Retargeting Funnel

Retargeting Funnel

Design a multi-stage retargeting sequence from the client's own conversion data: stage boundaries, per-stage message intent, exclusion logic, frequency caps, and a measurement plan.

Scope and handoffs

This skill decides stage boundaries, per-stage message intent, exclusion logic, caps, and the measurement plan. It does not build campaigns inside any ad platform, and it hands off neighbouring work:

• Working event tracking is a hard prerequisite - retargeting audiences cannot be built without it. Verify first via mbfinotti/advertising-skills@ad-conversion-tracking.
• The broader targeting plan across interest/lookalike/retargeting layers → mbfinotti/advertising-skills@ad-audience-targeting.
• Ongoing creative-decay monitoring after launch → mbfinotti/advertising-skills@ad-creative-fatigue.
• Writing the per-stage ads themselves → mbfinotti/advertising-skills@ad-copy-variants and mbfinotti/advertising-skills@ad-creative-brief.
• Splitting the total budget across campaigns and channels → mbfinotti/advertising-skills@ad-spend-allocation.

If you draft any customer-facing copy while illustrating a stage, route it through your preferred humanizer skill before the user ships it - this is not a copywriting skill.

Interview

Ask these questions one at a time, multiple-choice where possible. Stop asking as soon as you have enough to design; do not batch all of them into one wall of questions.

1. Business model: (a) B2C / e-commerce, (b) B2B, (c) hybrid (e.g. PLG SaaS with self-serve and sales-assisted tiers)?
2. Typical time from first visit to conversion: (a) under 3 days, (b) 3-14 days, (c) 2-8 weeks, (d) 2 months or more, (e) unknown?
3. Monthly site traffic and monthly conversions, roughly? (Order of magnitude is enough.)
4. Which behavioural events are actually tracked today: page views only, product/feature views, pricing views, cart or checkout starts, trial starts, purchases, CRM stages?
5. Which paid channels are in use or planned?
6. Offer type: (a) transactional purchase, (b) free trial, (c) demo request, (d) lead form / gated content?
7. Is a discount ever available as an incentive? (Many B2B businesses: no - that removes a rung of the ladder, not the ladder.)
8. Total monthly paid budget and the current prospecting-vs-retargeting split, if one exists?
9. Does a converter-suppression / exclusion audience already exist on any channel?
10. By what date must this be live and producing, and is that date hard? (The creative ladder in step 5 spans near-zero to a quarter in time-to-effect, so its ordering cannot be picked without this.)
11. Do you want the fastest revenue available this quarter, or durable assets that keep paying - a proof library, an incrementality baseline? (The first favours the cheap rungs, the second the slow ones.)
12. Effort ceiling: who can produce creative and how often, and can you actually obtain customer consent and legal sign-off? (A "no" removes the top rungs of the ladder rather than delaying them.)

Refuse to invent stage boundaries without time-lag data. If the user answered "unknown" to question 2, say explicitly: "I will not guess your windows - either export a time-to-conversion report, or I will use <named proxy> and flag every boundary as provisional." Acceptable proxies - always named as a proxy in the output, never silently substituted for real data:

• Ranked by evidence bought per unit of effort: analytics lag report > CRM sales-cycle export > sector-typical figure for the vertical.
• effort (heaviest first): CRM export (an hour, plus whoever owns the CRM) > analytics export (an hour, self-serve) > sector figure (near-zero, one search).
• The cheapest proxy is also the weakest. The CRM export overtakes the analytics one only when the cycle outruns what the pixel can see (B2B, multi-month).
• This is a default, not a law - re-rank against what the user already holds. A lag report exported last month beats a better one nobody will pull.

Workflow

1. Pull the funnel data

Gather these, in order:

1. Traffic by page type (content / product or feature / pricing / cart or signup)
2. Conversion rate by funnel stage
3. The time-to-conversion lag distribution
4. The main drop-off points
5. The abandonment rate at the deepest pre-conversion step

If your harness can read files, ask the user to export these reports and drop them where you can read them; otherwise ask them to paste the key numbers. If you can browse the web, you may fetch published benchmarks to sanity-check their figures - but the client's own numbers always win over benchmarks.

2. Set stage boundaries from converter recency percentiles

This is the core move: derive windows from the client's own converters, not from platform defaults. From the lag distribution, find the recency point by which roughly 80% of converters had converted - that is the edge of the hot window. Then:

• If 80%+ convert inside 7 days, collapse to fewer stages (often just hot / warm / win-back). More stages than the cycle supports produces empty stages.
• If the cycle runs 60-90 days (typical B2B, high-ticket), stretch every window and plan for list-based rather than pixel-based membership at the long end, since browser privacy limits make long cookie windows unreliable.
• Default retention windows on ad platforms are arbitrary relative to any specific business. A default-length window on a 90-day cycle silently drops most of the pipeline; the same default on a 2-day impulse cycle wastes weeks of spend on cold traffic.

3. Build the two-axis grid: recency × behavioural depth

Cross the recency windows with the behavioural-depth axis. Claim grid cells deepest first, in this order:

cart, checkout or trial start > pricing page viewer > product or feature page viewer > any visitor / content viewer

Past customer and churned customer sit outside the conversion ladder, as win-back tiers. Assign windows inversely to depth: deep actions get short windows (intent decays fast), shallow actions get long windows (they were never hot, so staleness costs less). Not every grid cell becomes a stage - take the 3-5 deepest cells where the client has volume and a distinct message.

That order is value, not effort: every tier costs the same to build - one audience definition, minutes - so nothing separates them but the intent they capture. Effort splits them only where the event is not tracked yet (question 4): an untracked depth signal costs an engineering ticket and a wait, which can drop it below a shallower tier for this quarter. Re-rank on what the account already owns - a CRM deal stage already synced outranks a pricing-view pixel that needs a sprint.

4. Size-check every stage after exclusions

Check each stage against the platform's documented minimum audience size, measured after the exclusions from step 6 are applied - a stage that clears the floor before exclusions can fall under it once fresher stages and converters are carved out. Current floors and retention ceilings are in references/platform-constraints.md; verify against official documentation before building, because these values change often.

Collapse rule: if a stage cannot clear its floor after exclusions, repair it in this order:

widen the window > broaden the trigger > merge adjacent stages > collapse to one combined warm pool

All four cost minutes in the audience builder, so effort does not separate them. The ordering is value: what each step spends is message distinctness (widening keeps the stage's job intact, merging fuses two jobs into one, the single pool gives up sequencing entirely).

Never ship a stage that will not deliver - an under-floor stage either serves nothing or serves at punitive CPMs while the learning phase never exits. Thin-traffic sites often end with two stages, or one (see the Ben Heath position below); that is a correct output, not a failure.

5. Assign message intent and offer per stage