paid media.the engine never sees your campaign, it sees every creative on its own.
Meta split the decision of which ad you see into two stages. The first one reads the creative, not the targeting. That is where creative volume stopped being a production luxury.
in three lines
- Andromeda shortlists which ads are eligible; ranking and the auction happen after that.
- Every genuinely distinct creative gets its own entity ID and competes on its own.
- The working threshold is 10-15 concepts a month, and a recolor is not one of them.
retrieval is a stage of its own
When someone opens Instagram, Meta does not pick one ad out of millions in a single pass. A retrieval engine called Andromeda shortlists a few thousand eligible candidates first. Ranking and the auction only work on that subset: GEM decides what should be shown next, Lattice settles the bid. Andromeda decides what can be shown; the other two decide what is.
Meta announced it in December 2024 and finished rolling it out across Instagram, Facebook and Messenger through 2025. It runs on NVIDIA hardware alongside Meta's own silicon, which lets it serve models up to 10,000 times more complex than the system it replaced. The figures Meta reports for it are a 6% improvement in retrieval recall and 8% in the quality of the segments it selects.
the creative is the targeting now
Here is the part that changes the daily work. Andromeda reads the creative itself, the image, the video and the copy, and decides who to show it to on that basis. Interests, lookalikes and custom audiences carry less weight than they used to. Every genuinely distinct creative is assigned its own entity ID and competes on its own inside the system.
That reorders where performance gets read. The signals live at the ad level, not the ad set. Pausing a whole ad set because it does not seem to work usually buries one or two creatives that were working. And Meta's own guidance moved from 3 to 5 variations per ad set to 10-15 distinct concepts per Advantage+ campaign.
distinct does not mean recolored
This is where most plans quietly fall apart. Changing the color, the soundtrack, the CTA or the headline does not produce a new concept. It produces a variant that competes against its own original. The same goes for cutting one video at 15 and at 30 seconds, or pulling another take of the same creator saying the same thing.
The test is simple: put the two pieces side by side, and if they tell you the same story, they tell Andromeda the same story. A distinct concept changes the hook, the person, the format, the angle or the narrative. Same testimonial from a different person counts. Same testimonial, different cut, does not.
what happens below the threshold
Ten to fifteen concepts a month is not a vanity number. It is the raw material the engine needs in order to have something to optimize with. Below that, the system runs out of variants to test and performance flattens. There is a second effect as well: once a creative fatigues, Meta stops giving it volume even though the budget is still sitting there.
The concepts also have to be spread out. The working rule is to distribute them across at least three to five different families, human voice, structured narratives, visual formats and culture, rather than shipping ten pieces from the same one. In regulated verticals the problem compounds, because Meta restricts Conversions API use in health and wellness: the engine receives fewer conversion signals and needs even more creative diversity to make up for it.
Andromeda does not evaluate your campaign. It evaluates each creative on its own. If two pieces tell the same story, the engine counts them as one, however your spreadsheet lists them.
what to do with this
count concepts, not files
Before signing off a monthly plan, run every piece through one question: does it tell a story another piece already tells? Twelve files in a folder can be four concepts and eight variants.
spread across families
Ten UGC pieces are a single kind of bet. Three to five different families, human voice, narrative, visual format and culture, cover more segments and give the engine something to test.
read reporting at ad level
If the dashboard only shows ad sets, you are not looking at where the decision happens. The unit of analysis is the ad. A weak ad set can be hiding the best creative of the month.
if you are under the threshold, say so
Four concepts a month can be run, but it is a floor, not a plan. The difference is stating the expected performance before you start instead of explaining it in the report.
questions we get
How many creatives do I need per month on Meta?
The threshold in use today is 10 to 15 distinct concepts per month per Advantage+ campaign. You can run fewer, but the engine has less material to optimize with and performance tends to flatten sooner.
Does changing the color or the copy count as a new creative?
No. A recolor, a new headline, different music or a different CTA are iterations of the same concept: they compete against their own original instead of adding diversity. To count as a new concept, the hook, the person, the format or the narrative has to change.
Is Andromeda the same thing as the Instagram algorithm?
No. Andromeda is the ad retrieval engine: it shortlists which ads are eligible to appear. The organic Instagram feed runs on a different recommendation system. Sharing a platform does not make them the same model.
where this comes from
- Engineering at Meta — official Andromeda announcement, December 2024.
- Adam Mosseri, Head of Instagram — public statements on the algorithm and its signals.
- Search Engine Land, Jon Loomer Digital, ALM Corp and the Logical Position 2026 Playbook.
- Champe Agency — internal Meta Ads knowledge base, re-verified against the primary source on 12 August 2026.
Sources are cited as text, with outlet and date. We don't link addresses we haven't verified.
and what it takes
The note explains the problem; the services explain the work.
the other notes
running on agents
Plenty gets said about agencies using AI, and very little about how one is actually run. This is the method from the inside: how knowledge is organised, where the human sits, and what breaks when nobody looks after it.
read the note → 05 · August 28, 2026reading year-over-year comparisons
Global CPM is up 32% since 2024 and US inventory is at an all-time high. A campaign can look worse in this year's report without having gotten worse for a single day.
read the note → 04 · August 28, 2026originality and reach
Instagram no longer rewards originality out of principle. It measures it. Reposting loses reach in ways you can verify, the penalty has moved beyond video, and the hashtag no longer makes up for anything. Here is what counts.
read the note → 03 · August 28, 2026measuring without e-commerce
Most real work doesn't end in a cart. It ends on a shelf, in a store, in a doctor's appointment. That can be measured. What can't be done is inventing the metric at month end.
read the note → 02 · August 28, 2026meta placement controls
Meta has finished retiring the manual placement selector. What remains are value rules with a hard floor of -90%. Brand safety stops being a list of unchecked boxes and becomes a production decision.
read the note →