Why Your Growth Target Is Failing: Moving Marketing Forecasting From Linear Trendlines to Business Inputs

Most marketing forecasting draws a linear trendline through past data and calls it a plan. The line projects the outcome a business produced last year without modelling any of the mechanics that produced it. Churn moves, retention curves bend, acquisition costs drift by channel, and the slope holds regardless. Your growth target does not fail at the end of the year. It fails on the day it is set.

• A linear trendline projects a result; a business-input model calculates one from the drivers that generate it.
• Churn, lifetime value and acquisition cost variance are the unit economics a trendline cannot represent.
• Input errors in a forecast multiply rather than add, so a single wrong assumption distorts every downstream number.
• Most forecasts are built on the data that is accessible rather than the data that is determinant.

Marketing Forecasting Fails Because a Trendline Extrapolates the Outcome, Not the Engine That Produces It

There is a difference between projecting a number and calculating one, and almost every failed growth target sits on the wrong side of it. Marketing forecasting is the practice of estimating future performance from data rather than judgement, but the definition says nothing about which data. That omission is where the discipline breaks. Last year plus 13% is not a forecast. It is a statement that nothing structural will change, expressed as a percentage.

The consequences reach the boardroom. The 2025 CMO Survey found that 64% of senior marketers consider demonstrating the impact of marketing actions on financial outcomes a major challenge. A forecast that cannot be traced back to business inputs cannot be defended in financial terms either, which is the same problem appearing at a different meeting.

The habit persists partly because of how long planning takes. The annual cycle is the format most teams still default to, even where rolling forecasts have been the stated best practice for years. The 2025 AFP Benchmarking Report puts the average mid-market budget cycle at four to six weeks, which means a meaningful share of plans are already describing a past state of the business by the time they reach the board. A method that assumes continuity is being applied to a window in which continuity has already broken.

Standard practice recognises four forecasting methods, and the blind spot is common to all four:

• Quantitative. Finds patterns in historical numbers.
• Qualitative. Substitutes expert judgement or survey input where data is missing.
• Causal. Tests how external forces such as price or ad spend move total sales.
• Time series. Tracks values across fixed intervals to detect seasonality and cycles.

Each is classified by the data it consumes, never by the business inputs it carries. A time series model of monthly revenue and a causal model of ad spend against sales can both be built without churn, cohort value or acquisition cost appearing anywhere. The methods are not wrong. They are silent on the variables that decide whether a growth target is reachable.

This is also why lowering a target rarely fixes anything. Reducing an output derived from a flawed model produces a smaller flawed number, not a more achievable one. The target was never the problem.

A useful test: ask what your forecast would do if churn improved by one point next quarter. If the answer is that the line keeps its slope, the forecast contains no business logic at all. It contains a slope.

This matters more now than it did five years ago because the inputs themselves have become less stable. Acquisition costs move with auction dynamics and platform changes. Retention curves shift with product and competitive pressure. A method that assumes stability performs worst precisely when planning matters most.

Churn, LTV and CAC Variance Are the Unit Economics a Linear Model Cannot See

Three units of business economics determine whether a growth target is achievable, and a trendline represents none of them. Unit economics describe what a single customer costs to acquire and returns over time, which is the only level at which a growth target can be validated as arithmetic rather than ambition. Each of the three breaks the forecast in a different way.

Churn is treated as a constant when it is a curve. The standard lifetime value calculation divides margin-adjusted revenue per customer by a churn rate, which assumes that rate holds across the customer base. It does not. Early-tenure customers churn at materially higher rates than late-tenure customers, so a single blended figure overstates value for young cohorts and understates it for mature ones. At company level the errors partly cancel. At segment level they do not, which is exactly where acquisition budget gets allocated.

Lifetime value is quoted before margin. The arithmetic is unforgiving here. At an 80% gross margin, calculating lifetime value on revenue rather than on margin inflates it by 25%, and the error grows as margins fall. A forecast built on the revenue version will justify acquisition spend the business cannot fund, and it will do so consistently rather than occasionally.

Acquisition cost is used as an average when it is a distribution. Blended CAC hides channel-level variance, and it hides the fact that acquisition cost is downstream of conversion rate rather than fixed. Improve conversion and effective CAC falls at identical spend. A model carrying one blended figure cannot represent that, so it cannot tell you whether a plan is achievable through efficiency or only through budget.

Get any one of the three wrong and the ratio between value and acquisition cost, the number most plans are ultimately defended on, stops describing the business.

Driver Based Forecasting Rebuilds the Number From Its Components Instead of Its History

Driver based forecasting replaces the projected result with a calculated one. Rather than forecasting revenue as last year plus a percentage, the model computes it from the operational variables that produce it: visitors multiplied by conversion rate multiplied by average order value, or new customers multiplied by margin-adjusted value net of churn.

This is the top down vs bottom up forecasting distinction, applied to marketing rather than to finance. A top-down forecast starts from a total and allocates it downward, which is what a trendline does by default. Bottom up forecasting starts from the granular inputs and aggregates upward, so the total is an output of the model rather than its premise. The difference is not presentational. A top-down number cannot be wrong in a diagnosable way, because it was never assembled from anything.

Driver based planning is the same discipline applied to budgets, and both come from FP&A rather than from marketing. The distinction that matters is the one FP&A Trends draws between consolidation and generation: consolidating your reporting aggregates outputs, it does not produce them. A dashboard that sums spend, sessions and revenue across channels is an aggregation of results. A forecast needs the variables upstream of those results.

Three things change once the model is built this way:

• Variance becomes diagnosable. A miss traces to a specific driver rather than showing up as an unexplained revenue gap, so the post-mortem discusses which assumption was wrong instead of who is responsible.
• Scenarios become instant. Changing one variable flows through the entire model, which turns planning into a conversation about ranges rather than a single committed number.
• Assumptions become visible. Every output is attached to a stated input, so the forecast can be argued with.

McKinsey reports forecasting error reductions of roughly 20% to 50% where modelling is applied to suitable demand problems. That range is not a guarantee, and the caveat matters more than the number: the gain depends entirely on whether the inputs feeding the model are worth the arithmetic performed on them. A driver based model fed with a blended churn figure is a more elaborate way to be wrong.

Small Input Errors Compound Faster Than Any Marketing Forecasting Model Admits

Forecast errors multiply rather than add, which is why a plan can be directionally sensible and still be wrong by a wide margin. This is the property that makes input quality more important than model sophistication.

Take the standard lifetime value calculation. A subscription product with $120 monthly revenue per account, an 80% gross margin and 3% monthly churn produces an LTV of $3,200. Move churn to 4%, a single percentage point, and the same customer is worth $2,400. One assumption, a 25% swing in the value of every customer the plan intends to acquire, and every acquisition budget downstream inherits the error.

Now compound it. That LTV feeds an acceptable acquisition cost, which feeds a channel budget, which feeds a volume assumption, which feeds the revenue target. A trendline never exposes this chain because it never contains it. The line moves smoothly whatever happens underneath.

The same sensitivity works in the other direction, which is the argument for measuring inputs properly rather than for pessimism. In subscription businesses, retention moves the forecast further than almost any acquisition lever available to the marketing team, because it acts on every future period rather than on one. A linear model gives that fact nowhere to appear, which is how retention ends up owned by nobody in the planning cycle.

A Revenue Forecasting Model Breaks Where the Data Sits, Not Where the Maths Does

Most failed forecasts are not modelling failures. They are data topology failures. A revenue forecasting model gets built on whatever inputs are reachable, and what is reachable is rarely what is determinant.

The inputs live in different places and belong to different teams. Media performance sits in the ad platforms. Sessions and conversion sit in analytics. Churn, cohort value and retention curves sit in the CRM or the warehouse. Margin sits in finance. Nobody owns the join, so the forecast defaults to the one series that exists end to end in a single system, which is historical revenue. The trendline is not a methodological choice. It is what remains when the joins are missing.

This is also why forecasting accuracy rarely improves through better modelling alone. When finance, sales and marketing forecast from different sources, the outputs disagree and the reconciliation meeting replaces the analysis. Consolidating the drivers into one system removes the argument before it starts, which is a governance benefit as much as an analytical one. It is the reason Primelis built Brand OS around consolidation first, bringing marketing data into a single layer before asking it to answer anything.

The practical test is not whether your model is sophisticated. It is whether you could change one assumption today and see it propagate. If updating a churn figure requires three exports and a manual paste, the model will not be updated, and an unupdated driver model is worse than an honest trendline because it carries false precision.

Brand Demand Is the Driver Almost No Model Carries

There is a fourth input, and it is missing from nearly every marketing forecast: the demand that exists before anyone enters a conversion path.

Brand demand determines the cost and volume of everything downstream. It sets what proportion of searches arrive already decided, how much of the funnel is navigational rather than persuasive, and what a given unit of media actually has to accomplish. Two businesses with identical spend and identical conversion rates will forecast differently if one has brand demand and the other is renting attention. Yet channel dashboards report clicks, sessions and conversions, which are all outputs of that demand rather than measures of it.

The measurement problem is that brand presence is now distributed across surfaces that do not report together, and part of it does not report at all. A brand shows up in search results, on marketplaces, in social feeds and inside AI-generated answers, and 70.6% of AI-driven visits arrive with no referrer header, landing in analytics as direct traffic with no attributable origin. Demand generated there is real, converts, and is invisible to the model that is supposed to predict it.

Brand OS exists for exactly that gap. It tracks where a brand appears across search, marketplaces, social and LLM answers, and connects that activity to business outcomes such as share of voice, conversion lift and sales growth rather than to platform vanity metrics. That last point is what makes it a forecasting input rather than a reporting layer: measuring conversion lift is measuring incremental contribution, which is the variable a media mix model exists to isolate and the one a trendline assumes away.

It was built with the Paid Media, Organic, Retail, Data and Tech teams together, because no single channel view contains the variable, and it sits alongside a stack that already arbitrates more than $1B in annual branded ad spend for over 200 brands. The same logic extends downstream: balancing excess stock against stock-outs is a demand forecasting problem, and it resolves correctly only when brand-driven demand is modelled rather than inferred after the fact.

The forecasting consequence is direct. Brand demand belongs on the input side of the model, alongside churn and acquisition cost. Treated as an output, it gets credited after the fact and budgeted last.

Rebuilding Marketing Forecasting Around Business Inputs Changes What You Commit To

A growth target derived from a trendline is a hope with a decimal point. Marketing forecasting becomes a planning instrument only when the number is calculated from churn, cohort value, acquisition cost variance and brand demand, and when those inputs live somewhere they can actually be updated.

The useful first move is not building a model. It is auditing which of those four inputs your current forecast can represent at all, and which it silently assumes to be constant. Most teams discover the answer is three out of four, and that the missing one is brand demand. Mapping those inputs into a single measurable layer is what Brand OS was designed to do, backed by 15 years of work with more than 300 global brands and over $1B in annual media managed across 200+ specialists.