How Perplexity Brought Voice Search to Millions Using the Realtime API

The Moment Voice Search Stopped Being a Gimmick

There is a moment that the team at Perplexity used as an internal benchmark during the development of their voice interface: a noisy bar in San Francisco. Someone pulls out their phone, asks Perplexity a question out loud over the chatter, and the answer comes back correctly. If it works there, it works anywhere.

That detail, shared by Perplexity engineer Paul Fryzel in a post on OpenAI's developer blog, says a lot about how seriously the company approached voice as a product surface. They weren't building a demo. They were building something that could survive real conditions background noise, imperfect microphones, mid-sentence pauses at scale.

The result? Perplexity now handles millions of voice sessions every month, built on top of OpenAI's Realtime API (Realtime-1.5). And the lessons they learned along the way matter far beyond one company's engineering blog. They signal a fundamental shift in how information is accessed, how brands get discovered, and how every ai automated marketing agency needs to think about search going forward.

The Scale of the Shift: AI Search Is Not a Trend, It's the New Normal

Before diving into the technical story, it helps to understand the scale of what's happening in AI-driven search. The numbers are genuinely staggering.

Perplexity processed approximately 780 million queries in May 2025 alone, with month-over-month growth running above 20%. The platform has grown from a niche research tool with 2 million users in early 2023 to 45 million active users by mid-2025 a trajectory that represents roughly 800% year-over-year growth. Its valuation reached $20 billion in early 2026 following a $200 million funding round.

Meanwhile, ChatGPT crossed 800 million weekly active users as of early 2026. According to industry data, traditional search volume could drop by 25% by the end of 2026 as chat and voice-based AI platforms capture more discovery intent. The share of search traffic referred from AI platforms is growing rapidly, and the platforms are diversifying Claude, Gemini, and Perplexity have all gained meaningful ground in AI referral traffic throughout 2025 and into 2026.

Voice is one of the main reasons that growth is accelerating. When people can simply speak their question and get a synthesized, accurate answer in return, the friction of searching drops to near zero. Perplexity's integration of the Realtime API is one of the most technically sophisticated attempts to make that experience work at production scale. Understanding how they did it offers a blueprint not just for engineers, but for marketers and brands trying to stay visible in the new search landscape.

What the Open AI Realtime API Actually Does (And Why It Matters)

The Realtime API from OpenAI enables low-latency, bidirectional audio interaction with a large language model. Unlike traditional approaches that require audio to be transcribed, passed to a text model, then converted back to speech, the Realtime API collapses that pipeline. The model processes audio directly, which dramatically reduces latency and allows for more natural conversation.

Think of it this way: when someone asks a traditional voice assistant "what's the best project management tool for a five-person remote team that's already using Slack?" the old pipeline would stitch together three separate systems. The Realtime API turns that into a single, continuous interaction. The response can come back in under two seconds, with context maintained throughout the conversation.

That responsiveness is not just a nice-to-have. It's the difference between an experience that feels like a conversation and one that feels like a command prompt with a voice layer on top. Perplexity, which already had a reputation for fast, cited responses in text format, saw voice as the natural next frontier for their product vision. The Realtime API gave them the technical foundation to build it properly.

The Four Engineering Lessons From Perplexity's Voice Build

What makes Perplexity's public case study unusually valuable is its candor. The team didn't just describe what they built they described what went wrong and what they had to rethink. There are four key lessons, and each one has direct implications for how brands and agencies should think about AI-driven voice experiences.

1. Context Management Requires a Strategy, Not Just a Buffer

One of the first serious challenges the team hit was long-form content. Perplexity wanted users to be able to interact with dense material multi-hour podcast transcripts, for example through voice. The naive approach was to load large chunks of that content into the model's context window. It failed quickly.

The issue was that large context updates fail in an all-or-nothing way. Attempting to push 10,000 tokens into a window with only 5,000 tokens of headroom doesn't gracefully trim the oldest content. It wipes out the preceding conversation history entirely. The workaround was to break everything into 2,000-token increments and feed them sequentially. More overhead, significantly more stable behavior.

The deeper insight was about message framing. In the conversation structure, different message "roles" system, user, and assistant tell the model how to interpret what it's reading. When too much contextual content was framed as user messages, the model started behaving as if the user had spoken every paragraph of a web page aloud. When too much came in as system, the model lost its ability to distinguish between its own instructions and the user's actual question.

Getting that distinction right is less about volume and more about semantic precision. The model behaves the way you describe the world to it.

2. Audio Consistency Across Platforms Is Harder Than It Sounds

Perplexity runs across multiple product surfaces: the core Ask product, the Comet browser, and the Computer interface. Each surface is built on a different client stack Swift for iOS, TypeScript for web, Rust and C++ for lower-level components. Native audio buffers across these stacks behave differently, and those differences introduced inconsistencies in how voice input reached the Realtime API.

The solution was to build a shared audio SDK in Rust that standardized the full preprocessing pipeline: resampling to 48 kHz mono, Opus codec encoding, and a complete audio processing chain covering echo cancellation, noise reduction, automatic gain control, and high-pass filtering. Every client now sends the same audio contract to the server.

This kind of infrastructure investment is invisible to users, but it is the foundation on which everything else depends. Inconsistent input produces inconsistent output. One poor audio experience can undermine trust in a product that works correctly 95% of the time.

3. Voice Activity Detection Needs to Be Calibrated for Real Environments

Voice Activity Detection (VAD) the part of a voice system that decides when the user has finished speaking is one of those problems that seems trivial until you try to make it work in the real world. In a clean environment, VAD is straightforward. In the world most users actually inhabit, it is significantly harder.

The team tuned specifically against the noisy bar scenario mentioned earlier. The failure mode they were trying to avoid: the model jumping in during a natural pause, treating silence as the end of a turn when the user is just thinking, pulling something up on screen, or about to read a passage aloud. This was particularly visible in technical use cases someone pausing mid-question to find the right formula, only for the model to respond before the question was complete.

The solution was a feature they called voice lock. Rather than push-to-talk (where voice is off by default and the user activates it), voice lock inverts the model. Interaction stays ambient and natural, but the user can explicitly lock the voice to claim the turn when they need it. It is a simple interaction pattern, but Perplexity believes it represents something more significant a standard design primitive for voice interfaces in complex, multi-step workflows.

4. Tools Should Be Minimal, Structured, and In Distribution

The team found that constraining the model's toolset to under ten core tools produced more stable and reliable behavior than offering a wider range. More importantly, how those tools were called and how their outputs were formatted had a significant effect on consistency.

The key principle: tool outputs should be formatted as structured data, not as assistant dialogue. When tool responses were written in natural language with embedded instructions, the model would mix spoken content with behavioral directives in unpredictable ways. When responses were structured as clean JSON separating the user-facing utterance from any behavioral flags the interaction pattern became much more consistent.

This is a useful reminder that language models behave more reliably when inputs resemble what they were trained on. Writing prompts, tool schemas, and tool outputs in ways that feel familiar to the model reduces edge cases and unexpected behavior.

What This Means for Search, Marketing, and Brand Visibility

The technical details above are not just interesting from an engineering perspective. They have direct implications for how brands get found, evaluated, and recommended in a world where voice is an increasingly common search interface.

The Query Has Changed

Traditional search behavior follows a fairly predictable pattern: a user types 2-5 words, scans a list of results, and clicks through to find what they need. Voice search, and AI search more broadly, inverts this. Users ask complete questions. Perplexity's data shows that its queries tend to be significantly longer and more research-intensive than traditional Google searches. Users are asking things like "what's the most reliable accounting software for a creative agency with a 15-person team that does both project billing and retainer work?" not "accounting software for agencies."

That shift in query structure changes what it means to be visible. You are no longer optimizing for keyword presence in isolation. You are optimizing for the relevance, authority, and specificity of your content against complete questions asked in natural language.

For example, consider two brands selling B2B CRM software. Brand A has a homepage that lists features and pricing. Brand B has published a detailed guide on how to choose a CRM for a distributed sales team, with real use cases, comparison frameworks, and cited data. When a voice user asks "what's the best CRM for a sales team that works across three time zones?", Brand B is far more likely to surface not because of keyword density, but because their content matches the structure of the question being asked.

AI Platforms Are Becoming Brand Discovery Engines

One of the most important behavioral shifts happening in search right now is that users are increasingly beginning their research journey inside AI platforms rather than on Google. According to industry data from 2026, AI search traffic converts at 14.2% compared to Google's 2.8% meaning users arriving from AI referrals are significantly further along in their decision-making process.

This has enormous implications for brand visibility. When someone asks Perplexity "which marketing agencies specialize in AI-driven lead generation for SaaS companies?", the platform synthesizes an answer from everything it can find about the topic. Brands that are well-documented, frequently cited, clearly categorized, and positively reviewed across credible sources are more likely to appear in that synthesis. Brands that exist primarily as a homepage with no supporting content infrastructure are largely invisible in these environments.

Every ai automated marketing agency operating today needs to ask whether its content strategy is built for this reality or for a search paradigm that is rapidly receding.

Voice Adds Another Layer of Urgency

Voice makes this dynamic even more acute. When a user is speaking to Perplexity or another AI assistant, they are receiving one answer not a list of ten results they can scan. The brand that doesn't appear in that single synthesized response doesn't just rank lower. It doesn't exist in that interaction at all.

According to research cited by Search Engine Journal, nearly half of all online searches are now conducted via voice, with that proportion higher among mobile users and younger demographics. As platforms like Perplexity scale their voice infrastructure using tools like the Realtime API, that behavior will only become more prevalent and the stakes of appearing in voice-delivered AI answers will rise accordingly.

The GEO Shift: Generative Engine Optimization Is No Longer Optional

Alongside traditional SEO, a new discipline has emerged that marketing professionals are increasingly treating as a distinct practice: Generative Engine Optimization, or GEO. Where SEO focuses on ranking well in traditional search results, GEO focuses on being cited, referenced, and recommended within AI-generated answers.

The market tells the story clearly. The GEO sector was valued at approximately $848 million in 2025 and is projected to reach $33.7 billion by 2034, growing at a compound annual rate of over 50%. Roughly 38% of business decision-makers have already allocated budget specifically to AI search optimization.

The principles of GEO are different from traditional SEO in a few important ways. Content needs to be genuinely authoritative and specific enough that an AI model would be likely to cite it as a source. Structured data and clear semantic markup help AI systems understand what a piece of content is about. Brand presence across multiple credible sources including industry publications, reviews, case studies, and partner mentions signals the kind of cross-platform authority that AI models use when deciding what to recommend.

For brands operating in the B2B space in particular, this represents a significant opportunity. Most competitors are still running SEO playbooks that were designed for a keyword-ranking world. The window to establish authority in AI-driven search while the competitive landscape is still relatively open is closing, but it has not closed yet.

What Comes Next: Voice, Agents, and the Always-On Search Experience

Perplexity's Realtime API integration is the clearest public example so far of what it looks like to build production-grade voice AI at scale, but it is not an isolated development. The broader trajectory of AI search is moving toward continuous, ambient interaction experiences where users don't initiate a search so much as stay in a state of ongoing conversation with an AI system that helps them navigate information in real time.

Products like Perplexity Computer and Comet, the company's agentic browser, represent the next stage of this evolution. Rather than answering discrete questions, these systems can take actions, browse the web, complete tasks, and maintain context across extended sessions. Voice is the natural interface for that kind of interaction because it allows users to direct the system while doing other things.

For businesses and marketers, this means that brand visibility will increasingly be determined not just by what appears in a search result, but by what an AI agent recommends when someone delegates a task to it. The question is no longer just "does my content rank?" but "does my brand get mentioned when an AI is helping someone make a decision in my category?"

That is a fundamentally different question, and it requires a fundamentally different content and visibility strategy.

Strategic Takeaways for Brands and Agencies

Pulling the technical and market perspectives together, several clear priorities emerge for brands and agencies trying to navigate this landscape.

Build content that answers real questions, not just keyword clusters. Voice and AI search reward specificity and genuine usefulness. A detailed, well-structured answer to a complex question is more valuable than ten pages of broadly optimized content.

Think about authority signals beyond backlinks. AI systems draw on reviews, citations, case studies, partner mentions, and presence in authoritative publications when evaluating what to recommend. Brand credibility needs to be built across the full ecosystem of places where AI models gather information.

Prepare for multi-modal search from a single consistent brand voice. As voice, text, and visual search converge on AI platforms, the brand's ability to communicate clearly and consistently across formats becomes a competitive differentiator.

Invest in GEO alongside SEO. The two disciplines are related but distinct. A brand that ranks well on Google but doesn't appear in AI-generated answers is leaving an increasing share of high-intent discovery traffic on the table.

Take voice UX seriously as a design discipline. As Perplexity's engineering experience shows, voice done well requires significant investment in audio quality, conversational design, and edge-case handling. The brands and platforms that invest in this now will have a meaningful head start as voice becomes a standard way to interact with information.

Conclusion: The Voice Era Is Here, and the Rules of Visibility Have Changed

Perplexity's work building voice search at scale with the Realtime API is more than an engineering case study. It is a clear signal about where information consumption is heading. Voice is not a novelty feature appended to a text search product. It is a distinct interface paradigm with its own technical requirements, user expectations, and discovery dynamics.

For businesses, the implication is straightforward: the way people find and evaluate brands is changing faster than most marketing strategies are adapting. The platforms handling millions of voice sessions per month are already deciding, in real time, which brands are authoritative enough to recommend. Every ai automated marketing agency that wants to stay relevant in this environment needs to understand not just how to optimize for traditional search, but how to build the kind of visible, credible, and well-documented brand presence that AI systems recognize and trust.

Perplexity proved that voice at scale is a solvable engineering problem. The question now is whether the brands and agencies they serve will adapt their visibility strategy to match.