How the audit works

Roast.AI fetches your store's HTML and extracts concrete technical signals: title, meta description, OG tags, viewport, h1/h2 counts, images without alt text, internal links, scripts, Shopify markers, trust keywords (guarantee, returns, reviews, SSL), and social proof mentions.

Those signals — plus the visible page copy — are passed to an LLM (Gemini Flash via Lovable AI Gateway) that assigns 0–100 scores across 8 dimensions: CRO, Trust, SEO, Mobile UX, Performance, Branding, Product Page, Checkout.

Scoring is strict: most real stores land between 30 and 60. An 80+ means measurable excellence, not merely "no obvious mistakes". The model is instructed to be brutally honest instead of polite and vague.

The number shown as "estimated revenue leak" is not invented by AI. It is calculated deterministically from your scores and two signal-based estimates: monthly traffic and AOV.

current_CR  = lerp(0.3%, 3.5%, score_conversion / 100)
target_CR   = 2.0%   // ecommerce benchmark
gap_CR      = max(0, target_CR - current_CR)

leak_CRO    = monthly_visitors × gap_CR × AOV

seo_norm          = max(0.5, score_seo / 100)
traffic_potential = monthly_visitors / seo_norm
lost_visitors     = traffic_potential - monthly_visitors
leak_SEO          = lost_visitors × current_CR × AOV × 0.3

revenue_leak = clamp(leak_CRO + leak_SEO, 0, 100_000) /month

where score_conversion = cro·0.5 + checkout·0.3 + product_page·0.2. The 0.3 factor on SEO leak is conservative: not all missing traffic converts at the same rate, and SEO score is not linearly proportional to real lost traffic.

• On the first audit we only inspect homepage HTML, not deep product or checkout flows (Pro does that).
• Traffic and AOV are estimates: the LLM anchors them to visible signals such as domain, prices, reviews, and Shopify Plus. You can replace them with your real data for a more precise calculation.
• The leak is capped at 100,000/month to avoid inflated outliers.
• We do not measure real Core Web Vitals in v1 (coming soon: PageSpeed integration).