Last updated 19 July 2026 · by Adrian Tan, SDM
You have a Meta campaign that works. Cost per lead is acceptable, return on ad spend is respectable, and the obvious next move is to spend more. So you raise the budget — and performance falls apart. Cost per result climbs, the ad set drops into learning again, and within a fortnight you are back where you started, having burned three weeks and a lot of money proving that the thing which worked at S$100 a day does not work at S$400.
Almost everything written about this repeats the same advice: raise budgets by 20% every 48 to 72 hours. That number is not Meta guidance. Meta has never published a percentage threshold for budget changes, and we can show you how we know.
This guide covers what Meta actually documents about scaling, what is practitioner convention dressed up as platform rule, and the constraint that matters more here than anywhere: there are only so many people in Singapore to reach.
The learning phase, as Meta actually documents it
Meta states that ad sets exit the learning phase “as soon as they can deliver stably”, usually “after about 50 results in the week after the ad set’s last significant edit”. Three details in that sentence get missed constantly:
- The threshold applies at ad set level, not ad level.
- The seven days is a rolling window anchored to your last significant edit, not a calendar week. Every significant edit restarts the clock.
- Meta’s own wording is inconsistent — the learning phase page says “about 50 results” while the learning limited page says “about 50 optimisation events”. Both are live. Treat 50 as an approximation, not a gate.
“Learning limited” is not a penalty. Meta is explicit: it “isn’t a penalty – it’s an indication that your budget isn’t being spent effectively”. Meta names six causes: audience too small, budget too low, bid or cost control too low, high auction overlap, an optimisation event that fires too rarely, and too many ads running at once.
What counts as a significant edit
Meta splits this into two tiers, and the second tier is where scaling lives.
| Always significant | May or may not be, “depending on the magnitude of the change” |
|---|---|
| Targeting change Creative change Optimisation event change Adding a new ad to the ad set Pausing the ad set for 7 days or longer Changing bid strategy |
Ad set spending limit amount Bid control / cost-per-result goal / ROAS goal Budget amount |
Notice that a budget change sits in the second column. It is conditional. And Meta’s only worked example for how conditional is a gift to anyone tired of being told the 20% rule is official:
“If you increase your budget from USD 100 to USD 101, that isn’t likely to cause one or more ad sets to re-enter the learning phase. However, if you change your budget from USD 100 to USD 1,000, one or more ad sets may re-enter the learning phase.”
That is +1% versus +900%. Meta brackets the answer and deliberately declines to locate it. There is no published threshold — not 20%, not 30%, not 50%. Every number you have read is somebody’s convention.
How much do practitioners disagree? Published 2026 recommendations range from 10% to 50% per step, and from every 48 hours to twice a week. That spread is the evidence that nobody knows.
The officially safe path, which nobody talks about
While Meta refuses to name a budget threshold, it is explicit about several things that do not reset learning — and they point somewhere useful:
- “Will Advantage+ campaign budget cause ad sets to re-enter the learning phase as it distributes budget? No.” Letting the campaign move money around is free.
- Adding a new ad set to an Advantage+ campaign does not reset the existing ad sets. This is the single most useful sanctioned fact for horizontal scaling.
- Ad-set-level edits stay contained to that ad set.
The counterweight: “adjusting your campaign budget might cause multiple ad sets within the campaign to re-enter the learning phase”. Under Advantage+ campaign budget, a campaign-level budget change has a campaign-wide blast radius. That is the real scaling risk, and it argues for adding capacity rather than turning the money dial.
Two more operational notes from Meta: batch your edits in bulk to minimise total learning time, and do not exceed seventy ad sets per campaign. After you add an active ad set, allow about two hours for budget distribution to re-adjust.
Auction overlap: what duplication actually costs you
The standard scaling playbook says duplicate your winning ad set. Before you do, understand what Meta says happens.
First, the reassuring part — your ads do not bid against each other. Meta: “we choose the ad with the highest total value to compete… other ads from that advertiser are not considered in this auction. This ensures that your ads will not bid against one another.”
The cost is subtler and worse: suppressed delivery. Meta states that auction overlap “can prevent an ad set from spending its full budget or achieving enough results to exit the learning phase”. Your duplicates do not inflate your CPMs; they starve each other of the events needed to stabilise. And Meta connects this directly to our topic: “too much auction overlap across your Page typically results in less predictable performance – especially when scaling your budget.”
Splitting across separate ad accounts does not help. Meta confirms overlap “can occur across these accounts”.
You can check overlap in Audiences → select up to five → Actions → Show audience overlap, but note two constraints. Useful data requires audiences of at least 10,000 accounts — a genuine limitation at Singapore niche sizes. And the percentage is asymmetric: Meta’s own example has 800,000 shared accounts reading as 80% overlap of a one-million audience but only 8% of a hundred-million one. In a market this small, overlap percentages look alarming by construction.
Meta publishes no overlap threshold. The 20%, 25% and 30% figures circulating are agency convention, and they contradict each other. Meta’s own framing is that overlapping audiences are “not necessarily a bad thing, but… can lead to poor delivery”.
The Singapore ceiling nobody calculates
Here is the constraint that makes Singapore different, and that we could not find addressed in any competing guide.
Per DataReportal’s Digital 2026 Singapore: total population 5.88 million, internet users 5.78 million, social media user identities 5.33 million. Facebook’s potential ad reach in Singapore is 3.80 million — 64.6% of the population. Instagram’s is 3.35 million.
Resist the temptation to add those together. DataReportal publishes no de-duplicated Meta figure for Singapore, and the true union is capped by the 5.33 million total social identities.
What that ceiling means practically: frequency is arithmetic, not opinion. If your targeting narrows 3.80 million to a realistic 400,000, then at a S$12 CPM, S$400 a day buys roughly 33,000 impressions daily. Inside a month that is around one million impressions against 400,000 people — an average frequency approaching 2.5 before you have done anything wrong. Double the budget and you are at 5.0 within the same month. In Indonesia the same spend disappears into a pool thirty times larger. In Singapore, budget converts to frequency far faster than most advertisers expect, and frequency is what kills performance.
Work out your own number before you scale: addressable audience, planned daily spend, prevailing CPM. If the frequency you land on in 30 days is above 3, more budget is not your next move.
What Singapore Meta costs actually look like
The best-documented Singapore source we found is Superads, publishing medians from over US$3 billion of anonymised Facebook spend:
| Metric | Singapore median | Peak observed | Note |
|---|---|---|---|
| CPM | US$15.10 | US$35.30 (Oct 2025) | SG month-to-month volatility ~US$5.00 vs US$1.90 globally |
| CPC | SGD 1.00 | SGD 2.02 (Oct 2025) | — |
| CTR | 1.48% | — | ~27% below the global median of ~2.04% |
That October 2025 spike is the 9.9 and 10.10 run-up showing up in real spend data — CPMs more than double the annual median. Singapore agencies commonly claim 20–60% seasonal inflation around Chinese New Year, the Great Singapore Sale, 11.11 and 12.12, but none publish methodology, so treat those as estimates. The Superads October figure is the one number here backed by actual spend.
The planning implication is straightforward: scale into the shoulders, not the peak. If your CPM doubles in October, the budget increase you scheduled for October is buying half the reach it would have bought in August, and your marginal cost per acquisition will look like a campaign failure when it is a calendar problem.
Creative volume: the real scaling lever
If budget is not the lever, what is? The most defensible dataset we found is Motion’s Creative Benchmarks 2026 — over 550,000 ads, 6,000+ advertisers and roughly US$1.3 billion of Facebook and Instagram spend between September 2025 and January 2026.
| Monthly spend tier | Avg new creatives per week (all accounts) | Top 25% of accounts |
|---|---|---|
| Under US$10K | 2.80 | 4.83 |
| US$10K–50K | 4.10 | 8.09 |
| US$50K–200K | 6.67 | 15.95 |
| US$200K–1M | 11.24 | 31.11 |
| US$1M+ | 18.85 | 54.64 |
The finding that should reorganise your budget: “even when budgets are the same, brands launching more creative get twice the number of winners.” Top-quartile accounts launch two to three times more creative than same-budget peers.
The supporting arithmetic is sobering. Only 5–8% of ads become winners (defined as spending at least ten times the account’s median single-ad spend). Roughly half of all ads never receive meaningful spend, and about 6% of ads drive the majority of an account’s spend. Twenty ads yields perhaps one or two winners. Fifty yields three or four.
One caveat in fairness: Motion sells creative analytics software and has a commercial interest in “make more creative”. It is also the only vendor with a verifiable 2026 dataset on this, and the ~5% win rate is independently corroborated.
Meta’s own December 2025 guidance points the same way but stops short of numbers. It defines creative diversification as “the creation of distinct campaign assets tailored to different personas or use cases”, explicitly distinct from iteration, and says ads should be “truly different in look, feel, storyline, and message”. Meta prescribes no specific creative volume — anyone quoting a “Meta-recommended number” is inventing it.
On fatigue, be equally careful about what is measured versus asserted. Ads Manager surfaces “Creative Fatigue” and “Creative Limited” delivery statuses. Frequency thresholds — decline above 2.5 weekly on prospecting, a cliff past 4.0, retargeting tolerating 8–10 — are practitioner benchmarks, not Meta figures. Useful, but not gospel.
Diagnosing a campaign that has stopped scaling
When performance degrades, the question is which layer broke. These signatures separate them:
| Symptom | Most likely cause | What to do |
|---|---|---|
| CPM rising, CTR stable | Auction competition or seasonality | Check the calendar before the creative; consider shifting spend timing |
| CTR falling, CPM stable | Creative fatigue | New creative into the same audience |
| Frequency and Audience Reached Ratio rising, Reach flattening, First Time Impression Ratio falling | Audience saturation | New audience — more creative will not fix this |
| Healthy CTR, falling conversion rate | Post-click: landing page, offer or price | Stop touching the ad account |
| Link clicks high, landing page views much lower | Page speed — or consent-driven tracking loss | Check page speed AND your consent banner before concluding |
The cleanest test to separate the two most-confused causes: if fresh creative in the same audience restores performance, it was creative fatigue. If performance only recovers on a new audience, it was saturation. They look identical in a dashboard and demand opposite responses.
Some cost pressure is simply market-wide and not your fault: Meta’s Q1 2026 results reported average price per ad up 12% year on year, with impressions up 19%.
Measurement, and why your ROAS is probably wrong
Install the Conversions API alongside the Pixel — Meta’s published figure is that advertisers with a web CAPI setup saw an average 17.8% lower cost per result. Note that “dataset” is a UI rename, not a replacement: the identifier is still your Pixel ID. Contrary to a persistent rumour, there is no Meta Pixel deprecation; no such notice exists in any changelog. Deduplication requires matching `event_id` values received within 48 hours.
Event Match Quality is scored out of 10, and the parameters that move it most are email, IP address, name and phone. Meta publishes no target EMQ score — the ubiquitous “aim for 7+” is agency convention.
Attribution changed materially in 2026, and most guides are stale. The old “7-day click plus 1-day view” combo is no longer a single selectable option: click-through (1 or 7 day), engage-through (1 day) and view-through (none or 1 day) are now chosen independently, under a model selector offering Standard, Incremental or Custom. From March 2026 engaged-view was renamed engage-through, click-through narrowed to link clicks only, and the video threshold moved from 10 seconds to 5 — Meta’s stated reason being that “46% of online purchase conversions with Reels now happen within the first 2 seconds”. Also commonly reported backwards: 28-day click survives; 28-day view was removed on 12 January 2026.
Now the part that inverts the usual narrative. The common claim is that Meta overstates ROAS. The two largest 2025 experiment sets found the opposite at account level. Haus, across 640 incrementality experiments with advertisers averaging US$14 million annual spend, found that “for every $100 of revenue the platform reports, Manual campaigns deliver roughly $12 more revenue to the business than Advantage+”. Stella’s 46 studies found average incremental ROAS of 2.87 against platform-reported 2.37 — an incrementality factor of 1.21.
The practical conclusion for scaling: platform-reported ROAS is a directional signal, not truth, and the error does not run in the direction most people assume. If you are scaling on reported ROAS alone, run a geo holdout before you commit serious budget. Note that a standard A/B test cannot answer this — every cell is exposed — and Meta’s Conversion Lift is access-gated through a Meta representative.
The compliance layer
Two Singapore rules bear directly on scaling.
PDPA and your pixel. The PDPC is explicit: “where targeting of advertisements involves the collection and use of personal data through cookies, the individual’s consent is required”, and “as good practice, organisations should provide individuals with the ability to set their cookie preferences within the website”. Consent is not needed for cookies that do not collect personal data, nor for authentication, security, user preference or network management. Critically, “the mere failure of an individual to actively manage his browser settings does not imply that the individual has consented”. Singapore’s position is materially weaker than GDPR — there is no requirement for a prior-blocking banner — so do not write your policy from European habits, but do not assume silence equals consent either. Penalties reach S$1 million, or 10% of Singapore annual turnover where that turnover exceeds S$10 million.
This connects to the diagnostics table above: consent rejections produce link clicks with no pixel-fired landing page view. A “broken landing page” is sometimes a working consent banner.
GST. The rate is 9%, effective 1 January 2024, and Budget 2026 announced no change. Meta’s rule: if your “Sold to” address is Singapore and you have not added your GST registration number, Meta adds GST at the local rate; add the number and Meta does not charge it. Reverse charge is widely misdescribed — per IRAS it applies to GST-registered businesses “which are not entitled to full input tax claims”, so a normally taxable Singapore SME does not apply it. Supplying your GST number is the cleaner route. If you are not GST-registered, that 9% is a real cost on every dollar of media, and it belongs in your scaling model.
When scaling Meta is the wrong answer
Three signals say the problem is not your Meta account: frequency above 3–4 per week on your core audience, spend increases no longer producing proportional revenue, and CAC rising quarter on quarter despite competent optimisation.
For Singapore B2B, one DataReportal figure should end a common argument: LinkedIn’s Singapore reach of 5.10 million exceeds Facebook’s 3.80 million. “Meta is where the audience is” is simply not true for B2B here — see our LinkedIn Ads guide for Singapore B2B. And TikTok’s adult reach of 3.80 million now matches Facebook’s headline reach, which is the subject of our TikTok versus Meta comparison.
The honest counter-position: exhaust Meta first. There is usually real scale curve left before diversification beats optimisation. When you do test a new channel, allocate around 10% of budget for a full quarter and judge it on blended marketing efficiency, not platform-reported numbers.
The other direction is geographic. Singapore’s Facebook ad reach is 3.80 million, full stop. Malaysia and Indonesia are order-of-magnitude larger pools. If your product travels, expanding the market is often cheaper than buying more frequency in this one.
Frequently asked questions
Is it true that increasing budget by more than 20% resets the learning phase?
No — not as a Meta rule. Meta classifies budget changes as only conditionally significant and publishes no percentage. Its sole worked example says +1% is unlikely to reset and +900% may, deliberately leaving everything between undocumented. The 20% figure is industry convention. Increase gradually because it is prudent, not because Meta said so.
Should I duplicate my winning ad set to scale?
Cautiously. Your ads will not bid against each other — Meta guarantees that. But auction overlap can stop an ad set spending its full budget or gathering enough events to exit learning, and Meta specifically warns it makes performance less predictable when scaling. Adding an ad set to an Advantage+ campaign is documented as not resetting its siblings, which makes it a safer route than raising campaign budget.
How many new creatives do I need per week to keep scaling?
Benchmark against your spend tier. Motion’s 2026 dataset shows accounts under US$10K/month average 2.8 new creatives weekly while top performers run 4.8; at US$50–200K it is 6.7 versus 16. With only 5–8% of ads becoming winners, volume is arithmetic. At equal budgets, brands launching more creative get roughly twice the winners.
Why does my Meta ROAS not match my actual revenue?
Partly modelled conversions and signal loss, partly attribution windows. But the direction may surprise you: the two largest 2025 incrementality studies found Meta under-reports at account level — Haus measured roughly $12 more real revenue per $100 reported for manual campaigns, and Stella found incremental ROAS of 2.87 against 2.37 reported. Treat platform ROAS as directional and validate with a geo holdout.
Is there a practical spend ceiling for a Singapore-only Meta campaign?
No published figure exists, and anyone quoting one is guessing. Calculate your own: Facebook’s Singapore ad reach is 3.80 million, so take your realistic addressable audience, your daily spend and the prevailing CPM, and work out the frequency you reach in 30 days. If it exceeds about 3, you have found your ceiling — and it is an audience ceiling, not a budget one.
Do CPMs really spike during Singapore’s shopping seasons?
Yes, and there is real data for it. Singapore CPM peaked at US$35.30 in October 2025 against a median of US$15.10 — the 9.9 and 10.10 run-up, in actual spend data. Local agencies claim 20–60% inflation around Chinese New Year, the Great Singapore Sale, 11.11 and 12.12, but publish no methodology. Plan increases into the shoulder months.
Where this leaves you
Scaling on Meta is less about the budget dial than the industry pretends. The documented levers are creative volume, adding capacity rather than raising campaign budget, keeping auction overlap out of your own way, and knowing when you have simply run out of Singapore.
The undocumented ones — the 20% rule, the overlap thresholds, the EMQ targets — are conventions. Some are reasonable. None are platform rules, and treating them as such is how accounts get managed by superstition.
If you want the arithmetic run against your own account — addressable audience, frequency trajectory, marginal cost of the next customer — our Meta ads team does exactly this, and our performance marketing practice handles incrementality testing. Our case studies show how we report it.
Related Meta guides
- Meta Ads Singapore: the complete guide — the cluster pillar
- Meta campaign structure: CBO vs ABO
- Meta Ads cost in Singapore
- Meta Pixel and Conversions API setup
- Meta ad formats and placements
- Google Ads Singapore guide — the other half of most budgets
Sources
- Meta Business Help Centre — About the learning phase, About learning limited, About significant edits, Understand auction overlap, Advantage+ campaign budget best practices
- Meta Q1 2026 results; Meta engineering blog (Andromeda); Meta creative diversification guidance (Dec 2025)
- DataReportal, Digital 2026 Singapore
- Motion, Creative Benchmarks 2026 (550,000+ ads, ~US$1.3B spend)
- Haus (640 incrementality experiments) and Stella (46 studies) on incremental vs reported ROAS
- Superads Singapore CPM/CPC/CTR medians
- PDPC, Advisory Guidelines on the PDPA for Selected Topics (rev. May 2024), Ch.7
- IRAS current GST rates and GST on imported services; Meta GST billing help
