Perplexity CEO Aravind Srinivas on the AI voice assistant race, Comet browser, and the TikTok bid

Summary

Perplexity CEO Aravind Srinivas sat down to discuss the company's push to become a native phone assistant, its browser ambitions, and how it thinks about monetisation — a conversation that reveals a company trying to move faster than Apple and out-manoeuvre Google at the application layer.

OEM deal and the assistant race

Perplexity is being pre-installed on phones from an OEM partner — Bloomberg had already reported it as Motorola's Moto AI lineup — with the ability to push-notify users to set Perplexity as their default Android assistant. Srinivas frames this as a significant unlock: OEMs had previously refused to take meetings on alternative assistants because Google's payments made any switch financially irrational. The same day, Perplexity posted that it had most of the assistant functionality working on iOS as well, relying on Apple's EventKit SDK to handle calendar, mail, reminders, Apple Music, and Maps — though native controls like alarms and volume remain off-limits.

On Apple opening up further, Srinivas is pessimistic. His realistic ask is that Siri be able to route to multiple AI apps based on user preferences, with Apple handling the simple, reliable tasks — setting alarms, sending texts — and third-party AIs handling anything multi-step.

Google antitrust

Perplexity's executive Dimitri testified at the Google DOJ trial the same day. Srinivas says he does not think Google should be broken up — not in America's interest, he argues, and it would just transfer a monopoly rather than create competition. He credits Google for maintaining Chromium, noting that every browser that has taken market share from Chrome, including Microsoft Edge and Brave, is built on Google's open-source work.

The specific behaviour Perplexity is pushing back on is Google's OEM coupling: if a manufacturer wants to ship an alternative AI assistant, Google can withhold approval of the Android version, which means losing access to the Play Store. Without Play Store, no OEM can sell a phone — Samsung couldn't make Galaxy Store work, and Meta and Amazon failed for the same reason. Srinivas is direct that Gemini the assistant is a poor product despite the underlying model being strong, and there should be no basis for locking it as default.

Models and the commodity question

Srinivas does not believe any lab will hold a model quality lead for more than a few weeks. The benchmark convergence — LM Arena, GPQA, Humanity's Last Exam — is pushing all frontier models toward similar outputs. His prediction is that in five years, consumers won't know or care which model is running beneath their app. Perplexity's strategy follows from that: use the best available model for any given task, including third-party models, and not let in-house model pride degrade the user experience.

For its own Sonar models, Perplexity is investing in post-training beyond instruction following — specifically reinforcement learning for native tool calls and task completion. The browser product, Comet, is central to this: agent failures inside the browser generate training trajectories that Perplexity plans to use to build eval suites and run RL post-training, using open-source implementations of GRPO and related methods.

Shopping and the hard operational reality

Perplexity launched in-app shopping and hotel booking toward the end of 2024, letting users transact directly without visiting a merchant site, including a Shopify API integration. The thesis was a clean innovator's dilemma play against Google, which cannot let users book on Google without destroying its own merchant referral revenue — which is why Google Flights and Google Hotels redirect to Expedia and Booking.com.

The product didn't land as intended. Users came for result quality, not checkout convenience, and the result quality wasn't good enough. Srinivas says Q1 was spent on the boring work: improving UI and latency, filtering low-quality merchants, fixing missing product images, and aggregating hotel review data across sources beyond TripAdvisor. Real-world fulfilment problems compounded this — hotel bookings sometimes never reached the property, and package tracking was impossible because merchants handled shipping.

The lesson he draws is stark: if you enter a vertical, you have to go all in. TikTok Shop, which he notes now has its own fulfilment operation in Seattle with ex-Amazon staff, is the reference point.

Monetisation and memory

Srinivas is firmly against traditional search advertising — injecting sponsored results into AI answers would undermine the core user value proposition. The subscription model is his primary focus, and he points to OpenAI's roughly $7–10 billion in annual revenue as evidence the TAM is real and growing. His argument is that AI assistants doing meaningful work are priced against personal assistants and chiefs of staff, not against a Google search CPM — positioning them at 100x lower cost but 10–100x more output value.

He leaves the door open to advertising, but only through hyper-personalised memory. Instagram's finding that engagement drops when ads are removed — because the ads themselves are that well-targeted — is his benchmark. To reach that, Perplexity needs data outside the app, which is part of the rationale for Comet: browser-level context on purchases, browsing habits, and location gives a far richer user profile than in-app queries alone, most of which are work-related and not personally revealing.

Comet and the ByteDance ambition

Comet, Perplexity's browser, was supposed to ship before this conversation. Srinivas says it was delayed by underestimating the project's difficulty and by trying to do too many things simultaneously. The target is now mid-May.

He describes Comet as Perplexity's first genuinely distinct second product — everything else, including the voice assistant and platform integrations, is essentially a variation of the core search product. Getting Comet right is the prerequisite before he earns the right to pursue the broader ambition he has articulated publicly: building a ByteDance of America.

After speaking with ByteDance's founder, Srinivas says the structural difference is that ByteDance runs shared infrastructure, growth, and front-end mobile teams across all its apps, allowing knowledge to compound across products. No US tech company operates that way. Google is the closest analogue — many apps under one roof — but the YouTube, Gmail, and Chrome teams operate in silos. Replicating ByteDance's model in the US would require a fundamentally different leadership structure and culture, and Perplexity has not earned that right yet.