Follow the purchase before you start moving ad budget.

A good answer starts with what the system can see.

When we discuss chatting with ad data, my objection is that the answer can come from a very limited part of the business. If the input is mainly Meta data, it can help us investigate what is there. That hardly means we have explained the company’s total sales.

Jacob points to the missing parts: the finances and what happens in the store need to sit alongside the advertising in the analysis. Each tool shows a part. We are trying to understand a purchase that passes through several of them.

We need to be a little difficult here. Which question are we trying to answer? What evidence is missing? A persuasive answer does not become more complete because it arrives quickly. I want to understand what the conclusion rests on before moving budgets around.

Customers were not rejecting invoice payments. They could not pay.

In the podcast, I describe a problem we had experienced in Såpa’s checkout. At first, it looked as though more customers were choosing Swish or dropping out in a way we did not understand. It was easy to describe the change as a behaviour of the traffic.

The problem turned out to involve a specific combination: people arriving from a Facebook ad on an iOS device and trying to pay by invoice. I describe how we found and fixed the fault with the payment partner. This is a historical incident, not a claim that the problem remains.

The interesting part is what we could have done if we had stopped at the total figure. Rewritten ads. Moved money. Looked for a new audience. None of those actions would have repaired the payment. When something looks strange, we sometimes need to follow the customer’s actual path, using the right device, traffic source and payment choice.

Before you buy more traffic: what is left on the shelf?

Jacob describes Juniper’s start to the year, with important bedding and towels out of stock. Some combinations were available, but not necessarily the products customers wanted together. He compares it to a clothing brand with mostly extreme sizes left.

That is a concrete analysis problem. The site may work, products may technically be in stock, and ads may keep getting clicks. Yet the offer can be considerably weaker for someone arriving to buy.

Look at what customers can buy together. Are the right sizes, colours and complementary products available? Jacob’s Juniper example shows why the inventory total can look reasonable while customers struggle to put their orders together.

High frequency does not mean every customer has seen you.

Jacob suggests investigating advertising to existing customers. I explain that we have already tried different approaches. The conversation is interesting precisely because we do not conclude that every company should adjust the same setting whenever growth feels slow.

His point is that frequency needs to be read alongside reach. A list can contain many customers while the ads reach only some of them. Repeated views among those reached do not prove that the rest of the customer base has been addressed.

I describe our situation and why, among other things, we want comments from people who know the products. Here too, the hypothesis needs to fit the evidence. What potential is there among existing customers? When did they last buy? Is there a relevant reason to buy again? Start there before turning general advice into a budget rule.

A bad day does not always need a new ad.

Jacob raises how difficult people find variation. When results are clearly worse than expected, we want to find a fault and act. We recognise the reaction in ourselves too.

We discuss how weekdays, seasons, campaigns, weather and product launches can overlap. That is a lot of signals to keep in your head. Experience helps us notice things, but hardly makes us immune to finding a pattern that is not particularly stable.

I want us to investigate the deviation properly. How large is it? Where in the flow does it occur? Has something specific stopped working, or do we need more evidence? Better analysis should also let us leave things alone when we have no good reason to change them.

AI can do the work. The numbers still need to be right.

Towards the end of the analysis discussion, I describe how answers to questions about store data had become more useful over time. I also explain that I still check the numbers, having previously seen tools return incorrect figures.

When an AI answer will affect the budget, I want to trace the figure back to its source. Which period does it cover, what is included, and is the calculation right? Otherwise we have merely made it faster to decide on the wrong evidence.