The pattern was obvious.
Every time we needed a coupon for something, we'd open three sites, click through six pop-ups, reveal twelve codes, and find one that worked. The category had completely abandoned the user.
Start typing to find a store.
Coupons. The category was poisoned by SEO farms publishing fifty codes per store, four of which work. We're trying to make this one corner of the web trustworthy again — using a thin team, a careful pipeline, and a lot of AI.
Not most of them. Not the popular ones. Every one. If a code is on this site, it was tested at checkout in the last 24 hours. If it doesn't work, we'd rather not show it.
That sounds like a small mission. In a category where 90% of listed codes are dead, it isn't.
"We took the category seriously enough to build the system above."
— pinned to the wall in our editor's room
Every time we needed a coupon for something, we'd open three sites, click through six pop-ups, reveal twelve codes, and find one that worked. The category had completely abandoned the user.
Codes fail because nobody tests them. We figured: if you actually test every code at checkout — with an automated cart simulator, AI cross-checks, and a human editor for the edge cases — the category becomes useful again.
The pipeline runs every 24 hours across 4,712 stores. Codes that fail twice disappear. The full methodology is documented on /how-we-verify — including the agents we use, the models, and a live verification log.
It sounds obvious. It's almost universally ignored. Codes that fail twice are removed the same day — no sad-face emoji, no "expired" tag at the bottom of the page. They just leave.
Every code has a structured verification log. Anyone can see when it was tested, by which agent, with what result. The methodology is documented in public — not behind a sales page.
No "Reveal!" buttons on dead codes. No fake countdown timers. The discount shown is the discount the merchant actually applied at checkout — not the one they advertised in a banner.
6 people. One Hetzner box. A SQLite database that fits on a phone. We didn't raise money. We don't need scale to be good — we need taste and a stubborn verification pipeline.
We use AI for what AI is good at: scanning thousands of merchant pages for new codes, cross-checking each one against published terms, drafting clean descriptions and FAQs. The humans on the team handle what AI is still bad at — taste, editorial standards, edge cases, and deciding what an "Editor's Pick" is.
The interesting engineering bit is the orchestration: routing each task to the right agent (or the right human) at the right point in the pipeline. That's documented in depth here:
Read the methodology →We're distributed across four time zones. The server lives in Ashburn, Virginia. We meet on Mondays.
Editorial Lead
Former pricing editor at a deal site you've heard of. Sets the bar on what counts as a working code.
AI Systems
Builds the verification agents, orchestrator, and the cross-check pipeline.
Operations
Keeps the cron healthy, the DB backed up, and the cart simulators behaving.
Cart Simulator Lead
Reverse-engineers checkout flows the moment they break and writes the resilient ones.
Editor
Reviews flagged codes, writes the edge-case notes, owns the style guide.
Engineering
Server, scheduler, observability. Wrote most of the structured-log layer.
We read every email. The editorial team responds within a day to anything about a specific code; the rest of the team takes a little longer.