Voice search got a decade of blog posts and one honest sentence never made it into them: voice was a delivery format, not a search engine.
You asked a speaker a question. It ran a normal search, took the result it was most confident in, and read it aloud. The index was the same index. The ranking was the same ranking. What changed was that ten results collapsed into one spoken answer, because you cannot read out a list to someone driving.
That collapse is the part worth carrying forward, because generative engines do the same thing for a completely different reason.
What voice search actually was
Google’s speakable structured data documentation is a good window into how narrow the real feature set was. Speakable marks the sections of an article best suited to text-to-speech playback. Google says the feature “is in beta and subject to change”, and that it “works for users in the U.S. that have Google Home devices set to English, and publishers that publish content in English”. When the Assistant reads a speakable section, it attributes the source and sends the full article URL to the user’s phone.
That is the reality behind years of voice search strategy: a beta news feature, one country, one language, one device family. Meanwhile the actual mechanism most voice answers used was the featured snippet, which was a search feature that happened to be readable aloud.
Where the old advice was right
Some of it was genuinely good, and it transfers to conversational AI almost intact.
Write the way people ask. Spoken queries are longer, more natural and more often phrased as questions. Retrieval today is embedding-based, so semantic match beats keyword match. Content written in natural question-and-answer form retrieves well in both worlds. This was right in 2018 and it is more right now.
Answer directly, early, in full. Voice advice said to put a complete answer in the first two sentences under the heading, because that is the chunk that gets read. Generative engines ground on passages, so the same structure gets you into the answer. Same tactic, different mechanism, still works.
Cover the long tail. Conversational phrasing generates enormous query variety. Google states that AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics. That is voice’s long-tail problem industrialised. See query fan-out.
Local and operational facts matter. A large share of assistant queries were local and practical. Hours, address, availability, price. Those facts still need to be on your site in plain text, because they are exactly what a grounded answer needs.
Where it was wrong, or is now
Position zero is gone as a target. Voice optimisation was largely featured-snippet optimisation. A generative answer is assembled from several sources, so there is no single winner slot to capture. Being the one snippet has been replaced by being one of the sources, which is a different and less binary goal.
“Optimise for voice” as a separate workstream never made sense. It was ordinary search with a speaker attached. Anyone who sold it as a discipline was selling formatting advice.
The statistics were never real. The famous voice search percentages circulate endlessly and trace back to forecasts, not measurements. Google does not publish voice query share. Neither do the assistant vendors. I have not been able to find a primary source for any of the commonly quoted numbers, so I do not use them and would be sceptical of anyone who does.
Speakable is not an AI visibility play. Read the eligibility restrictions above. If you are not a US English news publisher, it does nothing for you.
What conversational AI changed underneath
The input looks similar. The pipeline is not.
Classic voice: one query, one index lookup, one result, read aloud.
Conversational AI: query expansion into several machine queries, retrieval across them, grounding the generated text in what came back, then synthesis into prose that may cite nobody. It is a generative engine, and the answer is written rather than selected.
Two consequences follow. First, the follow-up question is now part of the interaction, so a session is a conversation with context rather than a series of independent queries. Second, the assistant can compose an answer that no single page on the web contains, which is not something a text-to-speech reader could ever do.
What to actually do
Keep the question-and-answer structure. It was the best part of voice-era advice and it maps directly onto passage-level retrieval.
Put your operational facts in HTML. Hours, coverage, pricing, limits, regions. Not in an image, not behind a script. Most AI crawlers do not document JavaScript execution, so check what they can see.
Stop optimising for the snippet and start being retrievable. Confirm the search-index crawlers can reach you, using the AI crawler directory to identify which agent does which job.
Ignore voice-specific tooling. There is no meaningful voice channel to measure separately. Measure AI visibility instead, which is the surface that actually absorbed the behaviour.
The short version
Voice search was speech attached to ordinary search. Conversational AI is a different retrieval and generation pipeline that happens to accept the same phrasing.
Keep the writing habits: natural questions, direct answers, facts in plain text. Drop the position-zero framing and the statistics, neither of which survives contact with how these systems now work.