Sample · representative data for format illustration
FaeWorks · Channel Causality Monitoring
Monthly Signal Report
Northwind B2B · B2B · Report month: August 2026 · Data window: rolling 12 months (Sep 2025 – Aug 2026) · Source: HubSpot full-history pull
1 · Surfaced findings
We only surface a finding when it clears every bar (enough data, low fluke-chance, real effect size, holds for 2 months). Most months that’s little or nothing — that’s correct.
1
surfaced (likely causal)
● Surfaced — likely causal
LinkedIn-sourced subscribers who engage before entering the funnel open deals at ~2.4× the base rate.
n = 61 · p = 0.02 · range 1.3–4.2×. This is the one relationship where the cause clearly comes before the effect — they engage on LinkedIn, then subscribe, then enter the funnel — so it isn’t just “people who were already interested.” Do this: fund the LinkedIn motion as a deal-development channel, not just awareness.
What this means: out of 61 real cases, LinkedIn-first people became deals more than twice as often — and there’s only a 2% chance that’s luck. That’s strong enough to act on. To make it airtight (“proven”), run the holdout test described below.
2 · Suggested actions
What to do with this month’s signal — where to spend, what to prove, and what to fix in the data.
💰 Investments
- Fund the LinkedIn motion as a deal-development channel — the one likely-causal driver this month.
- Hold paid search flat — it makes leads but shows no lift to opened deals.
🧪 Tests
- Run a newsletter holdout to turn its “watching” signal into proof.
- Instrument referral & social with UTMs so their real effect stops hiding.
🧹 Data hygiene
- 82% of contacts are OFFLINE/untracked — set Original Source on import- and integration-created contacts so attribution becomes possible.
- Audit integrations that create contacts with no source; enforce UTMs on every campaign.
Why this matters: the biggest lever this month isn’t a channel decision — it’s fixing the missing data. Until source capture improves, most of the database is invisible to this analysis.
3 · Lens A — What develops a Deal Opened?
Each channel’s lift on the chance a contact reaches Deal Opened, vs. those not exposed, across the rolling 12-month window. The vertical line is 1.0× (no effect). Bar color = confidence.
Likely causal — act
Worth watching
Inconclusive
LinkedIn exec → subscribe → funnel
Newsletter engaged subscriber
How to read it: longer bar past the line = stronger pull toward opening a deal. Green = trust it now; amber = promising but not certain; grey = no real signal. Paid search sitting left of the line means it isn’t moving deals, whatever its click volume looks like.
4 · Lens B — Where in the journey does each channel matter?
Lift at each funnel step, across the rolling 12-month window (all deals active Sep 2025 – Aug 2026 — not just last month, which is too small to judge). Blue = a real, significant effect (darker = stronger); grey “ns” = checked, but not distinguishable from luck. Hover a cell for n and p.
Swipe the table sideways for the rest of the columns →
How to read it: read across a row to see where a channel earns its keep. Paid search shows a real bump into MQL that vanishes by Deal Opened — it makes leads, not revenue. LinkedIn is blue the whole way across until the very end (where there just aren’t enough closed deals yet to call it).
5 · Change in significance since last report
How each relationship moved from last month’s run to this one — each run uses the full rolling 12-month window, so this is momentum, not a single-month snapshot.
Swipe the table sideways for the rest of the columns →
Plain English: LinkedIn crossed from “watch” to “act” this month. The newsletter keeps inching toward significance — worth the holdout test soon. Webinars drifted back; don’t read into a single soft month.
6 · Channels analyzed & how much data backs each
The standard HubSpot acquisition channels (Original Source), with the real data behind each — all within the rolling 12-month window. This is the coverage picture: which channels we can judge, and where a low count is a tracking gap, not low activity.
⚠ Missing-data alert: only ~18% of contacts carry a tracked source — 82% are “OFFLINE / untracked” (imports, integrations, list uploads, manual entry). Channel analysis runs only on the tracked slice, so several channels below are limited by missing data, not real performance.
Swipe the table sideways for the rest of the columns →
Plain English: the “outcome events” column is what matters — how many real deals each channel’s analysis rests on. Under ~30 is a hint; 30–60 we can start to judge; 100+ is solid. A red ⚠ means the count is low because the source wasn’t captured — a data problem we can fix, not proof the channel fails. We read both Original and Latest source (plus their drill-downs), and decompose the untracked bucket by drill-down (it's mostly bulk imports, CRM/integration syncs, and manual entry).
How to read this report
The four numbers we use, in plain language — so nothing here needs a statistics degree.
Lift (e.g. 2.4×) — how much more often people who saw this channel reach the outcome vs. those who didn’t. Above 1.0 helps, 1.0 is nothing, below 1.0 hurts. Bigger = stronger.
n (e.g. 61) — how many real cases the number is built on. Bigger = more trustworthy. Under ~30 is a hint, not proof.
p-value (e.g. 0.02) — the chance the pattern is just luck. Lower = realer. Below 0.05 is worth believing; above ~0.10 could easily be noise.
Range (e.g. 1.3–4.2×) — the believable span of the true effect. If the range crosses 1.0×, “no effect” can’t be ruled out yet.
Proven — bank on it (confirmed by a test)
Likely causal — act on it
Worth watching — a lead, not a verdict
Inconclusive — ignore for now
When to pay attention: act on anything Likely causal or Proven with n over ~30. Treat “worth watching” as a reason to keep looking, not a reason to spend money. Everything else is context.
7 · How these claims are kept honest
Guardrails on every run
- Cause before effect — exposure must precede the outcome before we imply direction.
- “Winner’s curse” checks — e.g. clients keep opening emails for years, which fakes a link; excluded.
- Confounders named — firm size, prior relationship, self- vs. list-added.
- Fluke correction — testing many channels inflates false wins; we correct for it.
- Silence by default — nothing surfaces unless it holds for 2 runs.
Turn a signal into proof
- Confidence climbs: Inconclusive → Watching → Likely causal → Proven.
- Historical data can reach Likely causal. Proven needs a real test.
- Recommended test: a newsletter holdout — withhold it from a random group, measure the funnel gap.
- That one test would move the newsletter from “watching” to a claim that survives a skeptic’s scrutiny.
- Holdout design is a premium add-on.