abhi.performs
Abhishek ThakralWhatsApp me
Audit · AI video SaaS · Meta Ads

An AI SaaS blamed its Meta ads for a ROAS collapse. The data said otherwise.

A subscription AI video tool saw Meta ROAS fall from 1.88 to 0.60 and cut spend from about €3,000 a day to €100. The team blamed weak creative. My Meta Ads audit of €256,555 of spend (1 March to 30 September 2026) found the creative was fine, most of the drop came from how purchases were counted, and 32% of the last quarter's budget went to countries that barely convert.

Chat on WhatsApp
Weekly ROAS · Mar to Sep 2026From Meta Ads Manager
1.88 → 0.60reported ROAS, early March vs July to September
Mar17 Mar: Meta changes reportingSep
+59%link CTR vs the "good" months
65%of purchases were view-through
32%of spend in no-signal countries
ROAS by week. The drop starts the week of 22 March, five weeks before the account's own tracking change. Weeks between 26 April and 30 June are left out to keep the bars readable.
Client
AI video creation SaaS, subscription, sold worldwide
Channel
Meta Ads, optimised to purchases
Period audited
1 March to 30 September 2026
Spend audited
€256,555 across 30 campaigns
My role
Account audit, diagnosis and recovery plan
The starting point

What the team believed was wrong

After about two years of profitable growth, the account had scaled to €3,000 to €4,000 a day worldwide. Then results slid, tests stopped working, and spend was cut to around €100 a day. Target cost per acquisition was under €40, against an entry plan of about $35 a month.

Our hooks and creative are weaker than competitors'.

The working theory: the ads stopped stopping the scroll, so new creative would fix it.

ROAS won't stay above 1, so we cut spend.

Budget went from about €3,000 a day to €100, which made every test slower and noisier.

Cost per new subscriber doubled in August.

The number the team felt most strongly about, and the one that needed checking first.

On 26 April we changed what counts as a purchase.

Renewals from existing customers stopped firing the purchase event. Only new subscriptions count now.

How I audited it

Understand the business, fix the ruler, then read the data in order

Five steps. The order matters: measuring before judging stops you from fixing the wrong thing.

Start with the business, not the dashboard

Pricing, plans, credits and who buys. A $32 to $39 entry plan against a sub-€40 CPA target means first-month ROAS near 1 is normal; the profit is in retention.

Lock the attribution window

Every export used the same setting so periods could be compared fairly. I pulled nine files: three periods at campaign level, three at ad level, plus country, age and age by gender.

Build a before and after truth table

Spend, purchases, CPA and ROAS for the period before the tracking change, the transition and the last quarter, then the same thing week by week.

Test the creative theory directly

Hook rate, hold rate and click-through rate per period. If the creative had broken, these would have fallen with ROAS.

Find where the remaining money leaks

Countries, ages and campaigns ranked by return, to separate what needs more data from what can be fixed today.

Name what Meta can't show

List the exact data needed to finish the diagnosis, rather than guessing: billing data, the date of the price change and whether the Conversions API is live.

Finding 1

The real drop started on 22 March, not 26 April

Over three periods, ROAS fell from 1.34 to 0.88 to 0.60 and CPA rose from €38 to €79. But the weekly view shows ROAS at 1.88, 1.86 and 1.73 for the first three weeks of March, then 0.86 the week of 22 March. That lines up with Meta's own reporting change on 17 March, five weeks before the account's tracking change.

Period (2026)SpendPurchasesCPAROASSpend per day
1 Mar to 25 Apr€108,7242,829€381.34€1,941
26 Apr to 30 Jun€75,8991,342€570.88€1,150
1 Jul to 30 Sep€71,932913€790.60€782

Weekly ROAS from Meta Ads Manager. Purple: above 1.5. Red: below 0.5. Weeks of 26 April to 30 June averaged 0.88. On a phone, swipe the chart to see July to September.

What this means: a large part of the "collapse" is two changes in how purchases were counted. Before 26 April every renewal fired a purchase, so Meta was credited with revenue from customers it had already paid to acquire. The true new-customer ROAS was always closer to 1. That is a measurement story, not a performance story.
Finding 2

Two thirds of "purchases" were view-through: people who never clicked

In every period, 63% to 69% of reported purchases were 1-day view: someone saw an ad, didn't click, and bought within a day. Only about a third were click-driven. On a low-price product sold worldwide, that share makes every ROAS number on the dashboard soft.

Meta Ads Manager Compare attribution settings panel with 7-day click selected and All conversions as the conversion count
The live account's Compare attribution settings panel. I used it to split click from view-through purchases before reading any ROAS.
Finding 3

The creative wasn't the problem

If the ads had stopped working, hook rate and click-through rate would have fallen with ROAS. They didn't. CTR in July to September was 3.12%, up 59% on the "profitable" early period. People were still watching and clicking. The top creative took 34% of last quarter's spend with a healthy 21.4% hook rate; it was the return after the click that didn't hold up.

Period (2026)Hook rateHold rateLink CTR
1 Mar to 25 Apr18.4%23.4%1.96%
26 Apr to 30 Jun23.9%27.2%3.23%
1 Jul to 30 Sep15.8%22.2%3.12%
So the break is after the click. That leaves two candidates Meta can't tell apart: a real conversion problem (landing page, offer, the August price change) or under-counting of real subscriptions. Billing data separates them, which is why it was the first thing I asked for.
Finding 4

A third of the budget was going nowhere

In July to September, €23,083, or 32% of spend, went to countries with fewer than 10 purchases each. Ages 18 to 24 and 65+ took another 16% of spend at a ROAS of 0.33 and 0.37. One worldwide campaign optimised to a different event burned €3,971 at 0.19. And 30 campaigns shared about €780 a day, too little for any of them to learn.

NetherlandsAbove break-even
1.52x
keep
GermanyNear break-even
0.99x
hold
CanadaNear break-even
0.98x
hold
United KingdomNear break-even
0.84x
hold
United States40% of spend, €28.7K
0.73x
tighten
Australia€2,632 spend
0.50x
cut
Spain€1,989 spend
0.38x
cut
Age 65+Part of 16% of spend
0.37x
cut
Age 18 to 24€134 per purchase
0.33x
cut
Poland€3,579 spend
0.29x
cut
Worldwide 'Subscription' campaign€3,971 spend, €180 CPA
0.19x
cut
1.0x line: break-evenkeep above break-evenhold near ittighten big spend, belowcut no return
Finding 5

"Cost per new subscriber doubled" was really a 15% rise

The number the team felt most strongly about didn't hold up in the ad data. July: €32,257 spend, 436 new subscribers, €74 each. August: €21,714, 256 subscribers, €85 each. Some single August days spiked near €120, which can feel like double, but the monthly figure rose 15%.

€74JulyCost per new subscriber, 436 subscribers.
€85AugustCost per new subscriber, 256 subscribers.
+15%Real changeNot 2x. Partly the metric, partly thin, choppy budget.
LTVThe real question€85 is cheap if a €35 a month subscriber stays four or five months.
What I recommended

Fix what's certain now, request the data that decides the rest

Do this week

  • Cut countries with no purchase signal and the 18 to 24 and 65+ age groups
  • Pause the worldwide campaign running at 0.19 ROAS
  • Consolidate 30 campaigns into a few with enough budget to learn
  • Report on 7-day click alongside the default, so view-through can't hide the trend

Need from the client

  • Billing data: real new subscribers and how long they stay
  • The exact date of the August price change
  • Whether the Conversions API is live, and its match quality
  • Confirmation that the optimised Purchase event matches real subscription payments

One more thing worth checking: in the pixel event list, the Purchase event the ads optimise to fired roughly a quarter as often as the subscription payment event. If those measure the same thing, Meta has been learning from a fraction of the real signal, which would explain unstable delivery at €100 a day.

For SaaS and app founders

Five checks before you blame your creative

Run these on your own account. They take an afternoon and they stop expensive rebuilds.

01

Check what your purchase event counts

If renewals fire it, your ROAS is flattered and Meta learns to find people who are already customers.

Renewals stopped counting on 26 April
02

Read ROAS week by week

Quarterly averages hide when the break happened, and the date usually tells you why.

The cliff was 22 March, not 26 April
03

Look at your view-through share

Split purchases into click and view. The more are view-through, the softer every ROAS number.

63% to 69% view-through in every period
04

Test the creative theory with numbers

Hook rate, hold rate and CTR tell you if the ads stopped working. Check them before ordering new creative.

CTR up 59% while ROAS fell
05

Judge acquisition cost against lifetime value

A subscription is profitable on retention, not the first payment. Pull how long subscribers stay before you cut spend.

€85 per subscriber vs ~€35 a month plan
Your account

Want this audit on your Meta account?

Send me a few details and I'll tell you whether your problem is creative, measurement or budget leaks.

Questions SaaS teams ask

Meta Ads for subscription SaaS: FAQ

Why did ROAS fall if the creative was fine?

Because most of the fall was in what Meta was counting, not in what customers were doing. Meta changed how it reports conversions on 17 March, and the account stopped counting renewals as purchases on 26 April. Underneath that there was a real but smaller decline: cost per new subscriber kept rising, budget was spread thin across 30 campaigns, and a third of spend sat in countries and ages that barely convert.

Should a subscription business count renewals as purchases in Meta?

Not for optimisation. A renewal is revenue from someone you already acquired, so counting it tells Meta to find people like your existing customers and makes ROAS look better than acquisition really is. Optimise and report on new subscriptions, then judge them against lifetime value from your billing system.

What does a high view-through share mean?

It means Meta is crediting the ads for purchases from people who saw an ad but didn't click it. Here about two thirds of all reported purchases were view-through. Some of those are real, many would have happened anyway, so the more of your conversions are view-through, the softer your reported ROAS is.

What do you need for a Meta Ads audit?

Read access to the ad account is enough to start. To finish properly I also ask for billing data (Stripe or similar) to compare what Meta claims with real new subscribers and their lifetime value, the dates of any price or tracking changes, and whether the Conversions API is live.

Is cost per new subscriber of €85 too high for a €35 a month product?

Not necessarily. It only looks unsustainable if customers cancel after one month. If the average subscriber stays four or five months, €85 is a good price to pay. That is why lifetime value from billing data, not first-month ROAS, decides whether to scale.

Why don't you name the company?

The account belongs to the client, so the name stays private. All figures come from the client's Meta Ads Manager exports for 1 March to 30 September 2026.

Free SaaS ads audit

Find out if it's really your creative

Send me a few details about your Meta account and what you're seeing. I'll tell you whether the problem is creative, measurement or budget going to the wrong places, whether or not we work together.

Chat on WhatsApp
Free · Reply within one business day · Month to month, no lock-in