Why it matters
Take a spec page where each fitting's dimensions sit in a labeled table. Thread size, body length, working pressure, and material each retrieve as one clean chunk, with the part number attached. Bury the same data in a prose paragraph and a retriever splits it mid-fact. The engine returns "3/8 inch" without the part it belongs to, or a pressure rating detached from its fitting. Wall-of-text pages chunk badly. Tables and short labeled blocks chunk well, because the split lands on a boundary that keeps the fact whole.
What chunks well vs badly
Chunks well: spec tables with one attribute per row, a Q&A block that answers one question completely, a cross-reference list pairing an OEM number with your SKU. Chunks badly: a long compatibility paragraph where the qualifying condition sits three sentences from the part it limits, or a feature wall that names a value in one place and its unit in another. The test is simple. Can a single passage stand alone and still be true?
In practice
A distributor with 40,000 hydraulic fittings moves the load rating, seal material, and temperature range out of the marketing paragraph and into a labeled table, one fact per row, part number repeated in the heading. Now each retrieved chunk carries enough context to be cited correctly. The buyer prompt "what is the max pressure on part 6801-6-6" returns the right row instead of a number sheared away from the part it describes.