Build log · SaaS web app
How I built ClearAudit.
In short
ClearAudit is a conversion audit for any web page. I built it on my own and run it in production. It measures what it can: Lighthouse runs three times and a real browser renders the page at phone width. Then it scores the page out of 10 across eight conversion areas and quotes the element behind every finding. Two audits a month are free, with no card.
The problem
Most AI audit tools hand back a score and generic advice, and I watched my own engine make the same mistake. An earlier version told a site owner they had no refund policy and no contact page, when both were linked in the footer. A founder can’t act on a report they can’t trust, and an agency can’t hand one to a client. I wanted an audit that only says what it actually saw.
What I built
- An audit engine, now v3.3 on Claude Haiku 5.5, that collects before it judges: the page text, facts read from the raw HTML, three Lighthouse runs and a real Chrome render at 412px.
- Scores computed by code from a published rubric: a score out of 10 across eight dimensions, from first impression and headline clarity to page speed and conversion flow.
- A claim guard that removes anything the page’s own markup refutes, and a second model that checks the claims that are left.
- Grading against the site’s own market (US, UK, EU, India and others), detected from the site or chosen by you, never from where the person running the audit sits.
- Three tiers: Standard, Advanced and Deep. Deep reads up to five pages and adds a funnel read. A Standard audit takes about 85 seconds and a Deep one about 150.
- The report as a document you can hand on: share links, a real PDF export, a tracker for each fix (shipped or skipped) and re-audits.
- Accounts, credits and billing, an app dashboard, first-party events and attribution, and an admin console with ⌘K search, a page per customer and email alerts.
Key decisions
I kept the arithmetic in code and gave the model only the judging.
A model is good at reading a page against clear criteria. It is bad at arithmetic and at rating its own confidence. So the rubric is published and the score is computed, not written by the model.
I run Lighthouse three times and use the median run.
One run decides too much by luck. Three runs of the same site in the same minute scored 75, 94 and 72.
Every finding quotes the element it is about, and the report says what the audit could not see.
Evidence next to the claim can be checked in seconds. A link the audit didn’t find is reported as “not found in the markup”, never “does not exist”, because pages that render in the browser can hide things.
The first audit runs before anyone is asked to log in.
People should see what the report is worth before they give anything for it.
Want something like this?
I build for clients the way I built this, and on client work you deal with one person from brief to launch. Tell me what you need, and the brief comes straight to me.