Trafilea PPC Handover

Carlos Carrero's last working day is Friday 4 September 2026. This page records what he ran, how often, and what stops running when he leaves. Part 0 is the only section a manager has to read.

0. Read this first

Dated obligations

Serving with no end date — act before Friday

Fourteen promo-bearing assets are enabled with no end date at all. An asset with an empty end date serves forever; on Microsoft Ads the redemption date (PromotionEndDate) does not stop delivery, only Scheduling.EndDate does. The full checklist is in reports/handover/data/MITIGATION.md. The ones customers can see today:

Four decisions Trafilea owes itself this week

1. Who runs the bi-weekly promo cycle. One script writes to SHM Bing ×4 markets, TK Bing, SHM Google and TK Google in a single run — so it now spans all three new owners. Splitting it by account is possible but the detection keyword lists are global, and they drift apart by the second cycle. That drift is the documented cause of "EOFY Sale" serving from June into August.
2. Who consolidates the weekly WoW deck. Fourteen segments, previously one person's Monday. It now splits three ways and has to converge before the deck ships.
3. Which mutating automations get migrated, and which are switched off on purpose. Four jobs write to Google Ads daily from a laptop that is being returned. Letting them die silently leaves nobody knowing they ever existed.
4. Who re-issues the credentials that authenticate as Carlos. Four n8n workflows keep running server-side but sign in with his tokens; every published report URL sits on his personal Cloudflare account.

1. Scope and ownership

Measured against the APIs on 1 September 2026, not from memory.

Who takes what

Built by Carlos, for other people's channels

These monitor or report on channels outside his scope. They need a decision of their own: adopt, or retire.

• com.ppc.gads-bid-change-history-shm — tracks bid changes on SHM USA Non-Brand Search, Shopping and PMax
• n8n SHM&TK vs Antares — Daily — covers every SHM channel and is consumed by the wider team, not just Carlos
• The SHM USA ROAS-drivers report and the shm-pacing-lever skill — both cover Shopping and PMax

2.1 Shapermint — Google Search Brand

~41 enabled campaigns across US, CA, UK, AU and OC, in two accounts (1037305439 for US/UK/AU, 9328705246 for Canada). The largest block of spend in this handover.

Reading a campaign name

SR_<country>_[WW_]BR_ShaperMint_<variant>. SR is Search, BR is brand, NB is non-brand (not yours here). The variants each catch a different query shape:

• Core_2 — the main exact-match brand campaign. The single most important campaign in the account.
• Modifiers / Modifiers_CPC — brand plus a qualifier ("shapermint reviews", "shapermint discount").
• Products, Bra_Products, Cami_Products, Swimsuits_Products, Tier2_Products — brand plus a product line.
• Core_Broad — the broad-match arm. Structurally the weakest performer; see below.
• Reviews — review intent.

Bidding philosophy — read this before "fixing" anything

Where it stands — 30 days to 31 August, ROAS CDP

Top Impression Share by campaign

Three live experiments — each needs a verdict

The account also carries roughly 25 zombie experiments dated 2020–2024 with end dates in 2037. They belong to old video and audience tests outside this scope. Ignore them; they are noise, not risk.

Open problems and what I would do

1. SR_US_BR_ShaperMint_Core_Broad is losing a third of its top-of-page to rank (34.7%) at 52.4% Top IS. It is the worst-performing brand campaign in the account and has been for months. Either commit to it with a real bid and tighter negatives, or fold its budget into Core_2 and Modifiers. Half-funding a broad brand arm is the worst of both.
2. AU Modifiers at 71.1% Top IS, 17% lost to rank, while AU Core_2 sits at 92.8%. Australia is also the fastest-growing market (+37% spend, ROAS CDP 16.33). Modifiers is the obvious next place to buy.
3. USA and Canada spend fell 16% and 24% month over month while UK and AU grew. Confirm this was deliberate pacing rather than a bid or budget drift — the bid-change history sheet covers SHM non-brand, not brand, so brand changes are not automatically logged anywhere except Notion.
4. Apply the LowCPC_CAP result rather than letting it expire. It is the clearest efficiency win available and it dies on 7 October by default.

2.2 Shapermint — Microsoft Ads

Account 150000172, 4 markets. 14 active campaigns — all Search, all brand. 148 paused Search campaigns and every Shopping, PerformanceMax and Audience campaign are paused as of 1 September.

Where it stands — 30 days to 31 August, ROAS CDP

Structure notes

• Every active campaign runs Enhanced CPC except SR_UK_WW_BR_ShaperMint_Core_Experiment (Max Conversions) and SR_US_BR_ShaperMint_Core_Broad, which reports no bid strategy at all — worth a look.
• Daily budgets total roughly $9,350 against ~$1,100/day of actual spend. Budget is not the constraint anywhere in this account; bids and impression share are.
• Of 448 ad extensions in the account library, only 56 are attached to the 14 active campaigns. Library status "Active" does not mean serving.

Open problems and what I would do

1. Six callouts are serving with no end date, including "Black Friday Early Access" in September and "Autumn Sale in UK". Fix these first — they are visible to customers today. Full list in MITIGATION.md.
2. Canada and Australia are underweight. ROAS CDP 15.91 and 13.38 against USA's 8.71, on $1,629 and $923 of spend. If the impression share headroom is there, these are the cheapest incremental orders in the portfolio.
3. Decide whether Bing Max stays dead. It was paused at ROAS CDP 1.67 USA and 0.97 CA. That was the right call on those numbers, but nothing documents whether it is a permanent structural decision or a pause pending a fix.
4. CDP is the source of truth here. Microsoft-reported ROAS is directional only; it disagrees with Tableau and always has.

2.3 Truekind

Google account 1871711997: 17 enabled campaigns across Search Brand and Non-Brand, Shopping and two Performance Max. Microsoft account 141454956: 2 active Search campaigns, everything else paused.

Where it stands — 30 days to 31 August, ROAS CDP

Structure notes

• Brand Search runs Manual CPC across US and CA (Core_2, Modifiers_CPC, Products_CPC, Coupons, Reviews), with Core_Broad on Max Conversion Value.
• Non-Brand: SR_US_TK_NB_Bras_2026 (Max Conv Value), SR_US_TK_NB_Competitor (tROAS), plus a broad-match experiment.
• Two Performance Max campaigns, brand and non-brand, both Max Conversion Value.
• Two daily negative-keyword automations write to this account — one for NB Search, one for Shopping NB. Both stop on Friday. See Part 5.1.

Open problems and what I would do

1. Bing Shopping at ROAS CDP 0.95 on $6,017. It is below break-even and it is the one line in this portfolio that loses money outright. It has since been paused — confirm that is intentional and permanent rather than an unlogged change.
2. SR_US_TK_NB_Bras_2026 holds 10.0% Top IS on $4,724 of spend. Either it is bid far below the auction or the keyword set is too broad for its budget. This is the clearest diagnostic to run first on this account.
3. Google Shopping fell 31% month over month while PMax grew 8% at a lower ROAS CDP (1.83 vs 1.96). If that shift was deliberate it is moving spend the wrong way; if it was not, PMax is absorbing Shopping's inventory on its own.
4. Brand Core_2 spends almost nothing ($136 in 30 days at 80.5% Top IS) while Products_CPC carries $5,880. Worth confirming the brand structure is splitting queries the way it is supposed to.

2.4 The Spa Dr

The smallest account and the one with the clearest open problem. Google 9662795380: 9 enabled campaigns. Microsoft 141948660: 3 active Search campaigns.

Where it stands — 30 days to 31 August, ROAS CDP

Structure notes

• Brand Search is Manual CPC (CPC_Core, Modifiers, Products, Reviews_Script) plus a Max Conv Value Reviews arm. SR_US_BR_SpaDr_CPC_Core holds 90.9% Top IS.
• Non-Brand: SR_US_NB_SpaDr_Reviews (tCPA, 65.2% Top IS) and SR_US_NB_SpaDr_Serum_Group1_Advertorial (Target Spend).
• Shopping Non-Brand was paused in April 2026; its negative-keyword automation still runs against paused campaigns.
• PMax excludes traffic through a campaign-attached shared negative list (sharedSets/12135164772), never per-term. Always add to that same shared set.

Open problems and what I would do

1. Cut or fix Demand Gen this month. $3,805 for two orders, growing. If it is being kept for reach, that decision needs to be written down somewhere and given a budget ceiling.
2. Shopping fell 59% but returns ROAS CDP 7.80 — the best return in the account, shrinking. Find out why before it goes to zero.
3. Google Search grew 30% at ROAS CDP 4.70. Healthy. This is where the Demand Gen budget most plausibly belongs.
4. Watch for zero-conversion days. This account has a known tracking anomaly — occasional Saturdays and mid-week days with normal impression share and clicks but no conversions. Verify in GA4 or Shopify before treating one as a performance problem.

3. Operating cadence

What ran, how often, and — the column that actually matters — what happens if it is skipped. Filter by a name above and the rows narrow to that person's work. Anything automated is covered in Part 5; this is the rhythm, including the manual work that no script does.

3.1 Daily

3.2 Weekly — Monday is the heavy day

3.3 Every two weeks

3.4 On demand

4. The promo cycle

Every ~2 weeks a new promotion goes live and sitelinks, promotion extensions and callouts have to change across six accounts. It is the highest-frequency manual process in this handover and the one with the worst failure record. It also has no owner after Friday.

4.1 The rhythm

1. Marketing announces the promo in a dedicated Slack channel, #mktcampaign_<brand>_<promo>_sale. Channels are revived, not recreated, so an old channel suddenly waking up is the signal.
2. Slack is the only source for the promo dates. The website never publishes them. Read them off the go-live post.
3. Read the live collection URLs off each market's homepage (see 4.2).
4. Edit the config block at the top of the script: name, start date, end date, per-market URLs, discount, and the detection lists.
5. Dry run. Read the diff line by line.
6. --push, then log to the Notion changelog.
7. Run the post-push audit. Not optional — see 4.4.

4.2 Read the URLs off the live site every cycle

Never reuse last cycle's slug. Each market homepage embeds a "saleUrl" naming the collection the site is currently pushing — public, no auth, no VPN:

Then validate each candidate: the <title> must name the promo and the page must carry a full product grid. An empty collection still returns HTTP 200.

4.3 The current cycle: Warehouse Clearance, and why it needs two pushes

Verified live 1 September 2026. Discounts: SHM 70%, TK 60%.

4.4 The detection-list trap

Two of the three keyword lists are not URL-guarded, so any keyword that matches the current cycle's own text makes the script re-flag and re-create the assets it just made, on every run.

• OLD_PROMO_URL_KEYWORDS — URL-guarded. The current cycle's slugs can safely live here.
• OLD_PROMO_TEXT_KEYWORDS — not guarded. Rotate every cycle.
• OLD_CALLOUT_KEYWORDS — never guarded, because callouts carry no URL. This list is the only detection path for them.

For this cycle: remove "warehouse sale" from both text lists before pushing, and add "labor day" / "labour day" back as those markets leave Labor Day. Keep "/warehouse-sale" in the URL list — that one is guarded. Ended promo names also belong in STALE_TO_PAUSE in apply_brand_promo_headlines.py; the two lists have drifted apart before.

4.5 The post-push audit is mandatory

The push log is blind to what the script never detected. Pull enabled campaign_asset and customer_asset by API and assert three things: zero prior-promo names still enabled, zero cross-country URL leaks, and every sitelink and callout carrying a non-empty end date.

4.6 The coverage gap

The updater only swaps assets that already carry a previous promo. A campaign that never had a promo sitelink or callout is skipped forever. Campaign-level assets override account-level ones per field type, so a campaign with its own evergreen sitelinks but no promo sitelink inherits nothing from the account.

Truekind had no promo callout anywhere until August 2026. Fill gaps with scripts/tk_promo_coverage_fill.py (dry-run by default, idempotent, reuses the live promo sitelink rather than duplicating it) immediately after the main push. Known open gap: SHM account 1037305439 has no account-level promo fallback, so its Shopping US/AU and Demand Gen campaigns receive nothing.

4.7 The safety net, and that it dies on Friday

promo_drift_watchdog.py runs every two hours during working hours and answers one question: do the landing pages our ads point at still match the promo the website is actually running? It is read-only, notifies only when the drift changes, and it exists because SHM UK moved to the Summer Bank Holiday Sale on 13 August 2026 and paid search found out on the 20th — seven days of UK brand ads pointing at the wrong collection.

It runs on the Mac. After Friday there is no watchdog. Whoever owns the promo cycle should either migrate it to n8n or add a manual mid-cycle check to their calendar.

5. Automations and continuity

Twenty Trafilea jobs are scheduled on Carlos's Mac; nineteen are loaded. Four workflows run server-side in n8n. Every job list below was read from launchctl and the n8n API on 1 September 2026 — not from the workspace documentation, which has been wrong about job state before.

5.1 The four that write to Google Ads

Each miner writes a staged candidate file that its apply job consumes the next morning. The miners are read-only; only the *-apply jobs mutate. Every applied batch is written to reports/<brand>-negatives/…-applied-YYYY-MM-DD.json, which is the rollback record: find the negatives there and remove them in the Google Ads UI.

5.2 Runs server-side in n8n — survives the Mac, not the credentials

Rotating the Google Ads credential without a human re-consent is possible today via scripts/n8n_gads_cred_rotate.py — but it injects the Mac's refresh token, so it stops being a fix the moment the Mac is gone. The durable answer is to re-mint both credentials under a Trafilea-owned service identity before Friday.

5.3 Reports and trackers that stop on Friday

5.4 Tied to personal accounts — Trafilea cannot administer these

The cheapest fix for the third row is to repoint the reports at a channel instead of a person. The first two require either transferring the projects or re-hosting them under Trafilea infrastructure.

5.5 Out of scope

The same machine runs scheduled jobs that have nothing to do with Trafilea. They are not listed here and nothing in this handover depends on them.

6. Data rules

Every rule here exists because someone got it wrong first. Each one is stated with the incident that produced it, because a rule without its reason gets dropped the first time it is inconvenient.

6.1 The CDP is the only source for business metrics

Say "ROAS CDP", never a bare "ROAS". The two numbers differ and readers need to know which one they are looking at.

6.2 source_category must be a list

The parameter became multi-value on 12 May 2026. Passing a bare string returns zero rows silently and any script polling for results simply times out.

The shared helper wow_lib.redash_fetch coerces a string into a list as a safety net. Any new script hitting 2635 directly must use the list form.

6.3 Join on campaign_id, never campaign_name

Rows also split by ad, adset and strategy (Acquisition and Retention). Sum them all.

6.4 Funnel definitions match Tableau, not intuition

• ATC% = add-to-carts ÷ users
• IC% = initiate-checkouts ÷ add-to-carts
• Checkout CVR = orders ÷ initiate-checkouts

The user column is users_1d. A bare users returns zero without erroring.

6.5 The platform-vs-CDP gap is attribution, not a bug

Platform-reported conversion value runs about +4% above CDP overall, roughly +2% on brand and +20% on non-brand. Spend reconciles almost exactly. The gap is the attribution window, and it is expected. Do not "fix" it, and do not mix the two sources inside one table.

6.6 CDP events settle late

A fresh pull of the last day or two will understate add-to-carts, checkouts and orders. Zeros in recent data usually mean lag, not a funnel collapse. Re-pull with max_age=0 before drawing a conclusion, and prefer windows that end yesterday rather than today.

6.7 Redash needs the VPN; the Antares MCP does not

Redash sits behind the Trafilea VPN, and VPN outages are the documented root cause of most stale-report incidents. The antares-performance-metrics MCP tool reads the same warehouse table over the public internet and was verified byte-identical to query 2635 on one brand-channel-week. It is a viable escape hatch — but re-verify equality per brand before moving any production job onto it.

6.8 Google Ads API notes that bite

• Pin REST callers to a supported version. The Python SDK negotiates its own; raw REST (all of n8n) does not. v21 was blocked progressively in August 2026 — the same query alternated success and 400 for hours, which reads like a flaky network unless you unpack the error body. A 400 is not an auth problem; a dead token returns 401.
• campaign.start_date does not exist in v23 — it is campaign.start_date_time.
• There is no flat campaign.target_roas. The real paths are campaign.maximize_conversion_value.target_roas and campaign.target_roas.target_roas.
• Impression-share metrics need campaign.advertising_channel_type = 'SEARCH' in the WHERE clause or they come back wrong on mixed-type accounts.
• Validate any new GAQL before running it: python3 scripts/lib/gads_fields.py check <resource> <fields>. campaign.name, for instance, is not selectable from campaign_asset.

6.9 Microsoft Ads read paths that return zero instead of erroring

Three calls in the extension API look like "this account is empty" when they are actually being read wrong. All three were hit while building this document:

• GetAdExtensionIdsByAccountId returns ArrayOflong — read .long, not AdExtensionIdentity.
• GetAdExtensionsByIds returns .AdExtensions, plural.
• GetAdExtensionsAssociations takes campaign ids as EntityIds plus AdExtensionType — passing AdExtensionIds raises a type error and omitting the type raises a bare SOAP fault. Never wrap it in a silent except.

And the one that matters operationally: Scheduling.EndDate stops delivery; PromotionEndDate does not.

Appendix — where everything lives

The workspace

Everything runs out of a local ppc-workspace directory, mirrored to a private GitHub repository. CLAUDE.md at the root is the operating manual: account registry, API patterns, and a section per automation with its schedule, log path and rollback procedure. It is long, it is current, and it is the first thing to read after this page. Treat its claims about job state with suspicion — it has been stale on that before.

Scripts by function

Registries worth knowing

• reports/links-registry.md — every published report URL.
• reports/wow-reports-registry.md — every WoW deck by week.
• client-context/trafilea/*.md — per-brand running context. The Spa Dr's file is the longest at ~160KB and carries most of the account's history.
• reference/gads-api/ — 1,412 annotated Google Ads fields; query it with gads_fields.py.

The Notion changelog

Every real platform change on a Trafilea account gets logged: bid changes, pauses and enables, budget changes, negative keywords, ad copy, new campaigns, asset additions or removals. Not logged: analysis, reports, audits, or anything touching the workspace itself. Database "Carlos Carrero - Changelog", id 1b5e19cdd4c580d492ddf9ac7a392567, posted through the REST API with NOTION_CHANGELOG_TOKEN — the Notion MCP does not have access to it.

This document

Source lives in reports/handover/: content fragments in parts/, styling and the owner filter in shell.html. Rebuild with python3 scripts/build_handover.py. The data behind every figure is in reports/handover/data/, regenerable with python3 scripts/handover_pull.py {bing|is|cdp|promo-assets|creds}.