Product / The engine spec

An inspection engine, not a checklist.

The full spec of what runs between “paste URL” and “here’s your report”: 43 deterministic checks, an AI-assisted review layer, and a scoring model you can check by hand. Plus an honest list of what a scan can’t see.

Run a sample audit View sample report

43 checks · ≤10 pages · results in ~90 seconds

First pass: 43 deterministic checks.

Every crawled page becomes a structured snapshot — title, meta tags, headings, links, image filenames, forms, JSON-LD. The checks read the snapshot and file findings with the evidence attached: the exact tag, filename, or missing line. Same input, same findings, no model in the loop.

missing <title>
links to nowhere
unsupported claim phrases
no privacy policy link
missing alt text
AI-cliché copy
no contact path
weak CTA ("Learn More")
…and 35 more

Ten dimensions, weighted into one score.

Each finding deducts from its category; each category carries a fixed weight in the Launch Readiness Score. The weights don’t move between scans, and the model has no opinion about how pretty the site is.

The ten scoring dimensionsWeightedfixed model

Trust & Credibility18%
Proof, credibility signals, believable claims, and the things that make a visitor trust the site enough to act.
SEO Readiness14%
Titles, descriptions, headings, canonicals, and the metadata search engines need to rank the site.
AI Search Readiness10%
Entity clarity, structured data, and whether AI assistants can tell who you are and what you do.
Technical QA14%
Broken links, empty hrefs, missing infrastructure files, and the plumbing that quietly fails after launch.
Privacy & Metadata10%
Image filenames, EXIF-style leaks, missing privacy policy, and data you didn't mean to publish.
Accessibility Basics8%
Alt text, language attributes, and the baseline checks that keep the site usable for everyone.
Performance Basics8%
Oversized images, script bloat, and first-pass weight checks before real users hit the site.
Content Quality8%
AI-cliché copy, hype language, and text that reads like a template instead of a business.
Conversion Clarity6%
Clear audience, clear offer, working contact paths, and CTAs that actually ask for something.
Compliance-Sensitive Copy4%
Language that may need review before use in regulated or claims-sensitive contexts. Not legal advice.

Weighted average → Launch Readiness Score100%

The deduction math per finding — critical −25 down to enhancement −2 — is published step by step on how it works.

Then the judgment layer reads it.

Deterministic checks can count your meta tags. They can’t tell you that “guaranteed 10x growth” is a lawsuit wearing a font. The AI audit reads like a skeptical buyer with a compliance lawyer on speed dial: unbacked claims, bot-flavored copy, language that may require review, and whether a stranger could tell what the business does.

The provider is pluggable — set AI_PROVIDER to openai or anthropic for a real model, or leave it unset and the default mock provider runs a heuristic audit with no API key and no page content leaving the building. Real providers fall back to the mock on any error, so the AI layer is never the reason a scan dies.

source: ai

Inspection ticket

High priorityContent

The homepage welcome copy is the page builder's untouched output

FILED ON · https://marlow-dental.example/

Evidence: The section under the hero reads: "Welcome to the future of dentistry, where cutting-edge technology meets a seamless experience designed around you." It says nothing specific about Marlow Dental — no doctors, no locations, no services, no reason to choose this practice.
Suggested rewrite: Marlow Dental has cared for families at our two neighborhood locations for over a decade. Same-week appointments, clear treatment plans, and a team that explains every step before it happens.

Every AI finding carries a source: ai chip — on the finding card in the rendered report and in the exported Markdown — so a judgment call never passes as a measurement. The demo scan has four of them.

What a scan looks like in the workspace.

Below: the findings queue and scan history from the demo workspace — Juniper Lane Studio auditing Marlow Dental, the sample client. Every row is seeded demo data, not a customer.

Findings · Marlow Dental Website21 open · showing 4

Critical blockerTrustsource: deterministic
Homepage headline guarantees a treatment outcome
/
Critical blockerPrivacysource: deterministic
Before/after photos are published with patient names in the filenames
/services
High priorityContentsource: ai
The homepage welcome copy is the page builder's untouched output
/
Medium priorityPerfsource: deterministic
The hero image is a 4.8 MB uncompressed PNG
/

17 more findings in the sample report

Scan history

Jun 28, 2026
3 pages · 88s · 21 findings · 3 critical
71
Jul 7, 2026
re-scan after fixes · 0 critical
94+23

“Launch readiness went from 71 to 94” is a sentence a client understands without a meeting. Every past scan stays stored, and the re-scan comparison is in the report.

See the full sample report

fix-prompt.md — trimmed

·You are a senior web developer and copy editor improving the launch readiness of a website based on an AgencyProof audit.

✓Site: https://marlow-dental.example

✓Current launch readiness score: 71/100. Resolve the issues below so a re-scan comes back clean.

·## Issues to fix (already sorted by severity — work top to bottom)

✗1. [CRITICAL] [Trust & Credibility] Homepage headline guarantees a treatment outcome

· Fix: Replace the guarantee with what the practice can stand behind…

✗2. [CRITICAL] [Technical QA] The “Book an appointment” button goes nowhere

✗3. [CRITICAL] [Privacy & Metadata] Patient names in before/after photo filenames

· … 18 more issues

Hand the fix list to the machine that made the mess.

Every report ships with a fix prompt: the complete issue list — sorted by severity, with evidence, fixes, tasks, and approved rewrites — formatted as instructions for Claude or any coding agent. Copy, paste, re-scan.

One click in the report viewer. No reformatting.

The scope, stated plainly.

What ships today, what doesn’t yet, and what a scan can miss — listed here so you don’t have to ask, and so we can’t pretend we never said it.

What it does today

43 deterministic checks across up to 10 public pages
AI-assisted review — mock by default, openai/anthropic configurable
Markdown report + print-to-PDF
Claude/Fable fix-prompt export
Re-scan comparison with stored score history
White-label reports on Studio

Not in the product yet

PDF file export — print-to-PDF works today
Scheduled scans — every scan is started by you
Login-gated scans — public pages only
Lighthouse / Core Web Vitals — perf checks are first-pass weight checks
Form submission testing — forms are read, never submitted
Multi-language audits — copy checks assume English

What a scan can miss

JavaScript never runs: The scanner reads server HTML only. A site rendered entirely client-side — which plenty of AI page builders produce — scans thin, and the report says so instead of guessing.
10 pages, depth 2: Big sites are sampled, not exhausted. A problem on page eleven is a problem the scan never met.
AI review can misjudge tone: That's why every judgment call is labeled source: ai — read it as a flag to check, not a fact.
Not a legal review: Compliance findings point at language that may need review. A lawyer is still a lawyer.
No ranking guarantees: The SEO score measures readiness — titles, metadata, structure — not where you'll rank.
Forms are read, not submitted: We verify the markup and the submit path exist. End-to-end delivery is yours to test.

That’s the whole engine.

The demo scan runs the same code paths as a paid one. Judge it on a real site.

Run a sample audit View sample report