Your Funnel Is Your Opinion
Every marketing team has drawn the funnel: strangers at the top, visitors in the middle, buyers at the bottom. Then the ad account gets built to match.
Ask who decided that order, and the honest answer is usually: we did. Not a test. A group of smart people in a room, deciding how customers should behave.
Meanwhile, the platforms you pay run some of the most advanced prediction systems ever built. They don’t follow your diagram. And every rule you add to script the route holds them back.
Think of it as a racehorse. Most advertisers buy the fastest horse in the world, then climb on and steer every stride. This article is about letting go of the reins.
How Most Accounts Are Built
A typical setup has three layers. Prospecting campaigns aimed at broad or interest audiences get the brand videos. Retargeting campaigns aimed at site visitors get product carousels and reviews. A bottom layer aimed at cart abandoners and the customer list gets the discount.
Each layer has its own budget, its own creative and its own rule about who it should reach. Every one of those choices was made by a person: which message suits a stranger, which suits a visitor, how many steps it takes and in what order.
The structure feels like strategy because it is tidy. Tidy is not the same as true.
It Is an Opinion, Not a Finding
Compare it with an email sequence. A good email team writes several versions of a welcome series. They change the order, the timing and the offer, then test the versions against each other. The sequence that wins is the one that sold more, not the one that looked sensible in a meeting.
The ad funnel rarely gets that treatment. Nobody tested whether the brand video works better before the review ad or after it. Nobody tested whether a stranger would buy straight from the discount ad. The order came from how people think customers should behave.
There is a second difference. In email, you send the next message. In paid social and search, you don’t. The platform decides who sees which ad, and when. Your funnel is a request you make to a system that has its own idea of the best order.
What the System Actually Does
Every time someone opens Instagram or searches on Google, an auction runs. In a fraction of a second, the system looks at every ad allowed to reach that person, from you and every other advertiser. It predicts how likely each ad is to get its advertiser’s result, weighs that against the bid, and shows the winner.
Meta says its first step narrows “tens of millions of ad candidates” down to “a few thousand relevant ad candidates.” In a separate post, it says each request has to “rank thousands of ads in a few hundred milliseconds.” All of that happens for one person, in one moment.
That prediction doesn’t ask which layer of your funnel someone belongs to. It asks what this person has done, in what order, and what usually happens next.
Meta’s models now learn from the sequence of “a person’s engagement and conversion events.” Its engineers call this “a departure from human-engineered feature aggregations.” Their examples of the old way are “ads a person clicked in the last N days” and “pages a person visited in the past M days.”
Read that again. That is a retargeting audience. Meta moved away from that kind of rule inside its own models, because the machine finds better patterns by itself.
Google’s Performance Max works the same way. Audience lists are optional signals, and Google says the campaign “may show ads to relevant audiences outside of your signals.”
Your Labels Don’t Hold
A broad prospecting campaign doesn’t only reach strangers. Unless you exclude them, it also reaches your site visitors and past customers. So the brand video you made for strangers gets shown to your best customer, because the system predicts it will get her to buy again.
Retargeting is looser than it looks, too. With Advantage+ audience switched on, Meta widens the search beyond your list. Its documentation says it creates “the broadest possible audience to search within.” The review ad you made for visitors can end up in front of someone who has never heard of you.
So a “cold” campaign can turn out to be the warmest one in the account. The path a buyer takes is built one impression at a time. It might be the discount first and the brand story second. It might be the same video three times. It can change from hour to hour, because the system is chasing the goal you gave it, not the structure you drew.
Here is how that looks for one person. She sees the discount ad from your “cold” campaign on Monday and a creator video on Wednesday. On Friday she searches your name and buys. Your analytics report gives the sale to branded search. Your funnel never saw it coming.
“But Retargeting Has Our Best ROAS”
Retargeting looks great because it reaches people who were already on their way to buying. They visited, they browsed, and some had a full cart. Many would have bought anyway. The ad showed up before the sale, so the platform gives it the credit.
That is credit, not cause. A high ROAS on warm audiences mostly tells you that warm audiences buy. It doesn’t tell you the funnel made them buy. When you protect that layer with its own budget, you pay to reach people who were already coming.
What the Game Computers Taught Us
In March 2016, DeepMind’s AlphaGo played Lee Sedol, one of the best Go players alive. In game two it played Move 37. DeepMind says that move had “a 1 in 10,000 chance” of being played by a human. Many experts first thought it was a mistake. AlphaGo won that game and the match, 4 to 1.
The next year came AlphaZero. It started “with no in-built domain knowledge but the basic rules of the game.” It never studied a human game. It played against itself, beginning completely at random.
After about nine hours of training, it beat the reigning computer chess champion, Stockfish. In 100 games it won 28, drew 72 and lost none. In a longer match published in 2018, it won 155 games and lost 6 out of 1,000.
What shocked people was how it played. DeepMind’s CEO, Demis Hassabis, called its style “alien.” Moves that looked wrong to experts kept turning out to be right. Human strategy had been built over centuries on what people believed worked. The machine, starting from the rules alone, found paths people had never tried.
There is one big difference. AlphaZero had a perfect scoreboard: win or lose. Your ad platform keeps its own score, and it over-claims. Give the horse the wrong finish line and it will brilliantly win the wrong race. That is why your real job is the goal, not the route.
Your Ad Account Is a Shell
You set the goal and the budget. You choose the countries, the exclusions and the creative. Those decisions matter.
Almost everything else happens behind the screen: who sees which ad, in what order, how many times, at what moment, and who gets the discount instead of the story. You don’t see those decisions. You see a report afterwards.
Your structure still shapes delivery, because every split divides budget and data. But the system doesn’t treat your campaign names as stages. It learns from everything it has: your ads, your website data and patterns across millions of other advertisers.
The machine isn’t perfect. It chases whatever result you tell it to count, and it will happily count sales that were coming anyway. That is a reason to give it the right goal, not to plan its route for it.
The Question That Settles It
If top-to-bottom were the best route to the goal you set, the system would find it on its own.
At every impression, it picks the ad most likely to lead to your result, given what this person has already seen and done. The path is the sum of those picks. If brand video, then reviews, then discount were the winning order, those picks would add up to it for each person, without being told.
So a hand-built funnel can only do one of two things. If it happens to be right, it adds nothing the right goal wouldn’t. If it is wrong, you have forced the machine to spend against its own predictions.
Either way, you put the weight on your bias instead of on its calculations. And you pay for that bias. Your budget and data get split into layers, so each campaign learns more slowly. Inside each ad set, the machine can only choose from the ads in that folder. And a human guess sits exactly where a prediction should be.
Set the Rules, Not the Route
Stop putting your human bias on the horse. Let it train itself how to run, the way AlphaZero did. Your job is to set the finish line and the fences.
Set the right goal. The machine chases exactly what you ask for. Ask for new customers and incremental results, not just any conversion. Otherwise it will find the cheapest people who were already buying.
Set hard boundaries. Exclusions are the fences: the people you won’t pay to reach, like recent buyers or people already in checkout. Meta’s documentation says these “non-negotiable business constraints are NOT expanded.”
Feed it real variety. Give it creative with different reasons to buy: price, occasion, proof, a problem solved. Different ideas let the machine build different paths. Ten versions of one idea give it nothing to choose from.
Consolidate. Run fewer campaigns with more data each, instead of a layer for every stage.
Judge it in the store. Look at new customers, total sales and profit, not at which layer the platform gives credit to. The platform grades its own homework.
People still move from not knowing you to buying from you. You just don’t get to choose their route anymore.
Your funnel was always a guess. Set the finish line, build the fences, and judge the race in the store.
Sources
Meta Engineering, “Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine,” December 2024
Meta Engineering, “Sequence learning: A paradigm shift for personalized ads recommendations,” November 2024
Meta for Developers, “Advantage+ Audience”
Google Ads Help, “About audience signals for Performance Max campaigns”
Google DeepMind, “AlphaGo”
Google DeepMind, “AlphaZero: Shedding new light on chess, shogi, and Go,” December 2018
Silver et al., “A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play,” Science, December 2018
MIT Technology Review, Demis Hassabis on AlphaZero’s “alien” style, as cited in Wikipedia’s AlphaZero entry
Comments