Email & CRM marketing: frameworks & methodology

You have ops capacity for one new flow this quarter, not five. Welcome, browse abandon, cart abandon, post-purchase reorder, win-back: the build queue is always longer than the team that has to wire it up. Choose by arithmetic. Size each candidate as monthly eligible population × trigger rate × expected conversion × average order value, then divide by build effort in engineering days. A browse-abandon flow on a site with 400,000 monthly sessions, 30% of visitors identifiable, reaches roughly 120,000 people a month. A win-back aimed at 8,000 lapsed buyers cannot catch that on volume even at three times the conversion rateconversion rateThe percentage of visitors or prospects who complete a desired action (purchase, sign-up, contact form), calculated as conversions divided by total opportunities.View full definition →. One rule overrides the maths: build welcome first, because it is the only flow whose output (a first order, a first product action) creates the eligible population every other flow feeds on.

Lifecycle mapping before campaign planning

Before a single sequence gets built, write the stages down with edges you can actually query. A usable mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → gives every stage four things: an entry condition expressed in data you hold, an exit condition, a maximum dwell time, and exactly one flow that owns the contact while they sit there. "Stage: registered, no order. Entry: account created, order count zero. Exit: first order, or day 45. Owner: onboarding stream." That is the level of precision that stops two teams messaging the same person about different things on the same Tuesday.

Four layers sit on top of the map: segmentationsegmentationDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition → (who qualifies), sequencing (order and interval), personalisation (which content block), measurement (which revenue metric, not which engagement metric). The layer teams skip is dwell time, which is why contacts rot in "new lead" for eleven months, still receiving week-two education emails.

RFM: scoring, tiers and where it breaks

RFM (recency, frequency, monetary value) was formalised by Arthur Hughes in database marketing work in the 1990s and it still outperforms most machine-learned segments on cost of maintenance. Score each contact into quintiles on all three axes and you get 5 × 5 × 5 = 125 cells. Nobody staffs 125 creative treatments. Collapse them into six or eight named tiers: champions (555, 554), loyal, promising, at risk (155, 145), hibernating, lost. Tier decides two things: cadence and offer depth. Discount the at-risk tier, never the champions who buy at full price anyway.

Amazon has run this logic at scale for years, with reorder prompts timed to the consumption interval of the item rather than to a marketing calendar. Subscribe & Save is that same maths made explicit to the customer.

Where RFM breaks matters more than where it works. If your purchase cycle is longer than your scoring window (mattresses, tyres, enterprise renewals), recency quintiles just sort contacts by acquisition date and tell you nothing. In subscription businesses, purchase recency is automatic and meaningless; substitute engagement or product-usage recency. In B2B the monetary value sits on the account, not the contact, so score the account and inherit the tier downward. And beware seasonality: score in January and everyone who bought a Christmas gift looks like a champion. Rescore monthly, keep the prior score, and act on movement between tiers. A champion sliding to "at risk" is a better churn signal than any absolute level.

Trigger logic and the suppression layer

A time-based sequence sends email 2 on day 3 whatever the contact does. A behaviour-based one sends email 2 only on a click, a pricing-page visit, a second session. The second converts better and costs more to maintain, because every trigger needs four parameters specified: the event, the qualifying condition, the delay, and the dedupe window that stops the same event firing twice in an hour.

Latency is where these break. Browse abandon fires well at 30 to 60 minutes and looks creepy at 5. Cart abandon at one hour, then 24, is the standard shape. The classic failure: the cart-abandon trigger fires on a session that actually ended in a purchase, because the order event arrived 90 seconds after the trigger evaluated. Fix it by re-checking the exclusion condition at send time, not at trigger time.

Suppression is the layer that keeps the whole apparatus from embarrassing you: unsubscribes and complaints, hard bounces, a per-person frequency cap, contacts sitting in an active sales sequence, anyone with an open support ticket or a pending refund, recent purchasers excluded from acquisition offers, and a quiet period after a service incident. Marketo, which sells marketing automation, ships communication limits that cap sends per person per day and per week, with operational emails flagged to bypass the cap. Copy the architecture even if you use something else: caps belong at platform level, above individual campaigns, because whoever built flow 7 cannot see flow 3. Without that, a customer who ordered, browsed, abandoned and opened a ticket in one evening receives four emails before breakfast.

The activation, retention and expansion arc

Map every track to one of three commercial moments: activation (first meaningful action), retention (continued use and renewal), expansion (upsell, cross-sell, higher tier). Most teams pour their effort into activation and run retention on autopilot.

Duolingo is the sharpest public example of retention messaging treated as product mechanics. Its growth team has written about choosing each learner's reminder time from that learner's own practice history, and about backing off when reminders go repeatedly ignored. The transferable rule: the value of a reminder decays with every one that is ignored, so count consecutive non-responses per contact and treat that count as a suppression variable rather than a reporting curiosity.

Second-order consequence worth budgeting for: retention and expansion tracks tend to be incremental, while acquisition promos frequently discount revenue you would have collected anyway. Hold out 5 to 10% of each track's eligible population as a control and report incremental revenue, not attributed revenue. The first time you do this on a "top-performing" win-back offer, expect the number to fall.

Send-frequency maths

Cadence is bounded from above by the engagement-based filtering the foundations lesson sets out, and from below by revenue targets. Find your own ceiling with a two-line model rather than a debate.

Take a 200,000-contact list. At two sends a week with 0.15% unsubscribing per send, you lose about 600 contacts a week. Push to four sends and unsubscribe rates on the extra volume typically climb; at 0.30% per send you lose about 2,400 a week, 1,800 more than before. Two extra sends at $0.04 revenue per delivered email produce $16,000 a week. If email-attributable value per subscriber is $12 a year, those 1,800 lost contacts destroy roughly $21,600 of future value in the same week. The extra sends lose money while looking like a win in the weekly report, because the cost lands in a different quarter than the revenue.

Then set caps by RFM tier rather than one number for the list: champions absorb four or five contacts a week, hibernating contacts get one a month or a single reactivation attempt before going into a sunset track. The Data & Marketing Association has for years put email's return in the region of $35 to $40 per dollar spent, which is exactly why marginal sends feel free. They are not.

Real-world cases

Amazon's reorder and recommendation emails key off purchase history and item consumption intervals, which is RFM automated: a customer with three electronics orders in six months enters a different post-purchase track from someone with one order two years ago.

Marketo (again, an automation vendor) built its data model around programs and streams with content exhaustion built in: when a contact has received everything in a stream, they stop rather than loop. Ask your platform what happens on day 400 of a nurture track; if the answer is "it repeats", that is a complaint generator.

Duolingo's retention loop turns a behavioural streak into the trigger, so the message has a reason to exist that the recipient recognises. Lifecycle messaging that references a real state (a streak, a cart, a renewal date) survives frequency pressure far better than editorial sends.

CMO action items

• Score the database on RFM this quarter. If you cannot produce quintiles today, that is a data problem to fix before any campaign brief gets written.
• Assign every live track to activation, retention or expansion, with one revenue metric each. Kill anything unassignable.
• Put a per-person frequency cap at platform level, with an explicit bypass list for transactional and service messages, and publish who can approve a bypass.
• Add a 5% holdout to your two highest-volume tracks and report incremental revenue at 90 days.

Common mistakes that kill results

List size obsession comes first. A list of 500,000 at 15% engagement is worth less than 100,000 at 45%, and the large one drags the small one down with it, because filtering decisions are made at domain level.

Second is overlapping triggers with no shared cap. Each flow tests fine in isolation; the damage only appears in the union of them, and it shows up as complaints rather than as unsubscribes, which is the expensive version.

Third is the sales-data gap. If reps do not log activity, behavioural triggers fire on incomplete history and a nurture email lands the morning after a lost-deal call.

Fourth is stopping measurement at engagement. Every track needs a downstream number: meeting booked, trial started, reorder placed, renewal signed. Without it you cannot defend the budget, and you cannot tell a cadence problem from a creative one.