Built in public by a solo founder and a fleet of agents. One entry per user-visible change — no filler commits.
NEW
We benchmarked our own model against the SOTA — and published it
UNISAL vs DeepGaze IIE on the Toronto dataset (unseen by both): DeepGaze wins accuracy by ~1-5%, at 27× the size and 14× the latency. The full table, the trade-off and the raw results are on the model page. SUM (WACV 2025) follows the same protocol next.
Detailed findings + PDF report on the free scanner
The free score now computes findings across 5 dimensions (attention, focus, hierarchy, coverage, hot zone). Your email unlocks all of them plus a one-page PDF report delivered to your inbox.
Scan your preview URL in CI — attention score, metrics, delta vs production, one sticky comment, and an optional min-score merge gate. Deterministic by design.
New app shell: overview strip, unified report lists, skeleton loading, proper hover/focus states, and the brand's display type on scores.
NEW
Fix re-score: prove the improvement
Accept AI fixes, then re-capture your page with those fixes applied and run the model again. A real before/after — "61 → 64 (+3)" — with the proof heatmap, not an estimate.
NEW
Dynamic OG cards for shared reports
Every shared report and score now renders its own social card: screenshot, heat overlay and score. 197 static cards cover the rest of the site.
NEW
Founding 100: Pro at €19/mo for life
The first 100 accounts lock Pro at €19/month forever. Public counter on the launch page.
npx heatpoints-mcp gives Claude Code, Cursor and any MCP client three tools: analyze_page, analyze_image, compare_pages. The only predictive-attention MCP.
We scanned 93 famous SaaS landing pages. 81% hide their hottest zone outside the hero; median score 41; zero pages above 80. Full methodology and data published.
Four free tools shipped: site-vs-site attention duels with shareable face-off cards, a "guess where the eye goes" quiz, and dedicated scorers for YouTube thumbnails and ad creatives.