Case study · Client not named · Utility software · Meta
US utility software brand: 13,639 purchases at $27 each from 30 months of AI ads on Meta.
AI Vidia made the ads and ran the Meta ad account for a US utility software company from May 2024 to November 2026. On $371,142 of Meta ad spend, Meta recorded 13,639 website purchases at $27.21 each and a 1.75x return on ad spend.
The method: about 22 new AI ads a month in English and Spanish, released in a batch every 14 days, so Meta Andromeda always had new ads to choose from. The client is not named.
- website purchases
- 13,639
- per purchase
- $27.21
- return on ad spend
- 1.75x
- Meta ad spend over 30 months
- $371,142
ROAS by campaignHighlighted: Spanish, 2.48x
On this page
How to read these figures
Source: the client's Meta Ads Manager
Every figure is what Meta reports for each campaign's lifetime. Purchases and revenue are the website purchase events and values Meta attributes to the ads under the account's attribution setting.
They are not checked against the client's payment data, and refunds are not deducted. ROAS is Meta-attributed revenue of about $648,700 divided by ad spend. 17 registrations from a small UK campaign are not counted as purchases.
What did each campaign return?
Eight campaigns ran in the account. The English core carried 84% of the spend; the rest were retargeting, a Spanish-language campaign and small tests.
Exhibit 1 · Results by campaign, Meta-attributed, campaign lifetime
| Campaign | Spend | Purchases | Cost per purchase | ROAS |
|---|---|---|---|---|
| Core prospecting, English | $312,679 | 11,369 | $27.50 | 1.73x |
| Retargeting | $19,802 | 886 | $22.35 | 2.10x |
| Target-ROAS bidding test | $10,385 | 434 | $23.93 | 1.70x |
| Prospecting, Spanish | $9,372 | 489 | $19.17 | 2.48x |
| Creator partnership ads | $5,703 | 230 | $24.80 | 1.90x |
| Creative testing setup | $4,964 | 105 | $47.28 | 1.04x |
| Budget test | $3,074 | 125 | $24.59 | 2.20x |
| Reach campaign and 3 pilots | $5,163 | 1 | n/a | 0.59x |
The core campaign's cost per purchase averaged $27.50 over its lifetime. It reached 38.9 million people at about $1.09 per 1,000 impressions, because the audience was worldwide and included many lower-cost markets.
How were the ads made and released?
One partner held both jobs: AI Vidia made the creative and ran the Meta account, so the client's team could stay on the product.
22
new AI ads a month
On average, in English and Spanish. Each batch brought new angles and concepts, not recoloured copies of the last one.
14 days
between batches
New creative went into Meta every two weeks.
Each refresh
weak ads switched off
Ads that had stopped performing were turned off and replaced, so the account never leaned on one ageing winner.
A new batch every 14 days, May 2024 to November 2026
Share of spend
- Core prospecting, English: 84%
- Seven other campaigns
AI Vidia and the client meet every 7 to 10 days for about an hour: what is scaling, what is tiring and which concepts go into the next batch. The client approves the direction without running the account or the production.
Why did Spanish ads cost less per purchase?
The Spanish-language campaign, added in November 2025, paid about 30% less per purchase than the English core. Reach cost half as much, and buyers spent the same.
English core
$27.50
Cost per purchase
Spanish
$19.17
Cost per purchase
30% lower cost per purchase
Exhibit 2 · English core against Spanish, Meta-attributed
- Cost per purchase (lower is better)
- English core$27.50
- Spanish$19.17
- ROAS (higher is better)
- English core1.73x
- Spanish2.48x
- Cost per 1,000 impressions (lower is better)
- English core$1.09
- Spanish$0.55
- Average order value (the same in both)
- English coreabout $47.50
- Spanishabout $47.50
The Spanish campaign ran for a shorter time on a smaller budget, with 489 purchases, and returns usually fall as spend grows.
How many new ads does Meta Andromeda need?
Andromeda is the retrieval stage of Meta's ad system. It narrows tens of millions of candidate ads to a few thousand for each person before ranking begins, as Meta Engineering described it on 2 December 2024. The ad itself now does much of the targeting, so distinct concepts give the system more chances to find buyers.
There is no fixed number. This account gives a working ratio, not a rule.
How Andromeda narrows the field
- Tens of millions of candidate ads
- A few thousand for each person
- Then ranking begins
The ratio in this account
- impressions per new ad
- 540,000
- purchases per new ad
- 21
About 660 new ads against about 355 million impressions and 13,639 purchases: one new ad for every 540,000 impressions, or every 21 purchases, released in two batches a month.
Enter your own monthly volume to see what this account's ratio would mean for you.
At this account's ratio
About 22 new ads a month, or about 11 in each two-week batch.
Stated in impressions and purchases, not dollars: this account ran worldwide at about $1 per 1,000 impressions, so its average $12,400 a month bought far more reach than the same budget would in the US.
What did not work?
Four results we would not repeat, and what each taught us.
Cheap clicks, few buyers
The creative testing setup had a 2.85% click-through rate at about a cent per click, yet returned 1.04x on 105 purchases.
Judge creative tests on cost per purchase, and keep test ads to the main Facebook and Instagram placements.
Reach did not buy sales
The reach campaign and three small pilots spent $5,163 and produced one purchase between them.
When the goal is a purchase, optimise for the purchase. A reach campaign buys attention, not buyers.
Retargeting looks better than it is
Retargeting returned 2.10x, but each person saw its ads about 44 times.
At that frequency many purchases would have happened anyway. A holdout test is the way to measure what it adds.
Bidding was not the lever
Target-ROAS bidding returned 1.70x, against 1.73x for the core on standard bidding.
A neutral result: the bid strategy alone did not change the return.
Test the method on your own product
The First Batch is one concept for one of your products, made as 4 videos and 4 image ads. Run it in your own Meta account and judge it on cost per purchase.
What this case does not show
- The client's name, its product or its campaign names.
- Profit. Refunds, renewals and the client's payment data are not part of these figures.
- Incremental lift. No holdout result is published, so the figures show what Meta attributes, not proof that each purchase came from an ad.
- A forecast for your brand. The ratio and the Spanish result belong to this account.
Questions
What results did AI Vidia get with AI ads on Meta for a utility software brand?
13,639 website purchases at $27.21 each and a 1.75x ROAS on $371,142 of Meta ad spend over 30 months. The figures are Meta-attributed and do not deduct refunds.
How many new ads does Meta Andromeda need?
There is no fixed number. This account received about 22 new ads a month, one for every 540,000 impressions or 21 purchases, in two batches a month. A larger account needs more distinct concepts, a smaller one fewer.
Do Spanish-language ads lower cost per purchase on Meta?
In this account they did: $19.17 per purchase against $27.50 for the English core, mainly because reach cost half as much while buyers spent the same. The Spanish campaign was smaller and shorter, so the gap may narrow as spend grows.
Who ran the Meta ad account?
AI Vidia ran it and made the creative, and met the client every 7 to 10 days to agree the next batch. On other accounts the client's own team launches and tests the ads; it is agreed per client.
Sources
- ResultsThe client's Meta Ads Manager, each campaign's lifetime, read by AI Vidia. Campaign names are left out.
- MetaMeta Engineering, Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine, 2 December 2024.
Cite this case
AI Vidia (2026). US utility software brand: 13,639 purchases at $27 each from 30 months of AI ads on Meta. https://ai-vidia.com/case-studies/utility-software-meta-andromeda