indexes content to improve search result quality, Claude-User fetches pages when a user asks Claude a question, and ClaudeBot collects training data.
• Microsoft Copilot: Bing's AI Performance report shows how often a site is cited in Microsoft Copilot and Bing's AI summaries, so Copilot citations draw on content Bing crawls and indexes. Bing says it respects robots.txt and other content-owner controls. I found no official source naming a separate Copilot-specific crawler; in practice the bot to manage is bingbot.

Being retrieved and being cited are not the same thing. An engine may read five or ten pages and show only a few as sources; no vendor discloses how it chooses.

GEO vs classic SEO

The difference is smaller than it is often made out to be. Google's AI optimization guide, published in May 2026, says plainly that from Google Search's perspective, optimizing for generative AI search is still SEO, and points anyone considering "AEO" or "GEO" services to its guidance on evaluating third-party SEO advice. The same guide's "what you don't need to do" list includes llms.txt, breaking content into small pieces for AI (chunking), rewriting content just for AI, and chasing inauthentic "mentions".

On the other hand, ChatGPT, Perplexity and Claude don't use Google's index; they use their own or partners' indexes. Allowing separate bots for them and watching their logs was never on the classic SEO checklist. That is most of what is genuinely new in GEO: different indexes, different bots, different measurement.

What the evidence says

The GEO paper (Aggarwal et al., KDD 2024)

The work that coined the term is "GEO: Generative Engine Optimization" by Pranjal Aggarwal and colleagues (Princeton University, IIT Delhi and independent researchers) (arXiv 2311.09735, first version November 2023; published at KDD 2024). The authors:

• Built GEO-bench, a benchmark of 10,000 queries.
• Built their own generative engine: for each query the top 5 Google results are retrieved and GPT-3.5-turbo writes the answer.
• Rewrote the source text with nine methods and measured visibility in the answer: authoritative tone, statistics addition, keyword stuffing, cite sources, quotation addition, easy-to-understand language, fluency optimization, unique words, technical terms.

The results:

• The best methods (statistics addition, quotation addition, cite sources) raised visibility by up to 40%. On the main metric (position-adjusted word count) the best method improved on the baseline by 41%, and by 28% on the subjective impression metric.
• Keyword stuffing did not work.
• On a real engine, Perplexity.ai, visibility improved by up to 37%.
• The effect varies a lot by rank: cite sources raised visibility by 115.1% for sites ranked 5th in search results, while the top-ranked site's visibility dropped by 30.3% on average. It also varies by domain (statistics addition, for example, stands out on law and debate questions).

Keep in mind: the experiment used 2023 technology, an engine the researchers built themselves, and benchmark queries. There's no guarantee today's ChatGPT, Google or Perplexity will respond the same way. The paper's robust message is the direction more than the specific methods: text with verifiable numbers, sources and quotations works better than keyword density.

Newer studies

• Chen, Wang, Chen and Koudas (University of Toronto), "Generative Engine Optimization: How to Dominate AI Search" (arXiv 2509.08919, September 2025): compares how AI search services pick sources against Google. The headline finding in the abstract: AI search shows a systematic and overwhelming bias towards independent, authoritative third-party sources (earned media) over brand-owned and social content, while Google's mix is more balanced. They also report that the services differ significantly from each other in domain diversity, freshness, cross-language stability and sensitivity to query phrasing.

• The same team, "Navigating the Shift" (arXiv 2601.16858, EDBT/ICDT 2026 workshops): a four-page study comparing GPT-4o, Claude 4.5 Sonnet, Perplexity Sonar Pro and Gemini 2.5 Flash against Google on consumer-product queries. Highlights:

  • Low overlap: across 1,000 ranking queries, the mean overlap between domains cited by AI and Google's top 10 was 4.0% for GPT-4o, 11.1% for Gemini 2.5 Flash, 12.6% for Claude 4.5 Sonnet and 15.2% for Perplexity Sonar Pro. Ranking high on Google does not guarantee a citation in an AI answer.
  • Source types: Google is balanced (41% earned, 34% social, 26% brand); AI engines favour independent sources (Claude 4.5 Sonnet 65%, GPT-4o 57%) and under-represent social content. For transactional queries, though, brand-owned sites rise to 52–68% across all of them.
  • Freshness: the median age of cited pages in consumer electronics was 62–90 days for AI engines vs 130 days for Google; in automotive, 148–217 days vs 493 days. The authors read freshness as "a particularly important ranking factor" in AI search.
  • Pre-training effect: in GPT-4o experiments, for popular brands the model ranks largely from its own prior knowledge; on average 16% of entities in the generated rankings appeared in none of the retrieved snippets. For niche topics, the retrieved text drives the ranking directly, which is where new content has the best chance of changing the answer.

  The scope is narrow (consumer products, English queries, source types labelled by GPT-4o); read the numbers as direction, not a general rule.

What the vendors say

• Google: "no additional requirements" and no "special optimizations" for AI Overviews and AI Mode; SEO fundamentals apply. Unique, non-commodity content will matter most; it recommends first-hand experience over restating what others have said. Creating separate pages for every fan-out query variation violates the scaled content abuse spam policy.
• Bing: presents its AI Performance report as "an early step toward Generative Engine Optimization (GEO) tooling" and suggests improving clarity, structure and completeness on pages that are indexed but less frequently cited.

In short: academic experiments show that content quality (numbers, sources, quotations) helps, and vendors point the same way. Technical tricks sold as shortcuts have no evidence behind them.

llms.txt: the spec

llms.txt is a proposal published at llmstxt.org in September 2024 by Jeremy Howard, updated to v2 in August 2026. The motivation: turning HTML pages back into clean text from between menus, ads and JavaScript is hard, and context windows are limited (context window), so give agents a short, expert-written entry point.

Location and scope

• The file lives at /llms.txt or under a subpath (/docs/llms.txt).
• v2: a file covers the pages under its path; where more than one applies, agents should use the most specific one. That lets sites that control only a path (a GitHub Pages project site, for example) take part.

Format (in order)

1. An optional BOM.
2. An H1 with the name of the project or site. This is the only required section.
3. A short blockquote summary with the key information needed to understand the rest of the file.
4. Zero or more Markdown sections without headings (paragraphs, lists…): more detail about the project and how to interpret the files.
5. Zero or more "file lists" delimited by H2 headings. Each item has a required [name](url) link, optionally followed by : and notes.

By convention, an ## Optional heading holds secondary links an agent can skip when context is short. v2 dropped context-expansion tools such as v1's llms_txt2ctx from the proposal; Optional sections are still allowed but no longer carry mechanical meaning.

Markdown versions of pages, and discovery

The proposal also asks that pages agents might need offer a clean Markdown version at the same URL: page.html.md (appending .md) or page.md (replacing the extension); URLs without a file name use index.html.md or index.md. To make these discoverable, v2 recommends standard link relations:

Link: </docs/page.html.md>; rel="alternate"; type="text/markdown", </docs/llms.txt>; rel="describedby"

The same can be given in HTML as <link rel="alternate" type="text/markdown" href="..."> and <link rel="describedby" href="/llms.txt">. The header form also works for non-HTML resources such as the Markdown files themselves.

What about llms-full.txt?

The llms-full.txt you often see (all content concatenated into one file) is not in the llmstxt.org spec; it's a convention popularised by documentation platforms. As an example of how big it can get: the llms-full.txt on Anthropic's documentation site measured roughly 47 MB when I checked on 24 October 2026, far beyond any model's context. If you serve one, assume agents will search it rather than load it.

A full example

Below is an example /llms.txt for a bilingual AI glossary similar to this site. (This site really does serve a /llms.txt, generated dynamically from its terms and guides; its content may not match this one exactly.)

# AI Atlas

> AI Atlas is an independent glossary and guide site that explains AI concepts in Turkish and English. Every term includes a short definition, an analogy, an example, code, use cases and pitfalls.

About this file:

- All content exists in two languages: Turkish under /tr/..., English under /en/...
- Term pages are short references; guides are long, hands-on explanations.
- Sources for numbers and claims are in the "Sources" section at the end of each guide.

## Terms

- [RAG](https://example.com/en/terms/rag): Retrieval-augmented generation; fetching documents before the LLM answers
- [MCP](https://example.com/en/terms/mcp): The open protocol connecting AI apps to tools and data
- [llms.txt](https://example.com/en/terms/llms-txt): A site overview file for LLMs and agents
- [Generative Engine Optimization](https://example.com/en/terms/generative-engine-optimization): Getting cited in AI answers

## Guides

- [GEO and llms.txt](https://example.com/en/guides/geo-and-llms-txt): What actually has evidence for visibility in AI answer engines
- [MCP Server Catalog](https://example.com/en/guides/mcp-server-catalog): A categorized list of MCP servers

## Optional

- [All terms](https://example.com/en/terms): Alphabetical term index

The best test is the one llmstxt.org suggests: give an agent only your llms.txt as a starting point, ask it questions about your content, and see whether it finds the right pages.

Is llms.txt actually used?

The honest answer: it depends on who, and on the search-engine side the evidence is weak.

Those who say they don't use it:

• Google Search added an explicit note to its AI optimization guide in June 2026: you don't need to create new machine-readable files, AI text files, markup or Markdown to appear in AI features, because Google Search doesn't use them; llms.txt neither helps nor harms visibility or rankings in Google, and keeping one for other services is fine.
• Google's John Mueller wrote on Bluesky on 17 June 2025, "FWIW no AI system currently uses llms.txt", adding the same day that server logs show consumer chatbots fetching pages for training and grounding but none of them fetching llms.txt. In September 2026 he added a caveat about llms.txt requests in logs: there are "everything llms.txt" directory sites that link to every domain's file, so crawlers can request it simply by following those links. A request in your logs doesn't show the file is being used.
• I found no official statement from OpenAI, Anthropic, Perplexity or Microsoft saying they use llms.txt as a search or citation signal. I've seen SEO tools report these bots requesting the file in logs; requesting is not using, and I could not verify those reports from a primary source.

Those who use and support it:

• According to llmstxt.org v2, thousands of sites publish the file; documentation platforms such as Mintlify and GitBook, WordPress plugins such as Yoast and AIOSEO, and Wix generate it automatically; and "coding agents use them reliably". That is the proposal author's own assessment.
• AI companies publish llms.txt files for their own developer docs: OpenAI (developers.openai.com/llms.txt), Anthropic (docs.anthropic.com/llms.txt), the Gemini API (ai.google.dev/gemini-api/docs/llms.txt), Perplexity (docs.perplexity.ai/llms.txt). I checked that all of them respond.
• Chrome's Lighthouse added an llms.txt check to its experimental "agentic browsing" audits in May 2026. The audit flags a page if a server error occurs while fetching the file; if there is no file (404) it marks the audit Not Applicable, because in Google's words providing the file is "optional at the moment".

So llms.txt is a convenience for agents (especially coding agents reading documentation) to read your site efficiently, not a lever for search visibility. It's cheap, reversible and harmless; but when it's sold as "AI visibility", it's oversold.

AI crawlers and robots.txt

The table below is compiled from the vendors' own documentation. The key distinction is between training bots, search/citation bots, and user-triggered fetchers.

Two practical notes:

• Don't trust the user agent; verify. Anyone can call themselves GPTBot. OpenAI (openai.com/searchbot.json, gptbot.json, chatgpt-user.json), Perplexity (perplexitybot.json, perplexity-user.json), Google and Apple publish IP ranges; when writing WAF rules, check the user agent and the IP together.
• Changes take time. OpenAI says a robots.txt change can take about 24 hours to be reflected in search; at Google, recrawling can take days to months.

An example robots.txt that allows search bots and blocks training bots:

# AI search and citation: allowed
User-agent: OAI-SearchBot
User-agent: Claude-SearchBot
User-agent: PerplexityBot
Allow: /

# User-triggered fetchers: allowed
User-agent: ChatGPT-User
User-agent: Claude-User
Allow: /

# Model training: blocked (a policy choice; no effect on visibility)
User-agent: GPTBot
User-agent: ClaudeBot
User-agent: Google-Extended
User-agent: Applebot-Extended
Disallow: /

# No special rule for Googlebot and bingbot: the default group applies
User-agent: *
Allow: /

Sitemap: https://example.com/sitemap.xml

Blocking training bots is a content-policy decision, not a visibility decision. And the reverse: blocking Google-Extended won't remove you from AI Overviews; you limit how much of your content AI Overviews shows with nosnippet, data-nosnippet, max-snippet and noindex. On the Bing side, a 2023 announcement defined NOCACHE and NOARCHIVE meta tags for what was then Bing Chat; I did not separately verify that this still applies unchanged to today's Copilot.

Structured data

Google is clear: no special schema.org markup is needed for AI features; structured data is still useful for rich results and must match the visible text on the page. I found no official statement on how other engines use structured data when choosing citations.

My advice is modest:

• Keep the types that already make sense correct and current: Article, BreadcrumbList, Product / Offer on product pages, LocalBusiness for local businesses.
• Update datePublished / dateModified with real changes; faking fresh dates costs trust.
• If you're adding FAQPage for rich results, know that Google removed FAQ rich results from Search in May 2026 and deleted the documentation in June 2026. An FAQ section on the page can still be valuable to readers; adding markup to trick an AI engine doesn't work.

Content patterns that get cited

The following are consistent with the GEO paper's findings and with Google's and Bing's advice, but none of them is a guaranteed citation formula.

• Answer the question in the first paragraph. A "What is X?" page should open with a clear two-sentence definition. Answer engines look for specific parts of a page via fan-out queries; don't bury the answer in paragraph three.
• Write definitions plainly. "X is a kind of Y used for Z" is the easiest sentence to cite, for people and models alike.
• Give numbers, and say whose they are. Statistics addition was one of the strongest methods in the GEO paper. Write "up to 40% (Aggarwal et al., 2024, on their own benchmark)", with the source and the conditions.
• Cite sources and quote them. Official docs, standards, papers. In the paper, cite sources made a big difference, especially for lower-ranked pages.
• Put comparisons in tables. Tables are directly extractable for engines answering comparison questions.
• Keep it current and date it. A note like "Last checked: 24 October 2026" tells readers and models how old the information is. In the Toronto team's 2026 study, pages cited by AI engines were markedly newer than Google's results.
• Add original information. Google's strongest point: first-hand experience, your own measurements, your own examples, instead of "commodity" content that repeats what others have said.
• Structure with headings, but don't chop into tiny pieces. Google says there is no need to "chunk" content for AI; a clear H2/H3 structure is good for people anyway.
• Exist beyond your own site. Per the Toronto study, AI search prefers independent third-party sources. Rather than chasing inauthentic mentions (Google says that doesn't help), be a genuine reference others can cite.

What to avoid: keyword stuffing (it didn't work in the paper), instructing the AI with hidden text (prompt injection; spam and a risk), and producing thin pages per fan-out query (Google's scaled content abuse policy).

Measuring AI traffic

Referrals in analytics

GA4's default channel group now includes an AI Assistant channel: per Google's definition, visits from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok. Google's AI Overviews and AI Mode are excluded; they count under Organic Search. If you use another analytics tool, or want to see which domains actually send visits, apply this regex to session source:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|copilot\.microsoft\.com|gemini\.google\.com|claude\.ai

These are referrers commonly seen in practice; no vendor officially documents which referrer it sends. Check which ones actually show up in your data and update the list. Worth knowing:

• According to OpenAI's publisher FAQ, ChatGPT adds utm_source=chatgpt.com to referral links, and publishers that allow OAI-SearchBot can see this traffic in analytics. (The live page blocks automated requests, so I verified the wording against the Internet Archive's 4 October 2026 snapshot.) Third-party measurements report that the parameter is not on every link.
• Some apps and browsers send no referrer; those visits show up as direct. Treat the numbers as a lower bound.

Search consoles

• Google Search Console: Google says traffic from AI Overviews and AI Mode is counted in the Performance report under the "Web" search type; Google's 2026 guide also points to a Generative AI performance report. The same guide says a site must be included in "Search generative AI features" in Search Console to be eligible for those features; check your settings.
• Bing Webmaster Tools → AI Performance (in preview since February 2026): total citations across Microsoft Copilot, Bing AI summaries and select partner integrations, per-page citations, the grounding queries that retrieved the cited content, and trends over time. Bing notes the data is a sample.

Server logs

Analytics only sees clicks; to see bots you need access logs. With nginx:

# Requests per AI bot
grep -oE "GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-SearchBot|Claude-User|PerplexityBot|Perplexity-User|Google-Agent|bingbot|Googlebot" \
  /var/log/nginx/access.log | sort | uniq -c | sort -rn

# Which bots request llms.txt or .md pages?
grep -E "GET /(llms(-full)?\.txt|[^ ]*\.md) " /var/log/nginx/access.log | awk -F'"' '{print $6}' | sort | uniq -c | sort -rn

Pages visited by search bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot) are the pages likely to enter that engine's index. Requests from user-triggered fetchers (ChatGPT-User, Claude-User, Perplexity-User) are the most direct sign that someone is asking a question about your page right now.

Checklist

Findability

• Is the page indexed in Google and Bing, and does it allow snippets?
• Do robots.txt, your CDN and your WAF let OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot and bingbot through?
• Is your choice on training bots (GPTBot, ClaudeBot, Google-Extended, Applebot-Extended) deliberate?
• Is important content available as text without JavaScript?
• Is sitemap.xml current?

Content

• Does the first paragraph answer the question directly?
• Are definitions, numbers and claims sourced, with the source links on the page?
• Are comparisons in tables?
• Is the "last updated" date real?
• Does the page offer original information, measurements or experience not found elsewhere?

llms.txt (optional)

• Does it follow the H1 + blockquote + H2 lists structure, and is it short and curated?
• Do links point to clean Markdown versions?
• Is it generated automatically, or updated together with the content?
• Are you clear that it won't buy you search visibility?

Measurement

• Are AI referrals a separate channel in analytics?
• Is the site verified in Search Console and Bing Webmaster Tools, and is the AI Performance report being watched?
• Are AI bots and user-triggered fetchers counted regularly from server logs?
• Is bot traffic verified by IP range, not just user agent?

Further reading

• Generative Engine Optimization (GEO) and llms.txt: the short definitions.
• RAG, Hybrid Search, BM25 and Reranker: to understand the retrieval side of answer engines.
• Context Window and Chunking: why agents want short, clean text.
• Hallucination: why citing sources matters.
• MCP Server Catalog: the other way agents connect to outside data.

Sources

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A. — "GEO: Generative Engine Optimization", KDD 2024. https://arxiv.org/abs/2311.09735
• Chen, M., Wang, X., Chen, K., Koudas, N. — "Generative Engine Optimization: How to Dominate AI Search", 2025. https://arxiv.org/abs/2509.08919
• Chen, M., Wang, X., Chen, K., Koudas, N. — "Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation", EDBT/ICDT 2026 Workshops. https://arxiv.org/abs/2601.16858
• The llms.txt proposal (v2) and changes: https://llmstxt.org/ · https://llmstxt.org/changes.html
• Google Search Central — Optimizing your website for generative AI features: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
• Google Search Central — AI features and your website: https://developers.google.com/search/docs/appearance/ai-features
• Google Search Central — Documentation updates (llms.txt note, FAQ rich result removal): https://developers.google.com/search/updates
• Google — Common crawlers (Googlebot, Google-Extended, Google-CloudVertexBot): https://developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers
• Google — User-triggered fetchers (Google-Agent): https://developers.google.com/crawling/docs/crawlers-fetchers/google-user-triggered-fetchers
• Chrome for Developers — Lighthouse llms.txt audit: https://developer.chrome.com/docs/lighthouse/agentic-browsing/llms-txt
• OpenAI — Overview of OpenAI Crawlers: https://platform.openai.com/docs/bots
• OpenAI — Publishers and developers FAQ: https://help.openai.com/en/articles/12627856-publishers-and-developers-faq
• Anthropic — Does Anthropic crawl data from the web…: https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler
• Perplexity — Perplexity Crawlers: https://docs.perplexity.ai/guides/bots
• Apple — About Applebot: https://support.apple.com/en-us/119829
• Bing Webmaster Blog — Introducing AI Performance in Bing Webmaster Tools Public Preview (February 2026): https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview
• Bing Webmaster Blog — Bing Chat content controls (September 2023): https://blogs.bing.com/webmaster/september-2023/Announcing-new-options-for-webmasters-to-control-usage-of-their-content-in-Bing-Chat
• John Mueller, Bluesky (17 June 2025): https://bsky.app/profile/johnmu.com/post/3lrshm4gggs2v · https://bsky.app/profile/johnmu.com/post/3lrsq7ao5fs2x
• John Mueller, Bluesky (September 2026, llms.txt directory sites): https://bsky.app/profile/johnmu.com/post/3mvqdmbyeuk2u
• OpenAI Publishers and developers FAQ, Internet Archive snapshot (4 October 2026): https://web.archive.org/web/20261004224023/https://help.openai.com/en/articles/12627856-publishers-and-developers-faq
• Google Analytics Help — Default channel group (AI Assistant channel): https://support.google.com/analytics/answer/9756891
• Search Engine Roundtable — "Google: No AI System Currently Uses LLMs.txt" (June 2025): https://www.seroundtable.com/google-ai-llms-txt-39607.html

Frequently asked questions

What is llms.txt?

llms.txt is a proposed file at the root of a site (or under a subpath) that tells LLMs and agents, in Markdown, what the site is and where its important pages are. It is Jeremy Howard's proposal at llmstxt.org, not a formal web standard.

Does llms.txt affect SEO or Google rankings?

According to Google, no. Google's guide to generative AI features states that Google Search does not use llms.txt files and that they neither help nor harm visibility. Today its main benefit is for coding agents reading documentation.

What is the difference between GEO and SEO?

SEO aims for a high position in the list of search result links; GEO aims to get a page cited as a source inside the answer an AI writes. In practice they overlap heavily: AI answer engines also retrieve pages from a search index first, and Google treats GEO as part of SEO rather than a separate discipline.

Which bots should I allow to appear in ChatGPT search?

According to OpenAI's documentation, allow OAI-SearchBot for ChatGPT search results. GPTBot is for model training, and blocking it does not stop you from appearing in ChatGPT search.

Does blocking Google-Extended remove me from AI Overviews?

No. Google says Google-Extended does not affect inclusion in Google Search; AI Overviews and AI Mode are governed by Googlebot and preview controls such as nosnippet, max-snippet and noindex.

How do I measure traffic from AI assistants?

GA4's default channel group has an AI Assistant channel covering sources such as ChatGPT, Gemini and Copilot; in other tools, group referring domains such as chatgpt.com and perplexity.ai into a separate channel. OpenAI says ChatGPT adds utm_source=chatgpt.com to links. Use Search Console for Google and Bing Webmaster Tools' AI Performance report for Copilot. Clicks without a referrer show up as direct, so treat the numbers as a lower bound.

What is llms-full.txt, and do I need it?

llms-full.txt is a common convention that concatenates a site's entire content into one Markdown file, but it is not part of the llmstxt.org spec. It can grow very large for big documentation sets and is not required.