Skip to main content
Back to AI Commerce Lab
AI·April 2025·9 min read

The Rise of Voice Search in E-commerce: How It Impacts Sales

Voice search never became the shopping interface 2019-era forecasts promised. What actually shipped: a narrow set of repeat-purchase use cases, a developer platform Google walked back in 2023, and a structured-data spec that's now mostly feeding AI answer engines instead of driving voice sales.

What the hype predicted

The 2019-2021 wave of voice commerce forecasts described a near future where shoppers spoke orders to Alexa and Google Assistant the way they now tap through an app, with billions of dollars in annual voice-driven sales by the mid-2020s. Retailers were told to optimize for "conversational keywords" and build voice-first shopping experiences.

None of the original forecasts this genre relied on were platform-disclosed figures. They were third-party projections, and neither Amazon nor Google has ever published a number confirming or denying them.

Why the prediction was structurally shaky from the start

Voice has no screen. A shopper comparing five pairs of running shoes by price, color, and review count is doing a browsing task, and a browsing task described entirely through audio is slower and harder than the same task on a screen, not more convenient. The original forecasts treated voice as a faster version of typing, when the actual constraint was never typing speed, it was the total absence of a visual comparison surface.

That structural limit doesn't disappear as voice recognition accuracy improves, which is likely why the platforms themselves narrowed scope over time instead of broadening it.

What Google actually built, then walked back

Conversational Actions: launched, then sunset

Google's third-party developer platform for building custom Assistant experiences, Conversational Actions, launched in 2016 and was deprecated on June 13, 2023, with Google redirecting developers to App Actions, which only integrate with Android apps rather than standalone voice experiences (Google for Developers — Conversational Actions sunset overview). A platform retiring its general-purpose third-party voice framework, mid-decade, is the opposite of what a maturing voice-commerce ecosystem looks like.

Speakable structured data: still in beta, repurposed

The schema.org speakable property, meant to flag content for text-to-speech playback on Assistant devices, remains explicitly labeled beta in Google's own documentation years after its introduction (Google Search Central — Speakable structured data). The specification itself, hosted at schema.org, defines the property generically enough to serve any consumer of structured content, not exclusively voice assistants (schema.org — speakable).

In practice, speakable markup today functions less as a voice-search ranking signal and more as one more structured input AI systems use to select and cite sources, a repurposing nobody was describing back when the property was introduced.

A specification staying in beta for years isn't a paperwork oversight. It's usually the platform's own signal about how much production weight to put on it.

What Amazon actually kept building

Alexa Shopping Actions: narrow, and still a developer preview

Amazon's own developer documentation for Alexa Shopping Kit describes the shopping actions, AMAZON.AddToList and AMAZON.RecommendShoppingProducts, as a developer preview that "might change as Amazon receives feedback and iterates," and every skill using them must explicitly ask the customer before delegating a shopping request to Alexa (Amazon Developer — Alexa Shopping Actions). This is recommendation and list-building, not full voice checkout, and it's explicitly still labeled preview rather than general availability.

Amazon also restricts the feature from child-directed and HIPAA-scoped skills, underscoring how narrow the sanctioned use case actually is compared to the original "order anything by voice" framing.

What Amazon's own analytics reveal about scope

Amazon's developer console reports two metrics for Alexa Shopping Actions: recommendations and conversions (Amazon Developer — Certify your skill with Alexa Shopping Actions). Amazon's own instrumentation, not a third party's, is scoped to recommendation and add-to-list behavior, not a full purchase funnel, which structurally confirms this never became a checkout channel.

Where "voice commerce" actually shows up in 2026

Replenishment is the one use case that stuck

The voice use case that survived is exactly the one the original hype treated as a footnote: reordering something already purchased before, where the product, size, and preference are already known, and voice just triggers a repeat transaction. That's a narrow, high-confidence, low-ambiguity task, which is precisely the profile voice interfaces are actually good at.

Why replenishment specifically, and not general shopping

A reorder has no browsing step, no comparison step, and no ambiguity about which product is meant, since the customer already bought that exact item before. Voice only has to resolve one thing: which previously purchased product the customer means, a closed, small vocabulary problem rather than an open-ended product-discovery problem.

General product search is the opposite: an open vocabulary, an unknown number of matching results, and a decision that usually benefits from a visual comparison. That mismatch, not a recognition-accuracy shortfall, is why replenishment scaled and general voice shopping didn't.

The real conversational shift happened in text, not voice

The interface that actually delivers the "describe what you want in natural language" experience the original voice forecasts imagined turned out to be AI chat, not a smart speaker. Google's own 2025 announcement of agentic checkout and conversational product discovery inside AI Mode and the Gemini app describes shoppers setting preferences like size, color, and budget in natural language, then having the AI complete the purchase (Google — Agentic checkout and AI shopping features). It's the same underlying capability the voice forecasts were reaching for, delivered through search and chat instead of a microphone.

2019 prediction versus what actually shipped

Category2019-era predictionWhat actually shipped by 2026
Voice checkoutFull purchases completed by voice commandNarrow recommend/add-to-list actions, still a developer preview (Amazon)
Third-party voice appsA growing ecosystem of custom voice shopping skillsGoogle's general-purpose framework deprecated in 2023
Voice SEOOptimize content for spoken featured-snippet answersspeakable markup still in beta, now an AI-citation signal
Natural-language product searchDelivered via smart speakersDelivered via AI chat and agentic checkout in search, not voice devices
Repeat/replenishment ordersOne use case among manyThe durable use case; the rest didn't scale the same way

What a minimal speakable implementation looks like

The specification asks for short, well-scoped sections, roughly 20 to 30 seconds of read-aloud content per entry, marked with CSS selectors or XPaths pointing at the actual HTML (Google Search Central — Speakable structured data). It's a small addition on top of existing article markup, not a separate content strategy.

{
  "@context": "https://schema.org",
  "@type": "Article",
  "speakable": {
    "@type": "SpeakableSpecification",
    "cssSelector": [".headline", ".answer-summary"]
  },
  "headline": "..."
}

The lesson this holds for the next hype cycle

Agentic checkout is early enough that it's worth applying the same skepticism to it that this piece just applied to voice. Google's own announcement is a real, shipped feature with named retail partners, which already puts it ahead of where voice commerce forecasts stood in 2019 (Google — Agentic checkout and AI shopping features).

The test that would have caught the voice-commerce overreach early still applies here: watch what the platform actually ships and instruments, not what a research firm projects it might be worth. A feature with named partners and a working demo is a different category of evidence than a market-size forecast, and that distinction is the one the original voice-commerce coverage collapsed.

What to actually build if you sell online

None of this means voice integration work was wasted, or that structured data investment should stop. It means the budget line labeled "voice commerce" should be relabeled around what actually shipped, replenishment and AI-citation signals, rather than kept pointed at a smart-speaker checkout experience that never reached general availability on either major platform.

  • Add speakable markup where it's cheap to add, since it now feeds AI answer and citation systems even without a voice-search payoff.
  • Treat replenishment as a genuine channel: predictable, known items, low ambiguity, worth a real reorder flow.
  • Don't build a standalone voice-first shopping app on a platform-specific framework; check the framework's own deprecation history first.
  • Invest the "conversational commerce" budget in AI chat and agentic checkout integrations, where the actual 2025-2026 platform investment has gone.
  1. Audit whether any existing voice integration depends on a deprecated framework (Conversational Actions is the most common dead dependency to find).
  2. Confirm structured data, including speakable, is implemented correctly regardless of voice-specific payoff, since AI systems now consume it too.
  3. Build or improve a replenishment/reorder flow, the one voice-adjacent use case with real staying power.
  4. Evaluate agentic checkout and AI Mode shopping integrations as the actual successor to the original voice-commerce vision.

FAQ

Why did smart-speaker voice shopping fail to scale while smart-speaker adoption itself grew?

Owning a device that can hear a request isn't the same as that request being a good fit for the device's output, audio only, no screen. General product shopping needs comparison and visual confirmation, which is exactly what a voice-only interface structurally can't provide, independent of how many households own a speaker.

Is voice search completely dead for e-commerce?

No, but it's narrower than predicted. Replenishment and simple recommendation flows have staying power; full voice checkout never reached general availability on Amazon's own platform and lost its general developer framework on Google's (Amazon Developer — Alexa Shopping Actions).

Should I still implement speakable structured data?

Yes, since it's low-cost and now functions as an input to AI citation and answer systems, even though Google's own documentation still labels it beta for its original voice-assistant purpose (Google Search Central — Speakable structured data).

What happened to Google's third-party voice shopping apps?

Google deprecated Conversational Actions, the framework that powered them, on June 13, 2023, shifting developer focus to Android-only App Actions (Google for Developers — Conversational Actions sunset overview).

What replaced the original voice-commerce vision?

AI-driven conversational shopping through chat and search interfaces, including agentic checkout, where an AI completes a purchase on a shopper's behalf after preferences are set in natural language (Google — Agentic checkout and AI shopping features).

Is Alexa Shopping Kit generally available for all sellers?

Amazon's own documentation describes its shopping actions as a developer preview, meaning behavior can still change, and excludes child-directed and HIPAA-scoped skills entirely (Amazon Developer — Alexa Shopping Actions).

What does Amazon's own analytics scope tell us about voice checkout?

The metrics Amazon exposes for Shopping Actions, recommendations and conversions on list-adds, don't include a purchase or checkout metric, which is a structural signal from Amazon's own tooling, not a third-party claim, that voice never became a checkout channel (Amazon Developer — Certify your skill with Alexa Shopping Actions).

References

From the Destm engineering archive. For current work on this topic, start at Solutions or the blog.