Memo

Growth Pod Memo — Expert Acquisition Funnel, 2024

Audience: Growth PM (analytical), Marketing Lead (less technical), VP Product (strategic) Author: Andrew Wilson · Date: 2026-05-10

1. Executive Summary

The single biggest opportunity is fixing the verification flow — ~1,925 experts a year drop off between Profile Completed and Verification Submitted (~38% step-down, the largest leak by count), and that's the highest-leverage A/B test we can run. On acquisition spend, LinkedIn outreach is the only paid channel above break-even (LTV:CAC ~1.34 at $515 CPA, vs ~0.18 and ~0.11 for paid_search and paid_social), and referral delivers the highest activation rate of any channel (~25%) at $0 spend — both worth scaling. The blended activation rate today is 12.1% on a like-for-like 60-day window; the funnel and channel opportunities above are how we move it. Top three actions: ship the verification-flow A/B (priority 1), reallocate paid spend from paid_search/paid_social to the top LinkedIn outreach campaigns, and pilot a referral incentive with adverse-selection guardrails. (LTV is gross expert payout per the assignment spec — read LTV:CAC as a payout-efficiency proxy across channels, not an absolute company return.)

2. Funnel opportunities — where to lift activation

Funnel reach on complete-window cohorts (signup ≤ 2024-11-01):

Two stages carry the most leverage:

1. Profile Completed → Verification Submitted (~38% step-down, ~1,925 experts). Experts spent 10–30 minutes on a profile and then chose not to submit verification documents. Hypotheses to test: (a) friction in document upload UX; (b) experts don't realise verification is required for paid engagements; (c) verification requirements aren't communicated up-front so the ask feels unexpected. This is a product lever, not an acquisition one — and the biggest single addressable opportunity in the funnel.
2. Verified → Activated (~55% step-down, ~1,471 experts). Verified experts who didn't see a paid engagement within 60 days. Hypotheses: (a) supply/demand mismatch (wrong industries or seniority for current client demand); (b) verified experts aren't surfacing in the matching algorithm; (c) "available for work" isn't on by default after verification. Recovering even 10% of this group would add ~150 activations against the current 1,219 (~12% lift) with zero additional acquisition spend — pure marginal upside.

Note on the Q3 → Q4 comparison. An unwindowed comparison shows Q4 activation rate ~30% below Q3 with z ≈ −4.6, which looks like a quarterly emergency. But that comparison includes Q4 cohorts that didn't yet have a full 60-day window to activate (the data ends 2024-12-31). Restricting both quarters to cohorts with a complete observation window (Q3 in full; Q4 through the first week of November) brings the gap to 4.5% relative with z ≈ −0.46 — not statistically significant at α = 0.05. The opportunity to focus on is the structural activation rate, not a Q4-specific intervention.

3. LinkedIn outreach is the only paid channel that works

Channel performance on complete-window cohorts:

Campaigns to scale. The top LinkedIn outreach campaigns by LTV:CAC are outreach_finance (1.47), outreach_healthcare (1.46), outreach_tech (1.40), and outreach_consulting (1.06) — all clear 1:1, and the first three clear the same threshold in volume and activation-rate rankings too. A reasonable next step is to roughly double monthly spend on the top three campaigns while pausing paid_search and paid_social, and watch incremental CPA over a four-week window. The data shows we can buy additional supply at the current CPA before saturation effects appear, but the bonus-test discipline from item 4 below applies here too.

On referral — promising but treat with adverse-selection care. Referral has the highest activation rate of any channel (25.2%) at $0 acquisition cost, but that depends critically on referrers self-selecting on high-intent friends. A blanket bonus changes the population: referrers may refer for cash rather than fit, and the activation rate of bonus-induced referrals will probably regress toward the blended mean. The test design (in §4 below) deliberately separates organic from bonus-induced referrals and gates scale-up on the bonus-cohort activation rate, not on the blended number.

On the unknown bucket. I initially flagged this as an attribution leak worth investigating; the investigation (Analysis 2.3) shows it's something else. Unknowns reach every funnel stage at roughly half the rate of known channels, and of the small slice that does submit verification, they are rejected at ~1.9× the base rate (with the important caveat that this is a within-submitters comparison, not a base-rate one — only ~12% of unknowns submit, vs ~35% of knowns). Zero of the 1,841 unknowns have ANY utm field populated, and the bucket is uniformly ~15% of every month, country, and industry. The conclusion is that unknown is best treated as a traffic-quality signal rather than as supply to reallocate: keep it separate in the channel reporting, and consider a referrer/IP filter at the top of the funnel so it doesn't pollute the verification queue.

4. Suggested Experiments

1. Verification-flow A/B (priority 1). Goal: lift Profile Completed → Verification Submitted from 62% to ≥70%. Variants: (a) inline verification requirements on the profile page so the ask doesn't surprise; (b) one-click "import credentials from LinkedIn" upload; (c) "skip for now" path with a day-3 retargeting email. Success metric: % of profile completers who submit verification within 7 days. Guardrails: verification approval rate must not fall, expert NPS must not fall.
2. Verified-but-not-activated reactivation. Goal: recover 10% of the ~1,471 verified-but-not-activated experts. Variants: (a) personalized "we have engagements in your industry" email; (b) self-service calendar scheduling for the first paid engagement; (c) Slack/SMS opt-in. Success metric: activation rate of treated verified experts within 30 days. Guardrails: unsubscribe rate, average payout per engagement (don't trade speed for quality).
3. Paid-channel reallocation. Cut paid_search and paid_social by 50% for four weeks, reallocate to the top LinkedIn outreach campaigns. Success metric: activated experts per dollar spend in the post-period vs the same period the prior year. Guardrails: total signup volume (don't starve the top of the funnel), share of unknown channel (an attribution leak would shift volume there rather than eliminate it).
4. Referral incentive program — adverse-selection-aware. Goal: test whether a payout-on-activation bonus to the referrer produces additional activated experts at acceptable quality. Three test cells at $100 / $200 / $300, randomized at the referrer level. Critical: track three groups separately — organic referrals (control), bonus-induced referrals (treatment), non-referred experts (baseline). Success metrics: (a) incremental referred-and-activated experts per dollar of bonus paid, AND (b) the activation rate of bonus-induced referrals must stay within ~5pp of the organic-referral baseline (currently ~25%) — if it doesn't, the bonus is attracting noise and we should stop. Guardrails: verification approval rate of treated referrals vs control (quality), self-referral fraud rate (monitor for circular email/IP patterns), and incremental cost-per-activation creep over a four-week window.

5. Data quality flags & what I'd do with more time

Data quality issues encountered and handled:

• Right-censored cohorts. Data ends 2024-12-31; signups after 2024-11-01 don't have a complete 60-day window. Every headline rate on the dashboard is restricted to cohorts with a complete observation window — without this, channels that ramped up in Q4 are penalized and the Q3-vs-Q4 z-test is heavily biased. The dashboard's methodology page spells this out and the dbt mart dim_data_window centralizes the cutoff.
• Event timestamps occasionally out of logical order. Three experts had profile_started_at before signup_at. Handled by enforcing monotonic stage timestamps in int_expert_funnel_stages (GREATEST(own_ts, prior_ts)). Would recommend adding a timestamp-ordering check to the event ingestion pipeline.
• Missing intermediate events. 236 experts have profile_completed (or a later event) without an explicit profile_started event. Handled by inferring upstream stage presence from later events. Worth auditing profile_started emission for silently dropped events.
• LTV as gross expert payout. Per the spec, the LTV proxy is the 12-month sum of payouts to the expert. Read the absolute LTV:CAC values as a payout-efficiency proxy; the relative ranking across channels is the load-bearing comparison.

Things I'd do with more time:

• Demand-side match analysis. The ~1,471-expert Verified → Activated drop is large enough that it's probably partly demand-side. Joining experts.industry to engagements.client_industry for the experts who DID activate would tell us whether some industries are demand-rich and others aren't. That changes the "fix verification" recommendation into "fix verification AND grow demand in industries X, Y".
• Statistical-significance flags on the dashboard's spend / CAC WoW deltas. Activation rate deltas already flag noise within ±2σ of the pooled SE, but spend / CAC don't yet. Adding bootstrap confidence intervals would keep reviewers from chasing weekly noise on those tiles too.
• Bot / device-fingerprint filter at the top of the funnel. The unknown-channel investigation suggests roughly 1,800 signups per year of low-quality traffic that survives all the way to the verification queue. Filtering these earlier would reduce reviewer load and tighten the rejection-rate signal.

Full model lineage, columns, tests, and SQL source for every model, macro, and analysis lives at /newtonx/dbt-docs/.