A CEO asks a fair question. What does one sale cost us in marketing?

You can answer that in ten seconds. Take the marketing spend, divide by the number of reservations. The number is real and it belongs on the slide.

The trouble starts with the next question, and it is always the same next question. Fine — so which part of it is working?

Because five sales a month spread across forty campaigns is not a dataset. It is five sales.

Between someone tapping an ad and someone collecting keys, a lot happens. Two weeks or two years. Two or three people who never appear in your CRM. A country border. At least four systems that do not agree with each other. Everything we call marketing measurement lives inside that gap. Not around it. Inside it.

This publication is about that gap. So it makes sense to start by describing it properly. Almost every argument in property marketing — budget splits, channel quality, creative testing, lead scoring — is really an argument about how to behave inside a measurement problem that most other industries do not have.

Part One — Three conditions that make this different

Property marketing is often discussed as if it were e-commerce with a bigger price tag. It is not, and the differences are structural rather than a matter of degree.

The sales cycle is longer than any attribution window that exists — and far less predictable than that sentence suggests.

Meta’s maximum click window is up to ninety days. Ninety days is the ceiling for the whole industry.

Plenty of buyers take a year to decide which country they want. But a reservation signed sixteen days after the first enquiry is not rare either — a couple who already visited last summer, already agreed with each other, and were waiting for a specific unit. Both journeys are normal. They are the same funnel.

So the average cycle length describes almost nobody. That is a nuisance for planning. The second consequence is worse: the fast buyers are the only ones who fit inside the attribution window. When a platform learns who is worth showing your ad to, it learns from a minority — and not a random one. It learns from whoever decides quickly. That might be your best customer. It might be someone with a small budget who needs no mortgage. The model cannot tell you which, because the slow buyers never enter its evidence at all.

And in the meantime, a dead lead and a slow lead look exactly the same. For months.

The product is finite, and it changes while you sell it.

You have 140 units, not unlimited stock. When the good ones go, “more leads” stops being a goal and becomes a cost. Price rises with construction stage, so the offer you advertised at foundation level does not exist at handover. A campaign can be judged excellent in March and be selling something unavailable by September.

The transaction does not happen online at all.

A form submission is not a purchase. It is a request to be contacted. The actual business event happens in a notary’s office, months later, often in a different country, and it enters a system your advertising platform will never see.

None of this is a complaint. It is the job. But it produces one consequence that runs through everything else: every number a property marketer manages is a substitute for a number that arrives too late to be useful.

Part Two — The ladder of substitutes

Those substitutes are not random. They form a sequence, and each step down is closer to money and thinner in data.

Cost per thousand impressions. Cost per click. Cost per lead. Cost per marketing-qualified lead. Cost per sales-qualified lead. Cost per reservation.

At the top you have thousands of events a day and almost no commercial meaning. At the bottom you have real commercial meaning and almost no events. Where on that ladder you choose to stand is, I think, the most consequential decision in the whole system — and it is usually made by habit rather than by choice.

To make the rest concrete, a worked example. The proportions are typical of a mid-sized developer selling to international buyers.

The worked example

Worked example, described above. Proportions are typical; the figures are illustrative.

Round numbers, chosen to keep the arithmetic readable. Sixty sales a year is a healthy business with real revenue. Now look at what it is as a measurement environment.

Part Three — What each rung can actually tell you

The usual way to describe this is to say the lower rungs are too slow. I used to say it that way, and it is not quite right. Good decisions do get made on one month of SQL data, and I will show one shortly.

The real question is not how long you wait. It is how big a gap has to be before you can believe it.

Everyone knows the principle already, just without the vocabulary. Flip a coin ten times, get seven heads, and nobody thinks the coin is loaded. Flip it a thousand times, get seven hundred heads, and everybody does. Same ratio. Different confidence. The second one had enough throws to rule out luck.

Campaigns work the same way. What matters is not the percentage printed on the report. It is how many real events sit behind that percentage.

So let us count them. Take one month — 1,000 leads — and split it evenly between two campaigns, 500 leads each. This is what is actually on each side of the comparison:

At MQL level, 100 against 100. At SQL level, 10 against 10. At reservation level, two or three against two or three.

Now the part that matters. For a difference between the two sides to be real rather than lucky, here is how far ahead the winning side has to be.

Worked example, one month, traffic split evenly. Calculated at the usual confidence settings for A/B testing (80% power, 5% significance).

Anything smaller than this is inside the range of ordinary luck. (One month, traffic split evenly; calculated at the usual confidence settings for A/B testing — 80% power, 5% significance.)

So at MQL level a campaign producing 138 against 100 is telling you something. At SQL level, 15 against 10 is not — you would need something closer to 27. At reservation level, one month can only show you a disaster or a miracle.

The fewer events you have, the bigger the gap has to be before it means anything.

That is the entire idea. Read it as a rule about how precise a question you are allowed to ask. At MQL level you can ask a fairly fine one: is this segment noticeably better than that one? At SQL level you can only ask a blunt one: is this producing anything at all, or is it dead?

Which is why the following two things are both true, and get confused with each other constantly.

This is a real signal.

One region sends 40 MQLs and produces zero SQLs. Another sends 40 and produces six. That is not a small edge, it is a difference in kind. The chance of a gap that large appearing by luck alone is about one in forty. One month is enough. Pause the first region and ask what is happening there — you are on solid ground.

This is not. Two campaigns each spent €2,000 and each produced 300 leads. Campaign A ended with 5 SQLs, campaign B with 3. Your report shows €400 per SQL against €667, a difference of 67%, and it looks decisive. It is not. The chance of a gap that size arising from luck alone is roughly three in four. Move budget on that and you are reallocating on a coin flip while feeling very data-driven about it.

Waiting longer helps, though less than people hope. A full quarter gives you 30 SQLs per side instead of 10, and the winner then needs about 55 against 30 rather than 27 against 10. Better. Still nowhere near the precision assumed when two campaigns get compared on a monthly report.

Part Four — The machine has the same problem

You might reasonably say: fine, humans cannot run that test, but the algorithm can. Let the platform optimise toward the real event.

It cannot, for a related reason. Meta needs roughly fifty conversion events per ad set per week to leave the learning phase and deliver predictably. Below that, an ad set sits in “learning limited” more or less permanently.

Look at the MQL row for a moment. Forty-six events a week, against a threshold of fifty. This developer could collapse the whole account into a single ad set — no geographic splits, no project splits, no audience tests — and still fall short of what Meta wants before it optimises reliably.

Two completely separate mechanisms — statistical power for the human, event volume for the model — hit the same wall within a few percent of each other.

There is a second thing hiding in that table. The number of campaigns you run decides which rung you are allowed to learn from at all. Split 1,000 monthly leads across forty campaigns and each ad set sees six leads a week — not enough to optimise even for leads. Consolidate to four and each sees fifty-eight, which clears the threshold for lead optimisation and nothing below it.

Account structure is usually discussed as a matter of control or tidiness. It is neither. It is the decision about whether your account can learn anything at all.

Part Five — Where that leaves us, which is not comfortable

I set out to argue that property marketers should simply move down the ladder, closer to money. That is too crude. Every rung is usable. The discipline is asking each one only the kind of question it has enough events to answer — and not letting a dashboard tempt you into precision that is not there.

But somewhere the weekly steering has to happen: precise enough to guide decisions, close enough to money to mean something. On these numbers that is MQL, and even it is uncomfortable. A few events short of Meta’s threshold, reachable only by giving up every split in your account. The floor is not a place to stand. It is an edge.

And here is the part I find genuinely awkward.

The only metric we can actually steer by is one we invented ourselves.

An MQL is not a market fact. There is no external definition. It is whatever your call centre decides it is, applying criteria your company wrote, recorded in a notes field by a person having a normal Tuesday. If two operators apply different standards, your channel comparison is partly a comparison of operators.

That is not an argument for giving up on it. It is an argument that the definition of MQL is the

load-bearing element

of the entire measurement system, and that most companies treat it as an administrative detail. If the thing you steer by is something you made, then the quality of the making is not a small matter.

Two honest caveats, because I do not think the case is closed.

There are methods that partly escape all of this — geographic holdout tests, incrementality experiments, marketing mix modelling. They work, and they need budget and volume most developers do not have. Worth doing when you can; not a general solution.

And the numbers here depend on the shape of the funnel. A developer with a stronger lead-to-MQL rate sits somewhere different. The method matters more than my figures: run it on your own shape and see what your data can actually support.

Finally — What this publication will be about

Each constraint in this article opens onto questions I want to spend time on.

If platforms only ever learn from buyers who decide quickly, what kind of audience are we slowly training them to find? If a cheap lead is cheap because it is unlikely to buy, what should replace cost per lead in a budget conversation? If your MQL definition holds up the whole system, what happens when the call centre quietly changes its standards? And if a dead lead and a slow lead look identical for months, when should anyone stop calling?

Those are the pieces I plan to write. Not tips, and not guides. Questions where the standard answer looks weaker than it first appears.

One thing to take away, if only one.

The distance between a click and a key is not a tracking problem waiting for a better tool. It is the actual shape of this industry, and the tools are not going to close it.

So the question worth asking about any number on your dashboard is not “how accurate is this?” It is: how many real events is this built on — and is the decision I am about to make bigger than what that many events can prove?