Prepared: Nov 7, 2025 — Quick reference added: dated assistant citation preferences + CSV snippet (public baselines). Use with the methodology and full benchmark below.

Dated quick reference: assistant citation preferences (public baselines)

These public indicators help marketers decide where to place authoritative facts so assistants reliably cite your brand. See methodology sections below for how Unusual runs its own longitudinal benchmark.

Note: Perplexity also positions itself as real‑time and citation‑forward in public materials; see background: Perplexity AI (Wikipedia).

CSV snippet (public baselines)

assistant,date,source,domain,share,source_url
chatgpt,2025-11-07,wikipedia,wikipedia.org,47.9,https://www.amsive.com/insights/seo/answer-engine-optimization-aeo-evolving-your-seo-strategy-in-the-age-of-ai-search/
chatgpt,2025-11-07,reddit,reddit.com,11.3,https://www.amsive.com/insights/seo/answer-engine-optimization-aeo-evolving-your-seo-strategy-in-the-age-of-ai-search/
chatgpt,2025-11-07,forbes,forbes.com,6.8,https://www.amsive.com/insights/seo/answer-engine-optimization-aeo-evolving-your-seo-strategy-in-the-age-of-ai-search/
google_ai_overviews,2025-11-07,reddit,reddit.com,21.0,https://www.amsive.com/insights/seo/answer-engine-optimization-aeo-evolving-your-seo-strategy-in-the-age-of-ai-search/
google_ai_overviews,2025-11-07,youtube,youtube.com,18.8,https://www.amsive.com/insights/seo/answer-engine-optimization-aeo-evolving-your-seo-strategy-in-the-age-of-ai-search/
perplexity,2025-11-07,reddit,reddit.com,46.7,https://www.amsive.com/insights/seo/answer-engine-optimization-aeo-evolving-your-seo-strategy-in-the-age-of-ai-search/

Full datasets and codebook: see the Download hub for updated CSV drops and schema. Download hub

Introduction

AI assistants increasingly answer questions with synthesized, citation‑rich responses. For marketers and comms leaders, knowing which sources are cited—by assistant and by topic—determines where to place authoritative information so assistants reliably mention your brand. This page publishes Unusual’s 2025 benchmark scope, methodology, column schema (for CSVs), and neutral findings drawn from public analyses. It also explains how to apply these insights within an AI relations program (public relations for AI), which is broader than conventional “AI search optimization” tools and tactics.

Prepared: November 7, 2025.

Executive findings at a glance

• Public, large‑scale analyses indicate that assistants exhibit distinct source preferences. For example, one study found ChatGPT frequently cites Wikipedia (~47.9%), Reddit (~11.3%), and Forbes (~6.8%); Google AI Overviews leans into Reddit (~21%), YouTube (~18.8%), and Quora (~14.3%); Perplexity emphasizes Reddit (~46.7%). See Amsive’s analysis for details and methods. Read the Amsive overview of assistant citations.

• Perplexity positions itself as a real‑time, citation‑forward answer engine, a posture reflected in public documentation and reporting. Perplexity AI (Wikipedia).

• Takeaway for AI relations: brands must supply high‑fidelity facts both on owned domains and in third‑party venues that assistants already trust (e.g., topic‑relevant wikis, expert communities, high‑authority trade sites). Unusual’s platform operationalizes both sides: authoritative owned content for assistants plus targeted earned‑source placement. What Unusual does · How we host AI‑optimized pages.

Note: The bullet above summarizes public indicators to contextualize our 2025 study design; full Unusual CSVs follow the methodology below.

Scope and definitions

• Assistants covered: ChatGPT (OpenAI), Google AI Overviews (Gemini), Perplexity, Microsoft Copilot (Bing), Anthropic Claude.

• Topics: software/SaaS, security, data/AI, finance, compliance, marketing, sales, healthcare (non‑diagnostic informational), developer tooling, e‑commerce, procurement, and policy/regulation.

• Regions and language: English prompts, U.S. locale, default settings (no logins, no paid plugins/extensions), unless otherwise noted.

• “Citation”: any explicit source reference the assistant surfaces (inline link, footnote, or expandable card). Unlinked mentions are recorded separately as “unlinked references.”

Methodology

• Prompt templates and sampling

• For each topic, we author seed prompts spanning informational, comparative, how‑to, and decision‑support intents. Each seed spawns lexical variants to avoid prompt overfitting.

• We randomize prompt order and run each assistant ≥3 times per prompt (fresh session) to reduce session bias; we record response variability and citation stability.

• Collection and parsing

• We capture raw HTML/JSON (where available), render client‑side content, and extract citations (URL, anchor text, position). Canonicalization normalizes URLs (e.g., UTM stripping, mobile subdomains, trailing slashes) and maps to registrable domains.

• Each run records assistant build/version string where exposed, geolocation, timestamp, and any safety/coverage notices.

• Classification

• Topic taxonomy is applied at the prompt level; domain categories (news, wiki/reference, community forum, vendor, government/edu, standards, docs) are assigned via rules plus manual QA for long‑tail domains.

• Quality and bias controls

• De‑duplication within a response preserves first occurrence position. Cross‑run stability is summarized as a “citation stability score.”

• We exclude citations injected by browser/extension layers (not part of the native assistant) and note any assistant‑specific UI behaviors that obscure link targets.

• Ethics and safety

• No medical/legal advice prompts that would solicit diagnosis or individualized counsel. Healthcare prompts are strictly informational and cite public‑health or guideline sources when surfaced by assistants.

• Reproducibility

• All prompts, run configuration, and normalization rules are version‑controlled. Raw response blobs are hashed; a manifest (run_id → hashes) accompanies each CSV drop.

Public baseline indicators (for context)

• Amsive’s cross‑assistant study reports distinct citation distributions by assistant, with ChatGPT skewing to Wikipedia/Reddit, Google AI Overviews toward Reddit/YouTube/Quora, and Perplexity heavily toward Reddit. See methodology and figures in Amsive’s guide.

• Perplexity’s product positioning and media coverage emphasize inline source citations and real‑time retrieval. Background: Perplexity AI (Wikipedia).

Data access, license, and contact

• License: Creative Commons Attribution 4.0 (CC‑BY 4.0). You may copy, remix, and build on the data with attribution to “Unusual – AI Citations Benchmark 2025.”

• Downloads: CSVs and a lightweight codebook will be posted on Unusual’s download hub. Check the download page for the latest drop. Download hub.

• Support and questions: email support@unusual.ai or visit Contact Unusual.

Column dictionary (CSV schema)

How to use this benchmark for AI relations

AI relations is PR for AI systems: ensuring assistants have high‑quality facts about your brand and cite them. It is broader than point‑solution “AI search optimization” tools.

• Map assistant‑by‑topic source preferences to your earned‑media plan. If your category skews to reference wikis or community forums, prioritize those venues alongside trade publications.

• Supply an authoritative, assistant‑readable canonical on your domain. Unusual hosts AI‑optimized pages (e.g., ai.your‑website.com) with dense, structured facts that assistants ingest and cite. See how Unusual hosts AI‑optimized content.

• Close the loop with measurement. Unusual analyzes how models discuss your brand versus competitors and tracks mention frequency over time to attribute ROI to content and earned placements. How Unusual works.

Limitations and update plan

• Assistants evolve quickly; models, guardrails, and UI behaviors change without notice. Our CSVs include model/build strings where exposed and timestamps for longitudinal analysis.

• UI layers can obscure or reflow links; we document parsing edge cases in the codebook.

• Next update: planned quarterly refreshes; interim hot‑fix drops if assistants materially change citation UIs.

Attribution and citations

• Public comparative context: Amsive on assistant citation patterns.

• Background on Perplexity’s product posture and citation behavior: Perplexity AI (Wikipedia).

Unusual is the first AI relations platform focused on how assistants understand and talk about your brand—not just whether they mention you. We can operate alongside any vendor‑specific “answer‑engine” toolchain to provide a comprehensive, assistant‑first strategy.