The problem
A sales brief goes out of date quickly. Redoing the research every week works for a few accounts, not fifty.
The useful signal isn't who a company is. It's what changed: a new pricing page, an enterprise tier, a wave of hiring. Those are good times to reach out, and you only see them if something re-reads the site.
What I built
Open any company's site and SignalizeAI writes a brief: what they do, who they sell to, who to contact, and a few outreach drafts grounded in that brief.
Save the company and it goes under watch. A background check re-reads the site, compares it with last time, and flags changes to pricing, hiring, messaging or new pages, with the before and after and a suggested opening line.
System design
Research once. Watch what changes
- Read the pageBrowser-side extraction
- Build the briefExtension + Workers API
- Watch & compareAlarms · snapshots · diff
- Surface changesAlerts + outreach context
Explore the full architecture
Research and watch paths, simplified. Watch checks require the browser to be running.
Design notes
A Manifest V3 extension with a browser-side watch loop, structured page snapshots, and a Cloudflare Workers backend for research and change summaries.
Keep fetching in the browser
The extension reads watched pages while the browser is running. The backend does not run a crawler.
Separate noise from change
Structured snapshot comparison suppresses whitespace, reordered navigation, tracking parameters, and failed price extraction.
Keep state across checks
Supabase handles authentication, saved prospects, and stored snapshots; the extension handles scheduling and extraction.
Calls I'd make again
- The AI never decides what changed. A plain differ finds the change and the model only writes one sentence about it, so it can't make changes up.
- Watching runs in your browser by default. Server-side checks need an explicit opt-in per company.
- Limits on page size, time, daily checks and AI spend, so one site can't run up costs.
The hard parts
- Telling real changes from noise. Reordered menus, tracking parameters and half-rendered pages are ignored, and tests replay saved pages from real sites to prove it.
- Getting clean JSON out of a model every time, with recovery when it doesn't.
- Being polite on the server: public addresses only, robots.txt respected, and a site that keeps failing gets paused.
What's still rough
Browser checks only run while the browser is open. Check times are a minimum, not a promise, and a company's fit is still a guess, so every draft stays editable.
