All writing
Case study
DeepSite — website audits that give you the fix, not just the diagnosis
Existing audit tools hand you a Lighthouse score and 47 tabs of docs. DeepSite scans your site once and returns a plain-English report you can actually act on.
- Client
- My own product
- Role
- Design · Full-stack · SEO · Performance
- Year
- 2025
- Stack
- Next.js · Playwright · Lighthouse · OpenAI
Why I built this
Every time I audited a client site pre-engagement, the feedback was the same: “OK, we know it scores 62. But what do we actually DO?”
Lighthouse and every existing audit tool is built for developers. It hands you a wall of red flags, links to docs, jargon like LCP, CLS, INP. If you're a designer, a PM, or a marketing lead, the report is unusable — you can't act on it, and you can't hand it to a dev without an interpretation layer.
I wanted an audit tool that closed the gap. Scan once, get a plain-English report ranked by impact, with the fix — not just the diagnosis — spelled out.
What already existed
- PageSpeed Insights — official, accurate, jargon-heavy.
- GTmetrix — nicer UI, same jargon problem.
- Screaming Frog — enterprise-grade, learning curve.
- Various “site checker” landing pages — mostly SEO agency lead magnets, low signal.
The gap: a tool that runs the same underlying scans but translates the output for humans who don't live in Chrome DevTools.
What I built
- One-click scan. Paste a URL, get a report. No account for the free tier.
- Four checks in one: performance (Lighthouse + Playwright), SEO (structured data, meta tags, sitemap, robots), accessibility (axe-core rules), and design consistency (font sizes, spacing scale, color usage across pages).
- Plain-English fixes. Every issue gets rewritten by an LLM pass to describe the impact (“this button is 4px off spec” instead of “flexbox item deviation on selector #cta-primary”), what to do, and why it matters — ranked by business impact.
- Visual-diff callouts. Screenshots with the problem highlighted, not a wall of text.
- PDF export. Send the report to a client, unbranded if you want.
- Free tier + premium continuous monitoring. One-off scans are free. Scheduled re-scans with alerts are paid.
The hard parts
Making the LLM output trustworthy. LLMs will confidently generate incorrect fixes if you let them. I anchored the LLM to the raw Lighthouse and axe output, then constrained it to rephrase — not synthesize new claims. Every fix in the output can be traced back to a specific rule ID from the underlying scanner. The LLM is a translation layer, not an inspection layer.
Design consistency detection. This didn't exist off the shelf. I built it by crawling every page, extracting computed styles for headings + buttons + spacing, then flagging outliers. “You use 12 different font sizes across your site” is a real finding that no other tool surfaces.
Runtime cost. Playwright + Lighthouse + LLM calls per scan is expensive. Cached scan results per URL for 24 hours, batched LLM calls per report instead of per issue. Cut cost per scan by roughly 70%.
Result
Used by designers, PMs, and dev leads — the audience I built it for. Most common feedback: “this is the first audit tool I've been able to hand directly to a client without editing.”
Zero telemetry beyond scan URLs (needed for caching). No user tracking, no email for the free tier.
What I'd do differently
Should've built the design-consistency check first, not last. It ended up being the most differentiating feature. I spent the first two weeks polishing the Lighthouse + a11y pipeline that everyone else already builds well.
Takeaway
Most audit tools optimize for the person running the scan. The real value is in optimizing for the person who has to act on the results — often not the same person. Building for that gap is where I found the angle.