For most of the last decade, voice assistants were a dead end for brands. They answered weather and timers, read one snippet aloud, and otherwise said "here is what I found on the web." That era is ending. Apple routes harder Siri questions to a large model. Amazon rebuilt Alexa on one. Google's assistant on Android is Gemini. The voice interface is now a front end to the same answer engines you see in chat.
The difference is the output. A chat answer can list five options with links. A spoken answer gives one, maybe two, and there is no link to click. If you are not the one name the assistant says, you were not in the answer at all.
How a voice answer is built
The pipeline is roughly the same as a chat answer: transcribe the question, decide whether to retrieve, pull from the web or a knowledge graph, and generate. Then the answer is compressed to something that can be spoken in a few seconds.
That compression is where brands get cut. A model that would write "three good options are A, B, and C, depending on your budget" tends to say "A is a popular choice." The entity that survives is the one the model was most confident about, which usually means the one with the most consistent, corroborated facts across the sources it retrieved.
Why consistency matters more here
In a written answer, a small inconsistency in your facts might cost you a mention in the second paragraph. In a spoken answer there is no second paragraph. When the model has to pick one name, it tends to pick the one it can describe without hedging.
Three things help most:
A clear one-sentence description that appears everywhere. Homepage, Organization schema, llms.txt, LinkedIn, G2, Wikidata if you have an entry. Same sentence, same facts. A model that reads the same claim in six places speaks it with confidence.
Correct local and operational facts. Voice queries skew practical: hours, location, phone, whether you deliver. If your Google Business Profile says one thing and your site another, the assistant may refuse to commit, or pick the wrong one.
An unambiguous name. If your brand name is a common word or shared with another company, make the disambiguating facts prominent: category, location, and what you are not.
The query shapes to prepare for
Voice questions are longer and more conversational than typed ones. "Best CRM for small business" becomes "what's a good CRM for a small landscaping company that does not want to pay much." Your content needs to contain the answer to the long version, which means pages that address specific situations rather than generic category pages.
Local intent is heavier. "Near me" and "open now" are a large share of voice queries. For any business with a physical presence, the operational data in your schema and business profiles is the content that gets spoken.
Follow-up questions are common. "Who makes that?" "How much is it?" Each one is a new retrieval, and the facts need to be findable for each.
What you can measure
Voice is harder to track than chat, because there is no results page to screenshot. The practical approach is to ask the same questions through the voice and chat interfaces of the same provider and compare. In our testing the chat answer is a reasonable proxy, with the list trimmed to one or two names. If you are consistently in the top two in chat, you tend to be the spoken answer. If you are fifth, you are not.
At AEO GEO Labs share of answer is measured across the chat engines weekly, and for clients where voice matters, mostly local and consumer businesses, we add a monthly voice spot check with the top ten questions read aloud to each assistant and the answers logged.
Where to start
Pick the ten questions a customer would actually say out loud to find a business like yours. Ask them to Siri, to Gemini, and to Alexa. Write down what each one says, including the cases where it declines to name anyone.
Then look at the sources. If the assistant named a competitor, find out where that competitor's facts live and whether yours live in the same places. The gap is almost always in the sources, not the model. For the full picture with sources mapped and fixes prioritized, AEO GEO Labs does that as part of a visibility audit, and the free answer-ready checker will tell you whether your key pages contain a sentence a model could say aloud.