| name | analytics-report |
| description | Use when the marketing team needs a weekly analytics rollup across GA4, Mixpanel, HubSpot, and social/email platforms. Produces a one-page digest with week-over-week changes, hypotheses for the changes, recommended next actions, and an explicit signal-vs-noise call on each metric. |
| safety | writes-local |
When to use
Trigger this skill when the request includes any of:
- "Run the weekly marketing analytics report"
- "What changed in our numbers last week?"
- "Why did organic traffic drop on Tuesday?"
- "Summarize email and social performance for the Friday standup"
- A scheduled Friday-afternoon job
Do not use for deep attribution analysis (separate multi-touch modeling project), real-time dashboards (live BI tool territory), or ad-hoc queries that would be faster answered directly in the source tool.
Inputs
Required:
- Reporting period — defaults to "last completed week, Monday through Sunday." Accepts custom ranges.
- Data sources — which of GA4, Mixpanel, HubSpot, LinkedIn, X, and email platform the report should cover.
Optional:
- Previous week's report — enables consistent format and trend continuity.
- Known events — launches, outages, holidays, or large campaigns that landed during the period. These are context that changes the interpretation.
- Focus metrics — if a specific KPI is in focus this quarter (e.g., "CVR on the pricing page"), flag it.
Outputs
A one-page Markdown report with the following sections:
- Header — reporting period, data sources, known context (launches, outages, campaigns).
- Top-line scoreboard — 6-10 KPIs with this week's value, previous week's value, absolute delta, and percentage delta.
- What changed — 3-5 bullets calling out the largest movements, each tagged as signal or noise.
- Why we think — a hypothesis for each signal-tagged change, with confidence level (high/medium/low) and the evidence behind it.
- What to do next — 1-3 specific actions for the coming week, each with an owner and a success criterion.
- Noise log — brief note on the noise-tagged movements and why they are being dismissed. Important for auditability.
Tool dependencies
- GA4 MCP (or manual export) — sessions, users, engagement rate, conversions.
- Mixpanel MCP (or manual export) — product-level events (signup, activation, feature usage).
- HubSpot MCP (or manual export) — email open/click, list growth, MQL/SQL counts, pipeline contribution.
- LinkedIn and X analytics — impressions, engagement rate, profile visits, follower growth.
- Email platform analytics — open rate, CTR, unsubscribe rate, deliverability.
- Optional: Slack MCP to post the report into
#marketing on Friday afternoon.
Procedure
1. Define the baseline
- Lock the date range. Default: previous Monday 00:00 through Sunday 23:59 in the primary reporting timezone.
- Pull the same range from the prior week for week-over-week comparison. For monthly views, pull the prior 4 weeks.
- Note any known context — a launch on Tuesday, an outage on Thursday, a public holiday — anything that will bend the data.
2. Pull the core metrics
At minimum, pull:
Traffic (GA4):
- Sessions
- Users
- Engagement rate
- Top 5 landing pages by sessions
- Top 5 sources (organic, direct, referral, paid, social, email)
Conversion (GA4 + HubSpot):
- Primary CVR (demo request, signup, whatever is the north star)
- MQL count
- SQL count
- Pipeline $ attributed to marketing
SEO (Search Console if available):
- Total clicks from search
- Average position for the top 10 tracked keywords
- Rank changes >3 positions on tracked keywords
Product activation (Mixpanel):
- New signups
- Activation rate (signup → first successful event)
- Week-1 retention
Email (ESP):
- Sends
- Open rate
- CTR
- Unsubscribe rate
- Top-performing subject by CTR
Social (LinkedIn + X):
- Impressions
- Engagement rate
- Follower growth
- Top-performing post by engagement
3. Calculate deltas
For every metric, compute:
- Absolute delta (this week − last week)
- Percentage delta ((this − last) / last × 100)
- 4-week trailing average, for context on whether this week is unusual
4. Signal vs noise
For each notable delta, apply this test:
- Signal if: the change is >15% AND larger than the typical week-over-week volatility for that metric AND not explained by a known calendar effect (holiday, launch, pause).
- Noise if: the change is within normal weekly variance, explainable by a one-off event, or based on a too-small sample to be significant.
- When in doubt, call it noise. False signals waste team cycles; real signals persist and will show up again next week.
5. Hypothesis for every signal
For each metric flagged as signal, write a hypothesis that includes:
- What changed (the metric and direction).
- Why we think (the most likely cause, based on evidence).
- Confidence level (high/medium/low).
- Evidence (what data supports the hypothesis).
- What would disprove it (what we would look for next week).
Never state a hypothesis as fact. "We think X caused Y, because Z" beats "X caused Y."
6. Recommend actions
For each signal, propose at most one action. Each action has:
- A specific deliverable ("draft a rebuttal post," "A/B test the pricing page hero").
- An owner (role or name).
- A success criterion ("CVR on pricing page improves 10% next week").
No action is a valid recommendation if the signal just needs another week of data.
7. Post and archive
- Post the report into
#marketing on Slack (Friday 4pm local).
- Archive in the weekly analytics folder.
- Link the previous week's report in the header for continuity.
Examples
Example 1: Standard weekly report
Output (abridged):
# Marketing analytics — week of 2026-04-13 to 2026-04-19
Sources: GA4, Mixpanel, HubSpot, LinkedIn, X, Customer.io.
Known context: Runners v2 launched Wednesday 2026-04-15.
Previous report: [link]
## Scoreboard
| Metric | This week | Last week | Δ | Δ% |
|---|---|---|---|---|
| Sessions | 48,210 | 39,540 | +8,670 | +21.9% |
| Signups | 312 | 268 | +44 | +16.4% |
| Activation rate | 58% | 61% | -3pp | -4.9% |
| Demo requests | 47 | 41 | +6 | +14.6% |
| MQLs | 182 | 165 | +17 | +10.3% |
| SQLs | 41 | 38 | +3 | +7.9% |
| Email open rate | 38.2% | 41.0% | -2.8pp | -6.8% |
| Email CTR | 4.1% | 3.8% | +0.3pp | +7.9% |
| LinkedIn engagement rate | 6.8% | 4.2% | +2.6pp | +61.9% |
| Organic search clicks | 6,420 | 6,380 | +40 | +0.6% |
## What changed
- **Sessions up 22% week over week** — **signal.** Launch traffic from Runners v2 on Wednesday drove most of the gain.
- **LinkedIn engagement rate up 62%** — **signal.** Launch-day post hit 280+ reactions, far above our 4-week average.
- **Activation rate down 3 percentage points** — **signal.** Inverse of signup surge. Worth investigating.
- **Email open rate down 2.8pp** — **noise.** Inside normal weekly variance (4-week stdev is 3.1pp).
- **Organic search clicks flat** — **noise.** Normal Tuesday-holiday week within tracking tolerance.
## Why we think
**Sessions up 22%**
- Hypothesis: Runners v2 launch drove a traffic spike on Wed-Thu that carried through the rest of the week.
- Confidence: **high**.
- Evidence: 68% of the session gain came from /runners/v2 landing page and /blog/runners-v2-launch. Referrer mix shows LinkedIn and Hacker News as top sources.
- Disproof: if next week's sessions return to the 39-40k baseline and no sustained SEO lift shows, the launch was a one-week event.
**LinkedIn engagement up 62%**
- Hypothesis: The founder's launch-day post outperformed because the mechanism (sidecar policy) was concretely novel.
- Confidence: **medium**.
- Evidence: post-level analytics show 82% of engagement came from the single launch post. Comments skew technical (platform engineers).
- Disproof: if next week's non-launch posts revert to 4% engagement, the lift is launch-specific not a sustained audience shift.
**Activation rate down 3pp**
- Hypothesis: Launch brought in a wider top-of-funnel including more window-shoppers who signed up but did not run a pipeline.
- Confidence: **medium**.
- Evidence: new signups from LinkedIn referrer activated at 48%, vs 62% for organic signups. Sample size: 312.
- Disproof: if activation stays depressed two more weeks with no launch effect, the drop is a product/onboarding issue, not a cohort quality issue.
## What to do next
1. **Watch activation by acquisition source for two more weeks.** Owner: growth PM. Success: we know whether the drop is cohort or onboarding by 2026-05-03.
2. **Ship a launch-week retrospective post next Tuesday.** Owner: content. Success: sustains LinkedIn engagement above 5% for the week.
3. **Do not touch email cadence or subject-line strategy this week.** Owner: lifecycle. Success: email open rate returns to baseline naturally.
## Noise log
- Email open rate -2.8pp: within 4-week stdev (3.1pp). Not acting.
- Organic search clicks flat: Tuesday was a US public holiday, depressing weekday search. Expected.
- X impressions down 12%: small sample, normal variance.
Example 2: Flat week (no material change)
Output (abridged):
# Marketing analytics — week of 2026-03-30 to 2026-04-05
Sources: GA4, HubSpot, Customer.io.
Known context: No launches, no campaigns. Spring break in most US school districts.
Previous report: [link]
## Scoreboard
[table showing all metrics within ±8% of prior week]
## What changed
- Nothing material. All metrics inside the 4-week weekly variance envelope.
## Why we think
- It was a quiet week. That is fine.
## What to do next
1. No new actions. Continue the planned content calendar.
2. Flag for next week: Runners v2 launch on 2026-04-15 will require careful
noise/signal calls. Plan to pull the prior 4-week baseline before launch.
## Noise log
- All movements within tolerance. No signals to log.
Constraints
- One page. If the report is longer than one page rendered, it will not get read. Cut.
- Every signal has a hypothesis. Never surface a change without attempting an explanation — even "we do not know yet" is an answer.
- Every hypothesis has a confidence level. High/medium/low. No weasel-worded "it seems that possibly."
- Every action has an owner. No floating recommendations.
- Signal vs noise is explicit. Every notable movement is labeled one or the other.
- Do not double-count metrics across sources. A session is not a user is not a visit. Pick the canonical definition and stick to it in every weekly report.
- Do not compare non-comparable ranges. A holiday week vs a normal week needs a caveat, not a silent comparison.
- No predictions dressed as observations. "CVR is trending down" is a hypothesis, not a metric.
Quality checks
Before returning the report, confirm:
