Web Text Fragments: Decoding Google AI Mode’s Hidden Citation Signal

Every Google AI Mode and Gemini citation URL contains a #:~:text= Web Text Fragment anchor. Decode it and you can see the exact sentence Google extracted from the source page. This is the most underused citation-intelligence signal in 2026 – and it has produced one of the most counterintuitive findings in AI search research.

What Web Text Fragments are

Web Text Fragments (the #:~:text= URL syntax) are a W3C-track browser feature that lets a URL link directly to a specific phrase on a page. When you click a Google AI Mode citation, the browser scrolls to and highlights the exact sentence Google quoted.

That sentence is the literal training-and-retrieval target of the AI engine. Reverse-engineering which sentences get fragment-targeted is reverse-engineering Google’s extraction policy.

The counterintuitive finding: position doesn’t matter

SALT Agency analyzed 2,318 unique URLs cited in AI Mode, recording the vertical pixel position of each highlighted fragment. The headline finding:

Researchers found no correlation between how high text appears on a page and whether Google’s AI selects it for citation. Average citation depth varied by vertical, from 2,400 pixels in travel to 4,600 pixels in SaaS – well below the traditional above-the-fold area.

SALT Agency, 2026

How to extract intelligence from text fragments at scale

Three practical workflows:

1. Manual reverse-engineering. Run your top 20 prompts in Google AI Mode. Inspect each citation URL, decode the #:~:text= fragment, and catalog which sentences Google chose. Patterns will emerge: definitions, statistics with named sources, lists with numerical structure.
2. Programmatic collection. Use a scraping tool to collect citation URLs at scale. Decode fragments with a small Python script (urllib.parse.unquote). Store in a database for trend analysis.
3. Open-source pipeline. The Grounding Citation Analysis project ships a complete pipeline that has analyzed 42,971 citations across 6 AI platforms. Reproducible, open-source, designed for this exact workflow.

What the extracted sentences consistently look like

After analyzing tens of thousands of fragment-extracted passages, recurring patterns:

• Definition-style sentences. “X is a Y that does Z.” Concise, complete, decontextualized.
• Statistics with sources. “According to [source], 47% of [thing] [does action].”
• Direct answers to implicit questions. Sentences that read like featured-snippet answers.
• List items. Especially numbered lists where each item is one complete thought.
• Quoted experts. Direct quotes from named individuals with sameAs-resolvable identity.

Conversely, what almost never gets extracted: long narrative paragraphs, content with heavy pronouns (“This means…”), sales copy, and AI-generated filler.