Abandoned cart recovery: what AI can and cannot fix

Every ecommerce tool sold in the last decade has quoted the same statistic at you, and it is roughly true. Baymard's compilation of 50 cart abandonment studies puts the average documented rate at 70.22 percent, with the individual studies spread between 55 percent and 84.27 percent. Seven carts in ten do not become orders.

What the tools do not quote is the sentence underneath it, and the sentence underneath it changes the entire economics of what you should build. Baymard is explicit that a large portion of those abandonments are a natural consequence of how people browse ecommerce sites, and it puts the share of US online shoppers who have abandoned because they were just browsing at 42 percent. You are not looking at seven lost sales. You are looking at three or four browsers, a couple of comparison shoppers and perhaps one genuine loss.

What does the 70 percent figure actually mean for a small shop?

It means your recoverable pool is far smaller than your dashboard implies, and that has a direct consequence for how much effort a recovery sequence deserves. If most of the abandonment is window shopping, the marginal return on a fourth reminder email is not small, it is negative, because the cost of that email is paid in deliverability across your whole list.

It also means the benchmark everybody compares themselves against is meaningless without knowing the traffic mix. A shop with heavy social traffic will sit far above 70 percent, because social traffic browses. A shop where most visitors arrive from a branded search will sit below it. Neither number tells you anything about how good the checkout is, which is the thing you actually wanted to know.

The useful metric is not the abandonment rate. It is the abandonment rate among people who reached the payment step, because reaching that step is the closest thing you have to a declaration of intent. That figure is usually a fraction of the headline and it moves when you fix something real.

Which abandonments are recoverable at all?

Baymard's list of reasons, once browsing is excluded, is the most useful thing in the whole subject. Extra costs at 40 percent lead it, followed by delivery being too slow at 20 percent, not trusting the site with card details at 19 percent, being asked to create an account at 18 percent, and a checkout that felt too complicated or too long at 17 percent. Website crashes and errors also sit at 17 percent, an unsatisfactory return policy at 13 percent, no visible order total before checkout at 12 percent, and a declined card at 10 percent.

Now sort those by whether an email can do anything about them. That sort is the piece of work no vendor will do for you, because the answer is unflattering to the product.

Two rows out of seven. That is the honest size of the email opportunity, and it is why shops that install a recovery app see a jump in the first month and nothing afterwards. The first month catches the backlog of errors and declines. There is no second backlog.

What can a model actually do here?

Three jobs, and none of them is writing a better subject line.

The first is classification. You have a cart, a session and a customer history, and the question is which of the rows above this abandonment belongs to. Some of that is deterministic: a declined payment leaves a record, an error leaves a log line. The rest is inference from behaviour, and it is exactly the sort of messy pattern matching that a model does better than a rule. Getting this right means you stop sending the same reminder to a browser and to somebody whose card failed, which are opposite situations.

The second is timing. The optimal delay is not a constant, and every guide that tells you one hour is quoting an average across shops that have nothing in common. For a considered purchase the useful email arrives days later; for a consumable it arrives the same evening. A model that learns your own repurchase and return timings, which is the same machinery described in the piece on predicting repeat purchases in a small shop, has a real edge over a fixed schedule.

The third is suppression, which nobody sells and everybody needs. The most valuable thing a model can tell you is which addresses not to email: the person who abandons weekly and never buys, the one who already bought the same item elsewhere in your catalogue, the one whose engagement has collapsed and whose next complaint will cost you inbox placement. Suppression protects the asset that the whole channel depends on.

When does a recovery email become a legal problem?

Sooner than most sellers expect, because the message sits in an awkward category. It is not a transactional email. The order does not exist. Under the American rules, the FTC's CAN-SPAM compliance guide distinguishes commercial messages, which promote a product or service, from transactional or relationship messages, which confirm a transaction, provide safety information, notify of a relationship change, deal with employment matters, or deliver goods the recipient already agreed to receive. A prompt to complete a purchase you never made is not on that second list.

Treating it as commercial email is the safe reading, and the requirements that follow are concrete. Accurate header information. A subject line that reflects the content. A valid physical postal address in the message. A clear explanation of how to opt out, honoured within 10 business days. The guide also notes you remain responsible for the compliance of whichever company handles your email marketing, which matters when the recovery sequence is running inside a tool you bought rather than one you built.

In Europe and the United Kingdom the question turns on consent, and the exception that most abandoned cart sequences rely on is narrower than people assume: it covers marketing to your own existing customers about similar products, where you offered a way to opt out when you collected the address and in every message since. An address typed into a checkout that was never completed is a genuinely contested case, and the conservative position, which is also the one that produces fewer complaints, is to treat an abandoned cart address as marketing consent only if the person actively gave it. The wider question of what personalisation you may run on customer data is set out in our piece on where consent limits ecommerce personalisation.

What does the inbox think of your recovery sequence?

Less of it than you do, and this is where the real risk sits for a small sender. Google's sender guidelines for bulk email ask senders to keep the spam rate reported in Postmaster Tools below 0.30 percent, with a recommendation to stay under 0.10 percent, and require marketing messages to support one click unsubscribe with a visible unsubscribe link in the body. Bulk senders, defined there as those sending more than 5,000 messages a day to Gmail accounts, also need SPF, DKIM and a DMARC policy in place.

A cart recovery sequence is unusually good at generating complaints, because it goes to people with the weakest relationship to you, at a moment when they have already decided not to buy, sometimes three times. A shop below the bulk threshold is not exempt from the consequences, only from the formal requirement: the complaint rate still shapes where every one of your emails lands, including the order confirmations that matter. Our notes on how AI spam filters read a small sender go through the mechanics of that.

Should the email carry a discount code?

Usually not, and the 40 percent figure is the reason people reach for one anyway. If extra costs are the leading objection, a discount does answer the objection, which makes it the most tempting lever on the board. It is also the one that costs you twice.

The first cost is obvious: margin, on an order you might have won without it. The second is slower and worse. A shop that reliably emails a code an hour after abandonment teaches its repeat customers that abandoning is how you get the code, and the behaviour spreads through exactly the segment you least want to train, the people who buy often enough to notice a pattern. You end up funding a discount programme you never decided to run.

Where a code does make sense is as a one off recovery on a specific failure, and only when the failure was yours. A checkout that crashed, a payment integration that rejected a valid card, an item that showed the wrong postage. In those cases the code is an apology with a purpose, and it is not conditional on abandoning, it is conditional on something going wrong.

If you do use one, the arithmetic worth doing first is not the conversion lift. It is the break even: at your margin, what fraction of the discounted orders had to be genuinely incremental for the campaign to have made money. For most small shops with a 10 percent code that fraction is high enough to be uncomfortable, and the exercise takes ten minutes on the back of an envelope.

Does SMS or a push notification change the arithmetic?

It changes the consent question much more than it changes the response rate, and small shops routinely get this the wrong way round. A text message reaches a device the person carries, at a moment they did not choose, about a purchase they declined to make. The tolerance for that is far lower than for an email that waits in a folder, and the complaint arrives as a lost customer rather than as a spam report you can see.

The regulatory position is also stricter in most markets, since text marketing generally requires clearer consent than email and the exceptions are narrower. If you are already unsure whether an abandoned cart address gives you an email basis, it almost certainly does not give you a texting one.

A sequence that works at 200 orders a month

1. Message one, within the hour, only for errors and declines. Say what happened. A card was declined, or the page failed. Include a link that restores the cart. This is the message with a real job.
2. Message two, the next day, only for carts above your average order value. One reminder, no discount, no countdown. If the objection was postage, address postage in the copy rather than pretending the objection was forgetfulness.
3. Stop. There is no message three for a first time visitor. The third email in these sequences is where the complaint rate comes from, and it buys a rounding error in revenue.
4. For known customers only, a fourth path exists. Someone who has bought twice and abandoned once is a different situation, with a different legal basis and a different tolerance. Segment them out and treat them properly.

Does personalisation beat fixing the checkout?

Not remotely, and the numbers above make the case without any argument from us. Four of the top five reasons people abandon are properties of your checkout, not properties of your email. Extra costs, delivery speed, a forced account and an over long form are all decided before anyone leaves, and they are all cheaper to fix once than to compensate for on every order forever.

There is a version of this that is genuinely hard to hear if you have just bought a recovery tool. The tool works. It just works on the smallest slice of the problem, and the reason it feels effective is that its revenue is attributed and the checkout fix is not. Nobody sends you a monthly report saying that showing shipping cost on the product page earned 4,000 pounds. The email tool absolutely sends you that report.

If the checkout itself is what needs changing and you are stuck inside a platform that will not let you change it, that is a build question rather than a marketing one, and it is the reason our AI ecommerce store builder hands you the code for the checkout rather than renting it to you. A field you can delete is worth more than a template you can theme.

The version we would run

Measure abandonment at the payment step rather than at the cart, because that is the number that responds to work. Fix the three cheapest checkout objections first: show delivery cost before the final step, offer a guest checkout, and cut any field you do not use. Then run two recovery emails, aimed at the two reasons an email can address, and put the effort you were going to spend on a five message sequence into suppression instead.

The 70.22 percent will barely move, because it never does and because most of it was never yours. What moves is the share of people who reached the payment page and left, and that is the only cart number worth putting on a dashboard. The way shoppers read the page before they get that far matters too, which is the subject of our piece on what an AI shopper can and cannot read on a product page.