Andromeda Creative Research: Find Different Angles, Not More Copies

MEASURED: AI-generated editorial illustration, not a product screenshot or observed competitor ads.

I am Agnes Maina, and I built the Meta Ad Library actor linked here. MEASURED means read in documentation or cited sources, not a new live test. INFERRED marks recommendations and example reasoning.

What the sources say

MEASURED: Meta Engineering described Andromeda on December 2, 2024 as an ads retrieval engine, selecting candidate ads before later ranking. Meta Engineering.

MEASURED: A Meta Blueprint lesson dated May 18, 2026 covers creative diversification across formats, visuals and messages. Meta Blueprint.

INFERRED: Two cautions before going further. Andromeda is not a new October update, and those sources do not establish why a particular account's results dropped. Advertiser discussion helped me choose this topic, but it is not controlled performance evidence.

INFERRED: My practical takeaway is narrower: when planning a creative batch, count distinct messages as well as files. Twenty versions of the same promise may leave important customer questions unexplored. That is a research recommendation, not a claim about how Meta will score your ads.

Concept or recolor

INFERRED: A recolor changes the background, swaps a headline word or crops the same footage differently. A concept changes who the ad is for, what it promises or what it offers as proof. A quick check: describe two ads in one sentence each without mentioning colors or layout. If the sentences match, you may have one concept.

The creative matrix

INFERRED: Use one row per concept, not per file. These examples are for an imaginary meal kit brand; they are not observed ads.

Audience problem	Promise	Proof	Format	CTA
No time to plan weeknight dinners	Dinner decided in advance	Customer cooking on camera	Short video	See this week's menu
Food wasted at the end of the week	Buy only what gets cooked	Comparison of two fridges	Static image	Build a box

INFERRED: Notice that the promise and proof change together. The first row speaks to planning. The second speaks to waste. Changing the background color of the first row would not answer the second customer's question.

Steps

INFERRED: Keep collection and judgment separate:

1. Put your own live ads into the matrix first. Several files may collapse into one row.
2. Pick advertisers in your category and a few neighboring ones selling to the same audience.
3. Open the Meta Ad Library actor. Enter them under advertisers, or use searchTerms for a problem phrase. Set country to US, activeStatus to active and maxAdsPerSearch to 20.
4. Read the available copy and creative. Open libraryUrl when the output alone is unclear.
5. Classify each ad into the matrix. When an ad repeats an existing concept, add a tally rather than another row.
6. Compare crowded rows and empty combinations within your sample.
7. Write three to five hypotheses that differ meaningfully from your current concepts. Explain the customer question each one answers.
8. Test them in your own campaigns under your own budget and measurement rules.

Limits of this research

MEASURED: The actor documents available copy, creative, libraryUrl and daysRunning, with a collection cap per search. The source determines what is available. See the actor page. I have not rechecked live pricing for this draft; review current Store prices before starting.

INFERRED: A long runtime does not prove profitability, ROAS or spend. An ad absent from a later limited sample does not prove it was stopped. Twenty ads per search are a reading list, not a census.

INFERRED: An empty cell might be an opportunity, or an angle others tried and dropped before you looked. The matrix cannot tell those apart. It helps you describe what is different about your next idea. Use other advertisers' ads for structure, then write your own words and make your own visuals. Nothing here promises better results.

INFERRED: For collection, use the Meta Ad Library Scraper. Classification stays manual. That is where you decide whether another file actually adds another idea.