Putting it together
Combining data & trend analysis
Individually, every source in this guide is partial. Combined carelessly, they produce confident nonsense. Combined well, they tell you something none of them knows alone.
Part one: reading a trend properly #
Most marketing decisions are made on a comparison of two numbers, and most of those comparisons are invalid. Before any modelling, it is worth being able to tell whether something actually changed.
Four questions before you react to a change #
- Is the data complete? Google Ads keeps filling in conversions for days after the click. Search Console lags two to three days. GA4's most recent day is partial. Never diagnose from the last 48 hours. This is the single most common false alarm in marketing reporting.
- Did the measurement change? A new consent banner, a tag deployment, a platform attribution change like Meta's in early 2026, a modified lookback window, a site migration. Check the change log before the campaign.
- Is it seasonal? Compare against the same period last year, not just last month. In South Africa, December and January distort almost every B2B dataset, and school holidays and payday cycles move consumer behaviour on a predictable rhythm.
- Is it bigger than the normal variation? If your weekly conversions have ranged between 40 and 60 all year, a week at 44 is not a decline. It is Tuesday.
Which comparison to use #
| Comparison | Good for | Fails when |
|---|---|---|
| Week on week | Spotting breakage fast | Public holidays, campaign launches, any weekly cycle |
| Month on month | Operational reporting | Months have different lengths and different numbers of weekends. February always looks bad. |
| Year on year | Genuine growth assessment. Controls for seasonality automatically. | The business changed materially, or the measurement did |
| Rolling 7 or 28 day | Seeing the underlying trend without daily noise. The best default for a dashboard. | You need to detect a sudden break quickly — smoothing hides it |
| Same period, previous year, indexed | Comparing this year's shape against last year's | A step change makes the index misleading |
Report a rolling 28-day figure against the same 28 days a year ago. It absorbs day-of-week noise, controls for seasonality, and is hard to game with a well-timed campaign. Add the raw daily line underneath for anyone who wants to see breakage.
Traps that produce confident wrong conclusions #
- Two lines that both went up. Sales rose, and so did your LinkedIn spend. So did the temperature. Correlation on a time series is nearly free — almost anything trending correlates with anything else trending.
- Reading a rate without its denominator. Conversion rate rose 20%, because traffic fell 30% and the remaining traffic was better qualified. Nothing improved.
- Averages hiding the distribution. Average order value rose because one enormous order landed. Use the median alongside the mean, always.
- Comparing a filtered number to an unfiltered one. Search Console query-level clicks against total clicks. GA4 sampled data against unsampled. This produces gaps that look like insights.
- Small numbers. Three conversions becoming five is a 67% improvement and it means nothing at all.
Keep a dated log of every site release, tag change, campaign launch, budget shift, price change and platform update. When a number moves in four months' time, this log answers the question in thirty seconds instead of three days. It is the cheapest analytics investment available and almost nobody does it.
Part two: joining data across sources #
Once you want one view rather than six, you need something to join on. There are only a few realistic keys.
| Join key | Joins | Reliability |
|---|---|---|
| Date + channel | Ad spend to outcomes, at aggregate level | High — it is just arithmetic. The basis of MMM. |
| UTM / campaign name | Ad platform to GA4 to CRM | Only as good as your taxonomy. This is why taxonomy matters so much. |
Click ID (gclid, fbclid) | A specific click to a specific conversion | High when it survives, and it often does not |
| Email address | CRM to email platform to ad platforms to finance | The strongest key you have, once someone identifies themselves |
| Transaction ID | GA4 to ecommerce platform to finance | High, and the basis of any credible revenue reconciliation |
| Company domain | Individual behaviour to account, for B2B | Moderate. The core of account-level attribution. |
Notice that the strongest keys all require someone to identify themselves. This is the fundamental asymmetry of marketing data: anonymous behaviour is abundant and hard to join; identified behaviour is scarce and joins cleanly. Everything in measurement is a negotiation between those two facts.
Part three: can attribution be solved? #
Not completely. It can be managed well. Three methods exist, they disagree with each other, and that disagreement is useful rather than a problem.
Multi-touch attribution (MTA) #
Assigns fractional credit across observed touchpoints. Intuitive, granular, and useful for optimisation.
Its limits are structural rather than fixable with better software. MTA can only credit touchpoints it can observe, so it systematically under-credits anything unobservable: a podcast mention, a conversation at a conference, a Slack community recommendation, a billboard, a colleague's referral. Privacy restrictions have shrunk the observable set further.
MTA also cannot distinguish influence from coincidence. Someone who was always going to buy still clicks a branded search ad on the way, and that ad receives credit for revenue it did not cause.
Marketing mix modelling (MMM) #
A statistical regression on aggregate spend and outcome data over time, with controls for seasonality, price, promotions and external factors.
MMM needs no user-level data at all, which makes it durable against every privacy change in this guide. It captures offline and unobservable channels. It handles diminishing returns and carryover effects that MTA cannot see.
It requires roughly two to three years of weekly data to be trustworthy, it operates at channel level rather than campaign or keyword level, and it produces confidence intervals rather than exact figures.
Google Meridian is the sensible default — actively invested in, with a no-code Scenario Planner shipped in February 2026. PyMC-Marketing is the strongest alternative if you have Python capability in-house, and the most flexible of the three. Meta's Robyn carries a caveat: no formal deprecation notice, but its last CRAN release was 3.12.1 in July 2025, and multiple agency and vendor sources reported through 2026 that Meta scaled back the team and stopped promoting it. Check its current status before starting a new build on it.
Incrementality testing #
The only method that establishes causation rather than correlation. Hold out a geography, an audience segment or a time period, run the campaign everywhere else, and measure the difference.
Results are routinely uncomfortable. Branded search and retargeting are the usual casualties, because both target people who were already going to convert. If a channel is genuinely driving new demand, a holdout test will show it. If it is harvesting demand created elsewhere, the test will show that too.
The cost is real: you are deliberately withholding spend, and tests take weeks to produce a readable result.
The triangulation approach #
No single method is sufficient. A working measurement stack uses all three, each for what it is good at:
MTA
Day-to-day optimisation within channels. Understood as directional, never as truth.
MMM
Quarterly and annual budget allocation across channels, including the ones nothing else can see.
Incrementality
Validating the other two, and settling arguments about specific channels once and for all.
Where the three methods agree, act with confidence. Where they disagree, you have found something worth investigating — and that is usually more valuable than another dashboard. A channel that MTA loves and incrementality testing cannot detect is almost always harvesting demand created elsewhere.
If this feels out of reach #
Most businesses do not need all three. The sequence that delivers most of the value:
- Fix data collection first. Server-side tagging, correct consent handling, Enhanced Conversions and platform Conversions APIs recover more signal than any modelling exercise.
- Make the CRM the reporting spine, and reconcile it to finance quarterly.
- Push closed-won revenue back into the ad platforms so they optimise toward money rather than form fills.
- Run one incrementality test a year on your largest or most-questioned channel.
- Add modelling only once the foundations are stable. Modelling built on broken collection produces confident, precise, wrong answers — and they are much harder to argue with than an obviously messy dashboard.
Common questions #
How long should an incrementality test run?
Long enough to cover your typical purchase cycle plus enough volume to detect the effect size you care about. For ecommerce that is often two to four weeks. For B2B with a six-month cycle, a clean holdout test on closed revenue is impractical, so most teams test against a mid-funnel proxy such as qualified pipeline and accept the compromise. Running a test too short is worse than not running it, because it produces a null result people mistake for evidence of no effect.
Do I need two to three years of data before starting MMM?
For a trustworthy model, broadly yes — you need enough weekly observations to separate seasonality from marketing effect. You can start earlier and learn the process, but treat early outputs as an exercise rather than a basis for reallocating budget. If you do not have the history, start logging weekly spend by channel now; that is the input, and it is easy to reconstruct badly and impossible to reconstruct well.
Our MTA and MMM disagree completely. Which is right?
Probably both, about different things. MTA over-credits observable, late-funnel, high-intent touchpoints. MMM sees aggregate effect including everything MTA is blind to. A common and entirely explicable pattern is MTA loving branded search while MMM barely notices it — because branded search captures demand that other channels created. An incrementality test settles it.