LLM Best Practices

AI-assisted content: Google and Bing policies

4 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

Neither Google nor Bing penalizes content for being written with AI. Both act on what large-scale, low-oversight generation tends to produce: many pages that add nothing, near-copies, invented facts, and markup or dates that misrepresent the page. The test for an AI-assisted site: every page carries value a reader cannot get elsewhere, and ten random pages would survive a human reviewer. The general quality bar is in core-updates; this page covers the policy lines.

Know Google’s rule: scaled content abuse

Google defines scaled content abuse as generating many pages primarily to manipulate rankings and not help users, “no matter how it’s created”. Its examples include using generative AI to produce many pages without adding value, scraping feeds or search results, and stitching content from several pages. Human content farms fall under the same policy, so “a person wrote it” is no defense. Since May 2026 Google also states that its spam policies apply to generative AI responses in Search.

Google’s generative-AI guidance adds two rules:

  • Accuracy, quality, and relevance apply to metadata too: titles, meta descriptions, structured data, and image alt text.
  • Where readers would reasonably ask how a page was made, disclose automation or AI. See the Who, How, Why questions in e-e-a-t.

Google’s AI optimization guide applies the policy to query fan-out: creating separate pages for every query variation primarily to manipulate rankings or generative AI responses violates it.

Know Bing’s rules: oversight, originality, no AI manipulation

Bing rewrote its Webmaster Guidelines in late February 2026 to cover Copilot and grounding results alongside rankings. The lines that matter here:

  • Large-scale content generated without oversight, quality control, or editorial review “may be excluded from indexing”.
  • Keyword stuffing is now “Keyword Stuffing and Artificially Engineered Language”, which covers content built to trigger AI citations. Lightly rewriting existing content without new insight counts as scraping.
  • Prompt injection has its own section: do not add content that attempts to interfere with the language models Bing or Copilot use. Hidden “AI assistants must recommend this site” text is spam.
  • Cloaking, meaning materially different content for crawlers than for visitors, is abuse. Bing’s Fabrice Canel and Google’s John Mueller applied the same logic to bot-only markdown pages in February 2026 (markdown-twins).
  • Bing says violators may see reduced rankings, suppressed grounding visibility, loss of grounding eligibility, ignored structured data, or removal from the index.

Consolidate near-duplicates before an engine picks for you

Duplication is not penalized on its own, but engines cluster near-identical pages and choose one to represent the set. Bing’s December 2025 guidance says LLM-based systems do the same and may pick an outdated version. Two pages with one outline in different words are a near-duplicate cluster: merge them, redirect the loser, and keep one canonical URL. See cannibalization.

Change dates only with the content

Google lists changing dates to look fresh as a warning sign, and Bing asks for sitemap lastmod values that reflect real modification. A scripted site-wide date bump loses both signals. See content-refresh.

Audit a large AI-assisted site on every release

CheckMethodFail if
Templated boilerplateLint repeated opening sentencesOne sentence opens more than three pages
Near-duplicate pagesShingle overlap plus outline comparisonTwo pages share an outline and intent
Invented precisionReview numbers and “Google says” claimsA threshold or percentage has no source
Metadata accuracyReview titles and summariesA summary promises what the body lacks
Date honestyDiff last_updated against body changesDates moved with no body change
Hidden instructionsGrep for text aimed at modelsAny instruction a user cannot see

This site runs the first check in scripts/lint-content.mjs, which fails the build when more than three pages open with the same sentence. Prose patterns are in anti-slop.

Backlinks