Meta is getting better at choosing which ads to show people. That does not mean audiences no longer matter. It means the way you build and test them needs to change.

In Malaysia and Singapore, Meta ad targeting is less about making the perfect list of interests. Set useful limits, pick a clear goal, and check if the people who respond are good prospects.

That does not mean you should hand every choice to a machine. Here is what Meta says, what its numbers leave out, and how to run a fair test.

What ‘stop building audiences’ really means for Meta ad targeting

“Stop building audiences” challenges an old habit. It does not mean you should delete every audience setting.

A small firm might once have made ad sets for several similar interests. Each had its own budget and results. The hope was that a neat label would reveal the best buyers. But labels do not always show intent to buy. Splitting a small budget also makes results harder to judge.

The better question is not “Can I build more audiences?” It is “Does this audience choice improve a result that matters to my business?” Sometimes a specific audience will help. Sometimes a wider eligible pool will work better. Neither answer is guaranteed.

Meta executive Matt Steiner put the shift this way in an AdExchanger interview:

“The importance of targeting is diminishing over time as ranking models get better at selecting who an ad would be relevant for.”

That is Meta’s view of its own systems. Your service area, legal duties, customer exclusions, and sales capacity still matter to Meta ad targeting. Nor does the quote prove that every broad audience beats a focused one. Start with the facts of your business, then test the settings that remain.

How Meta now chooses an ad

On 28 October 2026, Meta described upgrades to its ad system. First, the system narrows down eligible ads. A separate ranking stage then helps pick an ad to show.

Retrieval creates a shortlist

Andromeda is Meta’s retrieval system. It cuts a large set of eligible ads down to a shortlist. Ranking models then assess those ads more closely. Eligibility still matters: an ad must meet the rules and settings that apply before it can compete.

Meta also described a new retrieval model. It adds ads to the shortlist alongside an existing model. Our earlier Andromeda introduction covers the wider shift. This article focuses on how to make and test audience decisions under that shift.

GenRecs helps with new ads; GEM supports ranking

GenRecs uses concepts from ad text and images to help Meta grasp what an ad is about. It can compare a new item with similar ones, even if the new ad has few past interactions. This helps with the “cold-start” problem. The system still needs to learn.

Take a hypothetical Johor Bahru furniture shop. An ad for a small desk may share features with ads for other small workspaces. That context can help the system assess a new ad. It does not tell the shop which audience will lead to profitable orders.

GenRecs does not replace all engagement signals. Meta’s systems can still use past interactions. GEM is a foundation model that supports ranking models. It is not a switch you turn on in Ads Manager.

Updates may not work the same way for every campaign type or country. Check the options in your account before you plan a test.

What Meta’s performance numbers actually prove

In the 28 October report, Meta said its combined ranking improvements produced 8.3% more ad clicks and 15.7% more conversions on Facebook in Q2 2026. Those figures are Meta’s own reported results. They were not independently verified in the cited report.

The figures cover several changes working together. They are not a GenRecs-only result or a Malaysia or Singapore benchmark. They do not show the effect for every placement, advertiser, or campaign goal.

A conversion in an ad report may be a purchase, a form submission, or another tracked action. A business still needs to know what happened after that action. Did a person answer the phone? Could they be served? Did the sale cover its costs?

Imagine a Singapore training provider. One ad set gets more forms. Another gets fewer, but more people enrol. If the firm only looks at cost per form, it may make the wrong choice. Cost per qualified enrolment is more useful, if there are enough reliable data.

Meta’s figures are a reason to test, not a forecast for your account. Your test needs enough data, a useful outcome, and a clear view of how much doubt you can accept.

What advertisers still control

Automation changes how choices work together. You still need judgment. Start with what your business can deliver, then check which controls your campaign supports.

Business boundaries and audience settings

Set real location limits where the controls allow them. A clinic serving one part of Kuala Lumpur should not assume every enquiry across Malaysia is useful. A Singapore delivery service should check that its chosen area matches where it can fulfil orders.

Some audience inputs are hard controls; others may be suggestions that Meta can expand beyond. The distinction depends on the campaign setup and the current interface. Read the labels and help text in your account. Do not call a suggested age, interest, or audience a firm restriction without checking.

Keep real needs apart from preferences. Location, minimum age, or other rules may be vital for an offer. If a needed limit cannot be enforced in the interface, rethink the campaign or offer. Follow ad rules and privacy law, including consent and proper use of customer data. Sensitive categories may have extra limits.

Goals, signals, and the customer journey

Choose a goal that reflects the outcome you need and that your setup can measure. A click is not a qualified enquiry. A submitted form is not a signed contract. When possible, connect the ad result to a later business event using an appropriate, privacy-compliant process.

Check that your website events or lead forms work before you compare audiences. If an event is counted twice, missed, or counted differently between groups, results can mislead you. Use the same lead-status rules for both groups. Give each lead a similar chance of follow-up.

The offer and landing page also shape who responds. State the service area, price range, and next step clearly where useful. That can help people judge the offer before they fill in a form. It is not a substitute for a firm targeting limit.

Placement controls are separate from audience choices. Options can differ by account and setup; check your interface. Our guide to placement control changes offers further context. Likewise, the Ads Manager AI overview separates campaign assistance from the underlying delivery system. Verify current availability before acting on either guide.

A practical audience test for MY and Singapore

Good Meta ad targeting starts with a question you can test. The next example is hypothetical. Adapt it to your account, offer, budget, and local rules.

Start with one question and fixed conditions

Imagine a Johor Bahru renovation firm that only works in its service area. It wants to compare two audiences. One has a relevant interest input; the other is wider and has no interest input.

First, check that the account can create a real difference between those setups. If Meta treats the interest as a suggestion and expands both groups in much the same way, the test is not a clean comparison of strict audience boundaries. Document what each setting actually does.

Keep the service area, offer, ad, landing page, optimisation goal, dates, and lead process the same. Keep budgets similar and set limits you can afford. Do not change several settings during the test. If you do, you will not know which change affected the result.

If Meta offers a suitable randomised A/B test for your setup, use it to split eligible people between groups. Check its settings and reports before launch. A random split helps reduce overlap and bias. It does not guarantee a precise result.

Choose the outcome before launch

Define a qualified lead in advance. For the renovation firm, that might mean a real contact seeking work in its service area and within its project scope. Apply the same definition to both groups. Track spend, submitted leads, qualified leads, cost per qualified lead, and eventual sales where feasible.

Agree on a test period and decision rule before you look at results. Use past lead volume, sales cycles, and funds to judge if the test can yield enough data. No single spend or conversion threshold proves a winner for every business.

A Singapore firm can use the same method. For example, a hypothetical catering company could test two audience setups for one corporate lunch offer. Its sales team should use the same rules for event size, dates, and service area in both groups.

Meta ad targeting — read results with care

Small samples can swing on a few enquiries. If one group has three qualified leads and the other has five, the difference may reflect chance. Report the counts as well as the costs, and treat a weak result as inconclusive rather than declaring a winner.

Also check whether both groups had similar follow-up time. A lead that arrives on Friday may need longer to reach a sales decision than one received on Monday. Compare results over a consistent window and include the later sales outcome when it becomes available.

When feasible, add a separate randomised holdout that receives no campaign ads. Use a suitable lift study or controlled test. This asks a different question: which results did the ads cause? An audience A/B test alone cannot show that lift. A holdout may not be practical for a small account. Its design must account for spillover and small samples.

If the evidence is thin, keep the safer business limits and gather more data. Do not raise the budget automatically. For help setting up paid search and social reporting together, explore CODE/RAVEN’s IGNITE service. For lead qualification and follow-up, its PROSPECTOR service is another option.

Frequently asked questions

Does this mean I should use a broad audience for every campaign?

No. A wider eligible audience may be worth testing, but it is not always right. Start with real business and legal limits. Then compare the available settings against a useful outcome. Some offers, markets, and account setups may benefit from a more focused approach.

Is Meta ad targeting still useful for local businesses?

Yes. A local business still needs to reach people it can serve and comply with relevant rules. Audience inputs may also offer useful guidance or test ideas. Check which settings are enforceable and which are suggestions in your specific campaign.

Can I switch GenRecs or GEM on in Ads Manager?

Meta describes these as parts of its ad systems, not manual switches. GenRecs helps the system grasp ad concepts and find candidates. GEM supports ranking models. The options you see can vary by account and campaign.

Should I judge an audience by cost per lead?

Cost per lead is a useful starting point, but it can reward forms from people who will never buy. Also compare qualified leads, sales, and their costs over a consistent follow-up window. Use the same qualification rules for each group and respect privacy requirements.

What if my budget is too small for a clear audience test?

Keep the setup simple and avoid splitting spend across many tiny groups. Record results over a suitable period, but recognise that small numbers may remain inconclusive. Do not assume more spend will fix the issue. If you would like a second look at your setup, contact Xwork for a practical discussion.