What is Audience?
An audience is the set of people a campaign is allowed to reach, defined by who they are, what they have done, or what a platform infers about them. It sets eligibility rather than delivery: the bidder still chooses who inside it actually sees the ad.
Eligibility, not delivery
An audience is the set of people a campaign is permitted to reach. It is defined by attributes — location, age, language — by behaviour the advertiser observed, by data the advertiser holds, or by inferences the platform makes.
The distinction that matters most is between eligibility and delivery, and it is the one most often collapsed.
Targeting draws a boundary. Inside that boundary, the bidder decides which individual impressions to buy, on its own estimate of which are likely to produce the outcome the campaign was given. A campaign targeting two million people may deliver almost entirely to fifty thousand of them. A campaign targeting fifty thousand may never reach most of them.
This is why "who did we target" and "who saw the ad" are different questions with different answers, and why a delivery report is more informative than a targeting setting.
The kinds
Attribute audiences are defined by what people are: where they live, what language they use, what age bracket they fall into, what device they are on. These are the oldest form and the least contested.
Behavioural audiences are defined by what people did, either on the advertiser's property — visited, added to cart, purchased — or on the platform. The first kind is the advertiser's own data and is the most reliable thing in the account.
Uploaded audiences are built from lists the advertiser already holds: customers, subscribers, leads. Matched against platform accounts, they become targetable.
Modelled audiences are built by a platform from a seed: people who resemble a source audience by whatever similarity the platform computes.
Inferred audiences are the platform's own categories — interests, life events, purchase intent — assembled from behaviour the advertiser cannot see.
The first three are things the advertiser knows. The last two are things a platform believes, and the difference in reliability is worth keeping in mind when a campaign built on inference underperforms.
Size, and why it is really about frequency
The most common audience mistake is building one too small, and the reason it hurts is not precision but repetition.
Frequency is impressions divided by reach, so an audience does not absorb a budget the way a larger one does — it converts it into repetition. Halve the reach at the same spend and the average number of times each person sees the ad doubles. A small pool therefore reaches heavy repetition far sooner than the budget figure suggests, and the response falls, the complaints rise and the cost per result climbs for reasons that have nothing to do with bidding. This is the mechanism behind almost every fatiguing retargeting or custom-audience campaign, and it is arithmetic rather than a setting.
The second problem is learning. A bid model needs outcomes to predict from, and a small audience produces few. A campaign that cannot accumulate the platform's threshold of optimisation events in a week is a campaign whose strategy is guessing.
Both point the same way: where a campaign is fatiguing or failing to learn, widening the audience is the intervention that addresses the cause. Frequency capping suppresses the symptom by declining delivery, which on a saturated audience means unspent budget.
Why broad works better than it used to
For most of the history of digital advertising, the advertiser's job was to find the right people and the platform's job was to serve them. That has substantially reversed.
Delivery models now select impressions using signals no advertiser can see — recent behaviour, context, device, timing, patterns learned across enormous volumes of data. Given room to work, they routinely find pockets of demand that a manually defined audience would have excluded.
The consequence is that tight targeting is frequently a constraint rather than a refinement. It removes people the model might have chosen, on the basis of a hypothesis formed from a report.
This is not an argument for no targeting. Business constraints are real: a product cannot be shipped everywhere, a service is licensed in one jurisdiction, a minimum age applies. Those belong in the targeting because they are facts rather than hypotheses.
It is an argument against targeting as a substitute for creative. Which brings us to what actually signals who an ad is for.
Creative is targeting now
An ad that opens by naming its audience — the problem they have, the situation they are in, the words they would use — selects its own audience twice over.
It selects directly, because people who recognise themselves respond and people who do not scroll past. And it selects indirectly, because the platform learns from who responds and adjusts delivery toward more people like them.
On platforms that have removed detailed interest segments, this is not a stylistic preference; it is the mechanism that remains. An advertiser who wants to reach small-business owners will do better with an ad that speaks to running a small business than with an interest category that claims to identify them.
Exclusions carry more weight than inclusions
The one piece of audience work that has not been automated away is subtraction.
Excluding existing customers from prospecting, excluding converters from retargeting, excluding an audience that belongs to another campaign, excluding applicants and competitors and people who cannot buy — each of these is a fact the platform does not know and cannot infer, and each of them removes waste the bidder would otherwise happily buy.
An account that has spent an afternoon on its exclusions and left its inclusions broad is usually in better shape than one that did the reverse.
Overlap, and the campaigns that compete with themselves
Audiences defined separately frequently contain the same people, and when two campaigns or ad sets in the same account target overlapping audiences they can end up bidding against each other in the same auction.
The effects are real but usually smaller than the folklore suggests. The platform will generally choose between competing entries rather than letting them raise each other's prices indefinitely. What it will not do is tell you which audience actually produced a result, because both were eligible.
The reporting consequence matters more than the auction one. Two overlapping audiences compared side by side are not a clean test: whichever wins may have won because of the overlap rather than on merit, and a decision to pause the loser removes people who were never reached by the winner.
Where a comparison has to be clean, the audiences have to be made mutually exclusive with exclusions, which is more work than it sounds and is the only version of the test worth running.
What to decide before building one
Three questions settle most audience decisions, and they are worth answering explicitly rather than by default.
What does the business actually require? Geography it can serve, ages it may advertise to, languages it can support. These are constraints and they go in.
What does the platform know that I do not? Where the delivery model is strong and the data is rich, giving it room is usually better than narrowing. Where it is weak, more definition helps.
Who must never see this? Existing customers, recent converters, people in another campaign, anyone the business cannot serve. Exclusions are the work that stays the advertiser's.