Voice-to-Text Plus AI Is Rewriting Product Discovery

For most of the internet's history, finding a product has worked the same way: type a few words into a search box, then scroll. That's changing. Voice-to-text finally works well, and the AI on the other end understands what people say, so more of them are just saying what they want. Shopping starts to look less like a search and more like a conversation.

I see this shift from two sides: in my own habits, since I now dictate almost everything I write, and in the Amazon search data I work in every day. This week's Signals is about how people find products when they talk to AI instead of typing, and what brands can do about it.

For me, it started with a dictation app.

How I started talking to my computer#

In January I started using Wispr Flow. I talk, and it types what I say into whatever I'm using: email, Slack, ChatGPT, Claude, a search box. Within a few weeks I was using it for almost everything.

Wispr sends a monthly recap. September's said I'd spoken 151,786 words at 167 words a minute, across 2,616 separate sessions, which put me in the top 0.2% of its users. That's about 15 hours of talking to my computer in one month.

A different number stood out to me. Divide the words by the sessions and it comes to about 58 words every time I started talking. A typed search is a handful of words.

Speed isn't why I kept doing it. I've typed since middle school and I write code, so I was never slow at a keyboard. What kept me talking is that it changes what I say. When I talk to AI, I explain things the way I would to a friend: the background, what I've already tried, what I haven't figured out yet. I don't stop to fix the transcription, because the AI understands what I meant even when a word comes out wrong.

Shoppers are starting to do the same thing, and it changes what they ask for.

We type in keywords. We talk in paragraphs.#

Say you need a standing desk. Typed, the search is probably "standing desk 60 inch." Said out loud to an AI, it sounds more like this: "I work from a small home office, I want a standing desk that fits two monitors, a laptop and a mic, it can't wobble when I type, and I'd like to stay under about $700." That's 35 words instead of four. It tells the AI where the desk is going, what has to fit on it, how you'll use it, what would rule a desk out and what you can spend.

The words change too. People type the words they think a search engine wants, like "60 inch" or "dual motor." They talk the way they'd talk to a salesperson: "can't wobble," "a laptop and a mic," "two monitors." If your product page only lists specs, the AI has to connect "can't wobble" to your frame on its own, and it leans on reviews and other sites you don't control to do it.

Brand names are a risk of their own. Voice-to-text writes down what it hears, so an unusual brand name can come out as an ordinary word. I've told my AI tools that I dictate everything so they read past those slips, but most shoppers haven't. If someone says your brand name and the transcription gets it wrong, the AI goes looking for the wrong thing.

Last week I wrote about buying through an AI shopping agent (I Never Saw the Website. I Trusted the Agent.). I gave it long, spoken requests, and it used the detail in them. A few years ago, none of this would have worked.

Why this works now when Alexa didn't#

We've been promised this before. Dragon NaturallySpeaking brought dictation to the PC in 1997 (Nuance via PCR). The Boston Globe's reviewer found it still got a word wrong "in roughly every other sentence" (Boston Globe, 1997). Amazon launched the Echo in 2014 (Amazon), hoping people would shop by talking to Alexa: say what you want, and it shows up at your house. In 2018, The Information reported that only about 2% of Alexa owners had bought anything by voice that year (Gizmodo).

The problem was the software on the other end. The original Alexa handled what Amazon now calls "simple, formulaic requests" (About Amazon), so you had to know exactly what you wanted and say it the right way. Most Alexa devices had no screen, so there was nothing to compare. That works for reordering paper towels. It doesn't work when you need help choosing.

Large language models, the AI behind ChatGPT, Claude and Gemini, changed that. You can ramble for two minutes, change your mind halfway through and use the wrong word, and they still work out what you need, ask a follow-up question and show you options. The long, messy request that used to break a voice assistant is now the most useful thing you can give one.

Now the big platforms are building AI voice into everything:

• Phones. In May, Google announced Rambler, a Gemini-powered dictation feature for Gboard (Android's keyboard). It starts on the newest Pixel and Galaxy phones and "automatically removes filler words, formats grammar and punctuation" (Google; TechCrunch). In June, Apple announced new AI dictation in iOS 27, starting on its newest devices. It works in any app and handles "capitalization, punctuation, and formatting" as you speak (Apple).
• PCs. Microsoft says that "when people use voice, they engage with Copilot twice as much as when they use text" (Microsoft).
• Cars. In April, GM started bringing Gemini to about 4 million US vehicles so drivers can "speak naturally without memorizing commands" (GM).
• The kitchen. Amazon rebuilt Alexa on AI as Alexa+. In January it said Alexa+ users were having "twice the conversations, three times the purchases" (About Amazon). In March, an Amazon executive said that comparison was against the old Alexa (Modern Retail).

The people using Alexa+ so far are early adopters, so that last number probably looks better than it will at scale. Still, the same product that couldn't get people to buy by voice in 2018 now reports more purchases because it understands what people say.

How far along we are#

It's still early. Nobody publishes a clean number for how much people talk versus type. The closest I've found is a Twilio survey of 2,000 UK adults in August, where 22% said they now talk more than they type when using technology (Twilio). And in February, Google said nearly one in six searches in AI Mode, its conversational version of Google Search, "are now non-text, using voice or images," and that AI Mode searches run three times longer than traditional ones (Alphabet).

So most people still type. But requests are already getting longer, and an Amazon brand can see that in its own data.

Where Amazon brands can see it first#

If you run Sponsored Products, Amazon's pay-per-click product ads, your search term report shows the exact searches that triggered your ads, with the clicks and sales from each. If you're enrolled in Amazon Brand Registry, its Brand Analytics dashboard shows the searches that lead shoppers to your products.

To see the shift in your own account, work out what share of your search terms are six words or longer, and what share read like a question or a sentence. Then track both numbers every quarter. Amazon only keeps a couple of months of search term history in the ads console, so if nobody has been saving the reports, start saving them every month now. If you'd rather not do that by hand, it's one of the things we built Portal for. It keeps your Amazon ads history, search terms included, and connects it to Claude or ChatGPT, so you can ask which long searches are growing instead of building the spreadsheet yourself.

Then read the long ones slowly, the way you'd read customer interviews. They show the situations people are shopping for, in their own words, and they usually point straight at questions your listing doesn't answer yet.

One limit: this report only shows what people type or say into Amazon's own search box. Alexa for Shopping, the AI assistant in the Amazon app, used to be called Rufus. A long request to it, or to ChatGPT, may never show up in this report, so the report undercounts the shift. Even so, it's the earliest signal you can see in your own data.

What brands should do now#

Use the way customers talk as your keyword research. Sales and support calls, chat logs, reviews and those long search terms all show how people describe their problem out loud. Write those situations into your listings and pages in plain language, and say what the product doesn't do. A spoken request almost always includes a deal-breaker ("it can't wobble," "it has to fit under a 26-inch counter"), and a product that states its limits is one an AI can confidently recommend, or confidently rule out.

Keep a person in the loop. It's tempting to have AI write thousands of words of use-case filler. Google's spam policy specifically calls out "using generative AI tools or other similar tools to generate many pages without adding value for users" (Google). Draft with AI if you want, but have someone with real product knowledge check every claim. That's how we run content at Texin.

Check that your name survives dictation. Say your brand name and your best seller's name into your phone's dictation and see what comes out. Then describe a real buying problem in your category, out loud and in full, to Alexa for Shopping, ChatGPT and Gemini. See whether your product comes up, and how it's described.

Four predictions for the next two years#

1. Talking becomes the normal way to ask AI for things. As Apple and Google build AI dictation into their phone keyboards, talking to AI stops being a power-user habit. By the end of 2027, I expect most people who use AI regularly to talk to it for at least some of what they ask. Keyboards will stick around for editing things other people will read, like contracts and listings.

2. Search term reports get longer. Brands will see more six-plus-word, sentence-shaped searches in their search and ad reports. The ones that read those like customer interviews will learn faster than the ones that only mine them for keywords.

3. Voice shopping comes back, with a screen nearby. The more complicated the purchase, the more an AI helps, and talking is the fastest way to give it the detail it needs. Bigger purchases will start out loud, in the car, in the kitchen or on a walk, and finish on a screen where the shopper can compare.

4. B2B buyers talk to an AI before they talk to you. Hiring a service or choosing software is about as complicated as buying gets. Buyers will talk through their whole situation with an AI (team size, the tools they use, what went wrong with the last vendor) and have a shortlist before they ever fill out a contact form. Gartner predicts that "by 2028, 90% of B2B buying will be AI agent intermediated" (Gartner).

---

Texin.ai runs Amazon for brands and builds the content that helps them get found in AI search and by shopping agents. Talk to us.