Voice Search

What happens to a spoken question

The device transcribes the speech to text and hands that text to a search engine or an AI assistant. Spoken questions tend to be longer and phrased as full sentences, the way a person would ask a friend, so the query looks more like a question than a keyword. The engine then finds a passage that answers it and the device speaks that passage, usually naming the source it came from. There is no scrolling, and there is rarely a second result.

That makes voice search a narrow case of the problem citability describes: the winning page holds a self-contained passage that answers the question in one place. A heading phrased as the question with a plain answer beneath it is the shape a text-to-speech system can read aloud without stumbling.

Speakable markup

Schema.org defines a speakable property on Article and WebPage that points at the sections of a page best suited to being read aloud, located by a CSS selector or an XPath expression. Google’s documentation describes its support as a beta limited to English-language news content from publishers in the United States, so for most businesses the markup is a signal to know about rather than one to rely on. The passage itself does the work.

Voice search and AI assistants

The two have converged. An AI assistant asked a question by voice does what it does when asked by text: it searches, fetches a handful of pages, and writes one answer from what it found. The difference is only in the output channel. A business that an assistant names in its written answers is the one it will name aloud, and the same three conditions apply: the crawler received the page, the page contained a passage worth lifting, and the business behind it could be identified.