Why Facebook Ads Spend All Budget on One Ad (And How to Fix It)

Direct answer: When Meta spends almost all budget on one ad inside a set of “variations,” you usually have one of three things: (1) the system found a true relative winner and is allocating efficiently, (2) your “variations” are near-duplicates so there is only one real concept to learn, or (3) constraints (learning, audience size, bid/budget structure, creative fatigue) starve exploration. Treat “one ad takes all” as a symptom, not a single disease.

What you’re seeing in Ads Manager

• 10 ads in an ad set; 1 ad eats 70–95% spend.
• Siblings sit on crumbs of impressions or “learning.”
• You “tested” headlines/colors/crops and conclude creative testing is broken.

Sometimes the account is healthy: a real winner should get more budget. The failure mode is when spend concentration happens before meaningful exploration or across ads that were never distinct.

Mechanism stack (simple → deep)

1) Auction + learning dynamics (always on)

Delivery systems allocate impression opportunities toward ads expected to hit your optimization goal more efficiently. Early stochastic wins can snowball. If you reset learning constantly, you can also get weird concentration patterns. This layer is normal media buying—not unique to 2026.

2) Creative retrieval quality (Andromeda-era context)

Meta’s Engineering post on Andromeda describes a next-generation personalized ads retrieval engine—the stage that selects a shortlist of candidate ads from a huge eligible pool before later ranking. Meta reports large gains in model capacity and retrieval quality metrics in that post, motivated partly by Advantage+ and generative creative volume growth.¹

Buyer implication (inference, labeled): when the system can evaluate more creatives, true diversity matters more. Ten near-identical UGC cuts are not ten retrieval opportunities; they are closer to one idea wearing ten filenames.

3) Near-duplicate / “same concept” collapse (practitioner model)

Media buyers commonly observe that minor edits (color grade, 3-word headline swap, center crop) do not behave like independent ads. Delivery concentrates. Whether you name that “entity,” “clustering,” or “similarity,” the operational fix is the same: change visual DNA + hook + angle, not just metadata.

Decision tree: is concentration good or bad?

How to fix it (implementation)

1. Inventory the set: screenshot first frames side by side.
2. Score distinctness: different hook, different promise, different visual setting, different proof device. Need ≥2 hard differences to call it a new concept.
3. Kill fake variants: headline-only and recolor tests are usually analytics theater.
4. Rebuild 5 concepts from different angles (problem, demo, founder, social proof-as-story, offer math).
5. Match placement: don’t ship Feed 1:1 crops as “Reels tests.”
6. Read concentration weekly: % spend by ad inside ad set; alert if top ad >80% while siblings <N impressions.
7. Separate explore vs exploit: once a winner is real, scale it deliberately; don’t keep twelve twins “for learning.”

What not to do

• Don’t add 20 more micro-edits to “force” even spend.
• Don’t daily-restructure the account to chase equal delivery.
• Don’t assume equal spend = healthy test. Equal spend on junk is still junk.
• Don’t cite unverifiable “Entity ID threshold = 0.70” myths as platform fact.

Does Meta "sequence" your ads? What happens when you turn off the losers

A popular observation among media buyers in September 2026 goes like this: you launch five ads in one ad set. Ads 4 and 5 get most of the spend, but ad 1 is the one getting credit for the purchases. You turn off ads 4 and 5 to save money, and ad 1's cost per purchase suddenly gets worse. Turn them back on, and ad 1 recovers. The conclusion many draw is that "Facebook ads is sequencing": that Meta deliberately shows people ads in an order, with some ads warming people up and another closing the sale⁴.

The pattern is real and worth taking seriously. The explanation is simpler than a hidden sequencing feature, and it changes how you should manage ads. Four ordinary mechanisms produce exactly this effect.

1) Attribution gives one ad the credit for a journey that involved several

Meta reports a purchase against the ad that gets credit under your attribution setting, for example within 7 days of a click or 1 day of a view⁵. When someone sees ad 4 on Monday, ad 5 on Wednesday and then clicks ad 1 and buys on Friday, the report generally shows one purchase for ad 1 and none for ads 4 and 5. Ads 4 and 5 did real work, but it is invisible at the ad level. Switch them off and ad 1 loses the help it was getting, so its cost per purchase rises. It looks like sequencing. It is really shared work with single-ad credit.

2) Different ads reach different pockets of people

Meta's retrieval system picks which ads to consider for each person¹. Genuinely different ads get matched to different people. When you remove ads, you shrink the range of people the ad set can reach well, and the remaining ad is pushed toward people it suits less. Its costs rise, not because it changed, but because it is now doing a job the other ads used to do.

3) Turning ads off unsettles delivery

Meta treats several kinds of changes as significant edits that can put an ad set back into learning, including changes to creative⁶. Pausing several ads at once can cause a few days of unstable results. Some of the "ad 1 got worse" effect is simply the ad set re-settling.

4) The remaining ad wears out faster

With fewer ads, the survivor shows to the same people more often. Frequency rises, attention falls, and cost per purchase climbs. That is ordinary creative fatigue, sped up.

What to do about it

• Judge at the ad set level first. The question is not "which ad has the lowest cost per purchase?" but "does the ad set as a whole get cheaper or more expensive when I remove this ad?"
• Do not kill low-spend ads just because they show no purchases. An ad with few reported purchases may still be helping. If an ad has spent very little, it has not been tested either way.
• Test removals, do not assume them. When you want to cut ads, change one thing at a time and compare the ad set's cost per purchase for 7 days before and after. If it gets worse, put the ad back.
• Cut ads that clearly lose at the ad set level, meaning the whole ad set improves or holds steady without them.
• Keep genuine variety live. Several clearly different ads give the system more ways to reach and persuade different people. This is the same reason near-duplicates fail: they add no new reach.

Real winner or early favourite? A quick test

Concentrated spend has two very different causes. Either Meta found a real winner, or it picked an early favourite and never gave the others a fair chance. Here is how to tell them apart:

If it looks like an early favourite, the fix is not to force even spend. Launch new challengers that are genuinely different, and consider testing them in their own ad set so they get a fair budget before joining the main one.

Two 2026 changes that make concentration more likely

Placement controls are going away. Meta is removing placement, platform, device and OS exclusions from Sales and Leads ad sets. When every ad can run on every placement, an ad built natively for Reels and Stories can take most of the budget simply because it fits more of the inventory, while a feed-only design struggles. See Meta removed placement controls.

Small budgets exaggerate everything. On $30 to $100 a day, one ad getting 90% of spend can happen within hours, and the others may never collect enough data to be judged. Run fewer, more different ads on small budgets. See how many creatives to test at $50/day.

Where SellTheClick fits

If your pain is “we thought these were different,” run the pack through creative analysis for hook/angle overlap and similarity risk before the next spend wave. That does not replace offer testing or CAPI hygiene; it reduces self-competition inside the creative layer Meta’s retrieval stack cares about.¹

FAQ

Why does Facebook spend all budget on one ad?

Often because one ad is the only real concept, or it earned early efficiency and the system allocates toward it. Confirm whether siblings were ever distinct and whether they received meaningful delivery.

Is winner-take-most always bad?

No. A true winner should take more budget. It’s bad when concentration hides a failed test design.

Is this caused by Andromeda?

Andromeda is Meta’s retrieval engine upgrade for large-scale candidate selection.¹ It raises the importance of creative quality/volume. It is not a magic label for every delivery quirk.

Do headline tests create new ads?

Sometimes for learning signals; often not enough to behave as independent concepts if the visual story is identical. Validate with delivery reality, not filenames.

Should I force even spend across ads?

Usually no. Force better hypotheses. Artificial evenness can burn money on losers.

How many concepts should I test per wave?

Fewer truly different concepts beat dozens of twins. A practical wave is 3–7 distinct angles with enough budget to exit learning—exact numbers depend on CPA and vertical.

What’s the fastest diagnostic today?

Export top ad set by spend, sort ads by spend share, open first frames. If they look like siblings, rebuild concepts before adding budget.

Does Facebook show ads in a sequence?

Meta does not publish a feature that plays your ads in a fixed order. The "sequencing" effect media buyers see is usually several ads sharing the work of persuading someone, while attribution gives the purchase credit to only one of them.

Why does my best ad's CPA go up when I turn off other ads?

The other ads were likely helping: reaching different people, warming them up before they bought through your best ad, and keeping its frequency down. Removing them also unsettles delivery. Judge removals by the ad set's cost per purchase, not by one ad's numbers.

Worked example (illustrative structure, not fabricated ROAS)

Suppose an ad set contains:

• Ad A: UGC bathroom mirror demo, hook “Stop using X on damp skin”
• Ad B: same creator, same bathroom, hook text recolored
• Ad C: same demo, first 2 seconds trimmed
• Ad D: true different concept — founder in car, different promise, different proof

Buyers often see A take most spend, B/C starve, D sometimes explore. The mistake is concluding “Meta hates testing.” The cleaner read: A–C were one concept; D was the only alternate hypothesis. Your next wave should look more like D vs a new E/F, not B2/B3 crops.

Use this as a template for your own screenshots—do not invent performance numbers for A–D. Pull spend share from Ads Manager and paste real first frames into a slide. That artifact alone usually ends the argument in Slack.

Measurement checklist (copy into Notion)

• % of ad-set spend on top ad (7-day and 3-day)
• Impressions on ads 2–N (did they ever leave single digits?)
• First-frame collage attached
• Count of hard differences vs winner (hook / setting / talent / promise)
• Decision: scale winner / kill twins / new concept wave
• Owner + date of next concept drop

References

1. Engineering at Meta — Andromeda: next-gen personalized ads retrieval engine
2. Meta Business Help Center (verify current learning-phase / delivery documentation in Ads Manager help for your region)
3. SellTheClick — Reels vs Feed (placement-native creative)
4. Public post on X by @codyschneider (Sep 10, 2026) describing ad "sequencing" when pausing ads within an ad group. Practitioner observation.
5. Meta Business Help Center — About attribution models and attribution settings
6. Meta Business Help Center — Significant edits and the learning phase