Lesson 7 of 8 · 48 min

Attribution, experiments, and multi-touch honesty

Measure outbound without false precision: event taxonomy, multi-touch model views, experiment design with account contamination awareness, CRM field governance, and explicit dark-social unknowns.

Last-click theater → honest measurement

Outbound attribution is where good systems go to lie. Leadership wants a single number; reality is multi-touch, dark social, long cycles, and polluted UTM fields. This lesson trains honest measurement design: event instrumentation, model choices, experiment hygiene, and what you refuse to claim.
Attribution answers: which investments deserve more budget? It does not answer with perfect causality from CRM source fields alone. Senior GTM eng presents a measurement stack: raw events, identity stitching, model views (first/last/multi), and experiments for causal claims on sequence arms.
Tie to prior lessons: without L1 events, L4 arm ids, and L6 write-back, attribution is fanfiction. Do not re-teach full marketing analytics — focus on outbound-specific pitfalls and interview framing.

Event taxonomy for outbound

text
1OUTBOUND MEASUREMENT EVENTS23  contact.enrolled        {sequence_id, arm_id, icp_version}4  email.sent              {step_id, mailbox_id, domain}5  email.bounced           {type}6  email.replied           {reply_class}7  meeting.booked          {source_touch_id}8  opp.created             {amount, stage}9  opp.closed_won          {amount}1011  identity: contact_id, account_id, person_key12  store raw; compute models in warehouse views — don’t overwrite history
  1. 01First-touch — good for top-of-funnel discovery credit; rough for long ABM.
  2. 02Last-touch — simple; over-credits the closer email or inbound form.
  3. 03Linear multi-touch — spreads credit; still correlational.
  4. 04Position-based — weights first+last; common executive compromise.
  5. 05Experiment/holdout — causal for specific changes when designed well.

Outbound-specific pollution

Pollution sources: SDRs overwriting lead source on every touch; marketing UTMs on AE calendars; sequences without arm_id; meetings booked without contact match; recycled leads with ancient first-touch. Fix with locked fields, append-only activity timelines, and clear definitions docs.

Experiments that survive contact with sales

A/B subject lines are easy; structural experiments are hard. Design: unit of randomization (contact vs account), isolation (no dual enrollment), minimum sample, primary metric (positive rate, meeting rate — pick one), guardrails (bounce, unsub), and freeze period. Stop peeking daily without sequential testing discipline.
text
1EXPERIMENT CARD23  hypothesis: arm B (case study CTA) lifts meeting_rate by 20% rel4  unit: contact (account-cluster if multi-thread risk)5  allocation: 50/50 for 3 weeks or N=4k enrolls6  primary: meetings_booked / delivered7  guardrails: hard_bounce < 2%, complaint < 0.08%8  freeze: no copy edits mid-flight9  analysis: pre-registered; report CI not only point lift10  decision: ship / iterate / kill1112  Do not run 12 simultaneous micro-tweaks on one domain fleet.

Dashboards executives can trust

Offer layered views: (1) operational health (L1/L5), (2) conversion funnel by ICP band, (3) model-based pipeline credit with methodology footnote, (4) experiment board. Never show a single “ROI to 3 decimals” without assumptions. Honest footnotes are senior signal.
  1. 01Operational — sends, bounces, positives, SLA breaches.
  2. 02Efficiency — cost per meeting, cost per opp by band.
  3. 03Model credit — first/last/multi views side by side.
  4. 04Quality — win rate and cycle time by band (scoring calibration).
  5. 05Uncertainty — sample sizes, dark social caveat, CRM hygiene score.

Dark social and self-reported attribution

Buyers often find you via untracked paths. Use self-reported attribution on forms as a complementary signal, not a replacement for events. Triangulate: self-report + multi-touch + holdouts. Interviewers like candidates who name triangulation instead of worshipping one tool’s ROI dashboard.

Interview ways

  1. 01Q: How would you prove outbound works? Define success metric, instrument events, run geo/account holdouts or stepped-wedge if possible, report multi-touch alongside experiments, and include cost and brand-risk guardrails — not only opportunity count.
  2. 02Q: First-touch or last-touch? Show both plus a multi-touch view; explain bias of each. Use experiments for causal claims about sequence changes. Refuse single-number theater if asked to ‘just pick one’ without context.
  3. 03Q: What breaks attribution after a CRM migration? Lost activity history, remapped source fields, broken webhooks, duplicate contacts. Plan identity mapping and freeze experiment reads until timelines are repaired.

Connecting cost to outcomes

Unit economics from L1 only close when cost events join meeting/opp events. Meter enrichment $, send-seat amortization, and tool costs per tenant/arm. Without cost join keys, “ROI” becomes revenue folklore. Prefer cost-per-meeting bands over mythical multi-year LTV precision on day one.
  1. 01Cost events — enrich_attempt, enrich_success, send, seat_day.
  2. 02Join keys — tenant_id, contact_id, enroll_id, arm_id.
  3. 03Rollups — weekly cost/meeting by ICP band and sequence.
  4. 04Guardrail — pause arms that blow cost budgets even if reply rate looks fine.
text
1REPORTING CONTRACT (warehouse)23  fact_outbound_touch   enroll_id, arm_id, step_id, ts, cost_alloc4  fact_outbound_reply   message_id, class, ts5  fact_meeting          meeting_id, contact_id, ts, source_touch6  fact_opp_snapshot     opp_id, amount, stage, ts78  views:9    v_first_touch_credit10    v_last_touch_credit11    v_linear_multi_touch12    v_experiment_arm_stats  (pre-registered metrics only)
When leadership asks “what % of pipeline is outbound?”, answer with definition first: first-touch, last-touch, or touched-by-outbound-in-90d. Different definitions move the number 2×. Putting the definition in the chart title is a senior habit.

What to refuse in interviews

Refuse: single decimal ROI without methodology; open-rate as the experiment primary; overwriting Lead Source on every touch as “hygiene”; claiming multi-touch is causal. Offer instead: evented facts, dual model views, experiment cards, and cost-per-meeting with guardrails. That refusal is senior signal.
  1. 01Say — “Here are three views and their biases.”
  2. 02Say — “Causal claim only where we randomized.”
  3. 03Say — “Dark social is unmeasured; here is triangulation.”
  4. 04Don’t say — “Our dashboard proves outbound is 37.2% of revenue.”
Attribution is a model of reality, not reality. The senior skill is stating the model’s error bars out loud.
docsGoogle — Marketing attribution overviewGoogle AnalyticsdocsSegment — attribution and tracking conceptsSegmentarticleReforge-style growth measurement essays (multi-touch caveats)ReforgearticleClay blog — measuring GTM systemsClay

Checkpoint

CEO wants one outbound ROI number to 2 decimals for the board. Best response?

AInvent a precise last-touch ROI because boards hate uncertaintyBPresent layered metrics + methodology: cost/meeting, multi-touch views, experiment lifts, and explicit unknownsCRefuse all numbers until multi-touch is perfect in CRM
Sign up free to answer and see why

Checkpoint

You A/B two subject lines but AEs multi-thread accounts across arms. Issue?

ANo issue — contact-level randomization is always unbiasedBContamination risk: randomize by account or analyze with clustering; document bias if notCBan multi-threading forever to protect A/B purity
Sign up free to answer and see why

Checkpoint

Which instrumentation gap most breaks sequence experiments?

AMissing open pixels on every emailBMissing arm_id/enroll events on contacts so results cannot be joined to outcomesCMissing TikTok pixel
Sign up free to answer and see why

Checkpoint

Pipeline sourced “Outbound” fell 30% MoM after SDR training to “update Lead Source on every touch.” Diagnosis?

AOutbound collapsed — cut the channel immediatelyBLikely CRM hygiene/process pollution; audit field governance and activity timeline before concluding channel failureCAlways trust Lead Source over activities because it’s simpler
Sign up free to answer and see why

Checkpoint

When is a holdout the right tool for outbound proof?

ANever — holdouts are unethical in all B2B contextsBWhen you need causal evidence for investment and can randomize eligible accounts with guardrails and time boundsCOnly when you lack any CRM data at all
Sign up free to answer and see why

Can you design outbound events, experiment cards, and honest multi-touch reporting with explicit unknowns?

New to itGetting thereConfident

Takeaways

  • Instrument enroll/sent/reply/meeting events with arm and ICP versions.
  • Show multiple attribution views; use experiments for causal claims.
  • Govern CRM fields; document dark social and contamination risks.
  • Next: capstone multi-client outbound pipeline design (gos-capstone).

Next lesson: assemble the full system for multi-client production reality.

Sources

Free to read · better with Enzo

Learn it with Enzo

Save your progress, answer the checkpoints, and let Enzo quiz you on what you just read.

Continue with EnzoSign up free