Everyone keeps repeating that the open web is dead, strangled by AI summaries and Google’s shrinking referral traffic. Craig Campbell is quiet proof that the obituary is premature. As detailed in The Verge AI, the former Meta engineer turned down blank-check VC offers to chase the AI boom and instead built something almost nobody recommends in 2026: a website that lives on organic search.
His company, Past Maps, lets you overlay historical maps onto modern ones and fade between them. It started as a tool for his metal-detecting hobby, got shared on Reddit, and turned into a real business. The Verge AI reports traffic has climbed from 20,000 monthly active users to over 300,000 in year three.
Why this cuts against the narrative
The dominant story in AI right now is that search-driven publishing is finished. “Google Zero,” the point where the search engine answers everything itself and sends no clicks onward, is supposed to be the death of independent sites. Campbell’s experience complicates that.
He found Past Maps rising through search results as people hunted for hyper-specific local history: a church a grandmother attended, abandoned mines in one county. By structuring and tagging his data so Google could read it, he kicked off a flywheel. “This is how the web is supposed to work. This is actually the old school web,” he told The Verge AI. “It is alive and well, but only in these really, really small niches.”
That last clause is the whole thesis. The open web isn’t uniformly dying. It’s dying in the broad, commoditized middle, where AI can summarize and replace it. It’s surviving in narrow niches where the data is unique, structured, and not easily regurgitated.
The business model shift that matters
Here’s the part practitioners should study. An old-school publisher a decade ago would have lived on display ads, revenue controlled almost entirely by Google’s ad tech stack, which the DOJ ruled an illegal monopoly in 2025. Campbell skipped that trap.
Past Maps runs on subscriptions: a $9 weekly pass or $52 a year. That choice insulates him from ad-budget swings and from depending on the same company that controls his traffic. It’s a deliberate decoupling of how you get discovered from how you get paid.
He didn’t reject AI. He aimed it inward
The contrarian twist isn’t anti-AI. Campbell embraced the tools, just not as the product.
- A local agent on his desktop triages customer email once an hour, reading his Gmail, killing spam, and drafting replies. Service work dropped from one or two hours a day to about 10 minutes.
- When someone demands a refund, the agent queues the Stripe cancellation and pings him to approve. He stays in the loop on every send.
- He’s building an OCR tool for old maps, where curving labels and crowded text break off-the-shelf systems. Reasoning LLMs helped, but only paired with hands-on experimentation. “You have to still bring that human spark into the mix,” he said.
That’s a working template: AI as the back office, human judgment and a defensible dataset as the front.
What stands out
The honest footnote is income. Campbell says he earns about what he made as a mid-level Facebook engineer, and he admits he wonders what the VC route might have paid. This isn’t a unicorn story, and pretending otherwise would miss the point.
The point is durability. He owns his traffic source, his data, and his customer relationship, without a single VC board seat or an ad network deciding his fate.
Takeaways for builders
- Find the niche AI can’t flatten. Unique, structured data that resists summarization still earns search visibility.
- Sell access, not attention. Subscriptions beat ad revenue when one company controls both your traffic and your ad stack.
- Use AI to shrink your overhead, not to replace your product. The leverage is in operations.
- Stay in the approval loop. Campbell automates the draft, never the final decision.
The AI gold rush gets the headlines. Campbell’s bet suggests the quieter play, owning a small, defensible corner of the web, may outlast plenty of them. More on his approach is available at the original report from The Verge AI.