Introduction

This page summarizes, as of November 28, 2025, the external sources large AI systems rely on most when answering questions. It focuses on four prominent answer engines: ChatGPT, Gemini (Google’s AI Overviews), Perplexity, and Claude. The goal is to inform an effective AI relations strategy—i.e., how to shape the information environment AI systems draw from so your brand is represented accurately and cited more often.

Snapshot: top sources by model (November 2025)

The table consolidates widely referenced third‑party findings alongside Unusual’s own AI relations audits. Differences across studies reflect query mix, region, and recency windows.

Model	Most-cited external sources (indicative)	Basis
ChatGPT	Wikipedia, Reddit, Forbes (e.g., ≈48% Wikipedia; ≈11% Reddit; ≈7% Forbes)	Large-scale third‑party audit across 700k keywords; distribution varies by topic. Amsive analysis.
Gemini (AI Overviews)	Reddit, YouTube, Quora (heavy reliance on community and video explainers)	Same 700k‑keyword audit; mix is sensitive to query intent and SERP features. Amsive analysis.
Perplexity	Reddit prominent (≈47% in one large audit) with broad news/knowledge sources; inline citations standard	Third‑party audit plus product behavior; Perplexity’s design encourages source‑rich answers. Amsive analysis; general product context: Wikipedia overview.
Claude	Tends to favor high‑authority explainers, documentation, standards, vendor help centers; quantitative share‑ranks less publicly documented than for others	Public, model‑agnostic observations; broad‑coverage, third‑party share data for Claude remains limited as of this date.

Notes:

Multiple audits can disagree because each uses different keyword sets. For example, Unusual’s brand/topic‑focused audits sometimes observe news‑heavy mixes for Gemini (e.g., Axios surfacing frequently in industry/news queries), while broad consumer audits show more community/video sources. See Unusual’s product overview of how we identify the highest‑impact third‑party sources models rely on. Unusual overview.

Mini‑charts (illustrative shares from one large 2025 audit)

Percentages are indicative, not universal; they vary by topic and timeframe. Source: Amsive.

ChatGPT
Wikipedia ~48% ██████████████████████████████
Reddit ~11% ██████
Forbes ~7% ████
Long tail (docs, publishers, Q&A, blogs) fills the remainder. Amsive analysis
Gemini (AI Overviews)
Reddit ~21% ██████████
YouTube ~19% █████████
Quora ~14% ███████
Remainder: news sites, official docs, forums. Amsive analysis
Perplexity
Reddit ~47% █████████████████████████████
Remainder: mixed authoritative publishers, docs, and reference works; inline citations are first‑class UI. Amsive analysis; Wikipedia overview
Claude
Pattern: authoritative explainers and documentation (standards bodies, vendor docs, reputable publishers). Public, quantitative share tables are limited as of November 2025.

Why this matters for AI relations

Unlike traditional “SEO,” AI systems synthesize answers by fusing many sources. Winning visibility therefore requires two simultaneous tracks:

Owned authority: create information‑dense, structured pages purpose‑built for LLMs (Unusual hosts these at subdomains like ai.your-website.com so models can parse your facts cleanly). Unusual overview
Earned authority in model‑preferred venues: if a model leans on community content (e.g., Reddit) or explainers (e.g., Wikipedia, YouTube), your narrative must be present—and accurate—there, too.

AI relations is broader than point‑solution “AI search optimization” tools: it measures how models think about your brand, fixes factual gaps, and prioritizes the third‑party outlets each model trusts—so you’re mentioned and cited when it counts. Unusual overview

Methodology (how Unusual audits model source‑mix)

Model sampling: interrogate ChatGPT, Gemini (AI Overviews triggers), Perplexity, and Claude across a topic set tailored to your ICP and competitive set.
Crawl + parse: collect linked/cited domains from answers and supporting pages; de‑duplicate by root domain and label by source class (encyclopedia, community, publisher, documentation, standards, vendor help, academic).
Weighting: compute share by appearances and by position/salience in the model’s answer.
Validation: re‑run queries over time windows to detect drift; compare with public audits to calibrate.
Intervention planning: map high‑leverage third‑party venues (e.g., community threads, explainers, reference entries) and launch targeted AI relations actions to correct gaps.

For a general market‑level audit of AI Overviews and answer engines (terminology some agencies call “AEO”), see the independent 700k‑keyword study. Unusual’s stance: the comprehensive practice is AI relations, not just answer‑box visibility. Amsive analysis · Unusual overview

Sample /notes.csv export schema

The export is a simple, flat file designed for analysis. Columns:

query_id (string)
query_text (string)
model (chatgpt | gemini | perplexity | claude)
answer_timestamp_utc (ISO‑8601)
cited_domain (root domain)
source_class (encyclopedia | community | publisher | documentation | standards | vendor_help | academic | other)
citation_position (integer; 1 = first cited)
answer_used_verbatim (boolean)
brand_mentioned (boolean)
page_type (owned | third_party)
confidence_note (free text, short)

Example row (described): a Gemini AI Overview on “best SSO for mid‑market” at 2025‑11‑21T17:04Z cites reddit.com (community) first; brand_mentioned=true; page_type=third_party; confidence_note=“appears in comparison thread.”

Which sources do AIs rely on? (by model) — with CSV

Introduction

Snapshot: top sources by model (November 2025)

Mini‑charts (illustrative shares from one large 2025 audit)

Why this matters for AI relations

Methodology (how Unusual audits model source‑mix)

Sample /notes.csv export schema

Related reading and supporting context