Retargeting and Custom Audiences on Meta: A Singapore Guide (2026)

Last updated 20 July 2026 · by Adrian Tan, SDM

Buried in Meta’s own help documentation is a sentence that contradicts almost every Singapore agency deck you have ever been shown. Meta recommends A/B testing with Advantage+ Audience for almost all campaign types — except retargeting campaigns.

That is Meta telling you not to use its flagship AI targeting product for the exact thing this article is about. There is a good structural reason for it, and understanding that reason will tell you more about how retargeting actually works in 2026 than any list of audience types.

This guide covers what custom audiences genuinely do now, the arithmetic that makes retargeting behave differently in a market of 5.88 million people, the gap between the retention window you select and the one you actually get, and the PDPA rules that make uploading a customer list a very different risk from installing a pixel.

What a custom audience actually is

A custom audience is a list of people Meta can identify, built either from data you own or from interactions that happened on Meta’s own platforms. The distinction matters more than most people realise, because the two halves have completely different reliability profiles in 2026.

The critical difference: your-data audiences depend on browser and device signals that are actively degrading. Meta-engagement audiences happen inside Meta’s own apps, where no browser privacy control can interfere. In 2026 that makes engagement audiences considerably more dependable than website audiences — the reverse of the conventional ranking.

Two operational details that catch people out, both straight from Meta’s documentation:

• Website custom audiences do not backfill. Meta states plainly that the audience starts populating from the time you create it. If you launch a campaign and create the audience the same day, you are starting from zero. Create your audiences months before you need them.
• Engagement windows are rolling, not cumulative. Someone at day 29 of a 30-day window drops out tomorrow unless they re-engage. The audience refreshes itself, but it also drains itself.

On minimum sizes: Meta publishes no minimum for custom audiences. Its only guidance is soft — consider waiting until you have several hundred people before using a website audience. Every specific threshold you see quoted, whether 500,000 reach or a one-million potential audience, is vendor-invented. Do not let anyone attribute those to Meta.

The Singapore arithmetic nobody does

Singapore’s population was 5.88 million as of October 2025, with 5.78 million internet users. Facebook’s advertising reach here is 3.80 million people and Instagram’s is 3.35 million (DataReportal, Digital 2026 Singapore).

Now work down from a realistic website. Suppose you get 20,000 visitors a month. Meta’s own documentation walks through why the resulting audience is far smaller than that suggests: those visitors might fire 140,000 pixel events across seven pageviews each, but the audience is closer to the unique-visitor count — and shrinks further, because Meta only includes people it can match to an account and excludes anyone who has opted out of online behavioural advertising.

What you end up retargeting is frequently a four-figure pool. That has three consequences most Singapore advertisers never think through:

1. Frequency compounds fast. A modest daily budget against a few thousand people means the same person sees your ad repeatedly within days.
2. You cannot cap it. More on this below — it is the most important practical constraint in this article.
3. Your ad set may never exit the learning phase. Meta needs about 50 results in the week after the last significant edit. A small retargeting pool rarely delivers that, which is why retargeting ad sets so often sit permanently in the least efficient delivery state. We cover that in how to lower your cost per lead on Meta.

The 180-day window that is really seven days

Ads Manager lets you set a retention window of up to 180 days on a website custom audience. On a large share of your Singapore traffic, that number is fiction.

Apple’s Intelligent Tracking Prevention, documented publicly by WebKit, deletes all cookies created in JavaScript and all other script-writeable storage after seven days of no user interaction. Where someone arrives on a link-decorated landing page — which is what an ad click usually produces — that storage is capped at 24 hours. WebKit also states that ITP blocks all third-party cookies by default, with no exceptions.

Singapore has a high iOS share, so the gap between the window you select and the window you get is wider here than in most markets. A 180-day window on a Safari user is, in practice, closer to a week.

Two corrections while we are here, because both circulate widely in Singapore marketing content:

• Third-party cookies in Chrome are not going away. Google reversed course on 22 April 2025, confirming it will not roll out a standalone prompt and that third-party cookies remain on by default. Any local guide still warning you to prepare for their 2025 death is out of date.
• The Conversions API does not fix this. The Conversions API and Event Match Quality improve how well events match to Meta accounts for people who are already identifiable. They do not restore raw audience pool size and do not overturn Safari’s storage limits. Better matching, not more people. Our Pixel and Conversions API setup guide covers what it does and does not buy you.

Lookalikes in 2026: still here, quietly rebuilt

Lookalike audiences have not been deprecated. But one change has gone almost entirely unreported in Singapore, and it invalidates the mental model most local marketers still use.

You no longer select a country when creating a lookalike. Meta’s documentation states that new lookalikes will not show an estimated audience size in Audience Manager, and that Meta uses the ad set’s location settings to find similar people at delivery time.

The old reasoning — “a 1% Singapore lookalike is 1% of roughly 3.8 million reachable Singaporeans, so about 38,000 people” — no longer describes the product. The percentage is evaluated against whatever geography the ad set specifies, when the ad runs.

The working numbers that still hold:

• Minimum source audience: at least 100 people. Meta recommends 1,000 to 5,000.
• Up to 500 lookalikes from a single source audience.
• Source members are excluded from the resulting lookalike.
• A new lookalike takes up to three days to populate.

There is a second layer too. Advantage+ Lookalike is automatically enabled for new, duplicated and draft ad sets using lookalikes built from website, app or customer-list sources, and it lets Meta deliver outside your selected percentage. So the percentage is a hint rather than a boundary — twice over.

Why Meta says not to use Advantage+ Audience for retargeting

Back to the sentence we opened with. The reason is structural, and it comes down to suppression.

Advantage+ Audience is applied automatically wherever it is available. To turn it off you must actively choose to further limit your audience and uncheck the “use as a suggestion” boxes — it is opt-out, not opt-in. Any custom audience you add becomes a source to guide delivery rather than a boundary.

Then come the two lines that matter most:

• Custom audiences you exclude can still be added as audience suggestions. Your exclusion list can become a targeting list.
• You cannot exclude lookalike audiences under Advantage+ Audience at all, because exclusion constrains delivery.

Retargeting hygiene is mostly suppression: excluding existing customers, excluding recent converters, excluding people already being reached by another ad set. Advantage+ Audience structurally cannot do clean suppression. That is why Meta itself carves retargeting out of its recommendation, and it is the single most useful thing in this article.

For balance, Meta’s published performance claims for Advantage+ Audience are unusually well documented: a meta-analysis of 469 A/B studies run between January 2023 and August 2024, showing cost-per-result improvements of 14.8% on awareness, 9.7% on traffic, engagement and leads, and 7.2% on sales and app promotion. Note Meta’s own disclosed caveat — studies where the control already used broad targeting were excluded. So this measures Advantage+ against narrow manual targeting, not against broad.

The constraint nobody writes about: you cannot cap frequency

In a small market with a small retargeting pool, frequency is the thing that kills performance. And on the campaigns every SME actually runs, Meta gives you no lever for it.

Meta’s frequency controls — target frequency and frequency cap — are available only for reach and ThruPlay performance goals, within awareness and engagement objectives. If you are running retargeting optimised for conversions or leads, you cannot set a frequency cap at all.

Your remaining levers are audience construction and exclusion discipline: shorter windows, tighter pools, aggressive exclusions, and enough creative variety that repeated exposure is not repeated identical exposure. Which is precisely the toolkit Advantage+ Audience takes away. The two findings compound.

Customer lists: the CSV is riskier than the pixel

Uploading a customer list is mechanically simple and legally the most exposed thing in this article.

Mechanically, Meta hashes the list in your browser before it is sent, so raw records do not leave your machine. Meta publishes an identifier priority table worth following:

Two of Meta’s own tips matter disproportionately for Singapore lists: always include the country code on phone numbers (+65), and always put country in its own column, because Meta matches globally. Lists exported from local CRMs routinely fail both, and the match rate suffers for a reason that has nothing to do with data quality.

Also: NRIC must never be uploaded. It falls under Meta’s prohibition on national identifiers and separately fails Singapore’s own rules. There is no published match-rate benchmark, so treat any “typical 60–70% match rate” figure as folklore.

What the PDPA actually requires

Singapore is more permissive than the EU on cookies and stricter than most people assume on lists. Most SMEs have this backwards.

On pixels and cookies, per the PDPC’s Advisory Guidelines for Selected Topics (revised May 2024): there is no prior blocking-consent requirement here and no cookie-wall rule. Consent is not needed for cookies that do not collect personal data. Where targeted advertising does involve personal data collected through cookies, consent is required — and a cookie preference centre is described as good practice rather than a legal obligation. Importantly, the PDPC also states that an individual’s mere failure to manage their browser settings does not imply consent.

Whether pixel data is personal data is contextual, not categorical. A bare pixel hit on an anonymous visitor probably is not; a pixel feeding a profile tied to a logged-in customer, or a hashed email, almost certainly is.

On customer lists the position is much firmer, and this is where Singapore businesses get caught:

• The Legitimate Interests Exception does not cover direct marketing. The PDPC states organisations generally must obtain express consent to send direct marketing messages.
• Deemed consent by notification is excluded for direct marketing by regulation.
• Opt-out is not enough. The PDPC does not consider a pre-checked box appropriate.
• The Transfer Limitation Obligation applies, because uploading to Meta is a transfer to an overseas recipient.
• Purpose Limitation and Notification mean the purpose must have been disclosed on or before collection. A privacy policy written for order fulfilment does not authorise a custom audience upload.

There is one useful carve-out: the Business Improvement Exception can support using existing customer data for analytics and market research to understand your customers prior to marketing activity — what the PDPC calls preparatory activities. Segmentation analysis, yes. Sending them ads on that basis, no.

A live deadline worth diarising: private organisations have until 31 December 2026 to phase out NRIC use for authentication, with stepped-up enforcement from 1 January 2027. That is roughly five months away.

On penalties, one nuance almost every local blog misses: the PDPA allows financial penalties of up to S$1 million, or 10% of annual Singapore turnover where that is higher — but the 10% limb is confined to organisations with Singapore turnover above S$10 million. For a typical SME the practical ceiling is S$1 million, not 10% of revenue.

How much of the credit is real?

Retargeting reports beautifully because it takes credit for people who were already coming back. The academic evidence is more sobering than any dashboard.

• The best randomised study of retargeting (Sahni, Narayanan and Kalyanam, Journal of Marketing Research, 2019) found a genuine incremental effect — 14.6% more users returning within four weeks — but also that about a third of the first week’s effect landed on day one.
• A large field experiment at eBay (Blake, Nosko and Tadelis, Econometrica, 2015) found brand-keyword search ads produced no measurable short-term benefit, with the great majority of credited traffic arriving anyway.
• Facebook’s own 15-experiment study (Gordon and colleagues, Marketing Science, 2019) concluded that observational and attribution-based methods generally overestimate advertising effectiveness.
• Lambrecht and Tucker (JMR, 2013) found dynamic, product-specific retargeted ads were on average less effective than generic ones, until the consumer had already narrowed their preferences.

Retargeting works. It is simply smaller and shorter-lived than reported ROAS implies. In a small market the honest implication is short windows, tight pools and a modest budget share — the opposite of retargeting everyone for 180 days because the numbers look wonderful.

What to actually build

1. Create your audiences now, before the campaign that needs them — they do not backfill.
2. Lean on Meta-engagement audiences (video viewers, Instagram and Page engagers, lead form openers). They are unaffected by browser privacy controls.
3. Use short windows. 7, 14 and 30 days reflect reality far better than 180.
4. Turn Advantage+ Audience off for retargeting ad sets, per Meta’s own recommendation, so exclusions behave as exclusions.
5. Exclude deliberately: recent converters, existing customers, and anyone reached by another ad set.
6. Keep budget share modest and judge it on incrementality, not last-click ROAS.
7. Get consent right before uploading anything, and never upload NRIC.

Conclusion

Retargeting in Singapore is not the easy win the dashboards suggest. The audiences are smaller than your traffic implies, the retention windows are shorter than the ones you select, the credit is partly borrowed from visits that would have happened anyway, and the platform’s flagship targeting product is one Meta itself advises against here.

None of that makes retargeting a bad idea. It makes it a precision instrument rather than a volume one — best run with tight pools, short windows, disciplined exclusions and honest measurement.

Want your Meta audience setup reviewed properly? We check exclusion logic, retention windows, consent basis and incremental contribution rather than reported ROAS. Talk to us, see our Meta ads management or the complete guide to Meta ads in Singapore, and browse our case studies. Related: campaign structure, CBO and ABO, scaling a Meta campaign, what Meta ads cost in Singapore, and Meta ads versus Google Ads.

FAQ

Should I use Advantage+ Audience for retargeting?
Meta’s own guidance recommends A/B testing Advantage+ Audience for almost all campaign types except retargeting. The reason is that it cannot do clean suppression: excluded custom audiences can still be added back as suggestions, and lookalikes cannot be excluded at all. Since retargeting is mostly suppression, turn it off for those ad sets.

How long do Meta custom audiences last?
You can set retention up to 180 days for website audiences, with customer-list audiences configurable between 1 and 180 days and some engagement audiences running to 365 days. In practice Safari deletes script-written storage after seven days without interaction, and after 24 hours on link-decorated landing pages, so the effective window on much Singapore traffic is far shorter.

Why is my website custom audience so much smaller than my traffic?
Pixel events are not unique visitors, Meta only includes people it can match to an account, and it excludes anyone who has opted out of behavioural advertising. Audiences also do not backfill — they populate only from the moment you create them.

Do lookalike audiences still work in 2026?
Yes, but the product changed. You no longer choose a country when creating one; Meta uses the ad set’s location settings at delivery time, and new lookalikes show no estimated size in Audience Manager. You need at least 100 source people, with 1,000 to 5,000 recommended, and Advantage+ Lookalike may deliver outside your selected percentage.

Can I cap how often my retargeting ads are shown?
Not on conversion or lead campaigns. Meta’s frequency controls are limited to reach and ThruPlay performance goals within awareness and engagement objectives. For everything else your only levers are shorter windows, tighter audiences, exclusions and creative variety.

Is it legal to upload my customer list to Meta in Singapore?
Only with a proper consent basis. The PDPC’s Legitimate Interests Exception does not cover direct marketing, deemed consent by notification is excluded for it, and opt-out consent is not accepted. The Transfer Limitation Obligation also applies because Meta is overseas. Never include NRIC — and note that private organisations must phase out NRIC for authentication by 31 December 2026.