Telling an assistant which sentence to read.

When a voice assistant answers a question from a webpage, it has to choose what to say. Without guidance it takes a guess: the first paragraph, a meta description, sometimes a navigation label. You've seen the result: an answer that's technically from the right page and tells you nothing.

There's a standard way to mark the parts of a page written to be heard, so the assistant chooses from sentences you picked instead of guessing at your layout.

It costs almost nothing to add. The advantage is that nobody bothers.

What makes a sentence speakable

• It answers in the first clause. Not context, then answer. Answer, then context.
• It survives losing the page. No "as mentioned above", no "this", nothing that only makes sense with the layout in front of you.
• It's a full sentence. Headings and bullet points read badly aloud.
• It contains the thing somebody asked. If the question is what it costs, the number is in the sentence itself.

Where we use it

Across our largest build, and on every answer in our own FAQ. Those answers were written to be read aloud first, which is why they open with the answer and put the caveats after. Read one out loud and it holds up. That's the test.

The limits

Nobody publishes how much weight assistants give this, and it isn't a lever you can pull to manufacture a result. What it does reliably is remove any doubt about which sentence you'd want quoted. The hard part was never the markup. It's writing 4 sentences that survive being read aloud by a machine to somebody who can't see your page.