Most marketing measurement stops at“this might be the cause.”This is about getting past that.
I'm Andrei Popescu. Seven years buying paid media, on budgets from €500 to €200,000 a month. I write here about the methods that answer the questions a dashboard leaves open: which channels actually cause sales, whether a campaign changed anything, where the next euro stops paying off.
Every method gets the same treatment: how it works in plain words, the maths unpacked, where it lies to you, and a notebook you can run on your own data. The working is shown so it can be checked.
The writing
All posts →Search budgets and bids · 20 min read · Interactive · Video
Where am I losing money right now, and can I trust Google's own estimate of the fix? Impression share and a replay of logged auctions
Read the budget and rank leaks in your report first, then treat the simulator's promise as a replay of last week, close when nothing changes and wrong when competition moves.
Cross-platform reach · 24 min read · Interactive · Video
How many real people did I actually reach across platforms? Capture-recapture, the Petersen-Lincoln estimator
Adding two platforms' reach counts shared viewers twice. Your site's visitors bracket the real reach between a floor and a ceiling, and only a panel can give one number.
Frequency and audience decay · 20 min read · Interactive · Video
How many times is too many, and how long does an audience stay worth targeting? A Hill curve and a survival half-life
Cap where one more impression adds less to the chance of a sale than it costs, and read the audience's half-life off the people who were ever going to buy.
Forecasting · 17 min read · Interactive · Video
What will I get next month if nothing changes, and how far can I trust that number? A simplified Silverkite forecast
A forecast of next month's conversions with an honest range, checked on five hundred accounts: where it beats last month's total, where a ruler wins, and why the range runs narrow.
Seasonality · 11 min read · Interactive · Video
Why do some days always look worse, and when is it actually a problem? Seasonal indices
Most accounts run on a weekly rhythm, so compare each day with its own normal. One week tells weekends from weekdays, and a growing account can fake a weekend effect.
Anomaly detection · 19 min read · Interactive · Video
Something moved. Is it a blip, a trend, or a broken thing? A median score, the Theil-Sen slope and change-point detection
A blip, a trend and a break need three different tests. In my tests a step showed up at about a seventh of the size a single bad day needs.
Significance testing · 14 min read · Interactive · Video
Is this difference real, or am I looking at noise? A two-proportion test and the Benjamini-Hochberg correction
Two checks before you believe a gap between two rates: is it bigger than the clicks behind it allow, and did you find it by scanning a whole table?
Creative fatigue · 20 min read · Interactive · Video
Is my ad wearing out, or did something else change? Segmented regression on exposure
Fatigue runs on exposure, not days. Fit the bend against how many times people have seen the ad, and check that exposure rose before blaming the creative.
Conversion lag · 17 min read · Interactive · Video
Why do today's numbers look bad, and how long until I know the real result? Kaplan-Meier survival with a cure model
Recent days look worse because their conversions have not all arrived yet. Measure your own lag curve, divide young days by their share arrived, and wait out the window.
Search terms · 16 min read · Interactive · Video
Which search terms should I block, and which deserve a chance? The Wilson score interval
Block a term at once only if it has nothing to do with what you sell. Against a 4% benchmark, a term with no conversions needs 93 clicks before it can be judged.
Interrupted time series · 13 min read · Interactive · Video
I changed something. Did it actually work? An interrupted time series
Comparing before with after hands you the trend as your result. Project the line your metric was already on, then read the gap for a few days only.
Marketing mix modelling · 15 min read · Interactive · Video
How much should I spend on each channel, and where do the extra euros stop paying off? Marketing mix modelling
Put the next euro where it still earns the most. A mix model reads each channel's diminishing returns from weekly spend and revenue, and big moves still need an experiment.
Experiments · 14 min read · Interactive · Video
Is this channel actually driving sales, or would those sales have happened anyway? A geo-holdout read by difference in differences
A geo-holdout switches a channel off in random regions and compares them with the rest, the closest a marketer gets to proof that a channel causes sales.
Attribution · 16 min read · Interactive · Video
Which marketing channels actually drive sales, and which just show up at the close? The Markov chain removal effect
Last-click credits whoever closed the sale. The removal effect asks how many sales you would lose without a channel, and gives the answer as a range.
How I show a number
Illustrative: drag the slider
A figure without its range is an opinion wearing a decimal point. Every number in every post here arrives with how far it could be off, and where it came from.
8 weeks of data: too little to measure. I estimate, and keep the range wide and honest about it.
Did the in-store sales come from the paid campaigns?
If offline conversions are not connected, online spend and in-store revenue never meet in one place, and no amount of modelling fixes a join that does not exist.
- What would settle it
- Point-of-sale data joined on date and store, or a sales export you can line up against spend.
- Until then
- Report the online conversions with their range, and leave the in-store link open rather than guess at it.
Measured: straight from the data.Modelled: a model stands between, shown with its interval.Estimated: inferred from thin data, widest range.
About the name
Stochastic describes an outcome that carries randomness, something you can state as a distribution, never as one certain number. Strata are layers.
Marketing results arrive in layers: the channel, the audience, the week, the thing a competitor did on a Tuesday. The name is a standing reminder that any figure here is one layer read off a stack, and that quoting it without its range would be exactly the omission this site exists to argue against.
Get in touch
I take on freelance paid media and measurement work, and I answer questions about anything published here. Email is the whole system: no form, no newsletter, nothing that follows you around.
Also on LinkedIn.