Remarketing Playbook

1The play, in one page

Two channels, two different jobs. Meta has a switched-off campaign that was making money and should be turned back on. Google needs a small, tiered abandoner ladder built on audiences that already exist. Neither is expensive. Both were blocked on one measurement fix — which landed on 11 August and now needs one verification pass (§5).

The strategic read — remarketing is not the growth lever, it's the margin lever

Before the recommendation, the thing the account is actually doing. In the last nine days paid spend was £592.04 across both channels and produced 6 attributed sales — £98.67 a sale. But look at where the money sat:

Three quarters of the paid budget bought leads, not customers. That is not an accident and it is not a creative problem — it is what we instructed the platforms to do. All three live Meta campaigns are OUTCOME_LEADS, and Google's account carries seven biddable conversion goals of which four are not sales (§5). We asked both platforms for leads. They delivered: 306 Meta leads and 72 Google lead-form submissions in nine days. They are good at it.

So what the paid fleet actually recommends

Do not add remarketing on top of an account that is buying the wrong thing. Change what we are buying first, then reallocate into remarketing from money we are already spending. Net new budget: zero.

The logic in one line: we are already paying to create high-intent demand and then dropping it. 146 people opened the checkout in nine days and 129 walked away. Remarketing is the cheapest way to go back for them — but only once the account stops rewarding itself for generating leads it never converts.

What we expect — with the assumptions visible

No forecast is possible from the platforms today, because purchase values aren't passing through correctly (§5). So here is the arithmetic instead, and you can judge the assumption:

[MODELLED] — ~357 abandoners/month [ESTIMATE, 9-day base] × an assumed incremental recovery rate × $86.56, the blended amount quiz-rail buyers actually paid (n=15, 1–10 Aug, live Stripe). The recovery rate is pure assumption and nothing in our data supports a value. Gross, before ~3% Stripe fees and a 9.88% all-time refund rate.

Kill conditions — decided now, not later

• Meta retargeting: if 30 days at ~£5/day produces fewer than 3 purchases, it was a promo-period artefact and it goes back off.
• Google ladder: if after 30 days the touched group shows no measurable lift over the holdout, stop rung 2 first (it's the one costing margin) and keep rung 1 only if it is at least break-even.
• Either channel: cost per sale above £66.05. 🔴 Corrected 11 Aug — this page first said £98.67 and that number could not carry a kill condition. It divided BOTH channels' spend by only Google's sales (Meta made zero in that window), and one of the six "sales" was the phantom £297 promo conversion. Depending on which denominator you pick it lands anywhere between £23.93 and £118.41 — a threshold that would kill or spare the test arbitrarily. £66.05 = Google's own spend ÷ its 5 real purchases, which is the only version where the numerator and denominator describe the same thing. It is platform-claimed, not Stripe-reconciled — only 2 of August's quiz-rail subscriptions carry a gclid at all.

The two calls for Aga

Neither raises the budget envelope. Your cap is roughly $3,175/month across both channels; combined spend for 1–9 Aug was £592.04 (Google £330.25 + Meta £261.79) [LIVE 10 Aug]. Anything that would raise the cap is Aga's decision, and nothing here needs it.

2Who we remarket to — four tiers, not just abandoners

Checkout abandoners are the hottest audience we have, but at ~14 a day they cannot feed a campaign on their own — Demand Gen starves at that volume and frequency burns out inside a week. So we tier by how far down the funnel someone got, and treat each tier differently.

The funnel as it actually is, 1–9 August

The checkout converts at roughly 7.8% — about 9 real buyers of about 116 real arrivals, once the quiz_checkout_smoke test traffic (29 arrivals, 8 “successes”) is stripped. The page first published 11.6% (17 of 146) because that test traffic was counted. That is still a working checkout. The constraint is how few people reach it, not what happens when they do. [LIVE 10 Aug · GA4 pagePath /payments/quiz-checkout]

The four tiers, and what each one gets

[LIVE 10 Aug · GA4 quiz-host screen events, crawler removed · monthly figures are the 9-day count × 30/9, so [MODELLED] and will move]

What about the July list?

This changes what we target. The lifetime sale ran 2–12 July and is gone. The huge cart numbers from that period belong to it and to the remarketing campaign that ran alongside it — they are history, not inventory:

Cart activity fell about 95% when the sale ended. [LIVE 10 Aug · GA4 add_to_cart totalUsers] The Cart abandoners 60D audience still holds 1,900 people, but most of them were captured during that sale and age out of the 60-day window around 10 September. We can use the list — we just can't offer them what they walked away from, and we shouldn't plan around its current size.

3Google — the abandoner ladder

What already exists (don't rebuild it)

Audience size is no longer a blocker. Google dropped the minimum to 100 active users across Search, Display and YouTube in December 2025 — it used to be 1,000 for Search and YouTube. At 1,900 Display / 3,700 Search this list qualifies everywhere.

The ladder

[LIVE 10 Aug · both pages return HTTP 200]

Settings

Do not use Target ROAS. It needs roughly 30–50 valued conversions per 30 days to function and this lane will produce single digits. Revisit only if a campaign passes 30 purchases in 30 days.

The holdout — don't skip it

Hold back a random 15% of the audience from both campaigns. Without it we cannot separate recovered revenue from revenue that would have arrived anyway, and a meaningful share of abandoners return with no intervention at all. A ROAS number without a holdout is not evidence. If Demand Gen won't express the split cleanly, run two weeks on / two weeks off and compare.

Ad copy — paste-ready

Character counts are inclusive and inside Google's limits. Business name: The Movement Athlete. No star glyphs in ad text (policy risk) — "Rated 4.8". Injury wording stays generic, because Google's personalised-ads policy restricts health callouts in remarketing.

Rung 1 · long headlines (90)

Rung 1 · descriptions (90)

Rung 2 · long headlines + descriptions (90) — no trial language

Final URLs — keep the trailing slash, the redirect strips utm_* without it, and never hardcode a _gl= token:

4Meta — turn the paused campaigns back on

This is the highest-value item in the document and it needs no build at all.

Over the last 30 days [LIVE 10 Aug · Meta Marketing API, unified attribution, 11 Jul – 9 Aug]:

Five of the six Meta purchases in the last 30 days came from campaigns that are currently switched off — but read the next paragraph before drawing anything from that. In August the three active campaigns spent £261.79 and produced zero purchases (190 leads). [LIVE 10 Aug · Meta API, 1–9 Aug]

What to do, in order

1. Re-point and un-pause the three retargeting campaigns. They are already OUTCOME_SALES with custom_event_type: PURCHASE on warm audiences. Update creative and landing URLs to the current offer (the $297 lifetime is gone — send them to /start-training/), then switch on at a small budget.
2. Refresh the purchaser exclusion — it already exists. 🔴 Corrected 11 Aug: the ad sets DO exclude customers (lifetime_and_current_subscribers.csv, Purchase Last 730D, the lifetime CSVs). The defect is that those files are dated July 2024 and September 2025 — roughly a year stale. We are paying to advertise to people who already bought.
3. Leave cold prospecting on Lead for now. Cold Purchase optimisation needs roughly £13,800/month to exit the learning phase — about four times the entire envelope. The standard move is to step cold back to InitiateCheckout (~£1,200/month, which fits) while retargeting stays on Purchase. That is a separate decision from this playbook; flag it, don't bundle it.

Why Meta and Google don't do the same job here. Meta's warm pool is far larger than Google's abandoner list — CRM-matched audiences run to roughly 75,000 addressable people. So Meta carries broad warm retargeting, and Google runs the tight, high-intent abandoner ladder. That split is why we don't need both channels chasing the same 430 people.

5The measurement fix — two done, one that was never broken

What the account was doing — the before state

Google Ads reported 76 conversions for 1–9 August. Only 6 were sales:

[LIVE 10 Aug · Google Ads API, conversion_action_category, 1–9 Aug]

The cause was that ADD_TO_CART, BEGIN_CHECKOUT, SUBMIT_LEAD_FORM and DOWNLOAD actions were all Primary, so every Smart Bidding campaign was being told a lead is a conversion — and bought accordingly.

Done  1 · The non-sale actions are Secondary

All four are now primary_for_goal: false, so they keep reporting and stop driving bids:

[LIVE 10 Aug 21:2x UTC · Google Ads API v24, conversion_action]

Not one enabled action outside the PURCHASE category is Primary any more. Every remaining Primary action is a PURCHASE action. The few Primary lead actions still on the books are all status: HIDDEN app-side actions, which do not fire on the web.

Done  2 · The five $297 promo actions are Secondary — and a correction

All five are now Secondary and still enabled, so the history is preserved: 7663033877 · 7663405149 · 7663412083 · 7670450541 · 7670450544.

Reframed  3 · The value pass-through was never broken

That is the right read, and the data agrees. Purchase | Nic - Built on July 2026 | All Success Pages (7685639495) does pass value — when the page it fires on sends one:

[LIVE 10 Aug · Google Ads API, conversion_action 7685639495, 1 Jul – 10 Aug]

The reason it is intermittent rather than simply off: the action fires on "All Success Pages", and there are five different ones live. GA4 records purchase events on all of these in the last month:

[LIVE 10 Aug · GA4 purchase events by hostName + pagePath, 15 Jul – 10 Aug]

So the remaining job is not a tagging job. It is: get every buyer landing on a success page that passes a value — by retiring the old ones, repointing them, or adding the value push to whichever must stay. The legacy checkout.themovementathlete.com/success.html is the clearest case, still taking real purchases on a host outside the new estate.

Most of §5 was already found on 7 Aug and written into NIC_EXECUTION_PACK_2026-08-07.md. This pull confirms it on a fresh window — it had been running for another month before fixes 1 and 2 landed on 11 Aug.

6Your build order

Four jobs, in this order. Each one says what it is for, what happens if it is skipped, the exact steps, and how you know it worked. Job 1 is independent — you can ship it today without waiting for anything else.

Total time, and what it depends on

The validation column is your time checking what we've written against the real accounts, correcting it, and testing after it's live — not an afterthought. If any step here doesn't match what you actually see in the account, stop and say so rather than working around it; a step that's wrong on screen usually means we've misread something upstream.

7Every number, and how to re-derive it

Everything in this document regenerates with one read-only command:

It writes tools/analytics/data_remarketing_playbook.json. It issues SELECT and GET calls only — it never changes a campaign, budget, bid, audience or conversion action.

Rules this pull obeys

• GA4 is always host-filtered; totalUsers, never eventCount, for funnel steps; dimension_filter, never a top-N ordering (ordering plus a limit silently truncates and has produced confident wrong answers here before).
• Google Ads conversions (bid-counted) and all_conversions (includes engaged-view) are different columns. Both are carried and they are never summed.
• The YouTube crawler is reported separately and subtracted, never silently folded in.
• app-test purchases are excluded from every purchase count and reported on their own.
• Meta uses explicit time_range with unified attribution — date_preset excludes today and has caused a real reporting discrepancy before.
• 🔴 NEVER SUM META ACTION TYPES. Meta reports one conversion under seven labels — purchase, omni_purchase, offsite_conversion.fb_pixel_purchase, onsite_web_purchase, onsite_web_app_purchase, web_in_store_purchase, web_app_in_store_purchase — with identical counts and values. Take exactly ONE. lead is the aggregate and offsite_conversion.fb_pixel_lead is a SUBSET of it. Summing two of them is what published £2.41 and 306 leads on this page.
• 🔴 Quote the campaign's whole life, not the window that flatters it. The 11 Jul – 9 Aug window held 1.7% of retargeting spend and 3 of 7 campaigns.
• GA4 checkout counts strip quiz_checkout_smoke and test sources as well as the YouTube crawler — the smoke test alone was 29 arrivals and 8 "successes".

Cannot be verified from here

• Whether the July remarketing produced revenue. Purchase tracking was not capturing during the sale. Only a Stripe reconciliation against that window would settle it.
• Whether the app-test purchase events involved real money. Needs a live-mode Stripe check on those timestamps.
• Meta pixel deduplication. The pixel stats endpoint reports received events and cannot cross-filter dedupe status, so server-vs-browser counts are consistent with dedupe working but do not prove it.
• Forecast revenue for the ladder. At ~357 abandoners/month [ESTIMATE], with purchase values still only passing from some success pages, any figure would be invented. It becomes measurable once every buyer lands on a value-passing success page (§5 fix 3) and the holdout has run 30 days.