Retargeting and the honest limits of attribution

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.

---