meta-ads-framework

Meta Ads Framework

This skill makes you reason about Meta ads the way the team does: in tables, one decision per row, with every platform claim tagged by where it comes from. The tables are not decoration. Filling a table forces you to cover every case, expose missing inputs, and separate facts from rules of thumb. Build the tables first, then write at most one or two connecting sentences between them.

Platform facts are current as of September 2026. See section 8 before stating any of them as fact to someone else.

1. Table contract

Follow every rule on every output, including quick answers.

Rule	What it means in practice	Why it matters
Tables first	Decide which tables the answer needs (section 2), fill them, then add prose only as short connective lines	Structure exposes gaps that paragraphs hide
Label every table	The line above each table reads: Rows: what one row is. Columns: what the columns hold	A reader, or another model, must be able to use the table cold
One row, one object	Each row is one campaign, audience, ad, concept, signal or decision. No merged ideas, no lists inside cells	Rows become checkable actions
Short cells	Aim for under 25 words per cell. Numbers with units (%, x, days, EUR or USD)	Tables stay scannable and printable
Tag the evidence	Platform claims carry a tag from section 8: [M] Meta doc, [P] practitioner, [H] heuristic, [A] account data, [?] assumption	Heuristics must never pass as platform rules
Never invent account data	If an input is missing, write "Unknown" in the cell and add it to the Open questions table	A confident wrong number is worse than a gap
Close with decisions	Every answer ends with a Decisions table and, if anything is unknown, an Open questions table	The output is always actionable
House style	No em dashes. Plain words. Sentence case headers	Team standard
Deliverables	In chat, use markdown tables. If asked for a report, document or page, build a clean HTML page around the same tables	Same thinking, printable format

2. Pick the mode

Rows: the six request types this skill handles. Columns: how to recognise each, and the tables to produce, in order. Templates are in section 5.

Mode	Typical request	Tables to produce, in order
Build	"Set up Meta for X", "structure the account", "launch campaigns"	Intake, Campaign plan, Audience map, Concept matrix, Creative quota, Budget split, Decisions, Open questions
Audit	"Audit this account", "is our structure right"	Intake, Structure gap, Audience check, Creative health, Measurement check, Priority fixes, Open questions
Diagnose	"CPA went up", "ROAS dropped", "delivery stalled"	Intake, Symptoms, Hypotheses (in section 6 order), Saturation ladder check, Actions, Open questions
Creative plan	"Brief new ads", "what should we test", "concepts for Meta"	Persona x stage grid, Concept matrix, Creative quota, Refresh calendar, Decisions
Scale	"Can we increase budget", "how do we scale"	Readiness gates, Marginal CPA, Budget step schedule, Risks, Decisions
Weekly review	"Weekly Meta review", "what changed this week"	KPI summary, Ad-level saturation ladder, Rules fired, Next week actions

If a request mixes modes, run them in the order listed above and share the Intake table once.

3. Intake

Always start with this table. Fill what the user gave you or what connected tools return; mark the rest Unknown.

Rows: the account facts every recommendation depends on. Columns: the value, where it came from, and whether it blocks the answer.

Field	Value	Source	Blocking if unknown?
Business model and offer			Yes
Optimization event (purchase, qualified lead, trial)			Yes
Weekly volume of that event			Yes
Target CPA or ROAS			Yes
Monthly Meta spend			Yes
Markets and languages			No
Pixel plus Conversions API, deduplicated			Yes for Diagnose
Customer list size (for exclusions and seeds)			No
Warm pool size (monthly site visitors, engagers)			No
Current structure (campaigns, ad sets, ads live)			Yes for Audit
Creative production capacity per week			Yes for Creative plan
Special ad category (credit, housing, employment, social issues)			Yes
Attribution setting in use			No

4. The framework

These are the reference tables. Copy the relevant rows into your answer and adapt them to the account; do not paste them whole.

4.1 The two things that decide performance

Everything else serves these two. Meta matches each ad to each person at delivery time, so creative does most of the targeting, and every winning ad has a shelf life.

Creative matches the audience. Rows: the five ways a creative fits or misses its audience. Columns: what a match looks like, how to check, what to do on a miss.

Fit	Match looks like	How to check	On a miss
Persona	One ad speaks to one person with one problem; Scale runs 3-5 personas in one ad set	Age, gender, placement and Audience segments breakdowns per ad	Rewrite the hook for the segment it drifted to, or brief a new persona ad
Awareness stage	Cold ads name the problem or show the outcome; warm ads prove and handle objections; customer ads offer what comes next	Message vs the campaign it runs in	Move or rewrite
Distinctness	Each ad changes angle, persona, messenger or format [M]	First frames and first lines side by side	Keep the strongest near-duplicate, brief new angles
Placement fit	Native 9:16 and 4:5, captioned, inside safe zones	Placement preview, Safe Zone Guardrail	Re-export per placement
Promise continuity	Landing page repeats the ad's promise above the fold	Good CTR, weak conversion rate on one ad	Send that angle to a matching page

Creative is refreshed before it saturates. Meta flags fatigue only once cost per result has doubled, so act on the early rungs. Rows: saturation signals in the order they appear. Columns: threshold, where to see it, action.

Signal	Threshold	Where	Action
Frequency climbing	7-day frequency past ~2-2.5 cold, ~5 warm [P]	Ad level, 7 days	Brief 2-3 variants of the concept
Attention fading	CTR down 20-25%+ from the ad's first week while CPM rises [P]	Ad level, first vs last 7 days	Launch variants: new hook or first 3 seconds
Creative limited	Cost per result above the ad's past level, under 2x [M]	Delivery column	Rotate refresh batch in
Creative fatigue	Cost per result 2x past level or more [M]	Delivery column	Retire the ad; if its variants fade within 1-2 weeks, retire the concept
Concentration risk	1-2 ads carry most of an ad set's spend and their cost per result is rising [P]	Spend share by ad	Feed new concepts now

Plan production backwards from the hit rate: about 5% of ads become winners [P, Motion 2026], so one new winner takes roughly 20 launched ads.

4.2 Campaigns and budget

Rows: the four campaigns, each with one job. Columns: setup, funding and how each is judged.

Campaign	Job	Setup	Share of spend	Ad sets	Ads	Judge on
Test	Find winning concepts cheaply	Meta creative testing tool inside the Scale ad set (2-5 ads, up to 20% of budget suggested) [M], or a separate ABO campaign for hard spend control	15-20%; up to 30% on new accounts; less when Scale is on target [H]	None extra (tool) or one per concept, broad	2-5 per test	Cost per result and spend share; hook rate and CTR only diagnose [P]
Scale	New customers at target cost	Sales or Leads, campaign budget, highest volume; add a cost goal only once CPA is stable, loosely	55-70% [H]	1-3: Broad always, Lookalike signal, Interest signal; customer exclusion in audience controls	8-15, max 50 [M]	Blended CPA or ROAS, new-customer share, marginal CPA
Retarget	Convert engaged non-buyers	Original audience setup, Advantage+ audience off, so the custom audience binds [M]	0-15%, default 10% [H]	One consolidated warm ad set, recent buyers excluded	4-8	Cost per result, frequency with CTR, lift over Scale
Retain	Repeat, upsell, expansion	Original audience setup, small fixed budget	0-10% [H]	Customer list, purchasers 180 days	3-6	Repeat rate, revenue per customer

Rules that apply across the table:

Rule	Detail
Day one	Launch Test and Scale only. Add Retarget when Audience segments show Scale under-reaches warm users; add Retain when there is a repeat or expansion motion
Check before funding Retarget	Broad delivery often already spends 25-35% on engaged people and customers [P, Loomer]. If so and warm frequency is healthy, skip Retarget
Budget floor	Weekly ad set budget of about 50x target CPA [H]; if unaffordable, consolidate or optimize for a higher-volume event

4.3 Audiences

Rows: audience types from coldest to customer. Columns: what it is, how Meta treats it, how to build it, where it runs, when to use it.

Audience	What it is	How Meta treats it	Build rule	Runs in	Use when
Broad	Advantage+ audience, no suggestions	Only location, minimum age, language and custom audience exclusions bind [M]	Customer exclusion goes in audience controls, never suggestions	Test, Scale	Always the core ad set once the pixel has conversion history
Interest stack	3-5 related interests as one suggestion	Suggestion; Meta widens when it predicts better results [M]. Interests were consolidated in 2025 [M]	Group by theme, never one interest per ad set	Scale signal ad set	New pixel, niche, tiny budget, new geo
Lookalike	Similar to a seed of best customers	Suggestion under Advantage+ audience; hard only in original setup [M]	Seed min 100 people in one country, 1,000-5,000 recommended [M]; one tier per ad set, usually 1-3% [P]	Scale signal ad set	Strong first-party seed, new country
Custom: warm	Visitors 30-180 days, engagers 90 days, video viewers 50%+ 30 days	Suggestion under Advantage+ audience; hard with it switched off [M]	One combined audience, original setup, exclude purchasers 30 days	Retarget	Scale under-reaches warm users and the pool can spend without frequency climbing
Custom: hot	Cart, checkout, lead form open, pricing page, 7-14 days	Same as warm	Stay inside the warm ad set unless it can carry its own budget	Retarget	Thousands of abandoners a week
Custom: customers	CRM list, purchasers 180 days	Binding as an exclusion in audience controls [M]	Sync from CRM or re-upload monthly	Retain; exclusion everywhere; lookalike seed	Every account

4.4 Creative

Rows: the four levels at which a creative can change. Columns: what changes, whether Meta treats it as distinct, how many to make.

Level	What changes	Distinct for Meta?	How many
Concept	Pain point, persona, promise, emotional trigger	Yes [M]	4-6 live in Scale; new ones weekly in Test
Execution	Format and messenger: UGC, founder, static, carousel, catalog	Often, if visually distinct [P]	2+ formats per concept
Variant	Hook or first 3 seconds, headline, thumbnail, primary text	Usually grouped with its parent [P]	2-3 per winning execution
Micro-tweak	Colour, one word, music, button text	No [P]	None

Rows: campaigns. Columns: creative quota at any time and refresh cadence.

Campaign	Concepts live	Variants per concept	Ads per ad set	Format mix	Refresh
Test	1-5 per test	1-3	2-5 per test	Genuinely different ads only	Weekly batch
Scale	4-6	2-3	8-15	3+ of: 9:16 video, 4:5 video, static, carousel	Winners in one weekly batch; retire on the ladder
Retarget	3-4	1-2	4-8	Testimonials, demo, catalog, offer	Monthly
Retain	2-3	1-2	3-6	Catalog, launch video, offer	Per launch or season

Rows: monthly spend tiers (USD, Motion 2026 benchmarks [P]). Columns: average creatives tested per week, hit rate, winners per month that volume yields (derived).

Monthly spend	Tested per week	Hit rate	Winners per month
Under $10K	2.8	4.0%	~0.5
$10K-50K	4.1	6.4%	~1.1
$50K-200K	6.6	8.1%	~2.3
$200K-1M	11.2	8.6%	~4.2
Over $1M	18.8	8.8%	~7.2

Aim for the top quartile, which tests 1.5-3x its tier average [P].

4.5 Decision rules

Rows: recurring decisions. Columns: trigger, action, check cadence, evidence.

Decision	Trigger	Action	Check	Evidence
Protect learning	New ad set or significant edit: targeting, creative, optimization event, adding an ad, pausing the ad set; budget and bid changes depending on size	Leave until ~50 optimization events since the edit and at least a week; batch all edits into one weekly window	Daily, no edits	[M]
Kill a test	Spend 2x target CPA with no conversion, or CPA over 1.5x target after 3x target CPA spent	Pause, log the learning	Every 2-3 days	[H]
Graduate a winner	CPA at or under target over 3x target CPA of spend	Add to Scale in one weekly batch or as a new ad set in the Scale campaign; reuse the post ID	Weekly	[M] adding an ad resets learning
Scale budget	Scale on target 3+ days and marginal CPA under target	Raise 15-20% every 48-72 hours	Every 2-3 days	[H]; Meta publishes no safe %
Refresh creative	Any rung of the saturation ladder	Brief, launch variants, rotate, retire	Weekly	[M] flags, [P] early rungs
Consolidate	"Learning limited"	Merge ad sets, drop the weakest signal ad set, raise the funnel event, or loosen a tight cost control	Weekly	[M]
Cap retargeting	Warm 7-day frequency above ~5 with CTR falling	Move Retarget budget to Scale	Weekly	[P]

Marginal CPA = (this week's spend minus last week's) divided by (this week's conversions minus last week's). Average CPA hides the cost of the extra customers you are buying.

4.6 Naming

Rows: account levels. Columns: pattern and example.

Level	Pattern	Example
Campaign	Stage_Objective_Geo_Budget	Scale_Sales_IT_CBO
Ad set	Audience_Seed-or-window_Detail	LAL_Purch180_1-3pct
Ad	ConceptID_Format_Variant_Launch	C07_UGCvid_H2_260924

Keep the concept ID constant from Test to Scale so reporting can roll up by concept.

5. Output templates

Use these headers exactly. Add rows as needed; never drop a column. Each table still needs its Rows and Columns line.

Campaign plan (Build). Rows: campaigns. Columns as shown.

Campaign	Job	Objective and event	Budget control	Share	Daily budget	Ad sets	Ads	Judge on	Evidence

Audience map (Build, Audit). Rows: ad sets.

Ad set	Campaign	Audience type	Controls (hard)	Suggestions (soft)	Exclusions	Est. size	Evidence

Persona x stage grid (Creative plan). Rows: personas. Columns: awareness stages. Each cell holds the core message for that pair, or "Gap".

Persona	Cold: problem or outcome	Warm: proof and objections	Customer: what comes next

Concept matrix (Build, Creative plan). Rows: concepts.

Concept ID	Persona	Stage	Angle (pain or promise)	Messenger	Formats	Variants	Hypothesis	Campaign	Status

Structure gap (Audit). Rows: framework elements from section 4.

Element	Framework says	Account has	Gap	Impact (High, Med, Low)	Fix	Evidence

Creative health (Audit, Weekly review). Rows: live ads with meaningful spend.

Ad	Concept ID	Spend share	Cost per result vs target	7-day frequency	CTR vs first week	Delivery flag	Ladder rung	Action

Measurement check (Audit, Diagnose). Rows: tracking items from references/platform-details.md.

Item	Expected	Found	Pass or Fail	Fix

Hypotheses (Diagnose). Rows: candidate causes in section 6 order.

#	Cause	Evidence for	Evidence against	Test to confirm	Confidence

Readiness gates (Scale). Rows: gates.

Gate	Threshold	Current	Pass or Fail
On target	CPA or ROAS at target 3+ days
Learning exited	No ad set in Learning or Learning limited
Creative runway	2+ fresh concepts ready to launch
Saturation	No top-spend ad past rung 2 of the ladder
Tracking stable	No event drops or dedup issues this week
Marginal CPA	Last step's marginal CPA under target

Budget step schedule (Scale). Rows: planned steps.

Step	Date	Budget from	Budget to	Change %	Stop condition

Decisions (every answer). Rows: decisions or actions.

Decision	Why (evidence)	Owner	By when	Success looks like

Open questions (whenever anything is Unknown). Rows: missing inputs.

Question	Why it matters	Default assumed until answered

6. Diagnosis order

Work through causes in this order and stop only when evidence confirms one. Cheap, common causes come first.

Rows: candidate causes of a performance drop. Columns: what to check and the tell-tale sign.

#	Cause	Check	Tell-tale sign
1	Tracking	Events Manager volume, dedup, match quality vs last week	Conversions fell but backend sales did not
2	Attribution or reporting change	Attribution setting, Meta changelog	Reported results moved on a single date across all campaigns
3	Learning resets	Last significant edit column	Drop starts the day after edits or new ads
4	Creative saturation	Saturation ladder per top ad	Frequency up, CTR down on the ads carrying spend
5	Auction pressure	CPM trend vs seasonality	CPM up across all ads, CTR stable
6	Landing page or offer	Conversion rate by ad and page	CTR stable, conversion rate down
7	Audience or budget ceiling	Marginal CPA after recent budget steps	CPA rose right after a budget increase

7. Worked mini example

Request: "CPA on our Scale campaign went from 40 to 58 EUR this week, what happened?"

Rows: candidate causes. Columns: evidence and next step.

#	Cause	Evidence for	Evidence against	Test to confirm	Confidence
1	Tracking	Unknown	Unknown	Compare Events Manager purchases with backend orders, last 14 days	Unknown
2	Attribution or reporting change	Unknown	Unknown	Check whether all campaigns moved on the same date	Unknown
3	Learning resets	Unknown	Unknown	Check Last significant edit for the week	Unknown
4	Creative saturation	Unknown	Unknown	Build the Creative health table for the top 5 ads by spend	Unknown

Then the Open questions table asks for the backend order count, the edit log and ad-level data, rather than guessing a cause.

8. Evidence tags and freshness

Rows: evidence tags. Columns: meaning and how to use it.

Tag	Meaning	Use
[M]	Meta documentation or Meta engineering blog	State as platform fact, with the date checked
[P]	Published practitioner or benchmark (Jon Loomer, Motion, Common Thread Collective)	State as expert consensus or data, name the source
[H]	Heuristic used by practitioners with no Meta source	Say it is a rule of thumb; adjust to the account
[A]	The account's own data	Preferred basis for any decision
[?]	Assumption	Must also appear in Open questions

Platform facts in this skill were checked in September 2026. Meta changes features often. If web search is available and a [M] item matters to the decision, re-verify it before presenting it as current. If search is unavailable and the fact is load-bearing, say it was last checked September 2026.

9. References

Rows: bundled reference files. Columns: contents and when to read.

File	Contents	Read when
references/platform-details.md	Tracking, attribution, audience controls, creative specs, delivery settings, reporting hygiene	Audits, measurement checks, spec questions, any setup detail
references/sources.md	The vetting table (claim, Meta position, practitioner view, verdict) and source links	Someone challenges a rule, or you need to cite where a number comes from