Predicted customer lifetime value is one of the most quoted numbers in ecommerce email, and one of the least interrogated.
It appears on a profile. It has a currency symbol in front of it. It gets used to build VIP segments, to justify spend, occasionally to set acquisition targets. And in a large share of accounts it's either blank or built on evidence so thin it shouldn't be driving anything.
Nobody checks. It looks like a fact, so it gets treated as one.
What has to be true before the model will run
Klaviyo sets the thresholds out in its help centre, which is more than can be said for most of the people quoting the output.
For predictive analytics, an account needs at least 180 days of order history, orders placed within the last 30 days, and a minimum of 500 customers who have placed an order.
That last one is where most brands fall over. It means 500 people who have actually bought, not 500 profiles on your list.
A brand with 60,000 subscribers and 400 purchasers doesn't qualify. Neither does a brand that launched nine months ago, nor one that had a quiet month and has no recent orders feeding the model.
Klaviyo's RFM reporting has its own gate on top of that. The RFM analysis report sits inside Advanced KDP or Marketing Analytics, which are paid additions to the standard platform, and it also requires at least 500 customers who have ordered.
So there are two separate failure modes: you don't have the data, or you don't have the subscription. Either way the output isn't there, and building strategy on top of a blank field is how accounts end up with a VIP segment that contains eleven people.
The failure that's harder to spot
Not qualifying is the easy case. You look, the field is empty, you move on.
The dangerous case is qualifying by a whisker.
At 520 purchasing customers with 190 days of history, the model will produce numbers. They'll look exactly like the numbers a brand with 50,000 purchasers and four years of history gets. Same field, same formatting, same confident currency symbol.
They will not be equally reliable, and nothing in the interface tells you that.
There's a related problem with any model learning from history: it assumes the future resembles the past. Several ordinary business events break that assumption.
You ran a heavy discount period, which pulled purchases forward and compressed the intervals the model learned from. You changed your product range, so the consumption pattern behind your repeat behaviour is different now. You shifted acquisition channel, and paid social customers behave nothing like the organic ones the model was trained on. You put your prices up.
Every one of those makes historical patterns a worse guide, and none of them flag anything in the platform.
A prediction is a statement about the past dressed up as a statement about the future. Ask what the past looked like before you trust it.
What to do instead when you don't qualify
The good news is that everything predictive analytics gives you can be approximated with data you already have and a spreadsheet. It's less elegant. It's often more defensible.
Instead of predicted lifetime value, calculate historical value per customer, banded. Take everyone who's ordered, sum their spend, and split into quintiles. Your top 20% by actual spend is a real segment based on real money. It's backward looking rather than predictive, but it's true, which predicted CLV on 520 customers is not.
Instead of predicted churn risk, use your median purchase latency. Work out the median gap between orders. Anyone sitting at two to three times that gap with no order is drifting. Anyone past four times it is likely gone. This is a churn model. It's just one you built yourself out of arithmetic rather than one you're renting.
Instead of predicted next order date, use the same median applied to each customer's last order date. Less sophisticated than what Klaviyo does, considerably better than a blanket 30 day delay.
Instead of the RFM report, build RFM manually with segment conditions. Recency, frequency and total spend are all standard segment criteria in the base platform. You can construct the groups by hand without the paid add on.
None of this needs a data team. It needs an export, a couple of hours, and somebody willing to look.
What a bad prediction actually costs
Care about this because thin predictions produce expensive decisions.
Picture a predicted CLV built on 520 purchasers that overstates value for one group of customers. You build a VIP segment off it, then start giving that segment early access, free shipping and higher touch service. You're spending on people who aren't worth it and, worse, you're not spending it on people who are.
Or the churn risk score flags a group as at risk who simply have a longer natural purchase cycle. You fire a winback at them with 20% off. Some of them were going to buy anyway, at full price. You've just paid people to do something they were already doing, and taught them that waiting produces a discount.
That second one compounds. Discount a customer once when they didn't need it and you've moved their expectation permanently.
A thin prediction does not do nothing. It confidently points you at the wrong customers. A blank field at least tells you where you stand.
Where this shows up in a pitch deck
There's a category of agency that treats predictive fields as a differentiator. It gets put in the pitch deck. There's a slide about AI powered lifetime value segmentation, and it looks impressive next to a competitor talking about welcome flows.
Then the account gets built, the fields are empty or thin, and the segmentation quietly reverts to "purchased more than twice", which is what it was always going to be. Nobody mentions it again.
The honest position is that predictive analytics are genuinely useful once you're past a reasonable scale, and largely decorative below it. Knowing which side of that line you're on is worth more than any segment you'd build from the output.
How to audit it in ten minutes
Open a handful of profiles across different segments and check whether the predictive fields are populated. If they're blank on most, you have your answer and the rest of this doesn't apply.
If they are populated, the real work starts. List everything that's changed in the business over the period the model learned from. Price rises. Range changes. A heavy promotional quarter. A shift in acquisition channel. A supply issue that suppressed orders for six weeks.
The longer that list, the less weight the numbers deserve, and none of it appears anywhere in the interface.
Then look at what's currently built on top of those fields. Usually it's a VIP segment, a churn flow, and sometimes an acquisition target.
If the list of changes is long and the dependencies are load bearing, you've been making decisions on a number that describes a business you were running two years ago.
There's no shame in a model not applying to you. The manual alternatives are perfectly good and often more defensible.
Go and open five profiles. Ten minutes tells you whether any of this applies to your account.