The text-fragment phenomenon in Google AI Mode.

Nearly half of all AI Mode citations are highlighted passages — not plain links.

47.7% of 15.7M citations carry a #:~:text= fragment. Scroll-to-text highlighting is a default behavior of the engine, observed across all 148 industries. Prevalence varies by pillar — some show ~0% (a capture-timing difference worth noting), others exceed 60% — but the portfolio-wide figure is the headline.

Most pages get one highlight. A rare few become highlight magnets.

Across 2,738,150 cited pages, the median number of distinct variants per page is 1. But the distribution has a long tail: a top ~1% of pages carry 12 or more distinct highlighted passages. These "highlight magnets" are the pages worth studying.

A high-variant page (top ~1%) carries 12+ distinct highlighted passages — 415 pages carry 61 or more. The single most-highlighted page in the census holds 252 distinct variants.

An 80/20 long tail — with a recycled elite.

Of 4,612,626 distinct variants, 80.9% are used exactly once. But a tiny elite — about 2,300 passages reused 61 times or more — gets recycled relentlessly. The single most-reused passage in the portfolio was cited 661 times across one pillar's query set.

One passage answers dozens of different questions.

The recycled elite isn't one query repeated — it's one passage serving many distinct queries. A single well-formed paragraph becomes Google's reusable answer to an entire cluster of phrasings.

Authoritative census counts from the full variant table. Per-variant reuse counts in the client-flagged content analysis below can undercount a URL's true reuse (e.g. the HOA passage reads as 54× there vs 158× in the full census); directional trends hold, exact tier counts are approximate.

Variant accumulation is strongly associated with #1 rankings.

We joined each cited page's base URL to its organic Google rank. The more a passage is reused — and the more variants a page accumulates — the more likely that page ranks #1. These relationships are correlational (cross-sectional joins), not a demonstrated cause.¹

¹  Correlation, not causation. These are cross-sectional joins of cited URLs to their organic rank. They show that variant accumulation and recycling travel with ranking dominance — not that one causes the other. Ranking and recyclability most likely reinforce each other.

What the recycled passages have in common.

To study the passages themselves, we reconstructed the full highlighted text — because ~98% of fragments store only the first and last few words. We fetched the live pages and extracted the literal span: 528 of 650 client passages (81%), plus a 490-variant third-party "elite" benchmark of what wins across the open web. Each passage was independently classified by a 41-agent LLM analyst panel.

F1 · Extracted passages are full paragraphs, not snippets

Reconstructed client passages run a median of 117 words (mean 134; p90 197; max 639) — complete, multi-sentence answers, far larger than the truncated anchor or SERP snippet suggests.

F2 · The universal trait: answer-first and self-contained

Regardless of source, AI Mode overwhelmingly highlights passages that lead with the answer and make sense out of context.

F3 · The reuse signal: question-led passages get recycled far more

Within the client cohort, trait rates climb with reuse. High-reuse passages (≥6×) are question-led 48% of the time vs. 22% for single-use, and Q&A-format 33% vs 19%. Narrative passages get recycled less; one-off statistics essentially never recycle.²

²  Directional, small-n. The monotonic reuse→question-led trend rests on 28 and 12 passages in the top tiers, computed within the client cohort only (where reuse spans 1×–54× naturally). The direction is corroborated by the format split, but the top-tier percentages are directional, not precise. Pooling the ≥100×-selected third-party benchmark inverts the signal — purely a sampling artifact.

F4 · Pillarbase-shaped content skews toward the recyclable formats

Client pages over-index on exactly the question-led and how-to formats that F3 shows are most recyclable — while the open-web elite leans on definitions and statistics.

F5 · Targeted queries trigger citations

30.9% of client variants (201 of 650) were triggered by an AI Mode query matching a keyword we explicitly targeted — a lower bound, since AI Mode heavily reformulates queries. The fragment-cited client pages carry 2022–2024 publish dates: these are clients' existing question-formatted pages, so the finding is about content structure, not authorship.