Last year, Meta released incremental attribution. When we reviewed it last year, we were skeptical.

On paper, it was a great product. Meta has been way ahead of other platforms by offering incrementality testing products via Conversion Lift for many years and incremental Attribution was built on that.

But the first release was a damp squib. In one of our tests we saw our off-platform incremental CPA increase 5x when using it. It just wasn’t effective at finding the people on platform.

Fast-forward a year, and boy has it had it a glow-up.

How incremental attribution works

All attribution these days is modelled in some way. Ever since iOS14.5, platforms have had less “true” data than they had before. Thanks to cookies and ad blockers and privacy-modes, it’s harder to track a user journey deterministically.

And so even the “7 day click, 1 day view” attribution model we’ve all defaulted to over the last few years is a predictive model.

For businesses with no brand or organic or other presence, we find on average that model under-attributes the true performance. And for those with high volumes of external stuff happening it does the opposite.

So the old attribution model was always broken.

The new one is based on good stuff: all of the conversion lifts that Meta have run.

For those unaware, these are holdout tests that expose your ads to 80-90% of an audience and not the last 10-20%. You count the number of purchases in your unexposed audience and treat it as a baseline, then count the volume in the exposed audience. The difference is your incremental CPA.

Using that as a model made great sense. But last year, it just didn’t work out.

Subscribe now

California Wine Club: on-platform and off-platform improvements

We’ve been working with California Wine Club for over a year now.

Summer in DTC is never a peak period, and so we wanted to run a backlog of media buying tests. One of which was incremental attribution.

Now, given our history with IA, we ranked it with low confidence – however, we always like to revisit experiments every 6-12 months. Why? Meta’s algorithm is constantly changing. As well as the major updates like andromeda and gem and lattice, there’s also the smaller day to day updates they push all the time.

Not only that, but people change. Browsing habits, culture, usage of the platforms. With every change comes an opportunity to re-test something.

The results

We decided that the test needed to be a total account switchover and we’d measure before/after. In a dream world we’d run a conversion lift, but time and cost got in the way.

And so we ran a full switchover of our hero campaign setup.

Now, ignoring the fact that new campaigns usually take time to get up to speed, this straight away got off to a great strat.

The key findings were this:

• Platform CPA itself barely changed (3% up in one like for like period)

• Platform Incremental CPA went down a big chunk (29%)

• And off-platform new customer CPA went down even more (33%)

This was an account where spend was – owing to off-peak – static recently. But as a result of this we’re scaling spend and those ncCPA figures remain stronger than before.

Interestingly, before IA, most of the acquisitions attributed were viewthrough, whereas now they’re predomiantly click.

Testing on other clients

As you can imagine, a test like this created quite the excitement in our #growthstrategists chanel.

Since this July test, we’ve now begun rolling out incremental attribution tests with a series of other clients.

Other early results are positive:

• One client saw platform CPA come down 20% overnight

• Another we saw a 7% decrease in blend, while platform remained flat

• We’ve begun testing it with earlier stage clients where data is more volatile to good effect to.

But it’s not all singing and dancing.

We have also seen one instance of almost no impact, and at the same time an increase in ncCPA.

Outgoing thoughts

From automatically turning on AI optimisations to the buggy Ads Manager experience, it’s easy to get frustrated at Meta. But this is one of those great examples of algorithm improvements taking place in the background helping things along.

In the space of six weeks, IA has gone from a ‘steer clear of this’ to a position of almost being the new default.

It’s not 100% success rate (nor is anything). And so testing is vital. And we would highly recommend running our approach to testing which is:

Wholesale account switchover: before and after time window tests

Without this, our hypothesis is that Meta – given the choice of two ads to serve in the auction, will opt to serve the easier to attribute one (non-IA).

We have seen stronger impacts where accounts were viewthrough-driven, but it is impactful in most accounts even when that’s not the case.

How are you finding Incremental Attribution?

Leave a comment

Josh Lachkovic is the founder of Ballpoint. We’re the growth agency you hire when you want to dominate your competition. We find profitable growth through psychology-first creative. We’re AI-native, meaning you get the technology power of a tech startup powering a boutique agency. Get in touch if you want to scale spend from £100k to £500k.

Last Friday, Jones Road Beauty CEO Cody Plofker tweeted about Meta performance this year.

Jones Road Beauty spend an estimated eight-figures a year on Meta, and anyone who follows Cody will know they are on it when it comes to marketing evaluation.

Cody references Il Makiage, the brand who skyrocketed to $800m of revenue with a heavy weighting thanks to …

Read more

In 1991, David Ogilvy stood up at the Association of National Advertisers (ANA) Convention and said:

I bel…

Read more

Read more

In mid 2025, Haus released its meta-study of Meta ads1. The finding that stuck with me:

ASC performed 9% better than Manual campaigns at experiment midpoint. But 12% worse by the end.

In other words: ASC was quick to pick up low-hanging fruit, but had lower incrementality over time.

This matched what we were seeing. And it raised an obvious question: should…

Read more

Three weeks ago, we launched our Annual Report, which you can download now. We are publishing key stories from this in the run up to Christmas.

If you’ve been reading this newsletter for a while, you’ll know I’ve been banging this drum since 2023.

Clickthrough rates don’t correlate with CPA.

I wrote about it in March 2023. I wrote about it again in Septemb…

Read more

I hope you had an excellent Black Friday weekend, and now well into planning for the last stretch of the year.

A fortnight ago, we launched our second ever Annual Report.

What’s important this year?

Over the last 18 months, Meta has introduced four major algorithm updates.

1. Lattice

2. Andromeda

3. Sequence Learning

4. GEM

All four of these have changed how yo…

Read more

Last week, we launched our Annual Report, which we’ll be launching soon to a wider audience. For the next three weeks, we’ll be publishing key stories from that report followed by our usual Annual Book Review and reflection of the best articles of the year.

2025 is the year that AI ads started to become real alternatives.

Having an annual reflection is important because it highlights just how much (and how little) can change in a year. The production of ads is one of those key pillars.

Here are the learnings from our data, reflections on the year, and highlights of conversations we had.

Notes for references throughout
The data from our Annual Report is based on £11m of ad spend. The Annual Report launch was attended by 52 people – roughly a third of whom were clients.

Share

20% of teams still use no AI

At the Annual Report launch, I asked people for a show of hands as to how much AI they used in ads.

The results?

• 100% AI – 2% of the audience

• Majority AI (more than 50% of ads) – 10% of audience

• Some AI (up to 50%) – 68% of audience

• No AI – 20%

I wanted to dig into this and so have had multiple conversations over the last few weeks about this. Here are the core buckets of people:

AI doesn’t get approved because it’s not good enough

In many cases, these are the sorts of brands who get push back on creative anyway. Brands where creator content is still a discussion to be had rather than a default, and there is a push for brand consistency across social assets.

In this case, there is the view that AI creative is highly distinguishable, and not good enough. When pushed on whether if it was indistinguishable from non-AI creative whether it would still be used, the teams were mixed in terms of yeses and nos.

Ethical concerns

For some brands, the core issue is the big environmental cost associated with AI – and by actively using AI ads, you become an active contributor to that. For some of these brands they might either have the environment as core brand pillars or alternatively see it as in conflict of their B corp status.

Similarly, there are ethical concerns about a future creative joblessness created by AI.

Customer complaints

One thing we see a lot of is users on platforms being negative towards AI. Either ridiculing it for its silliness, or actively calling out the brand for its use. Common complaints are often those from the creative joblessness angle, but they are broad. One ad comment we saw this year called the brand “lazy at marketing” for using it.

The emergence of an AI-Free Corp

One of my predictions for 2026 is the emergence of a certificate or signal whereby you actively decide you won’t use AI.

The current global economy is full of businesses that offer premium or luxury experiences on top of otherwise commodity products. Take coffee as a core example, you can get coffee for very very little in a lot of places. Coffee is a commodity available to everyone, but over the last 15 years, the new coffee class has emerged.

Now, everyone wants something a little more premium, while there is still a coffee elite/snob that is lionised and catered for in a smaller selection of venues.

In short there is commodity + premium + luxury.

I think the same will become true for AI.

AI will become the de facto norm for a huge number of brands in advertising, and that will reach from commodity (likely the Meta’s offering) through to creative and well executed. And then alongside that, there will be a separate class of ‘No AI’ creative brands, who use that as a point of differentiation.

2026 prediction: The emergence of a certificate or organisation that shows you are “AI Free”

Either way, as it currently stands, the ‘no AI’ crew is bigger than the ‘all in’ crew. As for the middle: the ‘some’ vs ‘most’ I imagine this time next year, the ‘most ads are made with AI’ crew will become the norm.

Leave a comment

AI ads make up a small amount of overall spend

There are two charts in the Annual Report that I thought were telling.

The first looks at spend across the last 12 months by production type.

For reference, we view each of these categories as “what is the majority of the ad made by.”

• AI is made using an AI tool as a starting point

• Creator is individual creator/influencer on their phone

• Studio is person sat at a Mac with After Effects/Photoshop/etc

Now this reflects a few things. First, that AI has only become good enough this year. Second, that it’s still just a small percent of overall spend for the year.

When you look at this on a “spend per ad” basis however, it changes.

Spend per ad is useful because it removes other variables like ‘number of assets made for each type’.

In this case AI is now at just 75% of the spend per ad as Creator or Studio.

That’s an incredibly close gap considering how underinvested in the category is.

We’ve had over a decade working on studio ads. And five years working on creator ads. But AI ads have only had experimentation for the last year.

It’s still super early. But once our learning, creativity, and experimentation with AI catches up, it promises something pretty impactful.

Share

AI ads are some of our most creative

The most exciting thing from our side is that AI ads are very often where the best creativity is happening.

Take this example for our client YourParkingSpace.

Humour and satire are great in ads. When you think that ~99% of ads don’t add any humour to them, just a little can go a long way.

This was one of those curve ball ideas in a brainstorm that felt a little out there.

The premise?

YourParkingSpace helps people book parking spaces in advance of a trip. One of the potential ‘villains’ we came up with was the Traffic Warden, and soon the idea became ‘traffic wardens must hate it because it’s fewer tickets.’

This was all a bit of fun and a bit of comedy.

It’s also the sort of thing that over a year ago, would have been left in the brainstorm room.

Why? Because, to have made this ad would have probably cost £1,000, maybe a little more.

And on paper, the most likely thing that could happen is that this won’t work. And so even if we had taken it to the client, the majority of the times, it likely would have been left in the ideas column. (Regardless of who the client is I should say, this isn’t specific to them).

That’s smart – £1,000 for something with low probability, vs the other plethora of far more likely to succeed ads that cost nothing or just a few hundred quid is a no brainer.

But with AI, this ad was achievable for a few tokens while we perfected the scripting, the brief and the prompt. And it took a few hours to get it from idea to finished ad, and was live the following day.

This is just one example where AI has completely unlocked new levels of creativity.

Subscribe now

Reframing how you see AI in creative

AI can save some money on making creative, sure. But that’s not its power.

It’s real power is unlocking whole new levels of creativity by helping you come up with ideas you couldn’t do before.

And for all the teams who fall into the anti-AI camp, some of those reasons are going to be non-negotiables, especially those on ethical grounds. But for those where it is a question of preservation of Brand Image, I’d recommend getting those brand teams in the room to see what you can do.

And no it’s not about feeding a product url into an ad generator and making 1,000 executions. I don’t think that’s what the future of AI will be for any large aspiring company.

It’s about starting with the idea, and then letting AI help you bring that to life.

Ultimately at the moment, it is seriously outpacing its level of investment, which means there’s opportunity and arbitrage for those who are there early.

What are your most exciting uses of AI so far?

I’d love to see some people share their most creative AI ads below. What have you come up with? What have you been able to do that you wouldn’t have otherwise.

Leave a comment

Josh Lachkovic is the founder of Ballpoint, a growth marketing agency that helps brands scale from £1-10M ARR. Visit Ballpoint to learn more, or subscribe to Early Stage Growth for weekly insights on profitable scaling.

Note: Ballpoint works with brands looking to grow their spend from £30k per month to £300k per month.

TL;DR: We analysed £7.5m of ad experiments to understand how quickly you can make judgements on winners/losers.

We’ve all been there. It’s day two of a new ad test and there’s zero conversions. CPA is currently ‘infinite’ and your CMO is asking “why haven’t you turned that off yet, is it ever going to recover?”

Creative testing remains the most important lever you have on Meta. In the early days it is the difference between 0 and 1. And then during scale days, the rate at which you’re able to test and learn is the deciding factor in how quickly you can grow.

But testing costs money.

There are lots of ways to test. You can test in cost controlled campaigns, in flexible ad formats, in ABO ad sets per concept. You can have a budget assigned to testing like 15% or 20%, or you can have 60%+ effectively as your testing budget.

The cost to test is high in many directions. First you’ve got the cost of putting creative together. That’s a mixture of insight + strategy + briefwriting + creative execution. The AI gurus are promising that cost is 0, but if you’ve seriously pushed volume into the AI creative, you’ll know there’s cost in other places.

Second, then there’s the cost of testing the ad in platform. Depending on your testing approach this might eat up a huge amount of budget.

While the expected value for creative testing should be high – because if your Meta engine is profitable and efficiency, feeding it with good creative is what helps it deliver results – that doesn’t change the fact that a lot of spend goes into ads that won’t ever be seen again.

We’ve just analysed the last 12 months worth of data as part of our annual report, examining 4,000 experiments.

Here are a load of insights from that report.

As with all of these things, you should be analysing this on an account by account perspective. While there are good pieces of data here to average out, these do not mean we will uniformly apply these decisions.

Use this as inspo, rather than vibes decision making.

Subscribe now

By day 5, you can likely predict a good or a bad ad

This is one of our biggest learnings. We run seven day tests by default. We do this for a few reasons:

• We run most ads on 7 day click, 1 day view attribution windows and so it makes sense to judge against a window you’re asking

• In-week volatility exists a lot, and that can be true ad by ad, so we like to test for a whole 7 days.

When trying to identify at what point we can predict winners or losers, we looked at a few measures:

• Receiver operating character (ROC) curve
  ”The area under the ROC curve (AUC) represents the probability that the model, if given a randomly chosen positive and negative example, will rank the positive higher than the negative.”

• Correlation coefficient between day X CPA and day 7 CPA
  This looks at the correlation between two variables.

• % of ads that are – at day X – within Y% of the day 7 CPA

1. ROC Curve

First of all on the ROC curve, if we define an ad as good/bad at day 7, which in our analysis we defined as having intervened and turned off or kept on, we can usually see whether it will be a good/bad ad at day 5.

The ROC curve is at 0.98 by day 5, meaning 98% of the time we can predict whether we would make the choice we eventually made at day 7.

💡 ACTION: Potentially intervening earlier in tests meaning either: a 28% reduction in testing budget each week, or a 28% increase in creative testing.

Our next steps are to review these curves on a client by client basis, and then also examine days of the week into the mix. We usually have set creative test launch days which means we have less data here to see how it will do across the week.

2. Correlation coefficient

The next Q is how well do they correlate together.

Correlation coefficient shows how closely two variables correlate.

• At day 3, there is a .79 correlation coefficient: a strong positive correlation

• By day 5, it’s 0.94 – a very strong positive correlation

This shows that the CPAs at day 5 correlate very strongly to the CPAs at day 7 – it doesn’t mean they are the same, but that they align and therefore you can use day 5 CPA as a proxy for day 7 CPA. (All CPAs in our study are accumulative).

3. How close are the CPAs?

This is where there is a bigger difference that isn’t immaterial.

At day 3, we’re not that close at all to the day 7 CPAs: half of the ads are within a wide 30% margin. By day 5, half are closer to 10% of final day 7 CPA.

Subscribe now

What about the longer term view?

The second part of our analysis dug into the longer term.

Making decisions at day 7 is one thing, but how well are we making decisions and what happens to ads afterwards?

This analysis looks at our four biggest spending accounts, and looks at ads which are still live by day 30 (i.e. those which algorithmically have continued to spend).

For each of these accounts, we have successfully managed to scale these accounts by 2-10x over the last 12 months – at agreed efficiency targets.

One brand has a faster stabilisation rate, which maybe implies we’re better at identifying winners that scale. But even so being within 20% of day 30 CPA shows that there’s actually little stabilisation that takes even with those winners.

In fact:

• While only 32% of ‘winning’ ads go on to have a CPA which decreases by day 30.

• This is mostly tightly clustered with these four accounts: 28%, 28%, 31% and 39%.

Overall the median CPA change from day 7 to day 30 is a 19% increase in CPA.

We know that Facebook Ads follows a power curve. But this is a really good example of what you need to see for creative winners to sit comfortably in your campaigns.

It’s also to be expected. Most testing campaigns act as a fractionally smaller part of budget. When ads are served to wider audiences they hit diminishing returns. And so if an ad goes from £500 of testing budget to being put into a campaign spending £5k/day – and it goes on to be a : – that will likely hit more diminishing returns.

💡 You need to allow for increases in CPA in the short term

Spend dynamics in the day 7 to day 30 view

So what actually happens in that day 7 to day 30 view. As we promote winning ads after day 7 they go into campaigns that will allow them to spend more. But, importantly, we intervene less here and the spend is algorithmic rather than human-driven.

This is the stage when good ads start to separate from great ads.

The mean spend multiplier is 22x

The median spend multiplier is 4.4x

This means that most ads scale modestly, but then there is a small number that explode.

The differences between CPA improvers and worseners

Most ads increase in CPA during this period, and for those that do, the mean goes to 29x but median is at 4.82

For the ads that decrease in CPA, the mean is a far lower 7.3x and the median is a lower still 3.7x.

Our takeaway here is that this is for the most part the breakdown effect at play, and not a signal of the auction failing, but just one that ads differ from the isolated position of testing to that of being live in Hero campaigns.

Subscribe now

Closing thoughts

Identifying earlier winners is something we like to think a lot about, but so too is broader understanding of how ads behave after they become winners.

This should be a good logical reminder that a good CPA in testing does not mean CPA will stay the same at scale.

Ultimately, Meta is a volume channel rather than an efficiency channel. Your goal is to find ads that have strong volume potential with affordable change in CPA.

As pointed out in the intro, this data needs analysing yourself. But for those who aren’t in a position to do so, it should help provide some colour to what happens when a test goes live.

Leave a comment

🔗 When you’re ready, here’s how Ballpoint can help you

→ Profitably grow paid social spend from £20k/m → £300k/m
→ Create full funnel, jobs to be done-focused creative: Meta, TikTok, YouTube
→ Improve your conversion rate with landing pages and fully managed CRO
→ Maximise LTV through strategic retention and CRM - not just sending out your emails

Email me – or visit Ballpoint to find out more.

NB: We support brands spending above £20k/month.

❤️‍🔥 Subscribe to our Substack to learn how to grow yourself

… because agencies aren’t for everyone, but our mission is to help all exciting challenger brands succeed and so we give away learnings, advice, how-tos, and reflections on the industry every week here in Early Stage Growth.

Leave a comment

TL;DR

Small spenders on Meta often overcomplicate their account setups. You probably only need one campaign that excludes your existing customers. Focus all your time on creative experimentation.

We audit multiple businesses each week at Ballpoint. A lot of the time, these end up being too small for us to work with and we recommend they manage it in-house…

Read more

Meta Ads has gone through two revolutions over the last five years: one with iOS14.5 and the other with AI. Despite how well documented these are, these are the top 5 Meta lies we find startups still believe

1. Clickthrough rates can teach you how well an ad is doing

Meta has published data disproving this. Nielsen has disproven this. And in our annu…

Read more

Read more

We’d been testing it for the year leading up to it, and the headline conclusions were positive.

Where we were running ASC it was delivering on average a 33% reduction in CPA.

We’d run these tests in a variety of ways including isolated A/B tests using Meta’s inbuilt Experiment to…

Read more

Last week, Sean Frank, the CEO of wallet company Ridge posted some a downbeat end to 2024 for Meta Ads.

In his book, 2024 was as bad as the iOS14 year except today, there are now other opportunities around.

As an agency owner I am also highly exposed to people …

Read more

But no-one, and I mean no-one, has ever asked how Facebook sh…

Read more

One aspect of that is wanting to put some data behind things that often have assumptions to them. Things like thumbstop…

Read more

What happens when all these people go away? Ever since an early boss told a room of us ‘never to bring external factors as an excuse fo…

Read more

When I was head of growth at Thriva, I always had conversations amongst my network of ‘what’s working on Facebook?’ Or ‘how are you approaching horizontal scaling?’ ‘When was the last time you retrained your pixel?’1

When I was in-house I had a belief that there was a single truth that solved all the problems.

It was not long after starting Ballpoint that …

Read more

We first covered this last year when I mostly referenced it in relation to media spend. Then back in April as part of my Creative Diversity post, I shared our formula for how much creative we typically work towards.

Last week, I saw an extra piece of data that adds to this whole puzzle.

Varos is a …

Read more

On quick glance, does everything hold together in a coherent style? Does it feel and look similar? Is there consistency in how you’re approaching creatives?

If the answer is yes, then you are leaving huge amounts of growth on the table.

Read more