Four cross-source questions your growth team can't answer and the fifth one your AI tools just created
TL;DR — A cross-source question needs two or more systems joined together before it has an answer. Four of them decide many growth decisions: LTV by acquisition channel, retention by onboarding cohort, churn signal by support category, and expansion by product adoption tier. Your team can answer each one partway, inside one tool. AI tooling is adding a conclusions problem: when the product AI and the CS AI come to different conclusions, which one is right?
A cross-source question is one that no single system can answer, because every system involved holds only some of the input. Most growth teams can partially answer each of the four cross-source questions below inside one tool. None can answer any of them completely, and the AI tooling most teams bought this year is making the list longer. That last part is new: a year ago this post would have had four questions and no fifth problem.
What are the four cross-source questions a growth team should be able to answer?
Four questions describe the core of what a growth team does, and not one of them lives in a single system. I call these the Four Cross-Source Questions. Each one resolves to a cross-source metric — a number that cannot be computed inside any single system of record. The question is how a growth team phrases it before anyone tries to build it. Each is defined by the joined data it requires, not by the metric it produces:
LTV by acquisition channel — ad platforms + billing + product
Retention by onboarding cohort — product + CRM + billing
Churn signal by support category — support desk + billing + product
Expansion by product adoption tier — product + billing + CRM
Nothing on the right-hand side is exotic. Every company already runs these systems, and you can audit your own data stack against that list in an afternoon. No company assigns responsibility for the join itself by default, which is why these four get answered by whoever has time rather than by whoever has the data.
Gartner’s analysis of AI-ready data found 63% of organizations either lack the right data management practices for AI or are unsure whether they have them, from a Q3 2024 survey of 248 data management leaders. The four questions below need exactly those practices. The gap surfaces as unanswerable questions long before it surfaces as a failed AI project.
In my own market research work, most of the time goes to assembling inputs that were never designed to be joined, and the analysis is the short part at the end. The reconciliation is not a step before the work. It is most of the work, and no line in any budget describes it that way.
What is LTV by acquisition channel and why can’t most tools produce it?
Lifetime value by acquisition channel is the total revenue a customer produces over their whole relationship, attributed back to the channel that acquired them. Most tools can’t produce it because the channel attribution lives in the ad platforms and the CRM, the revenue lives in billing, and the retention behavior that separates a good channel from a bad one lives in the product.
The usual objection is that blended LTV is close enough. In my experience, it isn’t: blended LTV is not merely less precise than the per-channel figure — it moves in the wrong direction when channel mix shifts. Add a high-volume, low-retention channel and blended LTV falls even if every individual channel improved. The number reports a decline that did not happen anywhere in the business.
A consumer brand’s data stack shows these fault lines under different names. A DTC subscription brand runs paid social, affiliate and organic search. All three deliver first orders at a similar cost. Only billing knows which cohort is still paying in month nine, and only the product data knows which of them ever hit the usage that predicts month nine. The channel report says the three are equivalent. They are not, and no single system in that stack contains the disagreement.
What is retention by onboarding cohort?
Retention by onboarding cohort is the survival rate of customers grouped by when and how they were onboarded, rather than by when they signed. The question separates a product problem from an activation problem, and answering it needs three systems: the product for what they did in the first X days, the CRM for which onboarding track they were on, and billing for whether they are still paying.
Group by signup month and you get a chart. Group by onboarding path and you usually get a finding. That distinction shows up again in the metrics a Series C round gates on, where retention alone comes out of billing but retention by cohort does not come out of anywhere.
How do you find churn signals across data sources?
You find them by looking where the churn number isn’t. Billing knows a customer left and on what date. It has no idea why. The reasons are in churn signals sitting in support data: ticket volume, category, escalation, time-to-resolution. They are also in the product usage that fell off before anyone filed a ticket.
The join is the finding. A support category with average ticket volume but above-average churn ninety days later is invisible in the support tool, which has no churn data, and invisible in billing, which has no ticket categories. Neither system is broken, the finding simply has no home.
Why is your AI stack making this list longer?
Because each AI tool now produces its own conclusions, and conclusions are harder to reconcile than numbers. Another tool, another silo was the pattern when those tools produced numbers.
IBM’s findings on 2026 data trends puts the failure one level down. Edward Calvesbert, VP of Product Management for watsonx.data, argues AI pilots stall before production because the data underneath sits trapped across silos, missing the structure, metadata and governance agents need. Gartner sizes that gap at 60% of AI projects abandoned for want of AI-ready data.
I want to push that one level further, into territory IBM does not claim. The silos are no longer only in the data. They are forming in the conclusions. Fragmented data is a joining problem, and joining is a solved discipline with known tools. Fragmented intelligence is an adjudication problem: when your support AI and your product AI reach different conclusions about the same customer segment, no amount of pipeline work resolves it, because neither output is a number you can reconcile. Somebody has to decide which conclusion the company acts on. In most organizations that role does not exist: McKinsey’s 2026 AI trust survey puts only about a third of organizations at a mature level for governance and agentic controls, across roughly 500 organizations surveyed between December 2025 and January 2026. So the conclusion that wins is the one whose owner is in the room.
Adding an insight-generating tool to a stack with no adjudication layer adds a fifth cross-source question rather than answering the first four. That is what a customer intelligence layer is for — a governed place where the join and the judgement both live, instead of being reassembled under deadline by whoever is free.
What does this framework not cover?
The conventional answer to all four questions is a modeled warehouse: land every source in Snowflake or BigQuery, define metrics once in a semantic layer, and the joins stop being ad hoc. That answer is correct, and a team that has already done it can answer all four questions today.
The Four Cross-Source Questions describe teams that have not done it, which is most companies below roughly $50M in revenue and a fair number above. The boundary condition is ownership. If a data team owns the semantic definitions and is accountable for them, the join has an owner and this framework has little to tell you. The exception runs the other way too. A warehouse with no owner for cross-source definitions produces the same disagreement as no warehouse at all, more expensively. A dashboard displays what it is given, and so does a semantic layer nobody maintains.
—Steven Rencher, Founder of Monadux
Frequently asked questions
What are the four cross-source questions a growth team should be able to answer?
LTV by acquisition channel, retention by onboarding cohort, churn signal by support category, and expansion by product adoption tier. Each one requires joining two or three systems — typically some combination of ad platforms, CRM, billing, support desk and product analytics. A team can answer every one of them partially inside a single tool and none of them completely.
What is LTV by acquisition channel and why can’t most tools produce it?
LTV by acquisition channel is the revenue a customer generates over their full relationship, attributed to the channel that acquired them. Tools can’t produce it because channel attribution sits in the ad platforms and CRM while revenue sits in billing and retention behaviour sits in the product. Blended LTV is not a safe substitute: it moves in the wrong direction when channel mix shifts.
What is retention by onboarding cohort?
Retention by onboarding cohort is the survival rate of customers grouped by their onboarding path and timing instead of their signup date. It requires the product for first-fortnight behavior, the CRM for onboarding track, and billing for payment status. Grouping by signup month produces a chart, and grouping by onboarding path produces a finding.
How do you find churn signals across data sources?
By joining support data to billing outcomes rather than reading either alone. Billing records that a customer left and when. Support records ticket category, volume and escalation before they left. The signal is a support category with ordinary ticket volume and above-average churn ninety days later, which is invisible in both systems separately.
What happens when two AI tools reach different conclusions?
Somebody has to decide which conclusion the company acts on, and in most organizations no one holds that job. Each AI tool in a growth stack now produces its own conclusions, and conclusions are harder to reconcile than numbers. Fragmented data is a joining problem with known tools and a known discipline. Fragmented intelligence is an adjudication problem: when a support AI and a product AI disagree about one customer segment, no pipeline work settles it, because neither output is a number you can join.
How does a customer intelligence layer solve conclusion silos?
By giving the join and the judgement one governed home. A customer intelligence layer holds the cross-source definitions that make two tools comparable in the first place, and it names who decides when they disagree. That turns adjudication into an owned role with a defined process, so the conclusion a company acts on stops depending on who is in the room. Adding an insight-generating tool to a stack without that layer produces a fifth cross-source question on top of the first four.