A year-long study counted 15,699,298 Google AI Mode citations across 148 industries and found that 47.7% of them were scroll-to-text highlights: URLs ending in #:~:text= that point at one specific paragraph instead of at a page. The figures were published on 29 July 2026 in a contributed column on Search Engine Land, written by the content platform Pillarbase from its own citation dataset. The column's argument is that the unit AI Mode works with is the passage, not the URL.
The dataset belongs to a vendor that sells AI-citation tracking
Pillarbase published the study as a contributed column on Search Engine Land on 29 July 2026, and Pillarbase sells a content platform plus a paid AI citation readiness report built on the same kind of tracking the study describes. This is the vendor's data and the vendor's reading of it, not independent reporting by Search Engine Land. That does not make the numbers wrong, but nobody outside the company can audit the sample, and each recommendation points at work the company also sells.
47.7% of AI Mode citations pointed at a paragraph rather than a page
Pillarbase reported that 47.7% of the 15,699,298 citations it counted were scroll-to-text highlights, and that those citations resolved to 4.6 million unique highlighted passages across 2.7 million pages. The consequence for anyone doing AI SEO work is a change of unit: if roughly half of citations name a paragraph, the thing to audit is a paragraph. The column's advice follows: audit a page passage by passage, not as one block.
80.9% of highlighted passages were cited once, and about 2,300 were reused 61 or more times
The distribution is heavily skewed. 80.9% of the 4.6 million passages were cited a single time, while roughly 2,300 passages were reused 61 or more times. The most-recycled passage in the dataset was cited 661 times, and that one passage was reused across 483 distinct queries. One promotional-products retailer's best-performing paragraph answered 221 different queries on its own. A small number of paragraphs carry an outsized share of the answering, which is why the column's goal is reusable answers rather than more pages.
The median highlighted passage was 117 words and led with the answer
The median highlighted passage in the Pillarbase data was 117 words, which is a complete multi-sentence answer rather than a one-line snippet. 80% of extracted passages put the answer in the first sentence, rising to 81% on the author's benchmark of the web's most-recycled passages, and the column calls that the most consistent pattern in the whole study. The recommended paragraph length in the column is 75 to 150 words.
Pillarbase describes four traits shared by the passages that got extracted:
- They are full paragraphs, not clipped snippets.
- They lead with the answer.
- They are self-contained and readable without the surrounding page.
- They sit under a heading that matches the question a user would ask.
The column also advises favouring question-and-answer or how-to structure, since both formats naturally produce paragraphs shaped like answers.
Pages with 21 or more highlighted passages had a median rank of #1, and the author says that is correlation
Pillarbase reported that pages with 1 to 4 highlighted passages had a median organic rank of 11, while pages with 21 or more had a median organic rank of #1, with 67% of them ranking first outright. Among passages reused 100 or more times, 76% came from pages ranking #1.
The author disclaims causation directly, writing that this is correlation and not a proven mechanism, and concluding only that the passages AI Mode relies on overwhelmingly live on pages that already win in classic organic search. Adding well-formed paragraphs to a page sitting on position 40 is not a documented route to position 1, and the study does not claim it is. Passage quality is how an already-competitive page gets quoted, which keeps ordinary SEO work upstream of the whole exercise.
The headline figures in one place
These are the numbers Pillarbase reported from its year of AI Mode citation data.
| What was measured | Reported figure |
|---|---|
| AI Mode citations counted, across 148 industries | 15,699,298 |
| Citations that were scroll-to-text highlights | 47.7% |
| Unique highlighted passages, and pages behind them | 4.6 million passages on 2.7 million pages |
| Passages cited only once | 80.9% |
| Median length of a highlighted passage | 117 words |
The rank association reported in the same study, which the author labels correlation:
| Highlighted passages on the page | Median organic rank reported |
|---|---|
| 1 to 4 passages | Rank 11 |
| 21 or more passages | Rank 1, with 67% ranking first outright |
| Passages reused 100+ times | 76% came from pages ranking #1 |
What this means for marketers in Thailand
The column does not break its data out by country or by language, so the 148 industries it covers cannot be read as a statement about Thai-language results. What travels is the editing habit, not the percentages. Rewriting a service page so each heading matches a question a customer would type, and the paragraph under it answers in its first sentence at around 75 to 150 words, improves the page for human readers whether or not an engine ever quotes it.
Whether Thai-language passages get highlighted at the same 47.7% rate is speculation, and should be labelled as such: the study reports no split by language. For bilingual sites the defensible position is to apply the structure in both locales and measure your own citations rather than assume the global figures transfer. That measurement work is what generative engine optimisation covers, and it starts with knowing which of your paragraphs, not which of your pages, an engine is lifting.
Frequently asked questions
Is this study independent research?
No. The study was written and published by Pillarbase, a content platform that sells AI-citation tracking and a paid AI citation readiness report, and it appeared on Search Engine Land as a contributed column on 29 July 2026.
Does adding more highlighted passages make a page rank #1?
The study does not claim that, and the author explicitly calls the finding correlation rather than a proven mechanism. Pages with 21 or more highlighted passages had a median rank of #1, but the conclusion the author draws is that AI Mode leans on passages from pages that already rank well.
How long should a paragraph be if I want it quoted?
The column recommends 75 to 150 words, and reports a median highlighted passage length of 117 words. The pattern that mattered most in the data was position rather than length: 80% of extracted passages put the answer in the first sentence.
Do I need to add #:~:text= fragments to my own URLs?
No. Those fragments are what the study observed in the citation URLs recorded from AI Mode results, not markup a publisher adds to a page. The actionable part is how the paragraph is written and what heading sits above it.
Does the study say anything about Thailand or Thai-language pages?
No. The column reports 148 industries and does not state which markets or languages the citations came from, so any read-across to Thai-language results is an assumption rather than a finding.
Which paragraphs on your own site get lifted into AI answers is measurable rather than a guess. Our GEO and AI SEO work starts there.
