How B2B Marketers Use AI-Powered Recommendation Engines for Upsells

Lucas: You sign a seven-figure deal with an enterprise customer, the implementation goes smoothly, and then — nothing. No expansion, no upsell, just a flat line. That's the trap a lot of B2B marketers fall into. They're great at landing the whale, but terrible at fishing in the same pond afterward. Luna: Right, because the traditional playbook is all about new logo acquisition. But the real money, especially in SaaS, is in net revenue retention — getting existing accounts to buy more. Lucas: Exactly. And the most scalable way to do that, especially if you've got hundreds or thousands of accounts, is an ai powered recommendation engine. Not the Netflix-style 'you might also like' — I mean a system trained specifically on B2B purchase behavior, contract data, and usage signals. Luna: Before we dive into the mechanics — if these marketing conversations have sparked something you've actually used at work, and you want to keep the show ad-free and independent, you can support us at buy me a coffee dot com slash fexingo. It's just a small way to help us keep digging into topics like this. Lucas: Yeah, we really appreciate that. So — let's talk about how these engines actually work in a B2B context. The two main approaches are collaborative filtering and content-based filtering. Collaborative filtering looks at what similar accounts bought. If a manufacturing company of a certain size and industry bought your premium analytics module, the engine recommends that module to other manufacturing accounts with similar profiles. Luna: So it's like 'accounts like you also purchased.' That makes sense. But what about content-based filtering? Lucas: Content-based filtering looks at what a specific account already has and recommends complementary products. If they bought your data pipeline tool, the engine suggests your visualization layer. It's based on product attributes and historical bundle patterns. A lot of enterprise platforms — Salesforce, Adobe, AWS — have built this into their ecosystems. Luna: Let's talk about AWS specifically. They have this recommendation engine for their marketplace. When a customer provisions a certain service, like EC2, the engine suggests related services like load balancers or auto-scaling. But that's all based on infrastructure usage — how do you get that data into the system? Lucas: That's the critical part. The recommendation engine is only as good as the data you feed it. In B2B, you need more than just purchase history. You need usage data from the product — API calls, active users, storage consumed. You need contract metadata — renewal dates, seat counts, contract value. And you need unstructured data from support tickets and meeting notes. For example, if a customer's support tickets frequently mention 'scalability issues,' that's a signal they might need a premium tier. Luna: So it's about connecting CRM data with product usage data and support interactions. That's a pretty heavy integration project. Lucas: It is, but the payoff is huge. Adobe does this well with their Experience Cloud. They use a recommendation engine that analyzes which marketing automation tools a customer uses and suggests complementary products like analytics or audience manager. They've publicly said that customers who engage with their recommendations have a thirty percent higher attach rate — meaning they buy more products within the suite. Luna: Thirty percent higher attach rate — that's massive. But not every B2B company has the resources of Adobe. How can a mid-market company start? Lucas: Start small. Pick one product line and one customer segment. Build a simple model using just two data points: what they bought and what similar accounts bought. That's collaborative filtering at its most basic. You can implement it with a lookup table in your CRM, no machine learning required. Test it on a cohort of fifty accounts and measure the uplift in upsell conversion. Luna: So it's about proving the concept with a minimal viable version before investing in a full-blown AI infrastructure. Lucas: Exactly. And once you have the data, you need to think about the user experience. Where does the recommendation appear? In the product itself? In the customer portal? In a quarterly business review deck? The best B2B implementations I've seen embed recommendations into the product UI. For example, a project management tool might show a banner that says 'Your team has used five hundred automations this month — upgrade to the business plan for unlimited automations.' That's contextual and timely. Luna: Context is everything. If you recommend something at the wrong time — like during a contract negotiation — it feels pushy. But if you tie it to a usage milestone, it feels helpful. Lucas: Right. Timing is a feature of the recommendation engine itself. You can train the model to predict the optimal moment to present a recommendation based on historical patterns. For instance, many companies see a spike in upsell conversions within thirty days of a customer hitting eighty percent of their storage limit. The engine can flag that account and trigger a recommendation. Luna: Let's talk about measuring effectiveness. What metrics should a B2B marketer track? Lucas: The primary metric is incremental revenue attributed to the recommendation engine. You need a control group — accounts that don't see recommendations — and a test group. Then you measure the difference in upsell revenue over a quarter. Secondary metrics include recommendation click-through rate, conversion rate, and average deal size. And a leading indicator is the 'product affinity score' — how many complementary products an account has. If that score is trending up, net revenue retention will follow. Luna: So you're basically building a lead scoring model but for expansion opportunities instead of new logos. Lucas: Exactly. It's the same logic. But there's a pitfall: recommendation fatigue. If you bombard a customer with too many suggestions, they start ignoring them. One approach is to limit recommendations to one per session or one per week. Another is to use reinforcement learning — the engine learns which recommendations get ignored and stops showing them. HubSpot does this well with their 'suggested tools' feature in the CRM. Luna: What about data privacy? If you're using usage data and support interactions, you're treading into sensitive territory. Lucas: That's a real concern. You need to be transparent with customers about how their data is used. Ideally, the recommendation engine uses anonymized aggregate data, not individual user behavior. And you should always give customers an opt-out. Most enterprise buyers are okay with recommendations if they're clearly framed as value-added — 'here's how other customers like you get more out of the product' — not as surveillance. Luna: Any examples of companies that got this wrong? Lucas: I recall a case where a large software vendor deployed a recommendation engine that suggested premium add-ons based on support ticket keywords. Customers felt like their problems were being monetized. The backlash was strong enough that they had to redesign the entire system. The lesson is: recommendations should feel like a service, not a sales pitch. Luna: So the human element still matters. The engine can surface opportunities, but the sales team needs to deliver the recommendation with context and empathy. Lucas: Absolutely. The best B2B recommendation engines are decision support tools for sales reps, not automated upsell machines. They generate a prioritized list of expansion opportunities with a 'why now' rationale. The rep takes that into the conversation. One company I worked with saw a forty percent increase in upsell meeting acceptance rates when reps used ai generated talking points versus generic pitches. Luna: That's a huge leap. But let's talk about long-term maintenance. Once you've built the engine, how do you keep it from getting stale? Lucas: Model drift is real. Customer behavior changes, products get updated, market conditions shift. You need to retrain the model on a regular cadence — quarterly is a good baseline. Also monitor the recommendation conversion rate over time. If it starts declining, something is off. It could be data quality, a shift in market preference, or a product change that broke the recommendation logic. Luna: So it's not a set-and-forget tool. It requires ongoing investment. Lucas: It does, but the ROI can be enormous. If you increase net revenue retention by just five percent, that compounds significantly over a customer's lifetime. And in a world where acquiring new enterprise accounts is getting more expensive, expanding existing ones is one of the highest-leverage moves a B2B marketer can make. Luna: I'm curious about the next frontier. Where do you see this technology going in the next year or two? Lucas: I think we'll see more integration with generative AI. Instead of just listing recommendations, the engine will write a personalized email or in-app message explaining why a specific product is relevant. Adobe is already experimenting with this — their AI assistant can draft a recommendation based on account history. The other trend is multi-channel orchestration: the engine triggers a recommendation in the product, then follows up with an email, then surfaces a case study in the customer portal. It creates a coordinated expansion campaign. Luna: That sounds almost like an ABM play for existing accounts. Targeted, personalized, multi-touch. Lucas: Exactly. It's ABM applied to the install base. And the same principles apply — tiering accounts based on expansion potential, building segmented playbooks, measuring engagement and conversion. The only difference is you already have a relationship, so the bar for relevance is even higher. Luna: So if our listeners take one thing away from this episode, what should it be? Lucas: Start with the data you already have. You don't need a massive AI infrastructure to begin. Pick a high-value segment, build a simple collaborative filtering model using purchase history, and test it. The insights you get from just that first iteration will likely reveal more expansion opportunities than you expected. The technology is the enabler, but the real unlock is shifting your mindset from 'land and forget' to 'land and expand.'