Most 'you might also like' emails are lazy, and your customers can tell. Here's how to actually earn the second sale.
Upselling and cross-selling are not the same offer wearing different names, and AI only helps once you treat them differently. Segment customers by what they actually did (usage depth, replenishment timing, plan limits) instead of the lazy “they bought once” trigger. Use Klaviyo AI, HubSpot Breeze, or a ChatGPT/Claude drafting workflow to write copy fast, but never skip the human pass for brand voice and honesty about what the product does. Time the send to a real signal, a usage threshold, a replenishment window, a plan overage, not a calendar date. And watch the frequency: the fastest way to lose a repeat customer is to make every single email in their inbox a pitch.
Picture the inbox of anyone who bought a mattress eighteen months ago. They’re still getting “complete your bedroom” emails every few weeks, personalized only by first name, pushing a bed frame they already own or a pillow they already sent back. I’ve pulled enough of these flows apart to know exactly what this is: mail merge wearing a strategy costume. And it’s exactly why so many customers now delete anything a brand sends after the first purchase without opening it.
Let’s get the definitions straight, because most teams use “upsell” and “cross-sell” like they’re interchangeable, and that sloppiness is exactly where the targeting falls apart. An upsell moves a customer up within the thing they already bought: the better blender, the higher SaaS tier, the annual plan instead of monthly. A cross-sell moves them sideways into something related but separate: the filter cartridges for the blender, the integration add-on for the software, the conditioner that matches the shampoo they just reordered. The psychology behind each one is different, and so is the right moment to send it. Mixing them up is the first mistake I see on almost every audit, and no amount of AI horsepower fixes it.
Add AI to this equation and the targeting math stops being a guess, but only if you feed it the right question first. The old default, one segment called “purchased in last 90 days,” treats a gift buyer who will never order again exactly the same as a repeat customer three days from reordering. Both sit inside the same 90-day window; only one of them is an actual signal, and a static segment can’t tell them apart. Klaviyo AI and HubSpot Breeze can now score every customer on usage depth, purchase cadence, and predicted next-order date, and generate the segment plus the matching copy variant for each one in roughly the time it used to take to build that single static list.
That matters because upsell and cross-sell offers only work when they match a real signal. A customer who just bought their third bag of the same coffee is a great cross-sell candidate for a grinder. A customer who hasn’t reordered in four months is not, no matter how good the grinder copy is. AI doesn’t invent this insight. It just makes it cheap enough to act on for every customer instead of just your top 5%.
Here is the segmentation mistake I see most often, on ecommerce and SaaS teams alike: building the entire upsell program around a single trigger, “customer purchased,” and calling it done. That trigger tells you almost nothing about whether this is the right moment, or the right offer, for this specific person.
A better starting point, and the one I keep coming back to, is RFM segmentation (recency, frequency, monetary value), a decades-old retail framework that AI tools have finally made practical to run continuously instead of once a quarter in a spreadsheet. Klaviyo’s segmentation engine, for example, can build live segments on predicted next order date, predicted customer lifetime value, and churn risk, updating automatically as behavior changes instead of relying on static purchase tags.
For SaaS teams, this usually means pulling product usage data, API calls, seats filled, storage used, feature adoption, into whatever tool sends the email. HubSpot’s Smart CRM can hold these as custom properties and trigger workflows directly off them. If your usage data lives in a separate product analytics tool like Amplitude or Mixpanel, you need it flowing into your CRM or ESP before AI segmentation is worth building at all. Skip that step and you’re just personalizing the subject line on a still-generic offer. Customers notice every time.
This is the part teams underinvest in, because it’s invisible next to a slick email design, and the cost of skipping it doesn’t show up until weeks later. Blast an upsell offer to a segment that doesn’t match the signal and you don’t just waste that one send: the low opens and spam complaints it generates get read by mailbox providers as a signal about your sending domain, not just that campaign. That’s how a badly targeted upsell flow quietly drags down inbox placement for your welcome series and your receipt emails too. A well-defined offer to the right 500 people protects deliverability in a way a clever offer to 50,000 never will. It’s also the part where AI genuinely earns its keep at scale, since reviewing usage data on 50,000 accounts by hand was never happening anyway.
Once you know who gets the email and why, the copy itself is where most teams still default to a template that says “customers who bought X also bought Y” and call it personalization. It isn’t. Real personalization references what this specific customer did, not a generic co-purchase pattern pulled from your whole database.
Klaviyo AI (K:AI) can generate subject lines, full campaign copy, and product recommendation blocks that pull dynamically from a customer’s actual purchase and browse history, not a static “best sellers” list. The tell that a recommendation block is running on collaborative filtering alone: it recommends whatever most other customers bought next, not what this customer needs next. That’s how a woman who bought a men’s medium shirt as a gift ends up recommended more men’s mediums, or a customer who already owns the accessory gets pitched that same accessory again because it’s simply the most common add-on across the whole catalog. The model isn’t wrong, it’s answering the question it was actually trained on, aggregate co-purchase frequency, not this one person’s profile, and it’s the fastest way to make an “AI-personalized” email look exactly like the generic one it replaced. Its Smart Sending feature also caps how many emails a single customer receives across flows and campaigns in a given window, which matters more than people expect until you’ve watched five different upsell flows try to fire on the same person in the same week.
HubSpot’s Breeze Content Assistant drafts email copy from a plain-language prompt (“draft an upgrade email for accounts near their seat limit, friendly but direct tone”), and Breeze Intelligence can pull in CRM context, deal stage, plan tier, recent support tickets, so the draft is not generic even on the first pass. For B2B or SaaS teams already living in HubSpot, this keeps the personalization data and the copy generation in one place instead of exporting a CSV to a separate tool.
If your stack does not have built-in AI copy generation, or you want more control over tone, a custom GPT or a Claude project loaded with your brand voice guide and a handful of your best-performing past emails works well too. Feed it a short, specific brief: the segment, the trigger event, three real customer data points (not made-up ones), and what you want the customer to do next. What comes back is a first draft, never a finished email, and treating it as one is where most of the damage happens.
The AI draft is a first draft, full stop. I’ve reviewed enough of these before they went out the door to know the tell instantly: generic superlatives, a claim about the product that’s not quite accurate, a tone that doesn’t match anything else the brand has ever sent. Run a human pass for brand voice and factual accuracy every single time, no exceptions.
One workflow that actually holds up: export the segment’s real behavior data (last purchase, usage stat, plan tier), write a two-sentence brief describing the trigger and the offer, generate three subject line and body variants, then edit the winner by hand before it ever reaches a customer. That’s maybe fifteen minutes of work now for an email that used to eat an hour, and it still gets a human eye before it ships.
Bad timing kills more upsell emails than bad copy ever does, and it isn’t close. Send the pitch too early and it reads as pushy. Wait too long and the customer’s already solved the problem somewhere else, sometimes with a competitor you handed the sale to for free.
The difference between a well-timed trigger and a poorly-timed one comes down to what kind of signal it’s built on. A well-timed trigger fires on a leading indicator, a usage threshold crossing, a predicted reorder date calculated from that customer’s own cycle, something that describes an actual change in their need. A poorly-timed trigger fires on an administrative event, days since purchase, a fixed day-14 email sent to every SKU regardless of what it is, that has no causal link to whether this specific customer needs anything yet.
For consumable or replenishable products, the trigger should track the product’s actual usage cycle, not an arbitrary “7 days after purchase” default. A 30-day supply of vitamins doesn’t need a cross-sell email on day 3. It needs one around day 22 to 25, timed just ahead of when the customer would naturally run low. That’s exactly the kind of pattern AI-driven send-time and flow-timing tools are built to calculate from your own historical reorder data, instead of a guess dressed up as a rule.
The equivalent SaaS trigger skips the calendar entirely. What matters is a usage threshold: an account hitting 80% of its seat limit, a workspace approaching a storage cap, a user who’s adopted three out of five core features and looks like a good candidate for the tier that includes the other two. HubSpot workflows and most modern customer engagement platforms can fire off a custom property crossing a threshold, and that’s a far stronger signal than “it’s been 60 days since signup.”
The honest caveat here: none of this works if your usage data is stale or siloed in a tool nobody connected to your CRM. I’ve seen SaaS teams build a beautiful usage-based trigger on data that updates weekly, so by the time the email sends, the account already downgraded or churned. Get the data pipeline right before you get precious about the copy.
This is what it looks like end to end for a mid-size DTC skincare brand, the kind running Shopify plus Klaviyo, with a customer base that reorders on a fairly predictable cycle.
There’s no single “AI wrote our email” moment in this. It’s six small, boring decisions stacked on top of each other, and that’s honestly what most working AI marketing systems look like once you get past the demo. Less magic, more plumbing.
Swap the skincare brand for a project management SaaS tool and the workflow barely changes shape. The data pull becomes accounts nearing 80% of their seat limit in the last 30 days, filtered to exclude anyone already in an active sales conversation. The segment split separates accounts that are clearly outgrowing their plan (seats and storage both climbing) from accounts using one power feature heavily but nowhere near their seat cap, a cross-sell candidate for an add-on module instead of a plan upgrade.
The copy draft, brief, and human edit steps stay identical, only the trigger data changes: usage percentage instead of days since purchase. The send window becomes the moment the threshold crosses, not a calendar date, and the suppression check makes sure a customer already talking to a sales rep about expansion does not also get an automated upgrade nudge from marketing the same week. Same six steps, different data source.
Let’s be honest about the failure mode here, because it’s real, and I’ve watched it happen on teams that had every right tool in place. AI makes sending more emails to more finely sliced segments almost free. That capability alone isn’t a strategy, and used without restraint, it turns into exactly the kind of over-personalized, over-frequent pitching that makes customers unsubscribe from a brand they actually liked.
The tell is usually frequency, not content. One well-timed, well-targeted upsell email a customer half-expects reads as helpful. Three upsell-flavored emails in the same week, even if each one is individually well-written and accurately personalized, reads as a brand that only sees you as a wallet. Smart Sending-style frequency caps exist for exactly this reason. Use them, and set the cap lower than feels aggressive to your CFO.
The second tell is honesty about fit. An AI model trained purely to maximize click rate will happily recommend a product a customer doesn’t actually need, because a mismatched but curiosity-inducing recommendation still gets clicked, it just doesn’t convert, and the model reads that click as a win. It doesn’t know the difference between a click and a fit, it just optimizes for whichever metric you handed it, and click rate was never a measure of whether the offer was right. If the accessory genuinely won’t work with what they bought, or the upgrade tier genuinely won’t solve their actual usage pattern, don’t send the email just because the segmentation logic technically qualifies them. That’s a short-term win that costs you the long-term relationship, and customers remember being sold something wrong far longer than they remember a good recommendation.
My rule of thumb: if you would feel a little embarrassed explaining this specific offer, to this specific customer, on a phone call, do not send it as an email either. AI removes the friction that used to force a human to think twice before hitting send on a low-quality offer. That friction was doing more work than people gave it credit for.
This doesn’t mean slowing down on AI-assisted marketing. It means building the guardrails, frequency caps, fit checks, a human review pass, at the same time you build the automation, instead of bolting them on after customers start complaining. Every team I’ve seen get real results from AI upsell campaigns sends noticeably fewer emails than their competitors, and every one of those emails is better matched.
There’s also a quieter cost to getting this wrong that never shows up in an unsubscribe count. Brand trust erodes slowly, one slightly-off email at a time, long before anyone actually clicks unsubscribe. A customer who gets three mismatched offers in a row usually doesn’t leave your list right away. They just start assuming your emails aren’t worth opening, and that assumption is much harder to undo than a bad open rate.
An upsell moves a customer up within the category they already bought (a bigger size, a higher plan, an annual instead of monthly subscription). A cross-sell moves them sideways into a related but separate product or feature (accessories, an add-on, a complementary item). Each needs its own timing and its own copy, so lumping them into one generic “you might also like” email is usually the first thing worth fixing.
If you’re already on Klaviyo, use Klaviyo AI, it pulls directly from purchase and browse history so the drafts start more personalized. HubSpot users get similar value from Breeze Content Assistant paired with Breeze Intelligence for CRM context. If your stack doesn’t have built-in AI copy tools, a ChatGPT or Claude workflow loaded with your brand voice and real customer data points works well, but always add a human editing pass before sending.
It depends on the product, not a fixed rule. Consumables should trigger near the predicted reorder date based on past cycle length, not a flat number of days. Durable goods do better with a cross-sell in the first 1-2 weeks post-purchase, while satisfaction is high. For SaaS, skip dates entirely and trigger on usage thresholds, like an account nearing its seat or storage limit.
Yes, and it’s one of the more common mistakes I see. AI makes it cheap to send more finely targeted emails more often, which teams sometimes mistake for more effective marketing. In practice, more than one AI-personalized upsell touch inside the same 7-day window reads as pushy, not helpful, even when each individual email is well-targeted and accurately personalized, and it drives unsubscribes. Use frequency caps (Klaviyo’s Smart Sending is one option) and keep total send volume per customer sane.
You need enough purchase or usage history for patterns to show up, which for most ecommerce brands is a few hundred repeat customers, and for SaaS is a few months of usage data per account. Below that, AI segmentation mostly adds complexity without much signal to work from. Start with simple RFM-style segments by hand, and layer in AI-driven predictive segmentation once you have enough repeat behavior to train it on.
This guide draws on Klaviyo’s 2026 Email Marketing Benchmarks by Industry report, HubSpot’s 2025 sales statistics research, and current product documentation for Klaviyo AI, HubSpot Breeze, and Shopify’s post-purchase upsell app ecosystem (including Rebuy). All statistics and platform features were verified against live source pages before publishing. Sources are linked below.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.