Targeting is subtraction. Every filter you add removes people from the audience and, with them, removes the events the system learns from, so past a point refining makes a campaign more expensive rather than more precise. The limit is not measured in people but in results per week, and that number depends on your budget and your conversion rate.
Targeting is subtraction, and the subtraction has a price
There is a perfectly reasonable intuition behind fine targeting: if I only show the ad to people who fit, I am not wasting money on people who do not. It was very nearly true back when the delivery system knew nothing and you knew something. Today it runs the other way, and it is worth understanding why before accepting it out of obedience.
What the system does is spread your budget across the people in the audience looking for whoever is most likely to take the action you asked for. That needs two things: room to search in and examples to learn from. A filter takes away both at once, because it shrinks the ground and it shrinks the number of results per unit of time, which is what the system uses to correct itself.
There is a pricing effect on top. When many advertisers want the same people and the inventory for those people is limited, cost per thousand impressions goes up. A narrow audience is not just a smaller audience, it is a more contested one, and that is the part that rarely makes it into the conversation about targeting. How to read that cost and what to compare it against is in the ad metrics guide.
Minimum size is not measured in people
The question "how many people should an audience have?" has no answer because it is the wrong question. An audience of fifty thousand with a budget that reaches a thousand a day, and an audience of five hundred thousand on the same budget, behave almost identically: the system does not run out of people in either, it runs out of examples.
What is documented is the other side. Meta describes a learning phase and closes it when an ad set gathers in the region of fifty optimisation events within a rolling week. Below that rate, performance is unstable by design, and the instability is not fixed by waiting. It is fixed by giving the system more events.
Out of that comes a sum you can do in a minute, and it replaces any general advice. Take your weekly budget, divide it by your current cost per result, and see whether the number clears fifty. If it does not, you do not have a targeting problem, you have an ad set too small to learn, and there are three ways out: raise the budget, merge ad sets, or optimise for a more frequent event further up the funnel.
| Situation | What is happening | What to move first |
|---|---|---|
| Many ad sets, few results in each | None of them is learning | Merge ad sets |
| An ad set under fifty results a week | Permanent learning phase | Budget or a more frequent event |
| Frequency high and cost climbing | The audience is exhausted | Widen the audience |
| Cost per thousand impressions very high | Narrow and contested audience | Remove filters, do not add them |
| Enough results and stable cost | It is working | Leave it alone |
The last row is the hardest one to obey. An ad set that is doing well invites optimisation, and every significant edit sends it back into the learning phase. How many ad sets to open and how to split the money between them is the decision that comes before all of this, and it lives in the Meta campaign structure guide.
What interest targeting is still good for
Interests have not stopped working. They have stopped being where the intelligence of the campaign lives. They are still useful in three specific situations, and those are worth recognising, because outside them interests are usually an expensive way of guessing.
The first is the cold start. A new account with no history, no customer list and no traffic of its own: the system has nothing to begin from, and a sensible interest is a better starting point than nothing. It is scaffolding, and it comes down.
The second is a product with a strong context of use rather than a demographic profile. Selling climbing gear, invoicing software for freelancers or supplies for hair salons is nothing like selling clothes: the interest is not standing in for the system, it is naming a world the system cannot infer from your traffic because your traffic is still small.
The third is exclusion, which is the least discussed use of interests and the one that most often makes sense. Taking people who are already inside your business out of the audience is a business decision rather than an optimisation one, and the system will not make it for you.
A warning about stacking interests. Combining five interests with "and" instead of "or" produces tiny audiences that also do not mean what you think: belonging to an interest is an inference the platform makes, not a declared fact, and the error in each one compounds when you intersect them. If you are going to use interests, use them broad and in parallel.
Everything in this article happens in one place in GoFeed.
Try it freeThe audiences that are worth something: yours
The step change is not in refining interests. It is in leaving platform data behind and starting to use your own. There are three sources and they have almost nothing in common.
The first is your customer or contact list, which is the only audience built on facts rather than inferences. The second is people who have engaged with your accounts or watched your videos, which the platform knows first hand. The third is your website traffic, which needs the measurement set up properly; how links get tagged so that traffic is attributable at all is covered in the UTM and attribution guide.
Lookalike audiences are built on top of those sources, and there the size of the source does matter. Meta requires a minimum of one hundred people from a single country to build one at all, and recommends starting from a source of between one thousand and fifty thousand. Below that, the similarity is statistically imaginary, and a poor quality source produces a poor quality lookalike however uncomplainingly the interface accepts it.
What gets said far too rarely is that the source matters more than the percentage. A lookalike built on recent buyers and one built on visitors to a page are not wider and narrower versions of the same thing, they are two different audiences. Preparing the list is worth more than tuning the one per cent.
Exclusions and overlap are not the same thing
Excluding has two uses and they get confused often. One is not paying to reach people you have already converted, which in acquisition is almost always right and in repeat purchase depends on how often what you sell gets rebought. The other is stopping your own ad sets from competing with each other for the same people, which is a different and more expensive problem.
When two ad sets in the same campaign target overlapping people, the platform avoids showing both ads at once, but the internal split of the budget becomes opaque and the readings stop being comparable. The audience overlap tool exists precisely so you can see that before launching, and it is worth opening with more than two live ad sets on the same market.
The practical rule is short. Exclude customers in acquisition, exclude recent buyers of anything that does not get rebought, and do not exclude anything else just in case: every exclusion is another filter, and filters cost money. All of this is decided better with the budget in front of you, because a narrow audience and a broad one on little money fail for different reasons, and which one you get depends on what you can sustain per month: that is the conversation in the minimum ad budget guide.
Deciding it with your own numbers
All of the above turns into a short review you can run once a month, and it needs no special tooling. First, count the results per week of every live ad set and separate the ones that clear fifty from the ones that do not; the second group is not producing data, it is producing spend. Second, look at frequency and cost per thousand impressions on the ones that are learning, because that is where a narrowed audience shows up.
Third, and this is the step most people skip, write down what you decided and why. Without it, in March you will retest the audience you ruled out in January. Having spend and results from each brand's Meta, Google, LinkedIn, TikTok and Pinterest accounts in one list, next to that brand's content, is what the ads module does, and it also shows the audiences the account holds and, on Meta, the estimated reach the platform returns while you put the targeting together. It is the difference between reviewing with data and reviewing from memory.
One last warning. A well chosen audience does not rescue a bad creative, and a good creative survives a mediocre audience far better than the other way round. If you have an afternoon, spend it on the piece before the targeting.