AI Citation Matrix
How each answer engine finds the pages it cites, what you control, and what research says. Rows are based on vendor docs. Studies are snapshots, so re-check them before making big decisions.
Last reviewed: 2026-11-10
By Engine
| Engine | How it finds pages (vendor docs) | Your control | What studies found |
|---|---|---|---|
| Google AI Overviews, AI Mode | Google Search index. A page must be indexed and eligible for a snippet; there are no extra requirements. Both may use “query fan-out” across subtopics (Google) | Googlebot in robots.txt; nosnippet, data-nosnippet, max-snippet. Google-Extended does not affect Search | AI Overviews appeared on 13.7% of 55,393 queries; nearly 30% of cited domains were not in the first-page results (IMC 2026). AI Mode sidebar links had ~54% domain overlap with the top 10 (Semrush, June 2025) |
| ChatGPT search | OAI-SearchBot surfaces sites. Opted-out sites are not shown in answers, though they can still appear as navigational links (OpenAI). It sometimes sends rewritten queries to partner search providers, including Microsoft (OpenAI Help) | OAI-SearchBot for search, GPTBot for training, set independently | Every page ChatGPT cited was one its own search retrieved; retrieval explained 32% of variation against 2.8% for page properties (2610.15033). Pages retrieved more often are cited more often (2610.15382) |
| Perplexity | PerplexityBot surfaces and links sites in results and is not used for model training (Perplexity) | PerplexityBot; Perplexity-User for user-requested fetches | Over 91% domain overlap with Google’s top 10 (Semrush, June 2025) |
| Claude with web search | Claude-SearchBot indexes content to improve search results; blocking it may reduce visibility (Anthropic). Every response includes citations (Claude Help) | Claude-SearchBot, Claude-User, ClaudeBot (training) | No public study found that is specific to Claude citations |
| Microsoft Copilot, Bing | Bing and Copilot use the same crawling, indexing and ranking foundation as Bing search. Sitemaps and IndexNow support “grounding eligibility” (Bing Webmaster Guidelines) | robots.txt and meta directives, sitemaps, IndexNow | No public study checked |
| Gemini app | Google-Extended controls whether Google-crawled content is used for Gemini training and grounding (Google) | Google-Extended (robots.txt token only) | No public study checked |
Full crawler list: AI Crawler User-Agents.
Across Engines
- Engines disagree. On live pages, ChatGPT, Gemini and Grok picked the same top product for only 48.8–66.3% of brands (LIVE-GEO).
- AI search leans toward earned third-party media over brand-owned content (2509.08919).
- Brand mentions on Reddit (0.75) and on other webpages (0.70) correlate with how often AI answers name a brand. This is a correlation from a vendor study (Ahrefs, November 2026).
- Wikipedia, YouTube and Reddit were the most-linked sources in both Google AI summaries and standard results (Pew, 2025).
- The most-cited page type differs by industry. Listicles earned 61% of citations in B2B tech services and 0% in healthcare (DeltaV, 2026).
- Blocking only affects crawlers that honor it. One study counted 37,642 Grok citations to a publisher that blocks AI crawlers (Goodie, 2026).
What To Do
- Get indexed first. Every engine above cites from a search index or its own retrieval, so crawlable, indexable pages come before anything else.
- Allow the search crawlers you want to appear in (Googlebot,
OAI-SearchBot,PerplexityBot,Claude-SearchBot), and decide on the training tokens separately. - Earn mentions on third-party sites and communities where your topic is already discussed.
- Track cited URLs, not only brand mentions, and treat a single run as noise.
What Not To Infer
- Google says you don’t need new machine-readable files, AI text files or special schema to appear in AI Overviews or AI Mode (Google). One paper found no isolated evidence that
llms.txt-style files help AI retrieval (2610.38361). Shipllms.txtfor agents and docs, not as a citation lever. - A high citation share for one site in one study is a snapshot of that prompt set. The prompt corpus defines what a visibility score measures (2609.06811).
- Correlation studies don’t show what causes citations.