The Great Unification of Ad Algorithms: Meta, Google, and TikTok Built the Same Building

A while back, a friend who had spent nearly a decade running ad campaigns took me out to dinner. Before the dishes even hit the table, he let out a sigh.

"I think I'm about to be out of a job."

I was taken aback. A ten-year veteran — why would he say something like that?

His bread and butter, he explained, was slicing audiences into ever-finer layers. Gender, age, interest — cross them again and again, then factor in device type on top. It was a craft he had honed for nearly five years, and it had long been his team's trump card.

Then, last month, an intern launched an ad with no audience breakdown whatsoever — and it outperformed his.

I actually felt relieved hearing that. This wasn't his problem alone.

Over the past two or three years, a very quiet transformation has swept through the ad industry. Meta, Google, and TikTok — three rivals who had been at each other's throats for years — have, without conspiring, all built the same building.

Today, let me walk you through the blueprint.

What "Algorithm Convergence" Means

Let me first translate that phrase into plain language.

Each company gives its product its own name: Meta calls it Andromeda, Google calls it Performance Max, TikTok calls it Symphony. On the surface, they look like entirely different things.

But strip away the storefronts and you will find the same blueprint behind the load-bearing walls.

The blueprint rests on three pillars.

First, hand the job of finding people to the algorithm — stop defining "who is my target" yourself.

Second, broaden the audience range and give the algorithm room to maneuver.

Third, treat "can you produce creative at scale, and produce richly varied creative" as the single factor that decides who wins.

In one line: the algorithm finds the people; the creative makes them pay.

That audience net you have been clutching so tightly is failing — and across a wide swathe of the map.

First, Meta's Andromeda

Let's start with how the old system worked.

The old flow had three steps: filter the audience, score the people, then release budget according to the scores.

The choke point was step one. And that first filter was something the advertiser had to tune parameter by parameter — pick gender, pick age, tick interest boxes. The moment you locked in "ages 25 to 34, into fitness," everyone outside that circle who might genuinely buy from you could never see your ad.

That is the fatal flaw of manual targeting: you can reach the people you can imagine; the people you cannot imagine never even get near your ad.

Andromeda's move is to remove step one.

The name surfaced in Meta's own engineering posts from 2023 and 2024. What it does fits in a single line:

It turns "finding an audience" from "checking a roster" into "running a prediction."

What does "checking a roster" mean? You flip through a very thick book and look up who you are. Not in the book? Then you are not in.

What does "predicting" mean? You don't open the book. You pour the behavior of billions of people across the entire platform into one giant neural network and let it learn: what kind of behavior marks the most probable buyer for a product in my category?

You may never have searched for this brand, never clicked its ad, never favorited it. But over the past three weeks, you have kept hovering over similar products, comparing them, adding them to your cart again and again.

Its verdict? Scoop you up.

No labels — behavior alone. That is the entire gap between Andromeda and every legacy system.

Meta has published a number, too: its Advantage+ shopping campaigns deliver, on average, a 32% higher ROAS (Return on Ad Spend) than manual targeting. The figure comes from the company's own side-by-side tests, so take it with a grain of salt — but the direction is unmistakably right. Plenty of people in the DTC (direct-to-consumer) world who opened up their audiences and stacked their creative are seeing the same curve.

Which is how this line became the catchphrase of this round:

"Creative is the new targeting."

Google: Even "What You Bid On" Has Been Taken Away

In November 2021, Google launched Performance Max, which replaced the earlier Smart Shopping.

Back then, people took it for a convenience feature: it would auto-bid and auto-pool your budgets for you.

But looked at in hindsight, it becomes clear that Google swallowed the "channels" whole.

Google ads used to be run as separate buckets: search in one bucket, shopping in another, YouTube in a third, display in a fourth. Each bucket had its own audience and its own pacing, and each one needed a handler.

Performance Max will not let you manage any of that. It asks for exactly three things: creative assets, an objective, and a budget.

You hand it images, videos, headlines, and descriptions — it runs on those. You type "I want a ROAS of 4," or "each purchase has to land under this price cap" — it goes straight for it. The budget is the one thing you set yourself.

As for which cohort sees you in which context, at what pace your money is spent — it works all of that out on its own.

By 2024, Google had also wired semantic understanding into search. Instead of naive keyword matching, it now directly computes how many degrees of separation lie between "what the person behind this search really wants" and "what is inside my creative."

The wall of audience targeting has been dismantled, completely.

Here is a counterintuitive example: a person who has never searched your brand keyword. But in recent days, his searches, his browsing, his comparing behavior all carry a strong purchase intent and line up closely with your converting audience. The system said nothing, and quietly swept him into the Performance Max pool.

Smart operators have already switched their question. They no longer ask "how much should I bid, and on which keyword," but "which set of creative am I preparing, for which type of person?"

One question changed — and an entire era hands over with it.

Third, TikTok's Symphony: Creative Becomes an Assembly Line

TikTok's Symphony launched in 2024.

When it first arrived, it carried the "AI creative tool" label, and a lot of people figured: isn't this just a video-editing feature?

That reading isn't wrong — just too small. Symphony's real birthplace sits right next to it, in Smart+.

What is Smart+? A system that automates audience, bidding, and creative all at once. You no longer hand-carve audience profiles, and you no longer type bid amounts out line by line. You feed in the creative, and it goes out to find people using interest and behavior signals.

Symphony's job is to run a production line for creative: AI voiceover, AI scripts, AI copy rewrites, bulk remixing. It has exactly one goal: to let you put out 20, no — 50 distinct creatives at the same time.

Why did this become a need on TikTok specifically?

Because the logic of its recommendation engine is different from the ground up. When you scroll your "For You" page, it judges that content on the strength of the content itself: did you finish, did you engage, did you rewatch. Understand the content and you can understand "who" stood behind it.

An ad is itself a video. If you hand the system 20 different-looking ads, it gets 20 threads to test which style speaks to which group of people.

So on TikTok's turf, the level of your bid is not what decides the game:

It's whether you can put out 20 ads from 20 different angles all at once.

One Table, Three Platforms, Brought Into Line

Feeling tangled? Let me give you a table and line the three of them up side by side.

Dimension	Meta	Google	TikTok
How it finds people	Neural-network retrieval, broad by default	Fully automated, across all channels	Interest-graph retrieval, broad by default
Primary creative format	Image, video, carousel	Asset groups (mixed formats)	Short-form video
The advertiser's key lever	Creative variety + budget	Creative combinations + ROAS goal	Creative quantity + bidding method
Default attribution	7-day click / 1-day view	Data-driven attribution	Last-click, optional view-through attribution
Learning-period signal	50 conversions per ad set per week	Campaign-level conversion events	Campaign-level optimization events

Take a close look at the middle row — "the advertiser's key lever."

Put the three cells side by side, and the more you look, the more it feels like they were all written by the same hand.

One table, and the word "convergence" earns its keep.

Why Did All Three End Up on the Same Road?

Were the three companies in on it with each other? Of course not.

The real reason is three mountains, all pressing down on them at the same moment.

Mountain One: Apple Tore Down the Bridge First

In April 2021, Apple shipped iOS 14.5 and turned on ATT (App Tracking Transparency), its mechanism whereby an app that wants to track you across other apps first has to ask for your permission in a pop-up.

And most users picked "I don't agree." By most counts circulating at the time, the consent rate landed at only about two in ten.

Here is what that meant: the chain from an ad to a purchase was, in more than six in ten users, cut at the "link" point.

When conversion attribution can no longer reach a person, the audience packages you carved out by interest lose the ground they need to "see which spot broke and patch exactly there" — they become a blind box.

So the three platforms had to switch their playbook. They stopped chasing "which specific person, and what did he buy on which day," and instead pooled the behavior of billions of people into one aggregate model to predict, across a broad class, "who is most likely to buy." They went from reading individuals to predicting the whole.

And that job of "predicting the whole" is precisely what neural networks are best at.

Mountain Two: Regulation, Stricter Every Year

Privacy pressure has never relented.

From 2022 onward, GDPR enforcement visibly tightened. The EU's DMA (Digital Markets Act) took effect in November 2022. And California's CPRA (California Privacy Rights Act) had its reach expanded in early 2023.

Read those policies as one rope, and they all pull in the same direction: the amount of third-party audience data you can buy keeps shrinking.

Once you can't buy third-party data, the only path left is to dig inside your own first-party behavioral data. And mining behavioral data in order to predict — that, too, is exactly food for a neural network.

The Three: Tech Finally Crossed the Line

In 2017, "Attention Is All You Need" was published and pushed the attention mechanism into the spotlight.

By 2021, Google, Meta, and TikTok had all brought the Transformer onto the production line: Google put BERT in search, Meta put deep learning in its Feed sorting, and TikTok's recommendation system was standing on this foundation from the day it was born.

But there is a gate: the model has to be accurate to a certain degree before you dare hand over "finding people" to it. Below that gate, broad delivery did worse than manual targeting, because the model got too much wrong; once it crossed the line, broad delivery started winning — because the people the model can pull back are precisely the ones your manual lists could never reach.

All three crossed that line within almost the same window of time.

Which is why this thing was bound to happen — it was not luck. Convergence was the same answer, at the same moment, from three companies in the same era.

The Theory Done, Now the Doing

You can understand it all and still have to go out and run the campaigns. Jargon aside, the way all three platforms have landed right now is, at bottom, the same playbook.

Start with Meta: stand up a shopping-optimized Advantage+ campaign. No need to split the audience into layers; keep the budget as one piece. Put all your effort into creative: keep 10 to 20 assets in flight, and set aside one round of refresh each week. Bring the conversion signals back through a Conversions API. That's all.

Google next: run one Performance Max per product catalog. Don't build asset groups by audience — build them by "theme": each selling angle for a purchase is its own asset group, with every different headline, image, and video packed inside. Give the goal a planning ROAS or CPA (cost per acquisition) target.

TikTok last: launch Smart+, leave the audience open, and let Symphony do the scaling. If initial conversion signals look too thin, tack on a "native display" warm-up phase to cultivate the signal as you go.

Notice what the three landings share — it's a single sentence across the three platforms:

Campaign definition has moved from an "audience container" to a "creative container."

You used to think in terms of people. Now you manage a rack of creative. If there's one sentence to carve into the media buyer's desk, it is this one.

Okay, So How Do You Measure?

The world has become a black box, and with it a new problem.

The ROAS the platforms report is the product of their in-platform attribution: 7-day click, 1-day view, last-click — it only counts "what happened inside their house." It is not the same as one extra dollar spent necessarily landing one extra dollar earned.

So the smart moves come down to only a few.

One: run an incrementality test. By region, one batch keeps running, one stops. Watch what the "gap" turns out to be.

Two: hold all three platforms inside one brain. Where it's warranted, use a media-mix model — don't get dragged around by the numbers of any single platform.

Three: don't make your cycle too short. Finish a version, then tune the next. In a black box you don't make guesses — you read long horizons.

The Next Two or Three Years, in Three Steps

It is not going to stop moving. Further out, what I see are three steps.

Step one: delivery becomes more "self-governing." Google's Demand-object campaigns, Meta's budget automation — these are already prototypes of the system tuning its own parameters. Go one more turn and you will only need to answer three questions: what you sell, how much you can spend, and what your marketing goal is. The system covers the rest itself.

Step two: computing moves to the device. Apple's Core ML, Google's Privacy Sandbox — both are pushing "the math" down into the device so that the data never leaves it. When that path opens up, a new kind of signal appears: not "which person this is" but instead "this device has judged that this person in all likelihood wants this category." Your command console does not change: creative, a broad audience, one goal.

Step three: creative gets mass-produced by AI. Symphony is only the beginning. When the cost of creative approaches zero, the real scarce thing becomes "can the creative hold up?" The platform's creative-quality score, the asset-quality score, is already the first filter. The creative strategist is not going to get fired — this person is going to get more valuable. Because the machine can write a thousand versions, but it can't make the call for you: this time, whom to say it to, and how.

Wrapping Up

Back to the dinner table. My friend interrogated me for the whole evening, and his final question was:

"Okay, then, if that's what you're saying — did I just waste all those years of craft?"

I told him: you've been wasting nothing. The craft merely changes where it's aimed. What you know is "what makes people have a reaction" — and that answer is worth two lifetimes of money. It was once spent on "segmenting the audience." From now, it goes into "art of the creative."

"Besides," I said, "all those years you spent reading people — not a minute wasted. Starting today, just a different thing to read."

He sat quiet for two minutes, then raised his glass: "All right then. I'm going back this afternoon to build out the creative library."

I nodded. "Good. Hand the creative to the algorithm; keep the judgment to yourself."

Outside the window, the night sat squarely overhead, and a brand-new chess game had just opened.

You can keep up, too.