Meta Ads Audiences: Broad vs Narrow Targeting
Should I use broad or narrow targeting on Meta?
Test it, don't assume. Run Meta's A/B test with audience as the variable: it splits people so nobody sees both versions, which manual tests can't guarantee. Give both cells the same budget, aim for the 80% estimated power Meta recommends, and judge the winner on matured contribution, not reported ROAS.
Broad or narrow is a question your own account has to answer, and the usual way of answering it is broken. Two ad sets running side by side share people, so the result tells you which audience Meta chose to serve, not which one earns more. Meta’s A/B test fixes the split; the design below fixes the rest, from matched exclusions to a contribution metric your ledger can confirm.
If the narrow cell is built from a customer list, first read Is uploading a customer list to Meta compliant?, and for the broad side, Should I use Advantage+ audience or interest targeting?
Why does a side-by-side test give the wrong answer?
When ad sets from the same advertiser reach the same auction, Meta enters only the one with the highest total value, so your ads never bid against each other (auction overlap). That protects your costs, but it means two live ad sets do not get equal chances at the people they share: Meta picks. A “winner” from that setup reports Meta’s choice.
Meta says as much about informal tests: it does not recommend testing by turning ad sets on and off, because that “can lead to overlapping audiences”, while an A/B test splits your audience so that nobody sees both versions (about A/B testing).
How do I set up the test in Ads Manager?
- Write the hypothesis first, for example “broad beats the 1% lookalike on contribution per dollar”, as Meta advises before you choose a variable.
- Go to the Ads Manager main table, tick the narrow ad set and click A/B test in the toolbar (A/B test tools).
- Choose Audience as the variable, then change version B to the broad definition and nothing else.
- Give both versions the same budget and a schedule of 1 to 30 days, and aim for at least 80% estimated power (duplicating for an A/B test).
- Pick how Meta chooses the winner, such as cost per purchase, and record your own contribution metric beside it.
- Read the result in Experiments when the test ends, and treat a winner your order ledger contradicts as unproven.
What does a fair broad-versus-narrow test hold constant?
| Broad cell (A) | Narrow cell (B) | |
|---|---|---|
| Audience | Geography and age eligibility only | The declared interest, list or lookalike rule inside the same geography and age |
| Exclusions | Purchasers, employees and test IDs | The same purchasers, employees and test IDs |
| Budget | Half of the declared test budget | The other half |
| Offer, creative, placements, optimisation event | Identical | Identical |
| Main risk | Spend spread too thin to learn | Too few people to exit learning |
The rest of the design, written down before any money moves:
- Falsifiable expectation: broad produces a different matured contribution per assigned dollar, or a different CAC, than narrow; equality or reversal falsifies the directional claim.
- Held invariant: offer, creative IDs, destination, optimisation event, bid, placements, schedule, attribution setting and order-cost treatment.
- Budget: declare B and give each cell B ÷ 2; freeze reallocation and track spend imbalance against a tolerance you set in advance.
- Maturity window: concurrent delivery plus an observed conversion, cancellation and return cutoff.
- One primary metric: matured contribution per assigned dollar = (recognized revenue − variable order costs − media spend) ÷ (assigned test dollars). Secondary: CAC = acquisition spend ÷ first-party-verified new customers. These are our working definitions, built from Meta cell spend and your commerce and customer ledgers.
- Stop rule: stop for consent, policy, inventory, cash, claim or tracking-integrity risk.
- Inconclusive rule: call it inconclusive if delivery imbalance exceeds tolerance, exclusions fail, cross-cell contamination exceeds your limit, or outcomes are still immature.
How big does each cell need to be?
Meta divides your audience for an A/B test, so ad sets in a test are more prone to under-delivery from small audiences, and it says you may need to broaden more than usual (tips for A/B tests). Three documented limits apply to the narrow cell:
- Learning. An ad set becomes learning limited when it is unlikely to get about 50 optimisation events in the week after its last significant edit. Meta names small audience size, low budget, low bid or cost control, high auction overlap and an infrequent optimisation event among the causes (about learning limited).
- Lookalike seeds. A lookalike needs a source of at least 100 people, Meta generally recommends 1,000 to 5,000, and people in the source are excluded from the lookalike (create a lookalike audience).
- Difference. If the two audiences are too similar, Meta may not be able to declare a winner with confidence.
For what “too small” means in practice, see What does it mean when Meta says my audience is too small?
When is the result inconclusive?
When delivery was unbalanced, exclusions failed, contamination passed your preregistered limit, or outcomes had not matured. Meta may also declare no winner, which means it found no conclusive difference, not a tie you can act on. And an A/B test ranks two strategies against each other; it has no holdout group, so it cannot tell you how many sales either audience caused (lift and holdouts). Your thresholds stay inputs, not universal standards.
Shared audiences are the other trap; How do I check and fix audience overlap on Facebook? covers reading Meta’s overlap tool.
Can software help?
Bach.ai connects to your Meta ad account, audits it, and ranks the revenue leaks it finds by estimated dollar impact, each with a proposed fix. It applies a fix on Meta only after you approve it. Choosing broad or narrow is still a test you run.
For the full account review around audiences, see the Meta ads audit checklist.
FAQ
How do you design a fair broad-versus-narrow audience test?
Use Meta’s A/B test with audience as the only variable, so nobody sees both versions. Define both audiences and identical exclusions before spending, give each the same budget, hold creative, offer and optimisation event constant, and decide the winner on matured contribution per dollar from your own order ledger.
What invalidates a broad-versus-narrow comparison?
Unequal delivery, failed exclusions, cross-cell contamination above your limit, or outcomes that have not matured. Running the two audiences as ordinary side-by-side ad sets invalidates it from the start, because Meta enters only one of your ad sets into each shared auction.
How long should a broad-versus-narrow test run?
Meta lets an A/B test run for 1 to 30 days and recommends at least 80% estimated power, which depends on schedule and budget. Run long enough to reach that power and to let purchases mature past your cancellation and return cutoff, then stop on schedule.
Can this test show which audience caused more sales?
No. An A/B test compares two strategies against each other, but it has no holdout group of people kept from your ads, so it cannot measure incremental sales. For that, Meta’s conversion lift test splits people into test and holdout groups and compares their actions.
Sources: Meta Business Help Center, About A/B testing, Tools to create A/B tests, Create an A/B test by duplicating an ad set, Tips for improving A/B tests, Understand auction overlap, About learning limited, Create a lookalike audience, About lift and holdouts (all checked 1 Oct 2026).