AI Citation Format Specification by Engine: How ChatGPT, Perplexity, Gemini, and Claude Render Sources in 2026
ChatGPT
Claude
Gemini
PerplexityThis reference specifies how the four major answer engines render source citations in 2026 — inline anchor style, attribution placement, source-list density, hover and preview behavior, and structured signals — so content engineers can design extractable units that survive each engine's rendering pipeline.
TL;DR
- ChatGPT uses inline numeric superscripts that resolve to a hover preview card and a collapsible Sources list at the bottom of the answer.
- Perplexity binds every claim to a numbered bracket immediately after the sentence, plus a horizontal Sources strip above the answer body.
- Google AI Overviews and AI Mode attach a small link-icon chip to each cited claim and surface a vertical sources panel; AI Mode now embeds inline anchor links directly into prose.
- Claude.ai renders web-search citations as inline source badges with the publisher domain, while the Claude API returns a structured citations array with character or page ranges.
- All four engines reward content that is answer-first, schema-tagged, and front-loaded — 55% of AI Overviews citations come from the top 30% of a page, per CXL's 100-page study.
Why this specification matters
If you publish content meant to be cited by AI, you are not optimizing for one citation format — you are optimizing for four. Each engine renders sources differently, evaluates them with different ranking signals, and exposes different surface area to the end user. A page that is perfectly extractable for Perplexity may be entirely uncited by Gemini. This document defines, per engine, the exact fields that determine whether your URL appears in the rendered citation chrome.
It does not cover academic citation styles (APA, MLA, Chicago) for citing AI tools themselves — see standards bodies for those rules. This is a specification for how AI engines cite the open web back to users.
Spec format
Each engine block specifies six fields:
| Field | Description |
|---|---|
| inline_marker | Visual element placed inline in the answer body |
| anchor_text | Words or icons that become the clickable target |
| attribution_placement | Where in the rendered answer the citation appears |
| source_list | Whether and how a consolidated source list is rendered |
| hover_behavior | What appears on hover or tap of an inline marker |
| structured_signal | Programmatic surface (API field, OG tag, schema) the engine prefers |
ChatGPT (OpenAI, 2026)
| Field | Value |
|---|---|
| inline_marker | Numeric superscript chip appended to the cited sentence |
| anchor_text | Icon-only chip; tooltip shows publisher name and page title |
| attribution_placement | Immediately after the punctuation of the cited sentence |
| source_list | Collapsible Sources accordion at the bottom of the message; expands to a flat list of titles and URLs in citation order |
| hover_behavior | Hover or tap reveals a card with favicon, page title, publisher, and a 2-3 line excerpt of the cited passage |
| structured_signal | og:title, og:site_name, schema.org Article.headline, datePublished, and the OpenAI web.search retrieval index |
Behavior notes:
- ChatGPT now resolves citations to specific passages, not just URLs. Pages with semantically distinct sections cited by anchor (#section-id) get attributed to the correct subsection.
- When a single source supports multiple claims, the same numeric chip is reused — citation IDs are stable across the answer.
- The Reference section is reliably populated in 2026, a behavior that was inconsistent in early 2025 developer tests on the Responses API.
Perplexity (Sonar models, 2026)
| Field | Value |
|---|---|
| inline_marker | Square-bracket numeric markers, often back-to-back when multiple sources support one claim |
| anchor_text | Number only; the bracket itself is the link |
| attribution_placement | Immediately after the cited word or phrase, often mid-sentence rather than only sentence-end |
| source_list | Horizontal Sources strip rendered above the answer body, with favicon cards |
| hover_behavior | Hover preview card with publisher, full title, snippet, and access date |
| structured_signal | Live retrieval index (Perplexity claims a corpus approaching 100B pages); strong preference for sources with high factual density and explicit data |
Behavior notes:
- Citations in Perplexity are architecturally bound to retrieved chunks, not appended after generation. The numeric ID maps deterministically to a single retrieved document.
- Perplexity Pages and Collections inherit citation behavior from the answer engine — the same spec applies.
- Sources with explicit numeric claims and dated assertions are favored; bare opinion content is filtered out before ranking.
Google AI Overviews and AI Mode (Gemini, 2026)
| Field | AI Overviews | AI Mode |
|---|---|---|
| inline_marker | Small link-icon chip next to each cited claim | Inline anchor links rendered as standard underlined hyperlinks within the prose |
| anchor_text | Icon-only; the chip resolves to a single source | The anchor wraps a 1-4 word phrase from the cited sentence |
| attribution_placement | End of cited claim, before punctuation | Mid-sentence, on the most-extractable phrase |
| source_list | Vertical sources panel on the right (desktop) or expandable card (mobile) | Same panel plus inline links — two channels of attribution |
| hover_behavior | Hover shows publisher and page title; click navigates | Standard link tooltip plus Google's About this result panel |
| structured_signal | Standard Search index eligibility (must be indexed and snippet-eligible); schema.org FAQPage, HowTo, Article; Knowledge Graph entity recognition | Same as AI Overviews |
Behavior notes:
- AI Overviews now cite ~13.3 sources per response on average, up from ~6.8 in early 2024, reflecting a verification-bias shift.
- 55% of citations come from the top 30% of the page; only 21% come from the bottom 40%. Front-loaded answers win.
- AI Mode added inline anchor links in late 2025 — content with quotable, extractable noun phrases gets disproportionate inline visibility.
- Google's developer documentation states: any page indexed and eligible for a search snippet is eligible for AI Overviews citation. There are no extra technical requirements.
Claude (Anthropic, 2026)
Claude has two distinct citation surfaces that must be specified separately.
Claude.ai chat (consumer surface)
| Field | Value |
|---|---|
| inline_marker | Domain-cardified badge (favicon plus publisher domain) inline at the end of the cited claim |
| anchor_text | Publisher domain (for example nytimes.com) |
| attribution_placement | End of sentence, before punctuation |
| source_list | Compact source row near the top of the message; expands to titles and URLs |
| hover_behavior | Card with title, publisher, and accessed-on date |
| structured_signal | Anthropic web-search retrieval; preference for canonical, schema-marked, and recently updated pages |
Claude API (citations feature)
| Field | Value |
|---|---|
| inline_marker | Not visual — returned as JSON citations array attached to each text block |
| anchor_text | N/A — the API consumer renders its own UI |
| attribution_placement | Per text block; one block can have multiple citations |
| source_list | Caller decides; raw structure is [{ document_index, location, cited_text }] |
| hover_behavior | N/A |
| structured_signal | For PDFs: page-number range (1-indexed). For plain text: character-index range (0-indexed). Custom content documents allow caller-defined locators. |
Behavior notes:
- The Claude API requires citations.enabled=true to be set on all or none of the documents in a request — partial enabling is rejected.
- Only text citations are supported in 2026; image citations are not yet exposed.
- Claude Design (Anthropic Labs, 2026) renders citations as embedded source artifacts in slides and prototypes but uses the same underlying API structure.
Cross-engine comparison matrix
| Dimension | ChatGPT | Perplexity | AI Overviews | AI Mode | Claude.ai | Claude API |
|---|---|---|---|---|---|---|
| Inline style | Numeric superscript | Numeric bracket | Icon chip | Underlined link | Domain badge | JSON only |
| Avg sources per answer | 3-6 | 5-10 | ~13.3 | 4-8 | 2-5 | Caller-defined |
| Source list location | Bottom (collapsible) | Top (horizontal strip) | Right panel | Right panel | Top row | N/A |
| Mid-sentence anchors | No | Yes | No | Yes | No | Yes (char range) |
| Passage-level binding | Yes | Yes | Partial | Yes | Yes | Yes |
| Schema dependency | Medium | Low | High | High | Medium | None |
| Freshness signal | High | Very high | Medium | High | High | N/A |
Implementation rules for content engineers
- Front-load extractable claims. The top 30% of the page absorbs the majority of citations, especially in AI Overviews.
- Tag claims with schema. Article, FAQPage, HowTo, and Dataset schema improve eligibility for AI Overviews and AI Mode and indirectly improve ChatGPT and Claude retrieval.
- Keep canonical_url stable. All four engines deduplicate by canonical URL; mismatched canonicals split citation share.
- Use precise, declarative sentence shapes. Perplexity and AI Mode disproportionately cite short, fact-dense statements with explicit numbers or dated claims.
- Expose datePublished and dateModified. Freshness is a hard signal in Perplexity ranking and a soft signal everywhere else.
- Make titles and og:title extractable as anchor text. AI Mode and Claude.ai render publisher and title in the visible chrome.
- Avoid hiding claims behind interaction. Content gated by tabs, accordions, or JS-rendered blocks is reliably under-cited across all four engines.
Common misconceptions
- "Higher Google rank equals more AI citations." Partly true for AI Overviews, but Perplexity and Claude do not use Google's ranking. Per-engine optimization is required.
- "Citations are decorative." They are functional surfaces. They drive click-throughs, brand recognition, and downstream discovery. Citation share is now a tracked GEO KPI.
- "You can fingerprint your way into citations." None of the four engines documents a manual inclusion API. Inclusion is earned via retrieval ranking.
How to apply this spec
- Audit your top 20 pages against each engine row above. Which inline markers do they currently appear in? Use a citation monitor (ZipTie, Position.Digital, or a custom scraper with rotating queries) to sample.
- Build extractable units in the top 30% of every page: a 2-3 sentence definition, a numbered list, a comparison table, an FAQ block. These are the units citation engines lift.
- Track per-engine citation share as a primary KPI alongside organic rank. The four engines render differently, so they must be measured differently.
FAQ
Q: Do all four engines cite the same source for the same query?
A: Rarely. ChatGPT and Claude lean on their training-time corpora plus live retrieval; Perplexity and AI Overviews query live indices in real time. A 100-citation study (CXL, 2025) found less than 25% overlap across the four engines for identical prompts.
Q: Can I optimize once and rank in all four citation formats?
A: Partially. Schema, canonical URL, freshness, and front-loaded extractable claims help everywhere. But Perplexity rewards data density, AI Overviews rewards entity authority, and Claude rewards clean canonical structure. Per-engine tuning still matters.
Q: Are inline citation chips clickable?
A: Yes in all four engines. Click-through analytics differ widely — Perplexity drives the highest CTR per citation chip in 2026 benchmarks; AI Overviews drives the lowest because the answer often saturates user intent.
Q: Does Claude's API let me return Anthropic-style citations from any LLM?
A: Not natively. The structured citations field is Anthropic-only. Open-source approximations (custom
Q: How often does this spec change?
A: Major engine UI changes have shipped 2-3 times per year since 2023. This document is reviewed every 90 days; consult the updated_at frontmatter field for the latest revision date.
Related Articles
AI Answer Length Patterns: Word and Token Targets per Engine in 2026
Reference for AI answer lengths in 2026 — word and token targets for ChatGPT, Perplexity, and Google AI Overviews so writers format extractable answers.
AI Citation Confidence Scoring Framework: Predicting Source Inclusion Likelihood
AI citation confidence scoring framework: a predictive model that scores how likely generative engines are to cite a source based on retrieval, grounding, and trust signals.
AI Citation Half-Life Reference: Decay Patterns Across ChatGPT, Perplexity, Gemini, and Copilot
Reference table of AI citation half-life and decay patterns across ChatGPT, Perplexity, Gemini, and Copilot. Use to plan refresh cadence and forecast visibility loss.