LLM Citation Anchor Text Patterns: How Generative Engines Phrase Source Mentions

TL;DR

• Perplexity uses numbered superscripts ([1], [2]) inline with a numbered sources list at the bottom of the answer.
• ChatGPT (Search) uses inline clickable domain pills (example.com) and footnote-style links, with a Sources tray.
• Gemini / Google AI Overviews use pill-style chips and a Sources card carousel; phrases like "according to…" are common.
• Claude uses inline parenthetical attributions ((Source, 2026)) plus a sources tray, and exposes web_search_result_location spans via API.
• Copilot uses numbered superscripts similar to Bing/Perplexity, with hover preview cards.

Why anchor text patterns matter

AI citation surfaces are the new "snippet." The phrase, format, and location of an attribution determine whether a user clicks through, whether your brand name reaches the reader, and whether the engine's confidence model treats your content as a primary source. AI search has been shown to display citations inconsistently and inaccurately, which means structuring content for predictable extraction is part of optimization.

Per-engine reference

ChatGPT (Search and connectors)

• Inline anchor: Clickable domain pill rendered next to the cited claim, e.g. “…reduces latency by 27% (example.com).”
• Secondary anchors: Footnote-style superscript numbers in long answers; expandable Sources tray.
• Phrasing cues: “According to {brand}…”, “{brand} reports…”, “per {publication}…”
• What it rewards: Pages with strong brand entity in the page title and clean canonical URL. ChatGPT favors Wikipedia and large publishers (Wikipedia ~48% of citations in some datasets).

Perplexity

• Inline anchor: Numbered superscripts [1], [2], [3] directly after the claim sentence.
• Secondary anchors: Numbered sources list with thumbnail favicons, source titles, and short excerpts.
• Phrasing cues: Often answer-first without an explicit “according to,” relying on the superscript to carry attribution.
• What it rewards: Reddit (~47% of citations in some datasets), long-form blog posts, and content with extractable lists. Perplexity heavily quotes the first 2-3 sentences of well-structured paragraphs.

Gemini / Google AI Overviews

• Inline anchor: Pill-style chips embedded in the prose, sometimes with the source brand visible.
• Secondary anchors: Sources card carousel below the answer with thumbnail, title, snippet.
• Phrasing cues: “According to {brand}…”, “As explained on {publisher}…”, “{brand} notes that…”
• What it rewards: First-party official sites and content with strong on-page entity match (Yext data on 17.2M citations shows Gemini favors official websites more than the other engines).

Claude

• Inline anchor: Parenthetical (Source, YYYY) style attribution within the prose; clickable in claude.ai.
• Secondary anchors: Per-message sources tray; via API, web_search_result_location returns url, title, cited_text (≤150 chars), encrypted_index.
• Phrasing cues: “Per {publisher}…”, “{brand}’s documentation states…”, “As {author} notes…” — named-author phrasing is more common than on other engines.
• What it rewards: Smaller, niche outlets and standards-body documentation; sentence-span chunkability.

Microsoft Copilot

• Inline anchor: Numbered superscripts (¹1, ¹2) styled similarly to Bing.
• Secondary anchors: Hover preview cards on the superscript.
• Phrasing cues: Mix of “according to…” and unattributed sentences; relies on superscript.
• What it rewards: Content well-indexed in Bing with valid schema and crisp meta descriptions.

Anchor format reference table

Engine	Inline anchor	Secondary anchor	Common phrasing
ChatGPT (Search)	Domain pill (example.com)	Sources tray; footnote links	"According to {brand}…"
Perplexity	Superscript [1]	Numbered sources list	Often unattributed prose
Gemini / AI Overviews	Pill chip	Sources card carousel	"As {publisher} explains…"
Claude	Parenthetical (Source, 2026)	Sources tray; API spans	"Per {publisher}…" / "As {author} notes…"
Copilot	Superscript ¹1	Hover preview card	Mixed

Implications for content writers

1. Make brand entity unambiguous in the page title and first 100 words. ChatGPT and Gemini surface it in "according to…" phrasing; if the engine can’t find a clean entity, it falls back to bare domain pills.
2. Front-load the citable claim. Perplexity and Copilot quote near-verbatim from the first 2-3 sentences of well-structured paragraphs.
3. Date your claims inline. Claude prefers parenthetical (Source, 2026) spans; an explicit date is a magnet for that pattern.
4. Use named-author bylines and Person schema. Claude and increasingly Gemini cite “As {author} notes…” style; without a named author you forfeit that anchor.
5. Keep self-contained spans ≤ 150 characters. Claude’s cited_text cap is 150 chars; a tightly written answer-first sentence often becomes the literal citation.
6. Avoid clickbait or vague openers. Engines down-weight ambiguous openers when selecting which span to anchor a citation on.

Common misconceptions

• “Inline domain pills are stable.” They aren’t. ChatGPT regularly drops domain pills in favor of footnote links across UI revisions. Optimize for either format.
• “The [1] style is universal.” Only Perplexity and Copilot consistently use it. ChatGPT, Gemini, and Claude all render attributions differently.
• “AI engines always link the source.” No. Tow Center research found AI search tools frequently fabricate or misattribute citations. Brand strength reduces the risk of being misquoted without attribution.

FAQ

Q: Do AI engines link to specific page anchors?

Mostly to the page URL, occasionally to a #section anchor when the cited content lives in a clearly named section. Adding an id to each H2 increases the chance of a deep-linked citation, especially in Perplexity and Gemini.

Q: What is the difference between an AI citation and a brand mention?

A citation is a clickable link to your page; a mention is your brand name appearing in the answer text without a link. Both matter — mentions still drive recall and downstream branded search even when the user does not click.

Q: Do citations include the publication date?

Claude routinely includes the year in parenthetical citations. Gemini and ChatGPT sometimes display “published {date}” under the source card. Perplexity shows freshness color cues but no inline date.

Q: Should the citation phrase appear in my own content?

Yes — writing in a quote-ready style (answer-first, dated, attributed) increases the chance an engine will lift the sentence verbatim. Avoid first-person marketing voice in spans you want extracted.

Citations

: Columbia Journalism Review, AI Search Has a Citation Problem (2025).

: Discovered Labs, AI Citation Patterns: How ChatGPT, Claude, and Perplexity Choose Sources (2025).

: Yext Research, AI Citation Behavior Across Models: 17.2 Million Citations (2026).

: Anthropic, Web search tool — Claude API Docs. Retrieved 2026-04-28.