Briefing
How Simile combines purchase data and research responses
Purchase records show what customers chose. Interviews can help explain those choices.
Purchase behavior and customer feedback
Imagine a customer stops buying a product. Their purchase history shows the change, but it doesn't tell you why. Was it the price, the product, or something else? In Building the What-If Machine, Simile describes combining people's actions with surveys, interviews and experiments to model how they might respond to change. [1]
What to ask in a demo
For a buyer, the useful question is simple: can this approach help with the change you're considering? Bring an example, such as a new price or product. Ask which data the model uses and how the provider checked its predictions against real outcomes. You want to see how the different inputs help answer your question, not just a list of the data available.
A scenario needs a defined action space
Simile’s confidence-model description starts with a population, a situation and the actions available to that population. A price-change brief therefore needs to say which customers, products and alternative actions it covers. Returning a distribution across those actions is different from forecasting a named person’s next purchase. The scenario and available behavioral evidence determine what the simulated answer means. [2]
Confidence is a separate research question
The provider reports evaluating its confidence model on held-out questions, using error between simulated and observed distributions. That evaluation concerns whether a confidence signal identifies stronger and weaker simulated answers. It is not a universal accuracy percentage for every market scenario. The useful follow-up is whether the confidence signal and simulation have been tested on questions like the change your team is considering. [2]
How to compare a simulated decision with reality
Set the scenario and intended metric before viewing the output. If an observed outcome exists, keep it outside the model setup and compare the simulated distribution with that result. For a genuinely new intervention, a pilot or later observed result can provide the comparison. Document the assumptions that differ between the simulation and the real setting, including competing options and customer circumstances. A good average result over past questions can still conceal an important error on one new decision.
