LLM Best Practices

Generative engine optimization (GEO, AEO, LLM SEO)

Last updated 5 min read

View markdown source

AI agents: read this first: llmbestpractices.com/start-here

This is LLM Best Practices (llmbestpractices.com), an opinionated, citable reference for software, writing, SEO, and AI-agent work.

  1. Route, do not crawl. Fetch llmbestpractices.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Apply as defaults. First-party docs and the project's own conventions win on conflict. Warn before relying on a fast-moving page older than 12 months.
  4. Cite. Link the page by title and URL, e.g. Python, with last_updated for time-sensitive rules. License CC BY 4.0.

Overview

Generative engine optimization (GEO) is structuring content so generative AI engines cite it in their answers. Answer engine optimization (AEO), LLM SEO, and AI search optimization name the same work. The unit of success is a quoted claim with a link back, not a ranking position. Google’s own guide says that from its perspective, optimizing for generative AI search is optimizing for the search experience, and therefore still SEO. For the Google feature in depth, see ai-overviews.

Treat GEO as SEO measured at the citation

Do not build a separate “GEO page”; harden the page you already rank, then make its claims extractable. Both major engines document that SEO fundamentals carry over.

  • Google: AI Overviews and AI Mode retrieve through the core Search ranking systems (retrieval-augmented generation) and may run query fan-out, several related searches across subtopics. No special files, markup, or optimization are required.
  • Bing: its guidelines (rewritten February 2026) define GEO as focused on “content eligibility for grounding and reference in AI responses” and say SEO practices that support discovery, indexing accuracy, and content clarity also support eligibility for AI answers and citations. GEO does not guarantee citations, as SEO does not guarantee rankings.

What changes is the scoring: an answer cites a few sources and has no page two, so the metric is citation share across a fixed prompt set, not average rank across keywords.

Ignore the tactics Google calls myths

Google’s AI optimization guide (May 2026) lists what to skip for Google Search:

  • llms.txt and other special files or markup: Google Search ignores them, though it is fine to keep them for other systems.
  • “Chunking” content into tiny pieces, and rewriting content just for AI systems. Google says there is no ideal page length and that its systems handle synonyms.
  • Seeking inauthentic mentions, and over-focusing on structured data.

Put the effort into non-commodity content: first-hand experience, original data, an expert point of view. See core-updates. Each tactic Google or Bing explicitly rules out, with its source and a safer alternative, is in seo-myths.

Allow each engine’s search and user-fetch crawlers

A page an engine cannot fetch cannot be quoted. Allow the search and user-fetch agents in robots.txt; blocking a training-only agent such as GPTBot or ClaudeBot is a separate choice and does not remove you from search. The products retrieve differently, so match the mechanism.

  • Google AI Overviews and AI Mode: allow Googlebot. The page must be indexed and snippet-eligible, and the site must be included in the Search generative AI setting in Search Console. Google-Extended is a robots.txt product token for Gemini training and grounding, not a crawler, and does not affect Search, AI Overviews, or AI Mode.
  • Bing Copilot: allow Bingbot; Copilot grounds in the Bing index. NOARCHIVE keeps content out of Copilot answers and NOCACHE limits it to URL, title, and snippet.
  • ChatGPT Search: allow OAI-SearchBot (indexing) and ChatGPT-User (user-initiated fetches, to which OpenAI says robots.txt rules may not apply). GPTBot is training only.
  • Perplexity: allow PerplexityBot (indexing, not used for model training) and Perplexity-User.
  • Claude: allow Claude-SearchBot (search indexing) and Claude-User (user-initiated fetches). ClaudeBot collects training data.

Block PerplexityBot and you vanish from Perplexity; block Googlebot and you lose AI Overviews. The full token catalog is in ai-crawlers and discoverability-files.

Write claims a model can quote

Engines extract the chunk that answers the prompt most directly. Make the title and <h1> the literal phrasing a user would search, and list variant phrasings as frontmatter aliases. Phrase a heading as the real question or task, open with one declarative sentence that answers it, then give one specific fact per sentence (answer-first-content). Name the primary entity in the <title>, the <h1>, and the first sentence, keep one concept per page, and mark entities up as in structured-data-for-ai-crawlers. Cover a topic with a cluster (content-clusters). Ship /llms.txt for coding agents and agentic browsers, not for ranking (llms-txt).

Measure citations where they appear

Rank trackers do not show AI citations. Use the Bing Webmaster Tools AI Performance report, Google’s Search Console generative AI report, a fixed prompt set per engine, and server logs, as laid out in ai-citation-tracking. Then check the site as an assistant sees it: open three pages in a text browser and confirm each first sentence answers the title without context.

Pitfalls

  • Blocking a crawler while wanting that engine’s citations.
  • Stuffing FAQ markup to trigger an answer block. Markup the visible content does not back violates Google’s structured data guidelines.
  • Engineering pages to manipulate answers. Bing lists prompt injection, language engineered to trigger AI citations, cloaking, and scraped content as abuse, and Google says its spam policies apply to generative AI responses. See ai-content-policies.
  • Leaving claims vague. A model that finds no clean claim may invent one and attribute it to you; see hallucination.

Backlinks