In July 2017 we published a post on this page arguing that voice was about to reshape search. It cited a statistic you’ve probably seen a hundred times: according to comScore, half of all searches would be voice searches by 2020.
That didn’t happen. The statistic was also wrong and we repeated it anyway, along with most of the industry.
We’re leaving this URL where it is and rewriting what’s on it, because the way that prediction failed is more useful right now than anything the original post had to say about Alexa.
comScore never said it. When researchers eventually emailed the company to ask, they confirmed the figure wasn’t attributable to them.
The real source is Andrew Ng, then Chief Scientist at Baidu, in a 2014 interview with Fast Company. What he actually said was that roughly 10% of Baidu’s queries were already voice, and that within five years he expected at least half of all searches to be conducted through either speech or images.
So the citation was wrong about who said it, wrong about where it applied, and wrong about what it covered. It described Baidu, in China, where voice adoption was already years ahead of the US. It included image search. And it was one executive’s projection about his own company’s traffic, not a market forecast.
By the time it reached our blog and many others, it had become “comScore says 50% of all searches will be voice by 2020,” which is a completely different claim with a much more authoritative name attached.
Voice didn’t replace search. It became a command interface.
Today, people talk to devices constantly, and almost none of it is search in any sense a marketer would recognize. Timers, music, lights, thermostats, shopping list items, weather. The queries that do resolve like search skew heavily toward quick facts and directions, which are precisely the queries that were already resolving without a click on a results page.
There’s still no reliable public figure for voice as a share of search, and that absence tells you something. Neither Google nor Bing reports voice as a distinct query category, and they never built the segmentation the 2017 post was waiting for. Ten years on, you still can’t isolate voice traffic in a campaign. The category we were all preparing to optimize for never became a category the platforms bothered to measure.
Meanwhile every assistant that post named has been killed or rebuilt. Microsoft discontinued Cortana’s consumer assistant. Google is switching off the classic Assistant on mobile as of September 2026, handing phones over to Gemini. Apple shipped a rewritten Siri in June 2026 that runs on a Google Gemini backend with a ChatGPT handoff. Amazon opened Alexa+ to all US customers in February 2026.
Four assistants, four outcomes, none of them the outcome anyone was projecting in 2017.
The underlying intuition was sound. People really did stop typing three keywords into a box. Queries really did get longer, more conversational, more specific. The behavioral shift the 2017 post described is happening right now, at scale.
It just arrived through a text box instead of a microphone.
What changed search wasn’t people speaking to their phones. It was large language models giving people a reason to write full questions and expect full answers. The interface that finally broke keyword search was a chat window, which is about as low-tech as an interface gets.
Voice is now folding back in, but as a way to talk to a model rather than as a way to search. That’s a different product with different economics, and treating it as the fulfilment of the 2017 prediction would be its own kind of revisionism.
The AI search commentary you’re reading today has the same shape as the voice search commentary from 2017. Confident percentages, attributed vaguely, cited from each other rather than from source, describing an end state five years out.
Some of it will turn out right. Most of the specific numbers won’t, and the ones that fail will fail the same way: correct about direction, wrong about mechanism, wildly wrong about timing.
A few habits that would have saved us in 2017, and that we now apply to our own work:
Trace every statistic to its origin before repeating it. Not the article that cited it, the source. If the trail goes cold at a paywall or a dead link, the number doesn’t go in the deck.
Check the scope. The Ng prediction was about one search engine in one country. Plenty of AI search figures circulating now are single-vertical studies being quoted as universal.
Separate the behavioral claim from the channel claim. “People will ask longer, more conversational questions” was right in 2017 and is right now. “Therefore optimize for voice” was the part that wasted everyone’s time.
Be suspicious of any prediction that requires a specific interface to win. The behavior shows up somewhere. Which surface it lands on is much harder to call, and it’s usually the boring one.
The same thing we told you in 2017: optimize for the question, not the device.
Content that answers a specific question clearly, in a structure a machine can extract, works when the question is typed, spoken, or passed to a model. That was true for featured snippets, it’s true for AI Overviews, and it’ll be true for whatever comes after. It’s the one recommendation from the original post that survived, and it survived because it wasn’t tied to a channel.
Our guide to AI search content optimization covers how we approach that now, including a running section on what we’ve since got wrong. The off-page guide covers the entity and citation side, and we’ve written separately about what we’ve learned running this in practice.
If you want to talk about where your content actually stands in AI answers, get in touch. We’ll show you the data before we tell you what it means.