Key Takeaways
- Upside analyzed 102 buying signals across 19,459 deals and 205,036 contacts at ten B2B companies and found the same signal can deliver a 15x conversion lift at one company and zero predictive value at another.
- Generic buying signal predictive analytics software scores leads on patterns pulled from someone else's pipeline, which is why the dashboard looks confident and the pipeline doesn't move.
- The fix is not better signals, it's calibration: score every signal against your own closed-won and closed-lost data before you let it route a single lead.
- Start with the signals you already capture in HubSpot or Salesforce, measure lift against your own base rate, and only then decide whether to buy a third-party intent product.
- Off-the-shelf scoring is still useful as a cold-start default, but treat it as a hypothesis to test, not a model to trust.
Upside published research last week that I've been turning over in my head ever since. They looked at 102 buying signals across 19,459 deals and 205,036 contacts at ten B2B companies, and the headline finding is the kind of thing that should make every marketing leader pause: the exact same signal, defined the exact same way, produced a 15x lift in conversion at one company and zero predictive value at another. Same signal, same definition, opposite outcomes.
If you're running a modern revenue stack, you've probably bought into some version of the buying signals story in the last two years. Job change alerts, G2 category views, website visits from target accounts, hiring patterns, tech stack changes, the whole menu. You bought it because a vendor showed you a chart where the signal correlated with pipeline. The uncomfortable thing about the Upside data is that the chart was probably real, it just wasn't yours.
What I want to walk through here is why that happens, what it means for how you should evaluate these tools, and the one shift in how to buy them that falls out of the data.
The base rate is the thing you're actually buying
When a vendor tells you a signal produces a 3x lift, they're telling you something real about the dataset they measured it on. What they're not telling you is what the base rate was, what the sales motion looked like, what the ICP overlap was, or how the signal was defined inside that specific company's data model. All of those things matter more than the signal itself.
Think about a job change alert at a VP of Marketing. At a company selling to new VPs in their first 90 days (say, a CMS or a marketing analytics tool where the new leader is actively re-evaluating the stack), that signal is gold. At a company selling infrastructure that sits three layers below the VP's attention, the same signal is noise. The VP changed jobs. So what. Nobody in that buying committee cares, and the person who actually makes the decision hasn't moved.
This is what the Upside research is really showing. Signals are conditional on a buying process, and buying processes don't generalize across companies even when the signals look identical on paper.
Why generic scoring models feel like they're working when they're not
Here's the trap, and it's a subtle one. You turn on a generic signal product. It starts firing alerts. Your SDRs work the alerts. Some of those alerts convert. You look at conversion on alerted accounts versus non-alerted accounts, and alerted accounts convert better. You conclude the signals work.
Except what you've actually measured is a selection effect. Your SDRs prioritized those accounts, spent more time on them, wrote better emails because the alert gave them a hook. The lift you saw might be entirely about SDR effort reallocation rather than the signal predicting anything at all. What you bought as a prediction engine was really just an attention mechanism.
The way you'd test this properly is a holdout. Take a chunk of accounts that fire the signal and don't work them. Take a chunk that don't fire the signal and work them with the same intensity. Compare conversion. Almost nobody does this because it feels like leaving money on the table, and because the vendor's dashboard already tells you the answer you want to hear.
The Upside data is interesting precisely because it removes the SDR effort variable. They measured signals against closed deal outcomes across a shared methodology at ten companies, and the signals that worked at one company genuinely didn't work at another. That's a real finding, not a selection artifact.
What calibration actually looks like
The holdout is the diagnostic. Calibration is the fix. The move is to stop treating these tools as a scoring model and start treating them as a feature store. The vendor gives you the signals. You build the model.
Practically, that means a few things. First, every signal your vendor provides needs to be joined to your own closed-won and closed-lost deal history in HubSpot or Salesforce. Not at the account level in the abstract, at the specific deal level with timestamps. Did this signal fire before the deal closed, how far before, and did it fire more often on won deals than on lost deals of comparable size and segment.
Second, you need a base rate to compare against. If your overall opportunity-to-close rate is 22%, a signal that produces 25% close rate on alerted accounts is noise. A signal that produces 55% is interesting. A signal that produces 55% but only fires on 11 accounts a quarter is interesting but not operationally useful yet.
Third, and this is the part that surprises people, a lot of the signals that end up mattering most are first-party signals you already own. Someone from a target account looked at your pricing page twice in a week. A champion at a stalled opportunity opened a new LinkedIn role. A customer's parent company got acquired. These are not exotic third-party datasets, they're things sitting in your own systems that nobody has bothered to score.
When this category is the wrong call
Let me be honest about when this whole category is the wrong purchase, because I don't think it always is.
If your pipeline is under a few hundred opportunities a year, you don't have enough data to calibrate anything. The statistical noise swamps the signal, and you'd be better off having your SDRs read ten target accounts a day in depth than chasing alerts. Signal scoring is a volume game, and small-volume teams should buy judgment, not software.
If your sales cycle is long and your sample size of closed deals is thin (say you close 40 deals a year at six-figure ACV), the same problem applies. You cannot build a predictive model on 40 outcomes. You can have a point of view about which signals matter based on talking to customers, and that point of view will beat any generic model a vendor sells you.
And if you haven't gotten your first-party data house in order (clean account records, deal stages that mean something, closed-lost reasons that are actually populated), buying a third-party signal product is building a second floor on a foundation that's cracked. Fix the foundation. The signals will be more useful when you do.
The shift this should cause in how you buy
The thing I want marketing leaders to take away from the Upside research is a change in the question you ask vendors. Stop asking "what lift do your signals produce." Every vendor has a chart. Every chart is real somewhere. The chart is not your company.
Start asking: how do you help me calibrate your signals against my own deal history, and what does your product look like when a signal you provide turns out to not predict anything in my pipeline. If the answer is some version of "it will, trust us," you've learned something important. If the answer is "here's how we backtest against your closed-won data in the first 60 days and here's how we turn off signals that don't earn their keep," that's a different conversation.
What we're doing about it
The move we'd recommend to any client shopping signal tools is to pull the last 18 months of closed-won and closed-lost deals, tag the accounts, and run the vendor's signals retrospectively against that list before signing a contract. Most vendors will do this as part of a pilot if you push. The ones that won't are telling you something.
The pattern that tends to emerge is that two or three signals end up carrying almost all the predictive weight for a given company, and they're often not the signals the vendor leads with in the sales deck. That's fine. You bought a feature store. Use the features that work for you. Turn the rest off. Revisit in a quarter when your product or ICP has shifted and the weights have moved.
None of this is an argument against the category, it's an argument against buying it the way most people are buying it. The Upside data is the clearest evidence I've seen that the lazy version of this purchase, the one where you trust the vendor's model and route leads based on their scores, is actively harmful. The thoughtful version, where you treat their signals as inputs to your own model, is still one of the better bets a B2B marketing team can make right now.
If you want the longer version of how we work through this with clients, our growth blueprint walks through the full approach.
Frequently Asked Questions
What is buying signal predictive analytics software?
Buying signal predictive analytics software is a class of tool that captures behavioral and firmographic events (job changes, web visits, intent topic surges, hiring patterns, tech stack changes) and scores them to predict which accounts are likely to buy. Vendors in the space include 6sense, Demandbase, and UserGems, among others.
Why don't vendor scoring models work across companies?
Because the model was trained on someone else's pipeline. A signal's predictive value depends on your specific buying process, your ICP, your sales motion, and your base rate. Upside's research across ten B2B companies found the same signal can produce a 15x lift at one company and zero predictive value at another, which means the vendor's scoring weights are essentially a guess for your business until they're recalibrated on your data.
How do I calibrate signals to my own pipeline?
Join every signal to your closed-won and closed-lost deal history at the deal level with timestamps. Ask whether the signal fired before the deal closed, how far before, and whether it fired more often on won deals than comparable lost deals. Compare the alerted conversion rate against your base rate, and ignore any signal that doesn't clearly beat it. Most vendors will run this backtest against your CRM data during a pilot if you ask.
Should small teams buy buying signal predictive analytics software?
Probably not yet. If you close fewer than a few hundred opportunities a year, you don't have enough data to calibrate a model, and the statistical noise will swamp any signal. Smaller teams get more value from having reps read a small number of target accounts deeply than from chasing automated alerts.
What's the biggest mistake marketing teams make with intent data?
Measuring conversion on alerted accounts against non-alerted accounts without controlling for SDR effort. Alerted accounts get more attention, better outreach, and faster follow-up, so they convert better regardless of whether the signal itself predicts anything. The lift you're measuring is often the attention mechanism, not the signal.