Workflow

Step 1: Implement /llms.txt (recommended)

• Add a /llms.txt at the site root following the community proposal (https://llmstxt.org/). The file should be plain Markdown, human- and machine-readable, and point to canonical docs, API references, changelogs, and key long-form pages. The spec also recommends publishing .md versions of pages (URL.md) so tools can fetch clean text versions.
• Practical: include a short one-paragraph site description, a docs section with canonical links, and a prioritized list of pages (docs, API, major guides). Treat /llms.txt as additive — it helps AI retrieval but does not replace good structured content and metadata.

Example head of /llms.txt

AcmeCorp /llms.txt

AcmeCorp builds secure API tooling for payments and identity verification.

Priority docs

• https://acme.com/docs/getting-started
• https://acme.com/docs/api-reference

(See: llms.txt proposal — https://llmstxt.org/)

Step 2: Consider llms-full / .md pages and llms-ctx flows

• Where feasible, publish Markdown versions of documentation and long-form pages (page.html.md). Projects implementing the llms.txt recommendations often generate llms-ctx or llms-ctx-full context files for ingestion; this improves the quality of retrieved context for LLMs.
• If you have a docs build pipeline, add a step to emit .md artifacts and/or a llms-ctx bundle for core docs.

Reference: llms.txt proposal and implementation notes (https://llmstxt.org/).

Step 3: Structure content for AI extraction

• Write "structured answer" blocks — self-contained paragraphs that answer one question. Lead with a short answer sentence, then add numeric evidence, temporal anchors, and primary-source links.

Before (low citability):

We've improved our product's speed. Users like it.

After (high citability):

Acme v3.2 reduced API p99 latency from 340ms to 180ms (47% improvement) by migrating from PostgreSQL full-text search to a Typesense index; measured across 2.1M production queries (Jan 2025). Source: internal benchmark report (link).

Step 4: Score factual density (operational rubric)

Score each paragraph with signals and weights:

• Named entity: 2 points
• Statistic with anchor (date/metric): 3 points
• Explicit comparison: 2 points
• Source citation/link: 2 points
• Technical spec/version number: 1 point
• Temporal anchor (month/year): 1 point

Aim for paragraph scores >= 8 on priority pages. Paragraphs < 4 are unlikely to be cited.

Step 5: Platform-specific optimization (updated)

• ChatGPT / OpenAI: models and agent runtimes change rapidly. Recent OpenAI agent and Responses API work emphasizes agent environments and tool use (see OpenAI Signals) — maintain concise answers + strong primary links and test across model updates.
  • Evidence: OpenAI announced an agent runtime built on the Responses API and hosted containers (OpenAI News, Mar 2026).
• Perplexity: favors source diversity and inline citations — include multiple primary sources and clear inline links.
• Gemini / Google: structured data (JSON-LD), strong entity alignment, and canonical docs help. Google’s AI-related enhancements continue to evolve; prioritize schema and canonical structured signals.
• Bing / Microsoft-backed models: OG tags, FAQ schema, and accessible docs help retrieval. Platform behaviors differ; measure across engines.

Operational note: platform behavior can shift after model updates — e.g., independent data suggests ChatGPT Search cited fewer distinct sites after a change to GPT-5.3 Instant (see Search Engine Journal report). Track citation performance continuously across model versions and engines.

Sources: OpenAI News (Responses API, agent runtime) and Search Engine Journal (analysis of citation counts).

Step 6: Measure AI citability

• Periodically query target queries across AI engines (ChatGPT Browse, Perplexity, Gemini, Bing) and record whether your page is cited, paraphrased, or absent.
• Scoring: cited=3, paraphrased=1, absent=0. Citation rate = total score / (queries × 3).
• Track changes over time and correlate with model updates (major model releases or inference-tier changes reported by providers can alter citing behavior).

Examples (concise)

• Thin -> High-citability conversion: replace vague claims with date-anchored metrics, named technologies, and a primary-source link.
• Structured answer block for pricing, SLA, or API limits: one-line summary + bullets with exact numbers and dates.

Decision tree (operational)

• New content: design with structured answers and publish .md pages for docs → include links in /llms.txt
• Existing low-traffic content: rewrite top paragraphs to raise factual density to 8+
• Pages already ranking in Search: add GEO signals (llms.txt, entity density) but do not restructure titles/headlines aggressively
• Technical docs: prioritize emitted .md pages, version numbers, and code examples

Edge cases and gotchas (revised)

1. AI crawler blocking — verify robots.txt and any crawler allowlists. Many crawlers look for /llms.txt; do not mistakenly block popular AI crawler user agents.
2. Stale statistics — always attach dates to statistics and refresh benchmark pages. Model caches and retrieval indices may hold older snapshots.
3. Over-optimization — maintain readability; avoid stuffing entities at the cost of clarity.
4. llms.txt adoption is growing and a community proposal exists (https://llmstxt.org/). Adoption is not yet universal; treat it as high-value, additive metadata.
5. Attribution stripping — some systems paraphrase without link attribution; you cannot force citation, only increase probability by providing unique, verifiable data.
6. Platform divergence — model and index updates change behavior. Monitor the signals and adapt. Recent provider announcements (OpenAI, Google) and independent analysis (Search Engine Journal, Search Engine Land) show citation and crawling behaviors are in flux.

Evaluation checklist

• /llms.txt exists at site root and lists priority docs and canonical links
• Key docs are available as clean .md artifacts or llms-ctx bundles
• Average factual density per paragraph is 6+ on priority pages
• Each priority page contains ≥3 structured answer blocks
• robots.txt does not block AI crawlers you want to be crawled
• All statistics include temporal anchors (month/year)
• JSON-LD and OG tags are present for canonical pages
• Citation rate across AI platforms is tracked and monitored

Fresh signals to watch (examples)

• OpenAI — Responses API and agent runtime/hosted containers (see OpenAI News items from Mar 2026)
• llms.txt proposal and implementation notes (https://llmstxt.org/)
• Industry reporting on citation behavior after model updates (e.g., Search Engine Journal analysis on ChatGPT citation counts)
• Google Gemini / API updates (inference tiers) and Search Central documentation releases

Operating notes

• Add /llms.txt and .md artifacts to your docs CI pipeline. If feasible, generate llms-ctx bundles for the top 10 docs.
• Run a weekly or monthly citability check for top 25 queries and log provider, model, and citation outcome.
• Correlate major drops/lifts in citation rate with provider release notes and core algorithm updates.

Experiments to run next

1. Emit .md copies and a /llms.txt for a documentation site, then measure change in citation rate over 90 days across ChatGPT, Perplexity, and Gemini.
2. A/B test adding a dedicated structured-answer intro paragraph (scored ≥8) to 50 low-traffic pages and measure citation lift.
3. Track citation counts weekly and correlate with major provider release notes (OpenAI, Google) to quantify sensitivity to model updates.

Research-backed changes

• Fold community /llms.txt guidance into implementation steps and recommend .md artifacts and llms-ctx where possible (source: llms.txt proposal — https://llmstxt.org/).
• Add operational guidance to monitor citation rate continuously because provider model changes can materially alter citation behavior (industry reporting from Search Engine Journal and Search Engine Land; OpenAI agent/Responses API announcements).