How to get the most out of Meta Advantage+ and Google Performance Max: build the outer loop

Meta and Google now ship the thing that used to be a media-buying team. Turn on Meta's Advantage+ or Google's Performance Max and the platform decides who sees your ad, where it runs, what to bid, and how to pace your budget — a full optimization loop, learning from every conversion, included with your spend. That part of performance marketing is now free leverage — and none of it is a moat, because it ships to everyone.

Which raises the question this post answers: if everyone gets the same machine, what still separates the brands that compound from the brands that plateau? The answer isn't inside the ad account. It's in the loop a brand builds around the ad account — what it feeds the machine, and what it takes back out.

Key takeaways

• The platform now runs the routine buying — targeting, placement, bidding, pacing — and that leverage ships to every advertiser.
• Your side of the contract is the input list the platforms publish: structure, signals, creative, measurement.
• The real bottleneck is creative volume, not settings.
• The only part that compounds for you alone is your own record of what won and why — the outer loop.

What do Advantage+ and Performance Max actually do — and what do they ask for?

The deal the platforms offer is simple: they run the auction-side optimization, you supply the inputs. Meta's Advantage+ suite and Google's Performance Max both work the same way underneath — a model learns from every conversion you send it and continuously reallocates delivery toward whatever is converting. The routine decisions a paid team used to make daily — which audience, which placement, what bid — happen inside that loop now, at a speed and granularity no weekly optimization meeting can match.

And both platforms are unusually explicit about what they want from you in return. Meta publishes it as a named framework — the Performance 5: simplify your account structure so the model gets enough data to learn, lean on the Advantage+ automation instead of overriding it, diversify your creative, feed it conversion data through the Conversions API (the direct data connection between your store and Meta, so the model sees conversions the browser no longer reports), and validate results with real measurement. Google's Performance Max guidance makes the same asks in its own vocabulary: creative assets in volume — headlines, descriptions, images, video — plus audience signals to start the model in the right neighborhood, and clean conversion tracking so it optimizes toward the right outcome.

Read those two documents together and something useful falls out: they are a job description. Structure, signals, creative, measurement. Nothing on the list is a knob inside the auction — every item is an input the machine can't generate for itself. The teams getting the most out of the automation are the ones that took this reassignment seriously, and re-pointed the hours the machine freed up at the input list.

The bottleneck: the machine eats inputs faster than you make them

The machine's appetite for fresh creative outruns most brands' production cycle — and that, not settings, is where most setups underdeliver. The automation tests combinations constantly, fatigue arrives at machine speed, and repackaged variations don't count as new input: under Andromeda, Meta's overhauled ad-delivery ranking system, visually similar creatives are collapsed into a single entity, so near-identical ads compete as one. Resizes, recolors, and new headlines on the same concept are — to the model — the same ad. And the appetite has a real number attached: the DTC agency Pilothouse puts the Andromeda-era production target at 80–100 diverse creatives a week, built on a modular production system rather than one-off design work.

That changes the production math. A brand shipping one creative batch a month is feeding a machine that wanted new concepts weekly. The visible symptom is the one every growth lead running these campaigns has seen: strong first weeks, then decay, then a debate about whether the automation "stopped working." Usually it didn't stop working. It ran out of genuinely different things to test. The teams that adapted rebuilt their creative production pipelines to run at the machine's speed — and producing real variety at that cadence, without the cost scaling in step, is exactly the production problem AI-native systems exist to solve. (We mapped the buying-side half of that shift — agentic media buying underneath the platforms — earlier this year; this series is about the account-level work.)

The part you can't see: the platform keeps the learning

The platform's optimization loop is closed to you — that's the second asymmetry, and the one nobody's dashboard shows. Its models learn from your budget — every conversion sharpens their read of who buys from you — but you can't query that learning, it doesn't explain why a winning ad won, and none of it transfers when you pause, switch platforms, or leave. We've written about what a brand keeps when it rents intelligence; platform ad automation is the sharpest version of the trade. You're renting the best media buyer that has ever existed, and its notebook stays in its pocket.

YC partner Diana Hu argues that a company running self-improving agents should be built as a closed loop, and names three pillars: closed-loop coverage, queryable data, an intelligence layer on top. In paid media you cannot build the first pillar inside the platform, because the platform's loop is already closed and it isn't yours. What you can do is wrap your own loop around it.

The outer loop is the learning system a brand runs around platform ad automation: a queryable record of which ads ran and what happened, an intelligence layer that extracts why the winners won — angle, hook, offer — a file where those learnings are written down, and next week's creative generated from that file. The platform optimizes delivery inside its loop; the outer loop makes each batch of inputs smarter than the last. The line between automation and an AI-native system is whether a feedback loop closes on the system's own output — the outer loop is that line, drawn around a machine you don't own.

What one pass around the outer loop looks like

One honest caveat before the walkthrough: the ad example below is illustrative — the discipline is real, the specific numbers deliberately aren't.

Read. Monday, you pull the week's results at the level the platform does show you: which assets, which concepts. Suppose the read is that the price-led hook ("from $49, today") outperformed the destination-led hook ("your weekend in Lisbon") across most placements. The platform reports the what. It will not tell you the why — that's the intelligence-layer job, human or AI: the price-led concept is winning while the seasonal sale runs, so the offer is doing the work.

Write. That inference goes into the learnings ledger as a dated, one-line rule with its evidence attached: "Price-led hooks outperform destination-led while a discount is live — candidate, one week of evidence." This is the discipline we run on our own content engine, and here is a real line from our own file so you can see the shape: lead a hook with the reader's stake, never a concept label — across a 30-post sample on a founder's LinkedIn, concept-label hooks ran 0–2.3% engagement while the winners ran 4.9–8.3%. That rule entered our ledger as a candidate, was confirmed against another month of posts, and is binding on every hook we write today; it gets retired the week it stops being true. The mechanics carry over to paid unchanged — and the ledger is a file you own, which means it survives every platform change, tool swap, and contractor handoff.

Regenerate. Next week's brief opens from the ledger: five new concepts, each testing a different angle on the price-led premise the ledger just surfaced — each different enough, by design, to earn its own slot in the auction instead of being clustered with the last batch. The machine gets fresher inputs than your competitor's; the batch after this one starts from an even sharper file.

That's the whole loop: read what you can see, extract why, write it down, feed it back. None of it interferes with the machine — it makes you the advertiser the machine works best for.

The weekly cadence, as a checklist

The outer loop runs on a week. One honest boundary before the checklist: at low spend, a week of ad-level results is mostly noise — stretch the cadence until each read has enough conversions behind it to mean something. The loop's shape survives; the week doesn't have to. Here is the Monday version, with what healthy and starving look like for each step:

• FEED — ship genuinely different concepts. Healthy: every batch tests distinct angles, formats, offers. Starving: this month's batch is last month's winner in four aspect ratios.
• CHECK — signal health before creative verdicts. Healthy: Conversions API connected and event quality reviewed, so the machine optimizes on real conversions. Starving: judging creative on a week of undercounted data.
• READ — results at the level you can see. Healthy: a standing weekly read of concept- and asset-level results, kept in your own record. Starving: checking the dashboard when someone asks if ads "are working."
• WRITE — the why, into the ledger. Healthy: dated one-line learnings with evidence, candidate until repeated, binding once confirmed. Starving: insights that live in a Slack thread from March.
• REGENERATE — next batch from the ledger. Healthy: every brief opens with the current binding rules. Starving: every brief opens with a brainstorm.

If you run paid on Advantage+ or Performance Max today, one question tells you whether your outer loop exists at all: did last week's asset-level results change anything about this week's creative? If the answer is no, the loop is open — and the machine is optimizing inputs that never improve.

Where to start this week

Three moves, in order. First, stop fighting the machine where you still are: consolidate fragmented campaigns and give the automation the data volume it needs to learn — the platforms' own input lists above are the checklist. Second, treat that input list as the team's actual work now, and move the freed hours to creative, signals, and measurement. Third — before the next creative batch ships — open the ledger: one file, one dated learning per line. It's the least glamorous artifact in your stack and the only one that compounds for you alone, because the machinery improves every quarter for everyone. Your competitors get the same automation. They don't get your file.

The next post in this series takes the step the ledger depends on — reading results you can actually trust — and makes the measurement side concrete.

Book a strategy call →