M9.L5 · Paid Advertising
Retargeting and the honest limits of attribution
What you'll be able to doDesign a consent-compliant retargeting layer and critique attribution claims, adjusting decisions for what post-cookie measurement genuinely cannot know.
Learn
The warmest audience and the least honest number
Retargeting — showing ads to people who already visited, watched, or bought — is usually the cheapest, best-converting spend in any account. It is also the spend most likely to be lying to you, because it stands at the end of the queue taking credit for customers that everything else created. Both facts are true at once. This lesson teaches you to use the first without being fooled by the second.
Retargeting, plainly
Three pools you can advertise to:
- Site visitors — collected by a pixel or tag on your pages. Requires consent; more on that in a moment.
- Engagers — people who interacted with your content on the platform itself. Platform-side, no pixel needed.
- Customer-list matches — your Module 8 list, uploaded hashed. Using addresses for ad targeting is a processing purpose, and M8.L2's question applies: what is your lawful basis for this use?
The craft is short and mostly about restraint. Message people as warm: answer the objection that stopped them (your Module 2 research named it), show proof, offer the next step — don't replay the cold ad at someone who already saw it. Cap frequency: being followed by the same serving bowl for six weeks is brand damage, not marketing. Exclude recent buyers. Set a membership window — 30 to 90 days — which is also data minimisation doing its job. And size it honestly: the pool is only as large as the traffic feeding it, which is why retargeting is typically a small slice of budget — profitable, and structurally incapable of scaling.
Consent decides who your pixel can see
In the EU, advertising pixels and identifiers require prior consent — that's the ePrivacy rules plus GDPR, the same pair from M8.L2. A visitor who declines your banner never enters your retargeting pool. Not "enters it later", not "enters it anonymously". Never.
Size the effect with the cast. Ostara's banner is accepted by roughly 60% of visitors: of 6,500 monthly sessions, about 3,900 are pixel-visible — a workable pool. Herzog's ~800 visitors at an assumed 50% consent make a pool of about 400 people — too small to be worth more than a token campaign, which is exactly what Tomas runs: one modest objection-and-proof ad, about €25 of his monthly €150, energy spent on search instead. Right-sizing is the decision.
One posture point, stated once and plainly: a third to a half of your visitors being invisible to advertising systems is the law working as intended, not a problem to engineer around. Plan with it.
Modelled, not observed
The gaps consent creates don't show up as gaps. Platforms fill them by modelling: Google's Consent Mode, for instance, uses the behaviour of consenting visitors to statistically estimate conversions among the non-consenting. Apple's App Tracking Transparency prompt (2021) removed a large share of iPhone signal, and the platforms model around that too. So a growing fraction of the "conversions" on your dashboard were never observed — they are estimates, produced by the party selling you the ads.
Modelled numbers are not worthless — they're often the least-bad steering available. But there are now two kinds of numbers on your dashboard, and the marketers worth employing know which kind they're quoting: observed (a consented click, a tracked purchase) and modelled (a statistical guess wearing the same font). Your boss deserves to know too. That distinction is the heart of this module's honesty memo.
The 3.1 that isn't — Ostara, dismantled
The pack's canonical fact: Ostara's Meta ads report ROAS 3.1 — last-click attributed, and mostly retargeting (about €250 of the €300 monthly spend). Lesson 4 said 3.1 comfortably clears her 1.72 break-even. Now look closer, one line at a time.
What the dashboard claims:
€300 spend × 3.1 = €930 attributed revenue ≈ 12 orders (930 ÷ 78)
Twelve of the month's 91 orders, credited to ads — almost all by the retargeting line, because last-click attribution hands the credit to whoever touched the customer last. Retargeting, by construction, is always last: it only ever speaks to people already on their way. The prospecting, the organic Instagram that brought them, the gift-guide blog — all invisible at the moment of credit.
Then Mateus adds one crude, robust, GDPR-simple question to checkout: "How did you hear about us?" First month, 73 of 91 orders answer: Instagram 34, a friend or gift recipient 15, gift-guide blog 8, search 7, "saw an ad" 5, can't remember 4.
Twelve claimed. Five remembered. Both numbers are biased — dashboards over-credit (last-click, modelling), memory under-credits (people forget ads they acted on) — so the truth sits somewhere between. Now watch how much rides on which end is nearer:
If 12 orders are truly incremental: CPA = 300 ÷ 12 = €25
12 × €45.24 first-order margin = €543 — the spend earns its keep
If only 5 are incremental: CPA = 300 ÷ 5 = €60
€60 > €54 CLV — losing money on every customer, at lifetime value
The same €300 is either fine or forbidden, and no dashboard in her account says which. That is the honest state of attribution in a post-cookie world, and pretending otherwise is how budgets die politely.
Incrementality: the question behind every dashboard
The only question that matters is: what happened because of the ad that wouldn't have happened anyway? Large advertisers answer it with holdout experiments — geo splits, where matched regions get ads or silence and the difference is measured. You cannot run a proper geo-holdout on a €500 budget, and this course won't pretend you can. But you can hold the concept — every attribution claim should be read as "credited", never "caused" — and you can afford the SME toolkit:
- The blended check. Total marketing spend ÷ total new customers, set beside the platforms' claimed CPAs. When they disagree, the bank account is right.
- The checkout question. Crude, robust, free. It under-counts ads; say so when you quote it.
- The pause test. Meta claims 12 of Ostara's 91 orders. Pause it for two weeks: if orders barely move, the claim was mostly credit-taking; if they drop toward 13%, it was real. Cost of the test, at worst: a handful of genuinely incremental orders. One caveat a seasonal business must respect — never run it into November; the pack's own seasonality would swamp the signal.
Triangulate with two of these, pre-commit what would change your mind, and you are being more rigorous than most agencies. The durable posture: attribution precision was always partly an illusion — the assets that survive its loss are your first-party list (Module 8), consented data, and creative that works. Plan like it.
Do
Exercise 9.5.1 — The retargeting layer and the honesty caveats
Add the retargeting layer to your €500 plan — or justify skipping it — and write the three honesty caveats you would give your budget-holder about the results. Then assemble the Module 9 Project below.
Write these down — in your plan document, or on the worksheet at the end of this lesson.
| What to write | Guidance |
|---|---|
| Pool estimate | Monthly site visitors × an assumed consent rate (30/50/70%) = your retargetable pool. Do the multiplication. A pool under ~500 means skip, or token spend — and that's the right answer, not a failure |
| Your decision | Include · skip for now — both respected |
| If include | The budget in € from the €500 — at most 25%, and 15% is the habit — plus a message angle in 20–60 words that answers a named objection, not the cold ad replayed; a frequency cap per week; a membership window of 30/60/90 days; and your exclusions, converters at minimum |
| If skip | 30–80 words, referencing pool size or funnel readiness |
| Consent dependency | 20–60 words: how does your consent banner feed — or starve — this audience? |
| Three honesty caveats | 15–50 words each: what your reported results will and won't be able to claim — modelled conversions, last-click bias, invisible channels |
| Triangulation pick | One or two of: blended check · checkout question · pause test · geo split — with a 10–40 word implementation note. If you pick geo split, first ask honestly: feasible at your size? |
Where this goes: the Paid Advertising — Measurement & attribution — section of your Marketing Plan. It feeds Module 10's measurement plan, and together with your M9.L1–L4 artefacts it makes up the Module 9 Project.
Check
Rubric
Mark your own work against these criteria.
| Criterion | 8–10 | 5–7 | 1–4 |
|---|---|---|---|
| Right-sized decision | Include/skip follows from your own pool arithmetic | Defensible, but pool numbers not really used | Retargeting included because "it's cheap" against a 200-person pool |
| Consent dependency understood | States plainly who the pixel cannot see, and plans with it | Consent mentioned, consequence vague | Treats consent as a banner formality |
| Caveats are real | Three caveats a sceptical CFO couldn't dismiss | Caveats present but boilerplate | "Results may vary" |
| Triangulation implementable | Picked method could run this month, as described | Method right, implementation hand-waved | Promises a geo-holdout on €500 |
Pass: 5+ on every criterion.
Quiz — 4 questions
1. An EU visitor declines your cookie banner. For your pixel-based retargeting audience, they are…
- a) Added anyway — legitimate interest covers advertising
- b) Added after 30 days
- c) Never added — advertising trackers require prior consent, so refusers are invisible to the pixel ✔
- d) Added, but shown fewer ads
Why: ePrivacy plus GDPR make prior consent the gate for advertising identifiers. A refusing visitor never enters the pool — this is the law working as intended, and plans must be sized for it.
2. Ostara's dashboard credits Meta with 12 orders; her checkout question finds 5 buyers who recall an ad. The most defensible reading is…
- a) The dashboard is right — software doesn't forget
- b) The survey is right — customers don't lie
- c) The truth sits between: last-click and modelling inflate the 12, memory deflates the 5 — so judge the spend on blended numbers and test incrementality ✔
- d) Average them and report 8.5
Why: both instruments are biased in known directions. The honest move isn't picking a favourite — it's naming both biases and using a third check (blended CPA, pause test) to break the tie.
3. A "modelled conversion" on your dashboard is…
- a) A conversion verified twice
- b) A statistical estimate filling a consent- or signal-shaped gap — useful for steering, not proof ✔
- c) A conversion from a mobile device
- d) Fraud, and should be reported
Why: consent mode and post-ATT reporting fill unobservable gaps with estimates learned from observable users. Know which kind of number you're quoting — and make sure your boss does too.
4. Retargeting shows the best CPA in the account. Why is moving the whole budget into it a mistake?
- a) Platforms penalise single-campaign accounts
- b) Retargeting only harvests warm demand that other activity created — starve the top of the funnel and the warm pool empties ✔
- c) Retargeting is illegal above 20% of budget
- d) It isn't a mistake — always fund the best CPA
Why: retargeting's CPA looks good because it stands last in line, taking credit at the cheapest moment. It's a harvester, not a source. Judge prospecting and retargeting as one system, on blended numbers.
Advance
Module 9 complete. You can price attention, buy intent, create demand, forecast a customer's cost before spending a cent — and you know exactly how far to trust the numbers that come back. That last part is the rarest skill in this industry.
Next: Module 10 — measurement, properly. The pack's oldest arguments — the trade show nobody can prove, the churn numbers that disagree, the 3.1 that isn't — finally get settled with real instrumentation.
Mark your own work
| Good | Not yet | |
|---|---|---|
| Pool arithmetic done | Visitors × consent rate, written down, decision follows | Retargeting included on reputation |
| Warm message | Answers a named objection with proof | The cold ad, shown again |
| Caveats specific | Modelled vs observed, last-click bias, invisible channels — named | "Attribution isn't perfect" |
| Triangulation runnable | Your pick could start this month | Requires a data scientist you don't have |
Worksheet
THE SCHOOL OF NET MARKETING
Lesson 9.5 — Retargeting and honest attribution
MY POOL
Monthly site visitors ________
× consent rate (30/50/70%) ________ % ☐ E
= retargetable pool ________
Under ~500? Token spend or skip — and that's the
right answer, not a failure.
MY RETARGETING LAYER ☐ include ☐ skip (why: ________)
Budget € ______ (≤15% of the €500 is the habit)
Objection answered: _________________________________
Frequency cap ____ /week · window ☐30 ☐60 ☐90 days
Excluded: recent buyers + ___________________________
THREE CAVEATS FOR MY BUDGET-HOLDER
1 ___________________________________________________
2 ___________________________________________________
3 ___________________________________________________
(modelled vs observed · last-click bias · who the
pixel can't see · channels no model sees)
TRIANGULATION — pick one or two
☐ Blended check: total spend ÷ total new customers,
beside the platforms' claims. Disagreement? The
bank account is right.
☐ "How did you hear about us?" at checkout/booking
(under-counts ads — say so when quoting it)
☐ Two-week pause test (never during peak season)
☐ Geo split (be honest about whether you can)
REMEMBER
Credited ≠ caused. Retargeting stands last in line.
Judge prospecting + retargeting as one system.
Module 9 Project: assemble worksheet + this layer +
the ≤400-word honesty memo. Designed, never spent.
Next: Module 10 — measurement, properly.
theschoolofnetmarketing.com/learn/module-10