Perplexity Built an Amazing Product. It Won't Survive.

A LinkedIn post went around praising Perplexity as a $20 billion company that built zero AI models. The argument: they built the routing layer, the orchestration, the thing that makes the intelligence usable. Stripe didn't build the banks. Google didn't build the websites. The value is in making complexity disappear.

I get it. I've lived it. I've paid for Perplexity alongside ChatGPT, Claude, and Gemini for years — despite those same models being available inside Perplexity. That should tell you something about the product quality. The harness is legitimately best-in-class. But I believe Perplexity is doomed, and the reason is deceptively simple: they built something too good in exactly the space the major labs are racing to own.

The Case for Perplexity (It's Real)

Let me be fair before I'm critical. Perplexity earned my subscription. Multi-model routing, 400+ app connectors, a clean interface that makes complex orchestration feel like typing a question. For a while, this felt like the future of how we'd interact with AI — not picking a model, but describing what you need and letting the system figure it out.

The enterprise data backs it up. In January 2025, 90% of enterprise tasks ran on two models. By December, no single model held more than 25% of usage. The routing thesis is sound. The execution has been excellent.

So why am I bearish?

Search: The Original Moat Is Eroding

Search was the killer use case. Perplexity replaced Google for me — and I mean that literally. For over a year, my default search behavior was to open Perplexity instead of Google. Sourced answers, no SEO garbage, no scrolling past ads. It was a revelation.

But here's where we are now: Google's AI mode has gotten meaningfully better. ChatGPT's web search is solid and improving fast. Claude's search capabilities have caught up considerably. The gap hasn't closed entirely — Perplexity still has an edge in source quality and research depth — but it's narrowed from a canyon to a crack.

I still reach for Perplexity out of muscle memory, and I'll admit there's another factor: I like partitioning my usage. I don't want random internet searches clogging up my primary AI tools where I'm doing deep work. That's a real user behavior, but it's not a moat. It's a habit that erodes the moment someone builds a better tab or workspace model into a general-purpose tool.

The Browser Play: Smart Move, Same Problem

Perplexity saw search commoditizing, so they built Comet — an AI-native browser. There are mixed feelings on AI browsers, but I've found them genuinely useful.

I won't try to sell you on the generic "book a trip" demo. Instead, here's a real case that made me a believer: my daughter and I have celiac disease, and ordering food requires specific notes to be safe. I told Comet to find our local pizza place (they have three locations), order two pizzas with gluten-free crust, and let me know when it was ready. I switched back to working. It notified me when the order was done. It found the correct location, built the two pizzas with GF crust, and — without me asking — added a note about celiac in the comments field. I checked it over and clicked order. That's the kind of thing that makes you go okay, this is actually useful.

But here's the thing: I recently switched back to Chrome. Not because Comet got worse, but because Claude's Chrome plugin lets you schedule tasks, use skills, and do things Comet can't. Gemini has agentic tools baked into Chrome. ChatGPT launched Atlas (though I wasn't a fan — too much friction from normal browsing for it to stick). The point is, every major lab is building browser-level capabilities, and they're doing it with the weight of their full model ecosystems behind them.

Comet is a great product solving a real problem. But so was every feature that got absorbed by a platform.

A Side Note: AI-Powered Browser Migration

Speaking of switching browsers — I found an unexpectedly great use case for AI during the migration. My Chrome had years of bookmarks that I'd ported to Comet, but when I moved back, I didn't want to just dump the old Chrome bookmarks back. I exported both bookmark files, copy-pasted 90 days of browser history from Comet into a text file, and fed all three into an AI tool.

I asked it to categorize my bookmarks, analyze website usage frequency from the history, create an organizational structure, and build a "hot links" folder at the top for my most-visited sites. It spit out a clean bookmark file that I loaded into Chrome instead of the legacy mess.

This was a huge help — not just for migrating, but for organizing. I'd recommend it even if you're not switching browsers. Export your bookmarks and history, hand them to an AI, and ask it to build you a better system. Five minutes of work for a dramatically cleaner setup.

The Model Parity Problem

Here's the uncomfortable truth for Perplexity's thesis: at the margin, the models have a ton of nuance. But for the average user, they are reaching parity. Diminishing returns are kicking in hard. It matters much less today which model you pick than it did two years ago. GPT-4 vs. Claude 2 was a meaningful choice. GPT-5.4 vs. Claude Opus 4 vs. Gemini 2.5 Pro? For most tasks, any of them will do the job well.

That's actually good for the routing thesis in theory — if models are interchangeable, the routing layer adds value. But it's bad for Perplexity's business, because it means the labs don't need to be best at everything to satisfy their users. They just need to be good enough at routing within their own ecosystem, which is a much simpler problem than what Perplexity is solving.

The Stripe Analogy Doesn't Hold

The LinkedIn post compared Perplexity to Stripe: Stripe didn't build the banks, but it built the layer that made them usable. The difference is that banks had no interest in building developer-friendly payment APIs. That wasn't their business. The model labs, on the other hand, are extremely interested in building the harness layer. It's not adjacent to their business — it is their business. Every lab is shipping orchestration, tool use, multi-step agents, and app connectors.

When your moat is a harness that the platform providers are actively building themselves, you're not Stripe. You're a feature request on someone else's roadmap.

Where This Leaves Perplexity

I don't think Perplexity dies overnight. They have real users, real revenue, and real product quality. But the trajectory concerns me:

The search advantage is narrowing. The browser play faces competition from every lab. The routing thesis is being absorbed into the platforms themselves. And the brand awareness gap is real — most average consumers don't know Perplexity exists, while ChatGPT and Google are household names.

If I had to distill it to a single principle: building a harness that the labs could build themselves puts you in their crosshairs. Perplexity has executed brilliantly on a thesis that, unfortunately, lives in exactly the wrong competitive space.

In Part 2, I'll explore the flip side: if companies like Perplexity are in trouble, what does survive in the age of AI? The answer, I think, is less about building harnesses and more about building solutions to problems that require domain expertise the labs will never prioritize.

---

This is Part 1 of a two-part series. Part 2: Who Survives in the Age of AI? explores what actually works when the major labs are building everything.