01
Source
AI Discovery Agent
Scans merchant promo pages, affiliate feeds, partner APIs, and sitemap diffs every 15 minutes. Classifies discovered codes by likely structure (% off, $ off, BOGO, free-ship).
Methodology · Last updated 14:46 UTC
How we verify every code on this site.
A multi-agent verification pipeline that combines AI cross-checks, an automated cart simulator at checkout, a human editorial team, and live community signal — running every 24 hours across 4712 stores.
Four stages, four agents, one loop.
Every code passes through four discrete stages. Each stage is owned by a specific agent — AI, automation, editorial, or community — and writes a structured log entry that becomes part of the code's verification history.
→
02
Test
Cart Simulator + AI Check
Headless browser adds a qualifying item, applies the code, captures the actual discount. AI cross-checks the result against merchant's published terms (exclusions, geo, plan).
→
03
Publish
AI Editor + Human Reviewer
AI generates the title, description, and meta from the test transcript. A human editor reviews everything tagged "editor's pick" before publish. Auto-approve threshold: 96% AI confidence.
→
04
Re-verify
AI Watcher + Community
AI watches merchant pages for changes (price, terms, expiry). Re-runs the full test on every detected change AND on a 24h cadence regardless. User votes feed the same scoring loop.
Three AI subsystems power the verification stack.
None of them are general-purpose chatbots. Each is a narrow agent fine-tuned for one job — and the orchestrator routes work between them.
Discovery
Discovery Agent
Crawls merchant promo pages and affiliate feeds. Reads HTML, ad copy, footer banners, and email-subscription confirmations. Classifies any candidate code by likely structure before passing it downstream.
- Model
- Gemini 2.5 Flash
- Throughput
- ~620 pages / min
- Inputs
- HTML, sitemap diffs, RSS
- Outputs
- Candidate code + structural class
Verification
Cross-Check Agent
Reads merchant terms-of-service, exclusion language, and category restrictions. Validates that the discount returned by the cart simulator matches the merchant's stated promotion — and flags partial successes.
- Model
- Claude Sonnet 4.6
- Latency
- ~1.4s p50
- Inputs
- Cart transcript + ToS
- Outputs
- Pass / partial / fail + reason
Editorial
Editorial Agent
Rewrites the title, description, and FAQ block for each verified code. Tuned for clarity over copy — describes who the code is for, where it applies, and what limits to expect at checkout.
- Model
- Claude Sonnet 4.6
- Output
- Title, desc, 3-5 FAQs
- Style guide
- ~14 pages, versioned
- Human review
- Below 96% confidence
Where 14 editors fit into the loop.
AI does most of the discovery and verification work. Humans own the judgment calls — editorial picks, edge-case re-checks, exclusion notes, and the standards every agent is held to.
→
Editor's picks
The single highest-value code per store is manually re-tested and curated. These show up on the store hero with the lime "★ Editor's pick" tag.
→
Failed AI verification
When AI cross-check confidence drops below 96%, the code is queued for human review before publish. About 8% of new codes pass through this path daily.
→
Edge-case redemption
Codes that need unusual checkout flows (call-to-redeem, mail-in, app-only) are flagged by the cart simulator and routed to an editor for documentation.
→
Standards & style
The Editorial Agent runs against a 14-page versioned style guide written and maintained by the editorial team. AI doesn't drift; we update the guide.
A real verification log, in real time.
What follows is the structured log written when a code passes through the pipeline. Every code on the site has one of these — viewable on the code's detail view.
14:46:04
discovery
scrape
norton.com/promo
14:46:05
discovery
found
SAVE60 (new)
14:46:06
ai-check
parse
code structure → valid
14:46:07
cart-sim
add
Norton 360 Deluxe → cart
14:46:09
cart-sim
apply
SAVE60 → discount $83.99 → $33.99 (60%)
14:46:09
cart-sim
verify
matches advertised 60% ✓
14:46:10
ai-rewrite
title
"60% off Norton 360 Deluxe (first year)"
14:46:11
publish
live
/norton/save60
14:46:12
ai-watch
monitor
next re-test scheduled in 24h
The numbers the editorial team watches.
Aggregate success rate
92%
across 12847 verifications today
Avg code lifespan
23 days
from first verification to expiry
Categories covered
9
4712 stores tracked
Human review rate
8%
of new codes hit the editor queue
The stack that runs this.
We're a small infrastructure: a single Hetzner box in Ashburn, Virginia, a SQLite database with daily snapshots, a cron scheduler, and a pool of headless browsers. The architecture choice is deliberate — small, observable, and cheap to operate at the scale we need.
Verification
Playwright (headless Chromium)
Cart simulator. CAPTCHA-resistant. ~3.2s p50 / code.
AI agents
Anthropic Claude · Google Gemini
Discovery, Cross-Check, Editorial. Routed by task.
Storage
SQLite + WAL · Daily B2 snapshot
Single-file DB. 48MB at this writing. Restorable in seconds.
Orchestration
Node + systemd timers
15-min discovery, 24h re-verify, on-demand for editor pushes.
Delivery
Fastify + Nunjucks · Nginx
Server-rendered HTML. No JS framework. Cached at the edge.
Observability
Per-code structured logs
Every verification writes a log; logs surface on each code page.
Our commitment
If a code doesn't work, we'd rather not show it.
That's the only rule. The pipeline, the agents, the editors, the metrics — all of it serves that one principle. Coupons are supposed to save you money; a coupon that doesn't work wastes your time. We took the category seriously enough to build the system above.