Account Architecture
Architecture looks like organization and acts like physics: every cut you make in an account divides the same conversions across more learning cells, and the machine bids only as well as the cell it learns in. This module is the arithmetic of structure — the granularity dial, the SKAG extinction, the zombie audit, and the one-sentence test every campaign must pass.
What you will learn
Why architecture matters
Account architecture looks like organization and acts like physics. Every campaign or ad-set split divides the same conversions across more learning cells, and automated bidding only bids well in a cell with enough data. So structure is not tidiness — it is the allocation of statistical power. Get it right and the machine learns; slice it too finely and every cell guesses forever.
Most marketers treat account structure as housekeeping — a way to keep campaigns findable and reports tidy. That instinct quietly sabotages performance, because in the automated-bidding era structure does something far more consequential: it decides how much data each part of the account gets to learn from. The machine you met in module 3 is only as good as the conversion density in the cell it is learning inside. Architecture is the thing that sets that density.
Hold the hero figure above in your head for the rest of this module. Same business, same budget, same 600 monthly conversions — split one way, every cell starves on six conversions a month and nothing ever exits the learning phase; split another, every cell gets seventy-five and bids with confidence. Nothing changed except the lines on the org chart of the account. That is the entire argument: structure is how you ration the one thing the machine cannot do without.
The universal hierarchy — and what it allocates
Every ad platform uses the same four-level hierarchy under different names: Account (the business container — billing, pixels, history), Campaign (where budget and bid strategy live — a budget decision wearing a name), Ad set / ad group (the learning cell that pools conversions), and Ad (the creative that competes inside the cell). Knowing what each level is FOR tells you where every decision belongs.
Before you can structure an account well you have to know what each rung of the ladder actually controls — because most structural mistakes are decisions made at the wrong level. The platforms rename these rungs (ad set on Meta, ad group on Google) but the function is identical everywhere. Tap each level for its real job and the move it implies.
Billing, access, pixels/tags, policy history. One per P&L in most cases — splitting accounts splits data history and pixel learning permanently.
Budget and bid strategy attach here. A campaign is a BUDGET DECISION wearing a name — if two things deserve separate budgets or targets, they deserve separate campaigns. If not, they don’t.
Audience + optimization target pool conversions HERE (Meta ad set) or share keyword/intent themes (search ad groups). This is the unit that must clear the ~50/month line.
Creative variants compete inside the cell; the platform allocates impressions toward predicted winners. On social, creative is the targeting (module 5).
An ad set (Meta) or ad group (Google) is where conversions pool and the bidder learns. The most diagnostic number in any account is monthly conversions divided by the number of these cells. Above ~50 per cell, the machine has enough signal to bid well; far below it, the cell is guessing no matter how good your targets or creative are. Architecture’s whole job is keeping cells above that line.
Campaign themes: the four levels, decoded
A campaign is a budget decision wearing a name. The durable ways to group campaigns are by intent tier, by economics (margin or LTV bands), or by funnel job — because each of those genuinely deserves its own budget and target. Geography and the org chart are not on that list: they make reports tidy while taxing the data the bidder needs.
If the hierarchy tells you where decisions live, campaign themes tell you how to draw the lines between campaigns — and this is where most accounts go wrong. The reflex is to mirror something human-legible: a campaign per region, per product line, per manager. The auction is indifferent to all of it. The only splits worth their data tax are the ones that reflect a real difference in how you would fund or target the thing.
Sources: Google — Smart Bidding (data pooling) · Meta — campaign budget optimization. Heuristics labeled RGM analysis, argued at depth in the PSM and PSoc blueprint modules.
Granularity trade-offs
Granularity is a genuine trade, not a virtue. Slicing an account into more cells buys finer control but divides the same conversions into thinner data, so bidding destabilizes. The right level is the coarsest structure that still separates genuinely different budgets and economics — consolidate until cells learn, then split only where a real difference demands it.
There is no universally “correct” granularity; there is only the trade between control and data density, and where your account’s conversion volume puts the sweet spot. The dial below makes that trade physical. Drag it and watch the conversions-per-cell number — the one that decides whether anything learns — fall as the control granularity rises.
The trade is real on both ends (model is RGM analysis). Control rises as you slice; data density and stability fall. The 2016-era answer was maximum slicing (SKAGs); the Smart Bidding era inverted it: consolidate until cells learn, then add structure ONLY where budget or economics genuinely differ.
Inherited structures usually mirror someone’s spreadsheet: a campaign per region per product per manager. The auction does not care about your org chart — and every mirror-line costs data density.
Our test for any proposed split: “will these two things ever deserve different budgets or targets?” If no, they share a campaign. If yes, the split earns its data tax.
Geography almost never passes (one country’s economics rarely differ enough); margin bands almost always do; intent tiers pass everywhere we have tested.
Claim: Single-keyword ad groups (SKAGs) were dominant best practice in 2014-2018 and were dismantled industry-wide once auction-time bidding made pooled conversion data matter more than manual control. Source: Search Engine Land — the STAG/SKAG reversal. Context: Optimal account structure is a function of the era’s bidding technology — when the machine changes, the architecture answer changes.
The dial, and the SKAG extinction
Consolidate when splitting would starve cells of data; separate only when two things genuinely deserve different budgets, targets, or economics. The SKAG era — one keyword per ad group — was the high-water mark of over-slicing, and the industry reversed it when automated bidding made pooled data matter more than manual control. Doctrines age; the arithmetic of data density does not.
The single best case study in account architecture is the rise and fall of the SKAG, because it proves the deepest point in this module: the “right” structure is downstream of the bidding technology of its era. A structure that was genuinely best practice for years became an anti-pattern not because anyone was wrong, but because the machine underneath changed.
For half a decade the performance industry’s consensus structure was the SKAG: one keyword per ad group, thousands of cells, maximum control. It genuinely worked — under MANUAL bidding, where humans set every price and message-match was the only lever. Auction-time bidding inverted the math: the machine needs pooled data more than humans need granular control, and close variants erased the precision anyway. The industry that built SKAGs spent 2019-2022 dismantling them. The durable lesson is not “consolidation won” — it is that optimal structure is a function of the bidding technology of its era. When the machine changes, the architecture answer changes; doctrines age, arithmetic does not. (Era documented across Search Engine Land and PSM’s deep treatment.)
First week on any inherited account, we sort campaigns by last-edited date, oldest first. There is always a graveyard: seasonal pushes from two years ago, “tests” that never ended, migration duplicates.
Zombies cost three ways: they spend, they fragment data, and they bid against your live campaigns in the same auctions (the self-overlap tax).
Protocol: anything unedited for 6+ months gets justified in one sentence by a human or paused that day. Most cannot produce the sentence.
Two trees: inherited vs rebuilt
The platforms differ in names and budget controls (Google ad groups vs Meta ad sets, campaign budget optimization, etc.), but the structural principle is identical: cut once per genuine budget/economic difference, never to mirror an org chart. An inherited account’s tree usually reveals years of accreted splits that fragment data; the rebuilt tree makes every line earn its place.
Abstract principles land harder when you see them as account trees. Below are two structures for the same business: the one we inherited, and the one we rebuilt. Read them side by side — the difference is not aesthetic, it is the difference between cells that learn and cells that guess.
Device splits from 2015, color-level SKAGs, zombie seasonal campaigns still spending, names that document nothing. 31 conversions a month over 96 cells: nothing here has EVER exited learning.
Budget logic = campaign logic: brand fenced (module 3’s TIS quarantine), intent tiers separated because their economics differ, broad fenced as the learn-and-expand engine. Same keywords — 4× the conversions per cell.
The rebuild is not “fewer things” as a virtue — it is one structural cut per GENUINE difference in budget, target, or economics, and zero cuts for anything else. That sentence is the whole module.
Structure follows strategy.
Budget allocation by structure
Budget follows structure, and structure follows economics. Because budget and bid strategy attach at the campaign level, every campaign is implicitly a budget decision — so the campaigns you create should be exactly the things you would fund or target differently. Fence brand and exploration spend so they cannot quietly consume performance budget, and let consolidated cells pool the rest.
Once you accept that a campaign is a budget decision, allocation becomes a structural question rather than a spreadsheet one. The campaigns that deserve to exist are the ones whose money you would steer independently: brand defense fenced on its own (module 3’s impression-share quarantine), high-intent separated from research because their allowable economics differ, exploration capped so it learns without bleeding the account.
A bad system will beat a good person every time.
Architecture audit discipline
The architecture audit takes about 90 minutes and runs before you change a single setting: count learning cells and divide conversions into them, justify every campaign in one sentence, sort by last-edited to find zombies, hunt self-overlap, check one-primary-action hygiene, score names as documentation, and write the target structure before touching anything. The headline number is always conversions-per-cell.
Diagnosis precedes surgery. Before re-drawing a single line you run a fixed audit, because the temptation to improvise a “cleaner” structure on the fly is exactly how 96-cell monstrosities get built. The sequence below is the order we run it — and the first step alone usually settles the argument.
- Count the learning cells; do the division.Active campaigns × ad sets/groups. Monthly conversions ÷ cells. Under ~50 per cell on money campaigns = the headline finding before you read anything else.
- Map budget logic against structural logic.For every campaign: why does THIS deserve its own budget/target? No one-sentence answer = consolidation candidate.
- Run the zombie sort.Last-edited ascending. Six-months-untouched gets the one-sentence test. Log the pauses.
- Hunt self-overlap.Same keyword/audience reachable from two live cells? Check auction insights against yourself (search) and audience-overlap tools (social). Every collision is a tax with your name on both sides.
- Check conversion-action hygiene per campaign.One primary per goal (module 3). Mixed diets explain more “mystery” performance than any bid setting.
- Score names as documentation.Could a stranger reconstruct the budget logic from campaign names alone? If not, the account depends on tribal memory — rename to the logic.
- Write the target structure BEFORE touching anything.One page: campaigns, their budget logic, their targets, the migration sequence (parallel-run, module 3 steps). Restructure-by-improvisation is how 96-cell accounts happen.
A restructure resets learning (module 3’s red zone), so we treat it like a site migration: never all at once, never without a rollback path, never in Q4.
The sequence: build the new structure alongside the old at 20-30% of budget, let it exit learning, prove it on MER (never platform ROAS — module 1), then shift budget in steps while the old structure winds down.
The discipline that saves careers: keep the old campaigns paused, not deleted, for a quarter. History is your rollback; deletion is the only irreversible button in the interface.
Advanced playbook
Advanced architecture is migration discipline: a restructure resets learning, so it is run like a site migration — new structure alongside old at partial budget, proven on MER, shifted in steps, with the old campaigns paused (not deleted) as a rollback. And it is revisited whenever the bidding technology changes, because optimal structure has an expiry date.
The senior skill is not designing the perfect tree once; it is changing structure without paying for it in lost learning, and knowing when the right answer has expired. Both come down to treating architecture as a living system with a rollback path rather than a one-time setup — the migration trick above is the operational core, and the SKAG story is the reminder that today’s best structure is a hypothesis the next platform change can falsify.
Common mistakes
The classic architecture mistakes share one root: treating structure as tidiness instead of data allocation. Over-slicing into starved cells, mirroring the org chart, leaving zombie campaigns running, allowing self-overlap, mixing primary conversions, and big-bang restructures are the recurring six.
- Slicing into more cells for “control.” Below ~50 conversions/cell the control is an illusion — every cell guesses. Consolidate until cells learn.
- Structuring by org chart or geography. The auction ignores both; each mirror-line taxes data density. Solve reporting with labels and views.
- Leaving zombie campaigns running. They spend, fragment data, and bid against your live campaigns. Run the last-edited sort and prune.
- Allowing self-overlap. Two campaigns serving the same query/audience split its data and muddy ownership. One intent, one home.
- Mixing primary conversion actions in a campaign. Module 3 hygiene: one primary per goal, or the machine optimizes the cheapest one.
- Big-bang restructures. They reset learning everywhere at once with no rollback. Run a parallel migration in steps, old campaigns paused.
Quick answers
- How should I structure a Google Ads or Meta account?
- Group by what should learn together and what deserves its own budget or target — usually intent tiers, economic bands (margin/LTV), or funnel jobs — not by geography or org chart. Keep each learning cell (campaign × ad set/group) above roughly 50 conversions a month so bidding can stabilize. The test for any split: will these two things ever deserve different budgets or targets? If not, they belong together.
- Why does account structure affect performance?
- Because structure allocates data, and automated bidding needs data density to learn. Every campaign or ad-set split divides the same conversions across more learning cells; slice too finely and no cell gets enough events to escape the learning phase, so every cell bids badly. Consolidating starved cells is often the single highest-leverage change on an inherited account.
- Are single-keyword ad groups (SKAGs) still best practice?
- No. SKAGs made sense under manual bidding, where human control and exact message-match were the levers. Auction-time bidding inverted the math — the machine needs pooled data more than humans need granular control, and close-variant matching erased much of the precision — so the industry dismantled SKAGs between 2019 and 2022. The deeper lesson: optimal structure depends on the bidding technology of the era.
- How many campaigns should I have?
- As few as the genuine budget and economic differences require, and no fewer. Start by asking which things truly need separate budgets or targets (brand vs non-brand, high-intent vs research, distinct margin bands) and give each of those a campaign; everything else consolidates. The goal is cells above the ~50-conversions-a-month learning floor, not a tidy-looking tree.
- What is self-overlap and why does it matter?
- Self-overlap is when two of your own campaigns can serve the same query or audience, so they fragment that segment’s data across two learning cells and muddy which structure should own it. It is a tax you pay per click in lost data density. One intent or audience should have one home; quarterly audits check for collisions with auction-insights and audience-overlap tools.
- How do I restructure an account without tanking performance?
- Treat it as a migration, not an edit. Build the new structure alongside the old at 20-30% of budget, let it exit the learning period, prove it on MER rather than platform ROAS, then shift budget in steps as the old structure winds down. Keep the old campaigns paused (not deleted) for a quarter as a rollback path, and never run a big-bang restructure in Q4.
Operating checklist — score yourself
Use this as the operating standard for account architecture. None of it is about a tidy-looking tree — it is the discipline of allocating data so the machine can learn, cutting only where a real budget or economic difference earns the split.
Official documentation:
Google — Smart Bidding (data pooling)
Google — account structure guidance
Meta — campaign budget optimization
Meta — structure and the learning phase
The Deming Institute (quote source)
The structural reversal:
Search Engine Land — the SKAG-era reversal (fact-atom source)
Deeper treatment (blueprint editions):
PSM — Google Ads account architecture · PSoc — Meta ads architecture
RGM tools used in this module:
Campaign architecture builder · Audience & structure recommender · Campaign settings recommender
RGM glossary entries used in this module:
CBO · CAC · ROAS
Series: All modules in Performance Marketing Foundations.
Prove it. Earn your passcode.
Ten questions, CASE method (Context · Analysis · Strategy · Execution). Pass at 90% to unlock this module’s completion passcode — retake as many times as you like.