Meta Andromeda: How Many New Ads to Make Each Week, and How

Meta Andromeda shortened the life of a winning ad

Picture a winning ad. It sells for nine days. Then the cost per sale doubles, and nothing you touch brings it back.

The old fixes were easy. Show it to a new audience, try a new interest, or retarget the people who almost bought.

Meta now controls most of those choices. Its ad system, which includes a step called Meta Andromeda, decides which ads each person sees. Each ad gets a certain amount of life, runs a little longer if it does well, then fades.

Why would Meta do that? A feed that shows the same ad for a year is a bad feed. People scroll past it.

So the job changed. It used to be finding one great ad. Now it is building a line that makes new ones every week.

Most accounts are still hunting for one great ad. The ones that last build the line.

What is Meta Andromeda? The step that picks which ads get a chance

Meta Andromeda is the first step of Meta's ad delivery. It scans tens of millions of ads and keeps a few thousand worth showing to one person, as Meta's engineers describe it.

They built it for a flood of ads. In one month, more than a million advertisers made over 15 million ads with Meta's AI tools.

Here is the part that matters for you. Meta's system groups ads that look or feel alike and treats them as versions of one creative.

So ten small tweaks of one idea can count as one idea. A new button on the same photo still counts as the same ad.

Meta splits the problem in two. Creative fatigue is when people see one ad too often. Creative similarity is when your ads look too much alike, even if each file is new.

Both end the same way: a higher cost per result. Both have the same cure: ads that differ in look, story and message.

Our take: many accounts that say they test fifty ads are testing five ideas ten times each. That is not volume. That is one ad in ten outfits.

How delivery then matches each ad to a person is in how Meta finds your buyers.

How many new ads should you make each week? Let the budget decide

Make as many clearly different ads as your budget can test, and launch them every week. For most accounts that is 3 to 15 ads, not fifty.

Meta's own guide to creative diversification says the strongest advertisers launch new ads weekly. For its shopping campaigns, it suggests 20 to 30 percent of the budget for testing new ads.

Meta also gives advice that points the other way. Its learning phase page says to avoid high ad volumes, because the system learns less about each ad when there are many.

Both are true. Test many ideas over the month, but only a few ads at a time.

Our rule: spend about 25 percent of the budget on new ads. Give each new ad about one sale's worth of spend in its first week.

The math is short. Multiply your weekly budget by 0.25, then divide by your cost per sale. That is your batch.

Example: $100 a day is $700 a week. A quarter of that is $175. At $35 a sale, that buys five new ads.

At $1,000 a day, the same math gives fifty. Fifty ads a week only makes sense at about $1,000 a day.

Is your batch under three ads? Release every two weeks instead, with two weeks of that budget.

Do you collect leads instead of sales? Use your cost per lead, and set it with your max allowable cost per customer.

Put each batch in a new ad set. Meta treats adding an ad to a live ad set as a significant edit, which sends it back into the learning phase. That is the testing period, when results are less steady.

How to burn a month of ad budget in four easy steps:

1. Make ten copies of your winner with new button colors. Call it testing.

2. Drop all ten into the ad set that already works.

3. Let an AI write the claims. Skip the read-through.

4. When costs climb, blame the algorithm.

Brand-new offer with no sales yet? Size that test with how much to spend on a new offer first.

Station one: write the lines before you make any pictures

The Volume Line starts with words. A line bank is a list of 30 or more ad openings, written before anyone makes an image.

Sort each line by what it leads with: a pain, an outcome, an objection or a piece of proof. That sorting makes ads different. Colors do not.

Meta says the same thing in its own example: one ad on emotional appeal, another on product features.

The best openings name the exact moment your buyer is in. For example, an ads service might open: you doubled the budget on a winner, and your cost per sale tripled overnight.

Where do lines come from? Your buyers. Sales calls, reviews, refund notes, the questions people ask before they pay.

AI can sort, trim and vary those lines. It cannot know your buyer for you. Left alone, it writes the average of everything it has read.

To teach it your voice, use the rule-building loop in our AI marketing workflows guide. For which words to lead with, see before-words and after-words.

Pictures got cheap. Knowing your buyer never did.

Station two: build a scene bank and an AI actor library once

A scene bank is a small library of pictures and short clips that look nothing alike. A kitchen, a truck cab, a desk, a street at night.

For video ads with people, add an actor library. That is a set of made-up presenters you build once and reuse in every ad.

Build it in one sitting, as the steps below show. Then make versions of the same face: older, younger, a new outfit, a new room.

For example, one AI ad producer made more than 800 actors in a day for about $52. From one set of about 200, he kept about 180.

The same producer made 18 one-minute videos for $22 in generation fees. That is about $1 a finished minute. Prices move fast, so check today's rates.

The pieces are cheap now. Good judgment is the costly part.

One hard rule. An AI actor may show and explain. Never use one to fake a before-and-after or a result nobody got.

Does the actor have to look perfectly real? No. For example, one skin-care brand's AI ad scaled to about $100,000 a day and looked obviously AI.

A more realistic batch for that brand had a weaker hook rate. Hook rate is the share of viewers who keep watching past the first seconds.

What wins is the opening, not how real the actor looks. More on that in the first three seconds of an ad and our UGC ad format library.

Station three: assemble in numbered pieces so one bad clip costs one fix

Assembly pairs lines with scenes. Ten lines times five scenes makes fifty ads, and nobody had to write fifty ads.

For video, cut each script into short chunks of about 120 characters. Each chunk becomes one clip.

Then put a small number in the corner of every clip. Clip one, clip two, clip three.

Why bother? When clip three goes wrong, you ask for clip three again. The rest of the ad stays put.

Without numbers, one bad second means rebuilding the whole ad. That is how a fifty-ad week turns into a five-ad week.

Trim with word timings. A transcript with a time on every word shows where each word starts, so cuts land between words.

Keep a pronunciation list. Computer voices trip on some words. Spell them the way they sound, put a space inside joined brand names, and skip all caps.

Cutting real footage instead? The same numbered, word-timed method runs our clip pipeline.

Station four: a person checks every ad before it goes live

The gate is a person who looks at every ad before release. They check every ad, not just a few.

Volume multiplies mistakes as fast as it multiplies ads. One bad claim in a template ships fifty times.

Check five things: the facts, the picture, the sound, the rules and the people. The table below says what fails at each one.

Sound gets missed most. A clean studio voice over a shaky phone video on a busy street sounds wrong at once.

Rules matter at volume too. Meta's personal attributes policy applies to all fifty ads, not just the first.

The legal line is about fake people with fake stories. The FTC's rule on reviews and testimonials bans testimonials from people who do not exist or who never used the product.

The FTC's own answers say there is no blanket ban on AI avatars. The problem is a fake or false testimonial. Courts can impose civil penalties for knowing violations.

Our rule is stricter. AI presenters explain. Only real customers, with written permission, say it worked for them.

An invented customer is not a creative choice. It is a legal risk with a face.

Stations five and six: release on a schedule, retire on a rule

Release one batch on the same day each week, into a new ad set. That way, each weekly review looks at one full week of results.

The review asks two questions. Which ads are tired? What did the winners have in common?

Tired means the cost per result has climbed while the same people keep seeing the ad. Turn it off. Do not edit it, because a change to the creative restarts learning.

Then write one line about why the winners won. Write the reason, not a description of the ad. That line goes back to station one.

How to read each result without fooling yourself is in our creative testing engine.

Hiring this out? Pay for a quality bar, not a count. A fixed number of ads a week tempts anyone to cut corners, like dropping the extra footage that makes a video feel real.

Stations one to three are where AI saves time. Station four is where it never works alone.

Meta's settings change often, so they live in the dated box below.

Plan your Volume Line with AI: the copy-paste prompt

Paste the prompt below into your preferred AI. It asks what you sell, your budget and your cost per sale. Then it sizes your weekly batch and drafts your first line bank.

It never invents your numbers. If it can browse, it checks Meta's current help pages first and tells you what changed.

It writes the lines. You still run the gate.

Start the line this week

1. Run the prompt. Write down your batch size and your review day.

2. Write the line bank. Thirty openings in four piles.

3. Ship the first batch into a new ad set, after a person checks every ad.

That nine-day winner will not come back. The next one is already on the line.

Stop hunting for the perfect ad. Build the place where good ads come from.