Meta’s Andromeda update: everyone selling you the fix is the problem
Have you noticed that everyone explaining the Meta Andromeda update also happens to sell the cure?
Creative automation tools. Reporting dashboards. “Andromeda-proof” agency retainers. The scarier the explanation, the closer you are to a checkout page.
So let’s do something radical. Let’s read what Meta actually published, listen to the operators spending real money, and work out what — if anything — you should change.
What Andromeda actually is
Pin this definition to the wall: Andromeda is Meta’s machine learning system for the retrieval stage of ad delivery — the step that narrows tens of millions of eligible ads down to a few thousand candidates before the auction even happens. Announced December 2024, rolled out through 2025, built on NVIDIA Grace Hopper chips with a roughly 10,000x increase in retrieval model capacity (Meta Engineering, 2024).
In plain terms: before Andromeda, your ad’s biggest fight was the auction. Now there’s a bouncer at the door before the auction, and the bouncer reads your creative — the visuals, the copy, the offer, the persona it speaks to — and decides whether you’re even in the room.
Two things follow, and they’re the entire story.
First, near-identical ads get clustered and treated as one. Cody Plofker, CMO of Jones Road Beauty, had a Meta rep audit his account and found ads he considered different were flagged as similar — similarity is judged on “the whole package; visual, copy, offer, and most importantly persona” (Plofker, X, October 2025). Your fifteen headline variations? Meta sees one ad.
Second, your creative became your targeting. Different stories get retrieved for different pockets of people. Same story in a new colourway gets retrieved for nobody new.
The numbers Meta actually published (and the ones it didn’t)
Meta’s own reported gains from Andromeda: +6% recall improvement and +8% ads quality improvement “on selected segments” (Meta Engineering, 2024).
Six and eight percent. On selected segments. Hold the front page.
Now compare that to the numbers doing laps of your LinkedIn feed. A “60% similarity threshold” that triggers suppression. “65% higher ROAS” for brands testing 20+ ads a month. Fatigue windows collapsing to two weeks.
None of those numbers appear anywhere in Meta’s published material. I went looking. They originate from vendors selling creative automation and reporting tools, citing studies nobody can locate. Even “Entity ID” — the term every explainer thread treats as gospel — doesn’t exist in Meta’s public documentation.
The panic is the product. Fear of an algorithm you can’t see is a very effective sales mechanism, and 2025’s Andromeda discourse was a masterclass in it.
Meanwhile, here’s what operators with real budgets reported. Aazar Ali Shad, who manages over $7M a month in spend, called Andromeda “nothing more than a snake oil scheme” — his accounts, already running differentiated concepts, saw no performance change (LinkedIn, via Motion, November 2025). Marin Istvanic at Inflow: Andromeda changes how Meta selects ads but rewards exactly what it always rewarded — volume plus genuine diversification (X, October 2025). Jess Bachman at FireTeam, asked what her agency was doing about Andromeda before BFCM: “Absolutely nothing… what wins BFCM is offers, not creative” (Motion, November 2025).
So why does it still matter?
Because Andromeda is a mirror, not a mandate.
Nothing in Meta’s engineering disclosure requires new tactics. What it did was expose — mechanically, at the retrieval layer — that most of what brands called “creative testing” was recycling. If your last quarter’s testing log is one winning concept re-edited fourteen ways, Andromeda didn’t break your account. It just stopped paying you for pretending.
And the response from the sharpest operators wasn’t more ads. It was fewer.
Connor Rolain, Head of Growth at HexClad: “We used to make seventy-five ads a campaign. Now we make six that are genuinely different… you unlock new pockets of audience only when you actually change the story you’re telling” (Marketing Operators, via Motion, November 2025).
Seventy-five down to six. Read that again.
Dara Denney — who builds performance creative for a living — caps her teams at two variations per concept (Motion, November 2025). The unit of volume changed. It’s not ads any more. It’s stories.
What an AU brand should actually do
Most Andromeda advice is written for US accounts spending $1,000+ a day. If you’re an Australian DTC brand doing $200–500 a day, blindly running the “8–20 creatives per ad set” prescription — a vendor heuristic, incidentally, not Meta guidance — splits thin signal across too many entities and can make things worse.
The Australian reality: Meta reaches roughly 17.2 million Australians on Facebook and 14.3 million on Instagram (DataReportal, Digital 2025 Australia). The audience is there. The constraint is your signal density.
So here’s the tactical map, sized for the market we actually operate in:
Count concepts, not ads. A concept is a distinct persona × desire × message. New hook on the same story is a variation, not a concept. Audit last month’s ads honestly: most accounts I look at are running two concepts wearing twelve outfits.
Run 3–6 polar-opposite concepts. Different persona, different problem, different proof. A testimonial about sensitive skin and a founder rant about ingredient sourcing get retrieved for different people. That’s the diversity Andromeda prices — and it’s diversity of message, not look. New creator, same script, is the same ad.
Cap variations at two per concept. Denney’s rule. Spend the saved production budget on the next concept instead.
Ignore any number that doesn’t trace to Meta or a named study. If someone quotes a similarity threshold percentage, ask for the source. Watch what happens.
Offers still outrank all of it. Bachman’s point stands. Creative diversity gets you retrieved; a weak offer still loses the click.
In my experience running AU food and consumer brands through 2025’s rollout, accounts that consolidated into fewer, sharper concepts held or improved efficiency, while accounts chasing volume mostly funded their own confusion. That’s experience, not a study — weight it accordingly.
The signal worth watching isn’t a vendor threshold. It’s that Meta now ships Creative Fatigue and Creative Similarity metrics natively in Ads Manager (Meta Business Help Centre, 2026). When the platform builds the diagnosis into the dashboard, it’s telling you what it wants. Distinct stories. Fewer costumes.
If you want a second pair of eyes on whether your account is running six concepts or one concept in six outfits, that’s a conversation worth having — book a growth call at thrive-agency.co.
Your move.
FAQ
What is Meta’s Andromeda update?
Andromeda is Meta’s machine learning system for the retrieval stage of ad delivery, announced in December 2024. It narrows tens of millions of eligible ads down to a few thousand candidates per impression before the auction, using semantic understanding of the creative itself. Meta reported a 10,000x increase in retrieval model capacity, +6% recall and +8% ads quality on selected segments.
Does Andromeda penalise similar ads?
Penalise is the wrong word — it consolidates. Ads that are substantially similar across visuals, copy, offer and persona get clustered and treated as one candidate, so variations don’t buy additional retrieval. No published Meta threshold (such as a “60% similarity score”) exists.
How many ad creatives should I run after Andromeda?
There is no official Meta number. Operator practice from high-spend accounts points to 3–6 clearly distinct concepts with at most two variations each, rather than large volumes of similar ads. HexClad publicly moved from 75 ads per campaign to six distinct ones.
Did Andromeda change Meta ads for small advertisers?
The mechanics apply to every account, but prescriptions built for US-scale budgets don’t transfer directly. Smaller accounts (under ~$500/day) generally do better consolidating spend into fewer distinct concepts to preserve signal density.
