Research note

How major AI search platforms describe source use

A desk review of official Google, Bing, OpenAI and retrieval research documentation, with clear limits on what those documents can prove.

15 November 2026Desk review · 8 primary documentsMethod v0.1

This is a desk-research note. It summarizes primary documents and does not claim a live AI citation share.

Key findings

What this review establishes

  1. Google and Bing publish crawlability, quality and webmaster guidance without disclosing complete ranking or retrieval weights.
  2. Bing has introduced a Webmaster Tools surface for reporting AI citations across Copilot, Bing summaries and selected partner integrations.
  3. OpenAI's product announcement establishes that ChatGPT search can display linked web sources, but it does not provide a reproducible citation policy.
  4. Retrieval research explains source-grounded generation as a broader architecture, not a guarantee that a particular URL will be selected.
Research question

Scope and setup

This note compares first-party documentation from Google, Microsoft and OpenAI with foundational retrieval papers. The review asks what the documents actually state about source discovery, indexing and citation visibility.

The documents consistently support a cautious conclusion: publishers can improve clarity, accessibility and evidence quality, but no reviewed source publishes a complete rule that determines whether an individual AI answer will cite a page.

The review also separates product reporting from observed outcomes. Bing's AI Performance preview describes a measurement surface; it does not establish that every property has available citation data.

The sample is deliberately documentary. It does not claim a live citation share, a ranking, or a causal effect from a particular content change.

Method

How observations are assembled

  • Record the full question, locale and product surface.
  • Start a fresh conversation and retain one answer per run.
  • Capture every visible source URL and its position.
  • Repeat the question and keep each outcome separate.
  • Publish sample size, method version and unresolved limits.
Read the method
Data

Data availability

The production report will include a data dictionary, file format, license and permanent download path when the first verified sample is published.