heatmap-analyst

Click-engagement analyst that pulls element-level click data, a page-level scroll proxy, and bounce/exit signals from Humblytics to surface UX friction and ignored CTAs. Generates prioritized, data-ba

SKILL.md

GitHub

Heatmap Analyst

Purpose

Analyze Humblytics click-engagement data (element/target-level clicks, a single page-level scroll proxy, and bounce/exit signals) to diagnose UX friction and generate prioritized design recommendations. Live data comes from the Humblytics MCP (the click and page tools — get_clicks_details, get_clicks_breakdown, get_page_details, get_pages_breakdown, get_entry_exit_pages). This skill turns the interaction data Humblytics actually exposes into specific, ranked improvements for layout, CTAs, and content hierarchy.

Scope note — what Humblytics does and does not give you. Humblytics provides element/target-level click counts (with UTM breakdown), a single average scroll percentage per page, and bounce/exit signals. It does NOT provide pixel-level click heatmaps (x/y coordinates), a 25/50/75/100 scroll-depth distribution, rage-click detection, dead-zone maps, or per-device click segmentation. Anything in that second list requires a dedicated heatmap tool (e.g. Hotjar, Microsoft Clarity) — do not promise it from Humblytics. See "NOT available via Humblytics" below.

When to Use

  • A page has a high bounce rate and you need to understand why
  • CTAs are present but click-through rate is below benchmark
  • You want to verify that the important content is actually being seen
  • Users are reporting confusion or friction on a specific page
  • You're auditing a page before a redesign or A/B test
  • Investigating whether traffic from a specific source behaves differently on-page

Setup

This skill reads live data through the Humblytics MCP (server humblytics) — see the repo README to connect it. Once connected, the skill calls mcp__humblytics__* tools; the MCP handles auth, base URL, and property resolution, so there are no keys to paste or .env files to source here. Never paste API keys into chat — the key lives once in the MCP connection headers, not in transcripts.

The MCP auto-resolves the property for a single-property key (the common case). For a multi-property key, call list_properties and pass the chosen propertyId to each tool.

Before You Start

  1. Confirm the property — With a multi-property key, run list_properties and confirm which property to analyze (single-property keys auto-resolve)
  2. Identify the target page(s) — Which URL(s) are in scope
  3. Time range — Default to last 30 days; shorter windows are noisier
  4. Sample size check — Pages below ~500 sessions in the window produce unreliable heatmaps
  5. Context — Pull product/persona context if available so recommendations match the audience

Core Workflow

Step 1: Pull the Interaction Data

For each target page, fetch what the API actually returns:

  • Element-level clicks — clicks grouped by element target (and secondary), with clicks, unique_sessions, most_recent, a per-element trend, and a utm_breakdown (clicks + share by UTM source/medium/campaign). This is element-level, not an x/y coordinate map.
  • Scroll proxy — a single avg_scroll_percent for the page (one number, e.g. 23.7), plus bounce_rate, page_views, unique_visitors, avg_session_length. This is not a 25/50/75/100 depth distribution.
  • Cross-page click comparison — per-page total_clicks, unique_sessions, and top_targets[]{target, clicks, share}.
  • Entry/exit friction — entry and exit pages as a friction proxy.

Click CTR is not a field in the API — derive an engagement rate yourself as clicks / unique_sessions (or per-page top_target.share) when you need a CTR-like proxy.

Relevant Humblytics MCP tools (all take start, end as ISO-8601 and a timezone IANA name — there is no ?period= shorthand; scroll depth lives in the page tools):

  • get_clicks_details (page: "/path") — element/target-level clicks + UTM breakdown for one page
  • get_clicks_breakdown — cross-page top targets
  • get_page_details (page: "/path") — avg_scroll_percent (scroll proxy) + bounce_rate for one page
  • get_pages_breakdown — page-level views/bounce across pages
  • get_entry_exit_pages — entry/exit friction proxy

NOT available via Humblytics (needs a dedicated heatmap tool): pixel-level click coordinate heatmaps, scroll-depth distribution (25/50/75/100%), rage-click detection, dead-zone maps, and per-device click segmentation. If the user needs any of these, tell them Humblytics does not return them and point to a purpose-built heatmap tool (Hotjar, Microsoft Clarity, etc.).

Step 2: The Three Diagnostic Questions

Run each page through these three questions, using only data the API returns:

Q1 — Are visitors clicking what you want them to click?

  • Primary CTA click share: is the CTA target a meaningful fraction of total_clicks (use its share from get_clicks_breakdown or clicks from get_clicks_details)?
  • Secondary CTA click share: proportional to its importance?
  • Which target dominates clicks, and is it a high-value action or a low-value/navigation element?

Q2 — Are visitors engaging deeply enough to see the important content?

  • avg_scroll_percent: a low average (e.g. ~24%) suggests most visitors never reach below-fold content. This is a single average, not a depth distribution — do not claim "X% reached 50%".
  • Is the primary CTA likely above or below where that average scroll lands?
  • Cross-reference with bounce_rate from get_page_details.

Q3 — Where is the friction?

  • High bounce_rate / exit share (from get_page_details and get_entry_exit_pages) on a page that should convert = friction proxy.
  • Low scroll engagement on a long page where the CTA sits deep.
  • A CTA target that gets almost no clicks despite high page views = ignored CTA.

Frustration signals like rage clicks and clicks on non-interactive elements are not available from Humblytics — use the friction proxies above, and recommend a dedicated heatmap/session-replay tool if true rage-click detection is needed.

Step 3: Identify the Top 3 Issues

Rank all issues by expected conversion impact:

  1. Blocker — Primary CTA gets a negligible share of clicks, or low avg_scroll_percent suggests core content is rarely reached
  2. Friction — High bounce_rate / exit share on a page meant to convert; confusing affordances
  3. Waste — High click share on low-value elements (e.g., a Link/nav target dominating clicks instead of the CTA)

Always state the evidence: "avg_scroll_percent on the homepage is 24% and bounce_rate is 0.89, while the signup CTA hero-try-free took only 1.2% of clicks."

Step 4: Generate Recommendations

For each issue, provide:

  • Specific change — "Move CTA from below the pricing table to above the hero fold"
  • Expected lift — Estimate based on traffic volume and issue severity
  • Implementation difficulty — Copy change / layout change / redesign
  • How to verify — Which metric to watch; which follow-up A/B test validates the fix

Step 5: Output Format

Write a clean report with:

PAGE: [/path]
DATE RANGE: [window]
PAGE VIEWS / UNIQUE VISITORS: [page_views] / [unique_visitors]

HEADLINE FINDING:
[1 sentence capturing the biggest insight]

CLICK PATTERN SUMMARY (from get_clicks_details + get_clicks_breakdown):
- Primary CTA target + click share: [target] ([share]% of clicks)
- Highest-click element: [target] ([share]% of clicks)
- Total clicks / unique sessions: [total_clicks] / [unique_sessions]
- Notable UTM skew (if any): [utm_source/medium] drives [share]% of a target's clicks

SCROLL ENGAGEMENT (from get_page_details — single average, not a distribution):
- avg_scroll_percent: [N]%
- Implication: [most visitors likely do / do not reach below-fold content]

FRICTION PROXIES:
- bounce_rate: [N]
- Top exit pages (get_entry_exit_pages): [pages]

TOP 3 RECOMMENDATIONS (prioritized):
1. [Change] — Expected impact: [X] — Difficulty: [level]
2. [Change] — Expected impact: [X] — Difficulty: [level]
3. [Change] — Expected impact: [X] — Difficulty: [level]

SUGGESTED A/B TESTS:
- [Test hypothesis with clear control vs variant]

Interpretation Cheatsheet

Based only on Humblytics-available signals (element-level click shares, single avg_scroll_percent, bounce_rate/exit):

Pattern Likely Cause Action
A generic Link/nav target dominates clicks, CTA target near zero CTA invisible, weak, or out-competed by navigation Strengthen CTA prominence; reduce competing links
Low avg_scroll_percent on a long page Weak hook, above-fold doesn't earn attention Rewrite headline or move proof/CTA above the fold
CTA clicks concentrated on one variant Other CTAs are invisible or redundant Remove redundant CTAs; test single CTA variant
High bounce_rate + low scroll on a convert-intent page Above-fold fails to engage Audit hero copy/offer; move value prop up
Click share spread thinly across many targets No clear visual hierarchy Add hierarchy: emphasize primary action
One UTM source's clicks skew heavily to a low-value target Mismatched intent from that channel Align landing experience to that source's intent

Patterns that require pixel coordinates, rage-click detection, or per-device click maps (e.g. "rage clicks on image", "desktop clicks ≠ mobile clicks") are not diagnosable from Humblytics — use a dedicated heatmap tool.

Related Skills

  • cro-optimizer — Combines heatmap findings with funnel data for holistic CRO
  • page-cro — Full 10-point page audit; heatmap analysis is one dimension
  • ab-test-generator — Takes heatmap recommendations and launches them as tests

Shared Frameworks (REQUIRED reading)

Heatmap interpretation is highly context-dependent. The shared primitives in skills/_shared/ keep recommendations grounded.

  • _shared/frameworks/preflight-checklist.md — confirm minimum 500 sessions per page-period combo before drawing conclusions. Heatmap patterns on smaller samples are noise.
  • _shared/frameworks/anti-patterns.md — heatmap-relevant counter-evidence:
    • Mobile hamburger menu: NN/g says it hurts discoverability on task-oriented SaaS (Spotify hamburger → bottom-tab = +30% menu interactions). BUT Amazon's hamburger beat dropdown for browse-heavy ecom. Site_type is load-bearing — don't recommend bottom-tab universally.
    • Mobile exit-intent: architecturally broken (no cursor → no mouseleave event). If heatmap shows users leaving on mobile, the answer is not an exit modal.
    • Progress-bar velocity: NIH RCT shows slow-to-fast progress bars nearly double form abandonment. If your heatmap shows form-step drop-off, audit progress bar acceleration before redesigning fields.
  • _shared/benchmarks/patterns.json — when heatmap data confirms a problem (e.g., low scroll past 30%, CTA clicks dominated by a single variant), match to a pattern_id and quote the evidence-backed lift range for the fix. Most relevant categories: cta, navigation, above_fold.