The single-answer surface changes the economics
On a results page, position five still receives clicks. On a spoken answer, position five receives nothing at all. That winner-take-all property means voice and assistant optimization has a different risk profile: effort spent getting from tenth to third produces zero measurable return, while getting from second to first produces the entire prize.
The practical implication is selectivity. Pick the questions where you have a genuine chance - questions where you hold a strong ranking, own local relevance, or possess data nobody else has - and structure those to perfection rather than spreading spoken-answer formatting across every page.
Local queries deserve particular attention because the competitive set is small and the deciding data is factual rather than editorial. Correct hours, address, and service area frequently beat better content, because the assistant is answering a factual question and needs a source it can trust.
Writing that sounds right aloud
Text optimized for scanning and text optimized for speech diverge in specific ways. Parentheses, slashes, symbols, and nested clauses all read fine visually and become unintelligible when spoken. Numbers with units, currency symbols, and abbreviations need to be written the way they would be said if you want the reading to be smooth.
Sentence rhythm matters too. Assistants read at a steady pace with minimal pausing, so a forty-word sentence with three subordinate clauses arrives as an unbroken wall. One idea per sentence, two to three sentences per answer, is the shape that survives text-to-speech intact.
The cheapest quality control is literal: read the block out loud before publishing. Every problem in this category is audible immediately, and none of them are visible on screen.