Meta Creative Testing Matrix: 75 Ads in a Day (3 × 5 × 5)

By 4:17 PM on a Tuesday last March I had 75 ads uploaded to a single Advantage+ Shopping Campaign (ASC, Meta 的智能购物广告系列 — 一种由算法自动分配预算和受众的智能投放结构), a Google Sheet with 75 colored rows, and one cold coffee. The client — a B2B SaaS (Software as a Service, 软件服务) selling a $49/month scheduling tool to solopreneurs — was spending $2,100/day on Meta. By Friday of the same week, 61 of those 75 had spent exactly nothing. By the following Tuesday, 11 were getting traffic, 2 had become clear winners, and the kill-rate pattern was so consistent with every other test I'd run that year I stopped being surprised by it. Here's the exact 3 × 5 × 5 workflow that produced those 75 ads in 6 working hours.

Why 75, and why 3 × 5 × 5

The 6-ad-per-ad-set guidance Meta used to publish quietly disappeared from their documentation in 2025. The reason: their Andromeda update (a 2025 的 Meta 排序模型升级,核心逻辑从「匹配受众」转向「为不同人匹配不同广告」) needs creative volume, not audience precision, to find signal. Top spenders I talk to are now running 15-50 ads per ad set. But "50 random ads" is just a budget fire — you need a structure that lets the auction (the real-time ad-buying auction Meta runs every time it shows an ad) actually tell you why a winner won.

The matrix is the smallest structure that isolates three independent variables. Three hooks, five visuals per hook, five CTAs per visual. Every cell is a unique ad. The total — 75 — is the smallest number where Meta's algorithm has enough material to declare winners with statistical confidence, and where you have enough material to read the structure: which hook angle won, which visual concept won, which CTA pattern won. 5-variant A/B tests can't do this. 75 in a structured matrix can, in one week.

I picked 3 × 5 × 5 over 4 × 4 × 4 (also 64) or 5 × 5 × 3 (also 75) deliberately. Three hooks is the minimum for separating who's talking (problem-aware, solution-aware, brand-loyal). Five visuals is enough to cover the UGC / studio / lifestyle / before-after / founder archetypes without leaving gaps. Five CTAs is enough to test the offer ladder — free trial, $1 trial, demo call, lead magnet, hard-sell — without bloating the cells with noise. The cross-product of these three dimensions is the smallest matrix that lets you attribute a winner back to a cause, not just a correlation.

The day, time-blocked

8:00-11:00 AM — Three hooks via Claude

The brief is the only thing that has to be locked before this block starts. One product, one offer, one KPI (Key Performance Indicator, 关键绩效指标), one promise. The SaaS client's brief pinned: $49/month, 14-day free trial, KPI is activated trial signups, promise is "stop double-booking clients." That's the entire scaffold.

Three hook angles — problem-aware, solution-aware, brand-loyal — go into Claude with a single prompt: "Generate 25 distinct 1-2 sentence hook variations per angle. Each must work as a Meta ad primary text opening line. Vary tone: blunt, empathetic, contrarian, factual. No 'revolutionary,' no 'game-changing,' no clichés. Use the offer from the brief."

The output is 75 hook candidates. I score them on three criteria in 90 minutes: does it name a concrete pain, does it match the brief's voice, and would I keep reading if I saw it next to a photo of someone eating lunch. Roughly 25 survive per angle — a 75% pass rate that looks high until you remember how much of Claude's first pass is filler.

11:00 AM-1:00 PM — Five visuals per hook via Recraft / GPT-Image

Five visual archetypes. The five I use for almost every client test:

• UGC mirror — phone-shot, first-person, slightly imperfect lighting
• Product on white — clean studio, sharp shadow, the SaaS UI mocked up large
• Lifestyle — someone using the product in a real environment (home office, café, coworking)
• Before/after — split frame, "calendar in chaos" vs "calendar with the product"
• Founder talking-head — iPhone selfie, founder explains the product in 15 seconds

For a SaaS client I do 5 static variants per hook angle, not per ad — that's 15 visual sets, and within each set I keep the top 2 strongest compositions. 3 hooks × 5 visuals = 15 unique visual assets, not 75. The matrix gets filled in the next step.

Recraft handles the studio and lifestyle cells cleanly. GPT-Image-1 does better on UGC mirror and founder talking-head (the imperfect framing reads as "real" because it was rendered to look real). For each visual I spend roughly 8 minutes — generate 4 variations, pick the best, do one round of inpainting to fix the one thing the model got wrong (a misaligned button, a wrong-direction shadow). 15 visuals × 8 minutes is exactly 2 hours.

2:00-4:00 PM — Five CTAs, manually written

The CTAs are not generated. They are the part of the ad a human should still own, because the offer ladder is a strategic decision, not a generation decision. The five I write for almost every client:

1. Free trial, no card — "Start free, 14 days, no card required"
2. $1 trial — "Try the full product for $1, cancel anytime"
3. Demo call — "Book a 15-minute walkthrough with our team"
4. Low-commit lead magnet — "Download the free scheduling template (no signup)"
5. Hard-sell — "Get 50% off your first 3 months, this week only"

Each CTA pairs with each visual. Each hook runs against all 5 visuals, all 5 CTAs. The math: 3 × 5 × 5 = 75 cells, 75 complete ads.

I write the ad body in a single Google Sheet. Column A is the cell ID (H3-V2-CTA1), B is the hook, C is the visual, D is the primary text (the hook expanded into a 3-4 sentence ad), E is the headline, F is the description, G is the CTA button text, H is the asset URL, I is the predicted tier, J is the live cost-per-result at 72 hours, K is the live CTR (Click-Through Rate, 点击率), L is the verdict (kill / hold / scale). Predicted tier gets filled by a 60-second gut call — A, B, or C — so I can audit my own prediction against actual performance at 72 hours.

4:00-5:00 PM — Upload as one ASC, all 75 in one ad set

This is the part teams get wrong. The instinct is to split the 75 into multiple ad sets — one per hook, or one per audience, or one per visual. Don't. The whole point of the matrix is to let Meta's auction allocate spend across all 75 and find the winners. Fragmenting into multiple ad sets re-introduces the human bias the matrix was designed to remove.

I upload all 75 into a single ASC, one ad set, $2,100/day budget, lowest-cost bid, broad targeting (no interests, no lookalikes — let Andromeda do its job). Dynamic Creative (Meta 的动态创意功能,允许系统自动组合素材) is off for this test — I want each cell to be a discrete, attributable unit, not a recombined mashup. 75 distinct ad objects, 1 ad set, 1 campaign. Upload takes 40 minutes if I'm moving fast, including Meta's review queue.

The 72-hour kill: what Meta actually does

The first 24 hours are noise. The auction is in learning phase (the period when Meta's delivery system is calibrating who to show each ad to — typically needs ~50 conversions per ad set per week to exit cleanly), frequency is climbing, and spend distribution is roughly random across the 75. Don't read the dashboard more than once in the first day.

By 48 hours, a pattern starts. Roughly 25-30% of the cells (the predicted A-tier ones, mostly) start pulling disproportionate budget. The predicted C-tier cells are already at zero or near-zero spend. The middle is still noisy.

By 72 hours, the kill pattern is unmistakable. In the last eight 3×5×5 tests I've run, the median kill rate at 72 hours is 81% — between 60 and 65 of the 75 ads have stopped getting impressions entirely. Meta's auto-allocation has effectively cut them. Of the 10-15 that are still spending, 8-12 are getting meaningful traffic. 1-2 are runaway winners. The other 8-10 are "fine" — they're profitable, they just aren't exceptional.

This is the moment the test is actually won or lost. The temptation is to leave the 8-10 "fine" ads running because their CPA (Cost Per Acquisition, 单次获客成本) is acceptable. Kill them anyway. The 2 winners can absorb that budget and get 4-5x the impressions, which is what makes the next 30 days of scale work. Every soft-hold is a tax on your winners.

The auto-ranking spreadsheet

The Google Sheet does the kill work for me. At 72 hours I sort Column J (cost-per-result) ascending. Anything above the median CPA by 1.5x or more gets a red K in Column L and goes into a "kill tomorrow" list. Anything at 1.5x median or below gets a yellow H (hold). Anything at 0.5x median or better gets a green S (scale) and gets pulled out of ASC into a manual ad with a $400/day fresh budget.

This gives me a defended kill list, a hold list, and a scale list — all backed by 72 hours of auction data, all attributable back to a specific hook × visual × CTA cell. The CMO can argue with my creative opinions. They cannot argue with the spreadsheet. By Friday of test week, I have a defendable shortlist of what the brand should run for the next 30 days.

Why this beats 5-at-a-time testing

A 5-variant test needs roughly 2,000 impressions per cell to exit learning, which means ~10,000 impressions and 5-7 days to declare a winner. A 75-variant test needs the same 2,000 impressions per cell — but you only have to wait for one cell to clear that bar, which happens in 48 hours for the predicted A-tier cells. The "time to first decision" is faster, not slower, despite the volume.

The structural payoff is what 5-variant tests can't give you. With 75 ads built as a 3 × 5 × 5 matrix, by Friday I know that for this client, the problem-aware hook beat solution-aware by 2.4x, the UGC mirror visual beat founder talking-head by 1.8x, and the $1 trial CTA beat the free trial CTA by 1.3x. I can build the next test with that knowledge baked into the brief. A 5-variant test gives you a winner and a loser and almost nothing transferable.

That compounding is the actual point. One 3 × 5 × 5 test teaches you more about your offer than six months of 5-variant tests, in a single Tuesday.