Amazon Q, GTM growth metrics
Showing finance, product, and go-to-market which customers are arriving, staying, and leaving, month by month.
- Amazon Q Business
- Amazon Q Developer → now Kiro
- Amazon QuickSight → now Amazon Quick
SQL, Amazon Redshift, Power Query, Power Pivot, DAX, Excel modeling, Data validation, SOP writing
Illustrative dataEvery figure on this page is synthetic. The model ran on confidential Amazon data, which isn’t shown here.
Amazon Q was a new, high visibility investment: three AI products under one finance team. Finance could see revenue, but not the customers who drove it: how many were new each month, how many came back, and how many left.
I built the model that answers that, month by month, for all three products, with three groups in mind: finance, to explain what was driving revenue; product managers, to see whether new features kept people around; and go-to-market, to see where to focus the sales pipeline.
“Daniil established the foundational reporting and analytical framework that gave business partners clear visibility into product performance, key growth drivers, and emerging trends. He translated messy, complex data into actionable insights while navigating significant ambiguity.”
Angela Laluw, direct managerFinance Manager, AWS
Amazon Q is three products under one name, each sold by subscription.
Revenue shows how much a product is making, but the customers behind it explain why. Revenue can rise while a product quietly loses people, as long as the ones who stay pay more, so for a new product you need to see both to know whether growth is healthy.
The teams needed answers to three simple questions, every month.
- Are we gaining customers, or just revenue?
- Revenue can grow because more people are paying, or because fewer people are each paying more. Those mean very different things.
- Do we have a churn problem?
- If people leave as fast as they join, we have a stickiness problem.
- Are customers moving up or down?
- Moving from Lite to Pro means the product is earning trust. Moving back down means the opposite.
Some of the data I needed didn’t exist. Much of it did, but it sat in huge tables with hundreds of columns nobody used, because no one knew what most of them meant or how to turn them into actionable insights.
- Some of the data didn’t exist.
- Nothing in the tables said who was a new, returning, or canceled user, so every one of those labels had to be built. A user who moved from free to paid even got a new ID, which made one upgrade look like one person leaving and a stranger joining.
- Some of the data existed, but nobody used it.
- There were huge datasets with hundreds of columns, but little documentation on what those columns meant. The first job was working out which ones mattered.
- The database counted companies, not people.
- The database was organized around customer accounts, and one account can hold hundreds of users, so an account could look healthy while most of its people had quietly stopped using the product. Counting every person meant far more rows than the system was built to handle.
- Each product had its own dataset and definitions.
- Each product lived in its own dataset and counted users its own way. I built the model three times, once per product, using the same approach so the results could sit side by side.
I built the logic the model uses to sort users each month. Every paid user falls into one of the first four groups, and some also move between tiers. The charts below use the same words:
- New
- Paying this month for the first time.
- Retained
- Paying last month, still paying this month.
- Returning
- Paid before, left, and came back this month.
- Canceled
- Paying last month, gone this month. Also called churn.
- Upgrade
- Moved from a cheaper tier to a pricier one, like Lite to Pro.
- Downgrade
- Moved from a pricier tier to a cheaper one, like Pro to Lite.
That equation is the backbone of the model. The model checks it every month, and if the numbers don’t add up, something in the data is off.
I worked with more than five engineers and the BI teams to understand the system’s limits, then built the fixes into the model myself.
Learn the dataLearned Amazon Redshift, our database, and mapped which of its many columns described users.
Query itWrote SQL to pull each product’s monthly user data into Excel.
Load itExcel couldn’t hold every row, so I linked the SQL through Power Query and loaded it into the backend instead.
Model itPower Pivot measures label every user each month: new, retained, returning, canceled, upgraded, downgraded.
Summarize itA summary table pulls those measures by month. One “Select Month” cell sets the trailing 12 months.
Check itEvery month, retained plus new plus returning must equal the total. Any gap shows up as a variance.
I built this pipeline three times, once per product, because each product lived in a different dataset.
- Build the user logic from scratch.
- The data never said who was new or who canceled, so I wrote the logic that compares each person’s month to the month before and puts them in the right group.
- Count people, not companies.
- Looking at each person instead of each account took longer to build, but it was the only way to see who was staying.
- Make every cell dynamic.
- Change the “Select Month” cell and the whole model rolls forward, with no numbers typed in by hand.
- Document every step.
- Each product got an SOP that walks through the refresh, so anyone on the team can update the model each month.
QuestionAre we growing, or just replacing the people who leave?
Looking at growth is one thing, but knowing where it comes from and who’s coming back is another. Every month some people join for the first time, some come back after leaving, and some cancel. Growth is what’s left after all three.
Purple bars are people who joined or came back. Hatched bars below the line are people who canceled. In January fewer people joined and more canceled. A month like that got flagged: we would ask the product teams what they changed or shipped, and check with the go to market team whether sales outreach slowed.
Start with the users we had three months ago. Add the people who joined and the people who came back, subtract the people who canceled, and you land on where we ended.
QuestionOnce people start paying, do they stay?
The share of each month’s paying users who were also paying the month before. Higher means more people are staying.
The summary table finance used, for Amazon Q Business. Darker cells are larger numbers within each row.
QuestionAre customers moving up to Pro, or sliding back to Lite?
Upgrades from Lite to Pro sit above the line and downgrades sit below it, with the net change drawn as a line. Net dipped below zero once, in January, and grew steadily after.
Each card ranks the four strongest months for one type of movement, with the top month in dark ink.
Each rate is a share of the group it comes from: Lite users for upgrades, Pro users for downgrades, last month’s trials for conversions, and last month’s paid users for returning users, meaning people who had canceled and came back. Upgrades, downgrades, and trials are Amazon Q Business. Returning users are Amazon Q Developer.
The share of paid users on each tier. Pro dipped while Lite signups surged, then recovered as more people upgraded.
QuestionHow many free trials turn into paying users?
Bars show how many people started a free trial each month. The line shows what share of them became paying users the following month.
QuestionDo newer customers stick around longer than early ones?
Each row is a group of people who started paying in the same month. Reading across shows what share of that group was still paying one month later, two months later, and so on. Newer groups, lower down, keep more of their people, which is why the colors get darker toward the bottom.
The monthly summary tables finance used for all three products, from users to usage and revenue.
Illustrative data, generated for this case study. None of the figures on this page come from Amazon systems or reflect real Amazon users, customers, or results. The structure, metric definitions, and workflow match the original model, but every number has been replaced.
- Product teams could see what their features did.
- Product managers could line up each launch against retention, upgrades, and cancellations, and see which features kept customers paying and which ones didn’t move the numbers.
- Go-to-market knew where to invest.
- The sales team could see which customer types and tiers were growing fastest, so it knew where to put more of its pipeline and where growth was stalling.
- Finance could see the usage behind revenue.
- Token usage was part of the model, so finance could watch how usage per customer shifted as the mix moved between free, Lite, and Pro users, and plan costs around it.
- Revenue changes came with an explanation.
- I added revenue bridges for two of the products, breaking each month’s change into new, returning, and lost customers.
- Three reports became one.
- The three separate product reports turned into a single two page view, so leaders could compare the products side by side on the same definitions.
- A file 93% smaller.
- Moving the data into Power Query and Power Pivot took the model from about 350 MB down to 25 MB, with nothing lost.
- At Amazon, done means 100%.
- At a startup, the goal is to get something to 80% as fast as possible, then iterate to 100%. Amazon isn’t like that. Every number had to trace back to its source and every cell had to be dynamic, so the team could keep relying on the model after I left.
- The hard part is the data, not the formulas.
- Most of the work was getting comfortable with imperfect data: documenting every limitation, choosing a workaround, and saying clearly what the numbers could and couldn’t tell us.
- Once it works, make it faster.
- When the data outgrew Excel, Power Query was the fix. There’s almost always a more efficient way. Once the solution is clear, the next job is making it quick to run.
If I did it again, I’d start with the big picture before diving into the data. I went straight into the tables, and only later understood how each product was priced, why the tiers were set where they were, and what leadership was trying to grow. Learning that first would have shaped which metrics I prioritized and made the model sharper from day one.





enrich
