We began our journey with Google Ads, which has historically ranked second in terms of sales revenue for the online shop. In addition, search most directly represents the actual behaviour of the market, serving as the purest form of demand testing. In such a context, it is possible to honestly measure whether advertising genuinely influences a decision or simply confirms what would have happened regardless. An "honest test" means here that the channel does not conceal the impact of advertising — success cannot be attributed to visibility, emotion, or reach. If an advertisement works on Google, it is not because it was visually appealing or attention-grabbing, but because it helped a person make a decision they were already inclined to make.

It is important to emphasise at this point that the brand keyword ("goldtime") was excluded from the analysis and work during the first phase, as its incrementality is by its very nature minimal. Brand keyword bidding is employed by us solely for the purpose of "protecting" our own brand in situations where competitors are investing in it.

In the spring of 2025, we also commenced Marketing Mix Modelling, the results of which revealed that Google Ads was one of the most significant contributing factors. Given that the nature of always-on marketing makes it difficult to assess the brand's organic impact with sufficient statistical model accuracy, we undertook the present work as a follow-up project.

To bring about a change in results, we applied our EPE (Elevate Predictive Efficiency) framework, which encompasses behavioural analysis, segmentation and signal-based modelling, the creation of conversion value rules, and dynamic advertisements (ad messages that vary using IF functions).

By means of EPE, we excluded, for example, users who had visited within the last 24 hours and frequent repeat visitors from all search campaigns. A total of 7 similar audience exclusions and timing logic rules were created. In order to establish a feedback loop that develops the model, a custom EPE script has also been implemented on the account.

The EPE methodology can be applied across other channels as well, and this is a direction in which we are actively moving. We see a particularly significant role for it in Meta remarketing campaigns.

To understand the true incremental impact of advertising — rather than mere correlations — we employed a combination of a geo-experiment and a causal impact model (Bayesian Structural Time Series) approach.

The Elevate Predictive Efficiency (EPE) framework operates as "layers" of causal analysis: it combines micro-level data on clicks and conversions, signal correlations, and the probability of repeat purchase in order to assess the probable incremental value of each advertising activity. EPE is based on machine learning, which identifies behavioural patterns and optimises advertising bids in real time according to their estimated incrementality.

To validate the results, we employed control periods and a time-decay model, which assessed the impact of touchpoints over time — measuring how quickly the effect of an advertisement diminishes following the initial exposure. In this way, we eliminated incidental conversions and seasonal noise.