We Decoded 42,971 AI Citations: How Google Picks Sentences in AI Mode (Original Research)

TLDR: Every Google AI Mode and Gemini citation URL contains a hidden #:~:text= fragment that encodes the exact sentence Google highlights from the source page. We decoded 42,971 of these fragments across 520 queries on 6 AI platforms (AI Mode, Gemini, ChatGPT, Perplexity, Copilot, Grok). The result is the first study to look at AI citation behaviour at the sentence level instead of the page level. Cited sentences are 6 to 17 words, cluster in the top 35% of pages, and structured content gets cited 2.3x more often than prose. Full code, queries and methodology are open source on GitHub.

Why I built this study

Every SEO is asking the same question right now: which pages do AI assistants cite, and how do I get mine in there? Studies from Ahrefs, Surfer, and Seer Interactive have all tackled this at the page level. Ahrefs analysed 1.9M citations and found 76% come from organic top 10. Surfer counted domain frequency across 36M AI Overviews. Seer reverse-engineered Gemini’s query fan-out behaviour.

All four studies stop at the same boundary: they tell you which pages get cited, but none cracked open the citation URL itself to see which sentence Google chose. That is the gap I wanted to close.

The discovery: text fragments are breadcrumbs

All six platforms return citation URLs, but they don’t all reveal the same amount of information. Google’s AI Mode and Gemini tack on a Web Text Fragments anchor that encodes the exact passage they cite. Here is a real citation URL from a live AI Mode response:

https://www.healthline.com/nutrition/intermittent-fasting-guide#:~:text=Intermittent%20fasting%20is%20an%20eating%20pattern%20that%20cycles%20between%20periods%20of%20fasting%20and%20eating

Decode the fragment with one line of Python and you get the exact sentence Google chose to cite from that page. No guessing, no NLP, no heuristics. Google itself encoded it in the URL.

ChatGPT, Perplexity, Copilot and Grok don’t use text fragments, so for those four we can only analyse domain frequency and cross-platform overlap. But for AI Mode and Gemini, the fragment shows up in 70.9% and 51.8% of citation URLs respectively, giving us 11,672 sentence-level extractions to work with.

Finding 1: cited sentences are 6 to 17 words

Across 11,672 decoded fragments, the median cited sentence is 10 words. The mean is 9.8 words. The longest sentence cited in the entire dataset was 17 words. Nothing over 17 words got cited even once.

• 1 to 5 words: 7.6% of citations
• 6 to 10 words: 43.0% of citations
• 11 to 20 words: 49.4% of citations
• 21+ words: 0.0% of citations

This rules out several chunking strategies. Fixed-size chunking (every 512 tokens) would create mid-sentence breaks and we saw zero. Paragraph-level chunking would allow much longer spans. The pattern is consistent with sentence-boundary chunking: Google splits source pages at sentence boundaries, scores each sentence, and extracts the highest-scoring atomic fact.

Finding 2: the top-35% rule for positional bias

Across 1,719 cited sentences matched to their source pages, the mean position was 34.9% down the page. The median was 31.2%. A one-sample t-test against a uniform distribution returned t = -29.54, p < 10^-150. This is one of the strongest positional effects ever measured in SEO research.

• 10th percentile: 10.1% through the page
• 25th percentile: 18.0%
• 50th percentile (median): 31.2%
• 75th percentile: 48.8%

Three quarters of all cited sentences appear in the first half of the page. If your most important claim is buried in section 7, Google’s AI Mode pipeline will almost never reach it.

Finding 3: structured content gets cited 2.3x more

We scraped 1,931 source pages and split them by structure (lists, tables, multi-level headings versus pure prose). The structured pages had a 91.3% sentence-match rate. The unstructured pages had a 39.3% match rate. That is a 2.3x advantage for structured content.

The reason is simple: pages with headings and lists are already chunked by the author. Each list item is usually one atomic claim. Each H2 signals a topic shift that creates a natural chunk boundary. Pages with nothing but long prose force the algorithm to segment arbitrary text, which is a harder problem with more room for error.

Finding 4: AI Mode and Gemini share only 3.5% of cited domains

AI Mode and Gemini are both Google products, run on the same Gemini LLM family, and both use text fragments. You would expect them to overlap heavily. They do not.

AI Mode cited 5,428 unique domains. Gemini cited 1,876 unique domains. They share just 247 domains in common. Jaccard similarity: 0.035. They are clearly running different retrieval pipelines on different corpus snapshots.

The practical implication is that being visible in one Google AI product does not mean you are visible in the other. You need to optimise for both separately.

Finding 5: only 25.3% of AI citations rank in organic top-10

We compared every cited URL against Google’s organic SERP for the same query. Only 25.3% of AI cited URLs appeared in the organic top 10. ChatGPT was the most independent at 6.5%. Perplexity was the most aligned with traditional ranking at 43.5%.

Translation: you can be invisible in classic SEO and still be cited heavily by AI engines. Conversely, ranking #1 organically is no longer a guarantee of citation. The two pipelines have meaningfully diverged.

Finding 6: AI Mode tolerates older content

We extracted publication dates from 548 cited pages. The median age of cited content was 2.2 years. Over half (52.7%) was more than 2 years old. The single largest year was 2024 (26.5%) but content from 2013 and earlier still accounted for 7.1% of citations.

This contradicts the popular advice that AI search rewards freshness. It does for ChatGPT (which cites content ~458 days newer than organic per Ahrefs) and Perplexity. For AI Mode, evergreen authority matters more than recency.

Finding 7: readability is bimodal, not normal

Using textstat we scored Flesch Reading Ease across 11,411 cited sentences. Instead of a normal bell curve, the distribution split into two peaks: 23.5% Very Easy (Flesch 90 to 100) and 21.3% Very Confusing (Flesch under 30). The middle ground, Fairly Difficult content, was the least cited tier at just 5.0%.

Translation: Google cites both plain-language explainers and dense technical content at roughly equal rates. The losers are corporate jargon, hedged language, and mushy middle-of-the-road prose.

Try the research yourself

Everything is open source. Clone github.com/danishashko/grounding-citation-analysis, add your API key, and run the pipeline against your own queries. Estimated cost is $50 to $120 for 520 queries across all 6 platforms.

If you want to read the full article with all 9 findings, the platform-by-platform optimisation playbook, the chunking analysis, and the cheatsheet, the long-form version is in the repo’s article folder.