The advice is standard in SEO circles: do not open a paragraph with 'this', 'he' or 'therefore'. A search-driven model lifts a passage out of the page, the reasoning goes, so the passage has to stand on its own, and a pronoun at the front drags in context that is no longer there.
Tested against Perplexity answers, the rule falls apart. Across 60 answers, paragraphs the system cited and paragraphs it skipped opened with those words equally rarely. The opening did not mark the difference.
What the test found
Length did. Cited paragraphs were noticeably longer, and in those 60 answers Perplexity more often took fragments of 30 words or more.
The observation the tip grew from still holds. A search model pulls one or two paragraphs off a page rather than the whole article, so a fragment has to read without the paragraph above it, the heading over it or the footnote below. The mistake was the next step: stripping pronouns and connectives instead of building passages that carry a full thought.
What to do with it
Short paragraphs are often half a thought. One introduces, the next develops, and pulled out on its own the fragment breaks. A paragraph of 30 words or more more often closes the idea, with a claim and its explanation. Merging neighbouring short paragraphs on the same point raises the weight of the fragment without changing the sense. Reading each paragraph as though nothing else were on the page shows which ones fail.
Padding a paragraph to reach a word count does not make it self-contained, and headings do not help, because the model takes the paragraph rather than the heading and text together.
The measurement covers Perplexity only. There is no equivalent test for ChatGPT or Alisa AI, and the result should not be carried over untested. 60 answers are a slice, not a sample. Length may also be an effect rather than a cause: a 30-word paragraph is often simply better worked out, with a claim and an explanation behind it, in which case what decides is the quality of the thinking.
