1. What exactly Meta changed in 2026

Meta officially rewrote click attribution: click-through now includes only real link clicks, while likes/comments/shares were moved into a separate engage-through metric with a different attribution window.
This removes "fake clicks" and brings reports closer to Google Analytics, so for many it looks like ads have "dropped," while in reality some conversions simply changed buckets in reporting or dropped out of it.
At the same time, Meta is finishing the move toward full automation: Advantage+, broad, auto placements, AI creatives, auto budgets.
Manual control of targeting, micro-audiences, and bids is being systematically pushed into the background, while the algorithm's weight in decision-making keeps growing.

2. Why ROAS drops with the same ad quality

After the attribution change, most people see at once: CTR dropping, CPA rising, part of conversions disappearing from Ads Manager.
This is often not a deterioration in business results, but a "clearing up" of metrics: the system stopped drawing success from interactions that never led to the site.
The problem is that the average buyer reads this as "the algorithm is broken" and starts chaotically twisting campaign structure, budgets, and audiences, which makes the algorithm's relearning period harder.
Your post already contains this idea: most people are simply working by 2021 patterns, where the main lever was account structure, not signal quality and data volume.

3. The new role of the buyer

Meta in 2026 works like an AI traffic allocation system that expects three things from the advertiser: lots of creatives, clean conversion signals, and enough volume for learning.
If a campaign gives 1-2 conversions per day against a small budget and a bunch of fragmented ad sets, it will never exit the learning phase, no matter how much you "craft" targeting.
Hence the new role of the buyer:

not choose the audience, but build a structure where the algorithm finds the audience itself on broad targeting;
not micro-manage bids, but ensure a stable flow of relevant conversions;
not "draw" two banners, but systematically launch dozens of creative angle variants per month.

In essence, the buyer becomes a growth architect who designs an environment for AI learning, rather than manually controlling every screw.

4. Account structure 2021 vs 2026

Old logic:

10-50 ad sets, dozens of interests, lookalikes of different sizes, manual bids, duplicating campaigns for scaling.
Scale = multiply structure: copies of campaigns/ad sets, raising bids, rigid micro-control.

New logic:

1-3 broad audiences (geo + age + sometimes minimal filtering) and maximum automation: Advantage+ Shopping/App, broad, auto placements.
Scale = build a system for testing creatives and hypotheses, not a spreadsheet with 50 ad sets.

"3-level" model:

Testing Campaign - rough selection of creatives and angles.
Scaling Campaign - moving winners with a normal budget.
Evergreen Campaign - stable links that run for months.

It is exactly a structure where the number of campaigns is minimal, while internal variability (creatives, formats, messages) is maximal, that gives AI room for optimization.

5. Creative as the main lever in 2026

Meta's algorithm in 2026 is essentially a pattern-recognition system that looks for links between creative features, people's behavior, and conversions.
If you give it 2 banners with the same text, it simply has nothing to draw conclusions from, and all automation loses meaning.
Here it is logical to highlight what you call "scale through creative angles":

not images are tested, but hypotheses: pain, social proof, product demonstration, storytelling, before/after;
the best teams make dozens of variants per month, allowing AI itself to decide what "worked."

In the next 12 months, the trend will only strengthen: Advantage+ is taking over targeting and placements more deeply, so the difference between a good and a bad buyer will be almost entirely in creative thinking and building testing infrastructure.

6. Data as fuel: why everything breaks without 50+ conversions per week

Any AI model needs statistics: the more valid conversions and the cleaner the signals, the faster and more accurately the algorithm finds the right people.
If a campaign does not reach even the basic 50+ conversions per week (or at least the new lowered thresholds for modern Advantage+), it gets stuck in permanent "learning" and behaves randomly.
This aligns with your message: "scale is not about audiences, but about a testing system."
In reality, scale = enough budget on a small number of structures to give AI statistics for choosing the best combinations, and only then scaling those, not duplicating everything in a row.

Private survey of buyers on scaling methods: