The short version
- A creative testing matrix crosses two or three ad dimensions, usually angle, persona and format, so every cell is one possible ad.
- Plan on two axes with three to five values each, and keep hooks and offers out of the grid.
- Pick a few cells per round using one pattern, such as a row sweep or a diagonal spread, so each round answers one question.
- Never fill the whole grid. A full grid costs far more and leaves most cells with too little spend to judge.
- Put cell codes in your ad names and shade the grid after every round, so the next round almost picks itself.
The short answer
A creative testing matrix is a grid that crosses two or three things you could change in an ad, usually angle, persona and format, so that each cell is one possible ad. You do not make every cell. Each round you pick a handful of cells that answer one clear question, then move across the grid as the results come in.
The grid earns its keep by making gaps visible. Without one, it is easy to end up testing whatever was quickest to make, which tends to mean several versions of the same idea. With one, you can see at a glance that every ad speaks to the same buyer in the same format, and choose the next round on purpose. The glossary entry for a creative matrix has the short definition.
Choosing the axes
Each axis is one dimension of an ad, and each value on it is one option. Good axes change the idea rather than the polish. These are the usual candidates, with a rough sense of what each costs to change.
| Axis | Example values for a kids' bottle | Cost to change | Main axis? |
|---|---|---|---|
| Angle (why buy) | No leaks, cold until lunch, survives drops | Low | Yes, almost always |
| Persona (who it is for) | Parent of a preschooler, sporty 10-year-old, gifting grandparent | Medium, may need a new model or scene | Yes |
| Format | UGC-style photo, bold claim static, demo video, carousel | Low to medium | Yes, or as the third axis |
| Setting | School gate, football pitch, back seat of a car | Low | Sometimes |
| Proof type | Review quote, live demo, side-by-side comparison | Low | Usually folded into angle |
| Hook | Question, result, POV | Very low | No, keep it for iterations |
| Offer | None, bundle, free delivery | Low | No, keep it fixed |
Two axes are enough to plan with. A third earns its place once you have at least one winning cell. Keep each axis to three to five values: five angles by four personas is already twenty cells, which is more than many accounts can judge in a month.
Watch for axes that overlap. If one angle is "cheaper than buying bottled water" and the offer axis includes a bundle discount, both are pulling the price lever, and you will not be able to tell them apart afterwards.
A worked matrix
Frostbud, a fictional brand, sells an insulated children's bottle for $9.25. Its matrix crosses four angles with three personas, and every cell is written as a one-line scene:
| P1: parent of a 4-year-old | P2: sporty 10-year-old | P3: grandparent buying a gift | |
|---|---|---|---|
| A1: no leaks | A1·P1: a tipped-over school bag, dry drawings inside | A1·P2: a kit bag thrown into a car boot | A1·P3: wrapped beside a new school bag |
| A2: cold until lunch | A2·P1: ice still rattling at 12:30 | A2·P2: cold water at half-time | A2·P3: a summer holiday present |
| A3: survives drops | A3·P1: the bottle bouncing down the stairs | A3·P2: dropped on a gravel pitch | A3·P3: "lasts the whole school year" |
| A4: easy to clean | A4·P1: wide mouth, no hidden straw gunk | A4·P2: rinsed in seconds after training | A4·P3: one less chore for the parents |
Writing each cell as a scene does two jobs. It proves the cell is a real ad rather than a label, and it becomes the brief when someone makes the ad. If you cannot write a cell in one line, it is usually a weak cell, so leave it blank. The creative brief generator turns a chosen cell into a full brief.
Frostbud treats format as the third axis and tracks it in its test log rather than drawing a cube: F1 UGC-style photo, F2 bold claim static, F3 short product video.
Why a grid helps under Meta's current delivery
Meta's engineers describe Andromeda as the retrieval stage of its ad system, the step that narrows tens of millions of candidate ads to a few thousand for each person, and they wrote that the number of ad creatives in the system was expected to grow significantly as advertisers take up generative AI tools for making ads (Meta Engineering, December 2, 2024 (opens in a new tab)). The post does not say how different two ads need to be for the system to treat them as distinct. What a matrix does is make sure that the ads you add sit in different cells, rather than piling more versions into the one cell that happened to work last month. Our guide to Andromeda and creative diversity covers the background.
Why you never fill the whole grid
Every axis multiplies the count. Frostbud's grid is 4 × 3 = 12 cells. Add three formats and it becomes 36 ads. Give each of those three hooks and it is 108.
Money is only half the problem. A full grid is also hard to read. With 36 ads live, spend gathers on a few of them and most cells end the test with too little data to call, so you learn less from more ads.
Four ways to pick cells
The pattern you choose sets the question the round can answer.
| Pattern | Cells | Question it answers | Watch out for |
|---|---|---|---|
| Row sweep | One angle, every persona (3 cells) | Who does this argument land with? | Needs an angle worth sweeping |
| Column sweep | One persona, every angle (4 cells) | Which reason to buy works for this buyer? | Says nothing about other buyers |
| Diagonal spread | Each angle and persona used once or twice (4 to 6 cells) | Which broad direction is worth exploring? | A loss could come from the angle or the persona |
| Winner's neighbours | The winning cell and the cells beside it | Does the win travel to a new angle or buyer? | Only useful after a win |
A diagonal spread suits a first round in a new account, because it touches every row and every column without making every ad. For Frostbud, that could be A1·P1, A2·P2, A3·P3 and A4·P2, plus one wildcard off the diagonal such as A1·P3, the bottle wrapped beside a new school bag.
Use one pattern per round. Mixing a row sweep with a column sweep gives you a bundle of ads with no single question behind them, and the results will be hard to explain to anyone, including yourself.
Moving through the matrix round by round
A matrix is a map for several rounds, not a plan for one. A typical path, where each result decides the next move:
- Round 1, diagonal spread. Five or six cells across the grid, all in one format, to find a direction. Say A2·P2, cold water at half-time, sells at target and the rest do not.
- Round 2, row sweep on the winning angle. A2 shown to P1 and P3, with A2·P2 running as the yardstick. Now Frostbud learns whether "cold until lunch" is a sporty-kid message or one that works for every buyer.
- Round 3, format sweep on the best cell. The winning cell rebuilt as a UGC-style photo, a bold claim static and a short product video. This is where the third axis starts working.
- Round 4, iterations inside the winning cell. New hooks and opening frames for the best cell in its best format. Hook testing belongs here, once the grid has pointed at an idea. Our guide to concept vs iteration testing covers this step.
- Every round, one wildcard. One cell from a row or column nobody has tried keeps the next direction coming.
Shade the grid after each round: winning cells, losing cells, and cells that never got enough spend to judge. After a month the shaded grid is one of the most useful documents in the account. It shows which reasons to buy are proven, which buyers they work for, and where nobody has looked yet.
Frostbud's grid after two rounds
Round 1 was the diagonal spread above, with A1·P3 as its wildcard. Round 2 was the row sweep on A2, with A3·P2, the bottle dropped on a gravel pitch, as its wildcard. Every ad was a UGC-style photo (F1). Here is the shaded grid, with made-up results:
| P1: parent | P2: sporty kid | P3: grandparent | |
|---|---|---|---|
| A1: no leaks | Lost (round 1) | Untried | Lost (round 1) |
| A2: cold until lunch | Won (round 2) | Won (round 1), held in round 2 | Lost (round 2) |
| A3: survives drops | Untried | Won (round 2) | Lost (round 1) |
| A4: easy to clean | Untried | Too little spend (round 1) | Untried |
Three things stand out. A2 is the proven row, selling to parents and to kids. P3 has lost on three different angles, which says the gift buyer needs a different reason to buy altogether, such as a birthday or the first day of school, and that is a job for the next grid. And A3 won with sporty kids but lost with grandparents, so "survives drops" may be a sport message rather than a general one. Round 3 is the format sweep on A2·P2 as planned, and its wildcard is A4·P2 again, because that cell never got a fair go.
Naming ads so the grid reads itself
A matrix only works if every ad's results lead back to its cell. Put the coordinates in the ad name, such as FRB_A2-cold_P2-sport_F3-video_h1, and use the same codes in your test log. A filter on "A2" then shows every ad on that angle, whichever buyer it targeted. Our naming conventions guide sets out a fuller system, and the UTM builder keeps your tracking tags in step with the names.
Reading the results by row, column and cell
Read a matrix at three levels.
- By axis value. Add up every ad on each angle, then every ad on each persona. If the A2 ads together sell at target while the A3 ads do not, that is a stronger signal than any single ad gives you.
- By cell. A cell that wins along with its row and column is the clearest result you will get. A cell that wins while the rest of its row loses may be an interaction, an angle that only works for one kind of buyer, like Frostbud's A3·P2. Those are worth keeping, because they point at a narrow message for a specific person.
- By spend. Before you shade a cell as a loss, check it got enough money. In one ad set, cells compete for the same budget and some barely run. Shade those grey, not red, and bring them back later.
When one comparison really matters, such as two finalist cells, Meta offers controlled split tests. Its documentation says they run on audiences that do not overlap, and that tests with larger reach, longer schedules or higher budgets tend to give more statistically significant results (Meta Marketing API, split testing (opens in a new tab), as of October 2026). That makes a split test a tool for settling one important question, not for running a whole grid.
A matrix checklist
Run through this before each round:
- Two main axes, each with three to five values
- No two axes pulling the same lever, such as a price angle beside a discount offer
- Every chosen cell written as a one-line scene
- Offer, price and landing page fixed for the whole round
- One picking pattern, and the round's question written at the top of the log
- Cell codes in every ad name
- One wildcard cell from an untried row or column
- The grid shaded after the round: won, lost, or not enough spend
Matrix mistakes
- Filling the grid because you can. More cells feel thorough but leave each one underfunded. Pick the cells that answer this round's question.
- Using colour or font as an axis. These are polish. Test them, if at all, inside a winning cell.
- Too many values on one axis. Eight personas is a list, not an axis. Merge the similar ones.
- Letting the offer drift between cells. One cell with free delivery and one without turns the round into an offer test.
- Reading one ad as the whole row. A single losing ad does not sink an angle. Look at the row total and at how much each ad spent.
- Never retiring the grid. Once most cells are shaded, build a new one with fresh angles from reviews and fresh personas from the people who actually buy.
Building matrix rounds in BulkAds
- Write the cells as instructions. The AI ad generator can build a batch from your own instructions, so a round can be written as its list of cells: "Ad 1: cold until lunch, sporty 10-year-old at half-time. Ad 2: no leaks, parent of a 4-year-old, tipped-over school bag." A batch holds up to 12 ads, in 1:1, 4:5 or 9:16, as images or video, all made from your real product photo.
- Make the format sweep cheaply. A winning image can be turned into a short video ad without a shoot, which covers F3 in round 3. These are product clips with gentle motion, such as condensation running down the bottle, so a hands-on demo with a child still needs filming. BulkAds uploads take still images only, so run that filmed clip from Meta Ads Manager and use BulkAds for the image and AI video cells around it. Each batch is one size and either images or video, so the video cell is made separately from the two image cells. BulkAds does not make carousels.
- Launch one ad per cell. The bulk ad launcher creates one ad per creative in a single ad set, paused until you approve, with UTM tags that carry each creative's ID. Ads are named by angle and ID, so keep the persona and format codes against those IDs in your log.
- Read the axes. Creative analytics groups results by angle, hook style, format and size, which covers two of the three usual axes. Persona is not one of its built-in groupings, so match persona codes to ad IDs from a CSV export.
Questions people ask
What is a creative testing matrix?
A grid that crosses two or three ad dimensions, usually angle, persona and format, so each cell is one possible ad. You test a few chosen cells per round rather than every combination.
Which axes should a creative matrix use?
Angle and persona are the most useful main axes, with format as a third once something wins. Keep hooks and offers out of the grid: hooks belong in iteration rounds, and the offer should stay fixed so it does not skew the results.
How many cells should I test in one round?
As many as your spend can give a verdict to, which for many accounts is four to eight. At a $4.10 target cost per purchase and about 2.5 times that per ad, six cells need roughly $62.
Is a creative matrix the same as dynamic creative?
No. Dynamic creative takes several images, headlines and texts in one ad and mixes them into combinations, favouring the ones Meta predicts will do best. A matrix is your own plan for which complete ads to make and test, usually launched as one ad per creative so each cell has its own results.
How often should I rebuild the matrix?
When most cells are shaded as won, lost or not enough spend, usually after several rounds. Build the next grid with new angles from reviews and new personas from the people who buy.
Can a small account use a creative matrix?
Yes, with a smaller grid. Three angles by two personas gives six cells, and a round of three cells is enough to start. The grid still stops you from testing the same idea twice.
Keep going
- Features
AI ad generator
On-brand image ads in batches of up to 12
- Features
Creative analytics
Which ad actually sold — and why
- Free tools
Ad variation calculator
How many ads your hooks, visuals and copy make, and what testing costs
- Guides
Concept vs iteration testing: find the idea, then refine it
Concept tests find the idea that sells, iteration tests sharpen it. How to tell them apart, run each one, read the results and decide next week's mix.
- Guides
The 3-2-2 method: twelve ad combinations, one ad set
Three creatives, two primary texts, two headlines: how the 3-2-2 method works, how to set it up today, how to read the results and when to move beyond it.
- Glossary
Creative matrix
A creative matrix is a grid that plans a batch of ads by crossing the variables you want to test, such as angles against hooks or visual styles, so every cell is one ad and you can see what's covered and what's missing.
Written by the BulkAds team. Published , last checked . Meta changes its tools often — if something here no longer matches Ads Manager, tell us at contact@bulkadsai.com.
BulkAds AI is an independent product and is not affiliated with, endorsed or sponsored by Meta Platforms, Inc. Facebook and Instagram are trademarks of Meta Platforms, Inc.