Digital twins.

What a digital twin is

A digital twin is a virtual model of a real object or process that behaves roughly the same way as the original. The idea came from industry: a twin of a turbine lets you run the load in a calculation rather than on the real equipment, and see in advance where it will fail.

In marketing, the object being modelled is not machinery but people and campaigns. An audience twin is a model built on real data about your customers' behaviour, in which you can check what happens when the price, the offer or the channel changes.

The model consists of three parts: data on how the audience behaved before; rules describing the links between actions and results; an environment where you can run a scenario and see the calculation. Without the first part, the other two are meaningless - the forecast is built on history, not on assumptions.

Why marketing needs it

The usual way to test a hypothesis is to launch a campaign and see what happens. That is honest, but expensive: you pay for traffic to find out that the idea did not work. And slow: gathering meaningful statistics takes weeks.

Modelling lets you weed out obviously weak options before you have paid for them. Out of ten ideas, the calculation discards six and four go into a real test. The budget is spent on checking what actually has a chance.

The second use is scenarios that cannot be tested in reality, or are too risky. What happens if you raise the price by fifteen percent. What happens to revenue if you remove the discount for regular customers. How the load on managers changes if the flow of enquiries doubles. In real life such experiments are too expensive to run at random.

The third is speed. A calculation gives an answer in hours, not in a month. This changes the rhythm of work: hypotheses stop piling up in a queue.

Hypothesis testing and A/B tests

The model does not replace A/B testing; it comes before it. The division of labour looks like this: on the model you select candidates, on the live audience you confirm the winner.

The workflow is the usual one. Formulate the hypothesis in a testable form: not “improve the landing page” but “replacing the seven-field form with a three-field one will increase the share of submitted enquiries”. Run it through the model and see whether the calculation shows a noticeable difference. If it does, put the two variants in front of real traffic and wait for a statistically significant result.

The last step cannot be skipped. The model shows what is probable, not what has happened, and its error is only discovered on contact with reality. How to set up a live experiment properly is covered in our article on A/B testing.

A separate benefit of the model is that it forces you to formulate the hypothesis rigorously. Half of all marketing ideas fall apart at the stage where the expected effect has to be written down as a number.

What it takes to make this work

Data. The model is built on history: orders, enquiries, on-site behaviour, results of past campaigns. The shorter the history and the worse it is labelled, the closer the forecast is to guesswork. A company that started keeping records three months ago has nothing to model yet.

Volume. On a sample of a few dozen deals, patterns do not stand out - the scatter swamps any effect. This is an important limitation: for many small companies digital twins are simply premature.

Connected analytics. Data on advertising, enquiries and deals must be joined into a single chain. If you see clicks separately and revenue separately, modelling the effect of advertising on revenue is impossible. The basics of this linkage are in our article on calculating ROI.

A person who understands the model's limitations. A calculation always produces a number, even when there is not enough data for it. The confident look of the result is the most dangerous trap in this subject.

What not to expect

The model will not predict the reaction to something that has never happened. A new product for a new audience, entering an unfamiliar market, a campaign in a format you have not used - there is no history here, and nothing to build a forecast on.

The model does not account for external events. A change in legislation, a platform leaving the market, a competitor's actions, a seasonal shift - all of this stays outside it.

The model reproduces your past limitations. If you have always worked with one audience, the calculation will confidently prove it is the best one, simply because it has no other data.

And most importantly: the accuracy of a forecast is not the same as the correctness of a decision. The model answers the question “what happens if”, while the question “what should we do” remains yours.

Where to start

A small company does not need a dedicated platform. Start with a simple funnel model in a spreadsheet: how many impressions, what share clicks through, what share submits an enquiry, what the conversion to a deal is, what the average order value is. By changing one parameter you can already see how it affects the outcome - that is the first version of a twin.

Then refine the model with real figures for each stage and add a split by channel and segment. At this level it becomes clear which hypothesis is capable of moving the result at all, and which yields fractions of a percent and is not worth the time.

When the spreadsheet can no longer hold the required complexity, it makes sense to look at specialised tools and bring in an analyst. But the vast majority of companies have enough with the first two steps, and those are what deliver the main benefit.

An important part here is correctly collected data on advertising campaigns: without tracking and goals there is nothing to model. This is part of advertising setup and is done before any calculations begin.