Creative Entropy Is the Last Targeting Lever Left on Meta

Your audience settings stopped mattering somewhere around late 2025. If you run Advantage+ campaigns, Meta has been quietly ignoring most of your targeting inputs for a while now; what changed with Andromeda is that the machine started reading your creatives as the targeting layer instead. Different hook, different audience pocket. Same hook in a new font, same audience pocket, faster fatigue.

That flips the job description. You are no longer a media buyer who briefs creative. You are a creative portfolio manager who happens to touch Ads Manager.

What Andromeda actually changed under the hood

Quick mechanics, because the strategy only makes sense once you see the machine. Before your ad ever enters an auction, Meta runs a retrieval stage: billions of eligible ads get filtered down to a few thousand candidates per impression. Andromeda is the rebuilt retrieval engine, announced by Engineering at Meta in December 2024, running on NVIDIA Grace Hopper hardware with roughly 10,000x the model complexity of the system it replaced. Meta reported a +6% recall improvement and +8% ads quality lift on selected segments from the rollout.

Here is the part that matters for you: a retrieval system that complex is doing semantic and visual analysis of the creative itself. It knows your video opens with a talking head. It knows your static leads with a price claim. And it uses those features to decide which impressions your ad is even considered for.

So when you upload six variants that are the same UGC clip with different end cards, Andromeda sees one ad. It retrieves them for the same people. They saturate the same pocket. Then they die together, and your dashboard tells you "creative fatigue" as if it were weather.

The money confirms Meta is all-in on this architecture. On the Q3 2025 earnings call, Zuckerberg said the annual run rate for Meta's end-to-end AI ad solutions, Advantage+ included, had passed $60 billion, per Yahoo Finance. Nobody is rolling that back so you can have your detailed targeting fields again.

Creative entropy, defined properly

Entropy here means the amount of genuine informational difference across your live ad set. Two ads have high entropy when they differ on axes the retrieval model actually encodes: the angle of the argument, the format, who is on camera, what the first two seconds do, what kind of proof gets offered.

Meta has published the payoff. Campaigns with diversified creative showed up to 32% improvement on CPA and 9% better incremental reach in Meta's own research, as reported by B&T.

Now the uncomfortable bit. More than 4 million advertisers were already using at least one of Meta's generative AI creative tools by Q4 2024, up from 1 million just six months earlier, per AdExchanger. Everyone can flood the auction with variants now. When everyone has volume, volume stops being an edge. The auction is drowning in AI-resized, AI-recolored, AI-rephrased near-duplicates, and Andromeda's whole job is to cluster those and treat them as one.

One-liner to tape to the monitor: ten concepts beat a hundred variants.

The creative entropy matrix

This is the template we run at EGGKNITE's paid media practice for every Advantage+ account. Build your production queue against axes, never against "we need more ads." Every new ad must claim a cell combination that is not already live.

Where does AI generation fit? Downstream. Lock the concept-level entropy by hand, then let gen AI multiply within a cell: five hook variations of one UGC concept, aspect-ratio versions, headline swaps. AI is your intra-cell multiplier. Humans own the cells. The same logic applies whether you are prospecting cold or feeding remarketing pools: even warm audiences get retrieved against creative features now.

Naming conventions are your measurement system now

You cannot manage entropy you cannot query. With audience breakdowns gone, the ad name is the only reliable metadata field you fully control, and it survives every export.

Our convention encodes the matrix directly:

YYYY-MM_stage_angle-X_fmt-X_msgr-X_hook-X_vNN
2026-02_prosp_angle-price_fmt-ugc_msgr-creator_hook-question_v03

Six fields, underscore-delimited, no spaces, no creativity in the naming (save it for the ads). Now a pivot table on exported data answers questions Ads Manager never will: which angle is fatiguing account-wide, whether founder-face ads outlast creator ads, whether your live mix has collapsed to two cells without anyone noticing. If a name does not parse, the ad does not launch. Treat it with the same discipline you would apply to UTM parameters, because it is the same job: making machine decisions auditable.

Punchy version: your naming convention is your new breakdown menu.

Fatigue detection when the machine hides the levers

Fatigue in the Andromeda era is cluster-level, so watch it at the axis level, using your naming convention to aggregate. Remember that past a certain exposure count, repetition does not merely stop working; it actively annoys the exact people the model found for you, and the cluster's delivery costs climb to match.

Practical trip-wires we use, checked twice weekly on 7-day windows:

• CTR decay: a sustained 25–30% drop from the ad's 7-day peak, held for 3+ days. One bad day is noise.
• First-time impression ratio sliding under ~60% on prospecting campaigns while spend holds. Meta is recycling eyeballs.
• CPA drift with stable downstream conversion: if landing page CVR is flat but CPA is up 40%, the creative cluster is exhausted, and your funnel is fine.
• Axis-level correlation: three ads sharing an angle tag all decaying in the same week means retire the angle, and no amount of new executions of it will help.

When a cluster trips, you do a structured swap: replace it with a cell combination from the matrix that is currently dark. Random refreshes just re-enter the same crowded pocket. And resist the urge to fix fatigue by fragmenting campaigns; consolidated structure is still what feeds the model enough signal, a point that pairs with the consolidation logic in our guide to scaling Facebook ads.

The weekly pipeline that keeps entropy up

Cadence beats heroics. What survives contact with real accounts:

1. Monday: export, parse names, score entropy (how many matrix cells are live and spending). Flag tripped fatigue wires.
2. Tuesday: brief 3–5 new concepts, each claiming a dark cell. Gen AI multiplies each into 4–6 variants.
3. Thursday: launch into the existing consolidated campaign. No new ad sets, no audience edits.
4. Friday: kill anything past 2x target CPA with statistically real spend behind it, using the same hygiene rules from our Facebook ads optimization checklist.

Measure the system on blended CPA and account-level ROAS trend (run yours through our ROAS calculator monthly), never on individual ad win rates. In a high-entropy account most ads "lose," and that is the design: you are funding exploration so Andromeda can find pockets you could never have targeted by hand. Teams already doing AI-driven segmentation on the data side, like the approach in our ecommerce ad targeting playbook, will recognize the pattern: you shape the inputs, the model does the matching.

The targeting menu is gone and it is staying gone. The advertisers winning right now made peace with that early and reallocated the hours: less time in audience settings, all of it into a production system that keeps feeding the machine genuinely different arguments. Entropy is a supply problem. Build the supply chain.

Sources

• Engineering at Meta: Andromeda retrieval engine announcement, recall and quality figures
• Yahoo Finance: Meta Q3 2025 ad results coverage, $60B AI ad tools run rate
• AdExchanger: Q4 2024 earnings, generative AI advertiser adoption
• B&T: Meta research on creative diversification, CPA and reach lifts