Using data to drive decisions and uncover insights
The reference page every formula on this site links to. Percentages, averages, spread, probability, tests, regression, elasticity, compounding and lifetime value, each from scratch with one shop's numbers worked through.
Every SaaS makes a handful of decisions in its first month that it will live with for a decade, and most founders make them without noticing. This is a plain language walk through the ones that matter in a Rails product: how you separate one customer's data from another's, where the tenant lives in the code, who a user is when they belong to two companies, what a tenant is versus who pays the bill, and why the isolation model you pick is also your GDPR strategy. Written so a founder can follow the business consequence of each choice, with the honest cost of changing your mind later.
Every ecommerce owner wants to know what the competition charges, and every one of them asks the same question first: are we allowed to? A plain language map of the rules that actually apply in the EU, Germany and Austria: why prices are facts and facts belong to nobody, where the database right does and does not reach, what the Ryanair cases really decided, when scraping becomes unfair competition, where GDPR starts, the competition law trap nobody warns you about, and what a crawler has to look like so that I am willing to run it.
Sample sizes, spend per platform, time at a small daily budget, concurrent tests, visitor bands for A/B tests, kill rules, channel fit by stage and what a forecast needs.
One loyalty project, four methods: behavioural segments, elasticity per segment, a loyalty loop and an experiences questionnaire, planned month by month.
How structural equation modelling shows which guest experiences actually drive loyalty: a hotel path model, total effects, fit in plain words and an investment decision.
How to tell whether a loyalty programme changes behaviour or just pays for it: the loyalty spectrum, the three Rs, earn and burn design, incremental margin against a control group, breakage by cohort, the points liability three ways, and the legal limits on expiry and rule changes.
How media mix modelling uses adstock and saturation curves to estimate the marginal ROI of every channel and where the next pound should go, worked through on an invented hamper shop with the formulas, the numbers and the checks.
What a tagging server on your own domain fixes and what it does not: Google Tag Manager server side, the Meta Conversions API, GA4 Measurement Protocol, enhanced conversions and deduplication, built so consent still decides every forward.
How to build a nightly loop in a Solidus shop that reads competitors' public prices, stock, costs and elasticity, sets each product's price inside limits, moves ad budget towards profit, and stays inside UK consumer and competition law. Formulas, worked examples and Rails code.
Chrome kept its third party cookies, and it barely matters. What Safari, Firefox and the consent banner really remove, what first party data is and is not, and how to measure without following people.
Fifteen concrete jobs that data engineers, analysts and data scientists do for a SaaS business and an online shop, in the order they should be done, each rated for importance and worked through with invented numbers you can check.
Three job titles that sound alike and are three different jobs. One coffee subscription business followed through all three roles, what each costs to hire in the UK, the order to hire them in, and what goes wrong when you skip one.
How to put an honest number on one more email, webinar invitation or retargeting ad: response curves, marginal value, the full cost of a contact and holdout groups, worked through for an invented Manchester SaaS.
k means, latent class analysis and RFM compared on an invented hamper shop: how to choose the number of segments, profile and name them, check that they are stable and put them to work.
A/B tests, A/B/n tests, bandits, multivariate tests, holdouts, geographic tests, audience splits and time based tests: what each one answers, what it needs and where it breaks.
What a segment really is, the five tests it must pass, a priori versus post hoc splits, and why one churn model per segment beats one model for everyone.
How PCA and factor analysis turn 24 survey questions or 18 behavioural variables into a few dimensions you can name, score and act on, with a hotel survey and a hamper shop worked through.
When the ad budget is set from sales, sales and spend cause each other. Why one regression then overstates the ad effect, and how two stage least squares corrects it.
How to find the early action that predicts retention, clear the path to it, and build habits that last: cohort analysis, the Hook model, what the 66 days really mean, habit stacking, checklists, head starts, the goal gradient and the UK rules on onboarding emails.
How to forecast weekly demand with seasonal dummies, handle autocorrelation, judge models on a holdout and plan stock and staff against an 80% error band.
Hick, Fitts, Miller, the peak end rule, the serial position effect, cognitive load, friction and framing: what each piece of research actually found, what it does not prove, and how to use it to review a page, with worked examples.
The two systems of thinking, the biases behind a purchase, defaults, friction, hyperbolic discounting, mental accounting, choice overload and status quo bias: what the original studies found, how well they have held up, where UK law draws the line, and how to measure each one.
The maths and psychology of referral programmes: the viral coefficient, what a value below one is still worth, why cycle time is about timing until you pass one, and the UK rules every refer a friend scheme has to follow.
Why people copy other people when they are unsure, the six kinds of social proof, what the research says about reviews, and the UK law on fake reviews, with worked examples and code to test it on your own pages.
Ten pieces of maths that turn growth data into decisions: lifetime value against acquisition cost, payback, cohorts, survival curves, Bayesian tests, bandits, attribution and mix models, power laws, price elasticity and retention, each with a worked example.
Goal gradients, fresh starts, implementation intentions, defaults, status tiers, social comparison and risk reversal: the original studies, how well each has held up, worked examples, and the tactics UK law now bans.
Scarcity, urgency, the fear of missing out, the endowment effect and the IKEA effect: what the research found, when each one helps, when it destroys trust, which tactics UK law now bans, and how to measure them past the click.
Anchoring, decoys, charm prices, tiers, reference prices, partitioned prices and loss aversion: what each original study found, how well it has held up, how to test it on your own shop, and which tactics UK law now bans.
How panel regression separates a good shop from a good quarter, why pooled data lies about marketing, and how fixed effects give you an honest like for like sales figure.
How survival analysis tells you when customers leave, not just whether, and how to turn Kaplan Meier curves and a Cox model into a lifetime value you can defend.
Seven in 10 carts are abandoned and a sixth of US retail sales come back. Two predictive systems, an abandonment risk scorer and a return risk predictor, built in Python and wired into Solidus, with the business case worked through properly.
How dynamic pricing actually works for a shop: estimating elasticity from your own orders, the markup rule, a demand model, the guardrails, and the UK price transparency and competition rules, with Solidus and Rails code.
SARIMA, Prophet and gradient boosted trees compared for a shop's weekly sales: when each earns its place, how to test them honestly, and how a forecast becomes a reorder point in Rails and Solidus.
How logistic regression ranks customers, where the next brochure stops paying for itself, and how the same model turns baskets into cross sell recommendations.
Orders, tickets, visits: when the outcome is a count, ordinary regression misbehaves. Poisson regression, rate ratios, exposure and overdispersion, worked through on a Leeds hamper shop.
Deleting a customer's rows is easy. Their influence on a trained model is not. What Article 17 says in the UK and the EU, what regulators have actually said about models, what machine unlearning research can do, and the Rails architecture that keeps you out of the problem.
Most AI projects stall on messy data, old systems and no clear first step. The right order of work, three patterns for adding AI to a Rails and Solidus stack without replacing it, what MCP changes for legacy systems, the data protection basics and five realistic places to start.
Why a newsletter keeps producing orders for weeks, how the Almon polynomial lag model measures that curve, and how an invented Leeds hamper shop could use it to pick a three weekly send cadence.
How to add product recommendations to a Solidus shop: co occurrence, collaborative filtering with ALS, content based filtering with TF IDF and a hybrid of the two, with working Python and Rails code, an A/B test and the UK consent rules, using an invented Bristol tea merchant.
What Test Driven and Behaviour Driven Development are, which Rails tools do which job, what to test before every release, and how to use AI for tests without handing it the judgement.
A year of hamper sales, one log log regression and an elasticity of minus 1.6: how to model demand with OLS, read the output and find the price that pays.
A finding says what happened. An insight names the decision it changes, prices it and states the risk. How to tell them apart, score them and write one.
From the naive benchmark to exponential smoothing, SARIMA, Prophet, regression and structured judgement: which forecasting method fits which product, how to give a range instead of a number, and how to check the forecast honestly, with worked examples and Python.
AI and machine learning get used as synonyms, but they are not. What each one is, how a model actually learns, what each cannot do, the Python toolkit, and why the label matters for your projects and under the EU AI Act.
What happens when rival shops run the same clever algorithm against each other? Bidding wars, the Nash equilibrium, why learning agents chase a moving target, and where optimisation turns into algorithmic collusion.
How lookalike audiences work on Meta and Google today: seeds, sizes, setup, the 2026 shift to lookalikes as suggestions, what happens under the hood, and the UK rules for pixels and customer lists.
Why strategy has to start from how people actually decide, why what customers say and what they do disagree, and how to turn a founder's belief into a question your data can answer.
A low hourly rate is one line in a longer sum. How to price the hours, the rework, your own time and the risk, with a worked example and the story of a client who chose on rate alone.
Cloud AI, self hosted models or a hybrid of both: what each can do, what goes wrong when staff paste the wrong thing, what the GDPR and the EU AI Act actually require, and how to route sensitive work so it never leaves the building.
Mean, median, standard deviation, confidence intervals, p values and correlation, worked through on an invented hamper shop's orders so you can read marketing numbers honestly.
What a unified marketing stack really is for a small shop: one customer and event record, consent carried on every row, and the costs of building or buying each piece.
Why I left growth work on content sites for shops and subscription apps, and the four things that change when growth moves from tactics to measurement: margin, incrementality, cohorts and disciplined experiments.
How to let your email list shape your adverts and your adverts shape your emails: customer lists on Google and Meta, suppressing recent buyers, match rates, refresh timing and the UK law behind it.
Performance Max and Meta's Advantage+ took the old levers away, but not the control. What still steers them, what the reports now show, and how a holdout tells you what the campaign really adds, shown on an invented Bristol coffee roaster.
What a p value really tells you, how wide the margin of error on a conversion rate is, and why a significant lift can still lose you money, worked through for an invented Leeds candle shop.
Your rival is doing the maths too. Bertrand, Cournot, Stackelberg, the prisoner's dilemma in brand bidding, repeated games, focal points and signalling, each worked through with invented British businesses and every step shown.
How to split an ad budget and set a price when you cannot know how the month will go: scenarios, diminishing returns, the shadow price of budget, the flaw of averages and the markup rule, worked step by step for an invented candle shop in Bristol.
How to copy a Solidus shop's orders into BigQuery every night with the Ruby gem, keep personal data out, tag every order with its campaign, and measure sales and customer value by campaign without fooling yourself.
How to let your shop tell your adverts and emails when something changes: signed webhooks, a queue, actions that are safe to repeat, and a schedule for what no webhook will ever tell you.
Building many pages from one template is fine. Building the same page 400 times with a different town in the headline is spam. How to tell the two apart, and how to build the first kind.
How to spot the people still deciding, turn what they do on your site into an honest purchase probability, and send the message only where it changes the outcome, within UK profiling law.
Dave McClure's five pirate metrics turn the vague question of growth into five counts. How to define each one, read the rates between them, and fix the stage worth the most.
Chrome kept its third party cookies, but most of your visitors are already out of reach. How to turn organic search into first party data you can use lawfully, and measure content by what each click is worth.
Every UTM parameter GA4 recognises, the channel rules that sort your words into buckets, one naming vocabulary, what breaks, and how the labels feed attribution. With an invented Leeds candle shop and three campaigns, in pounds.
Your price is only half of the decision, because your rival is choosing too. Payoff matrices, dominant strategies, Nash equilibrium, repeated games, credible threats and reaction functions, worked through for two invented Bristol coffee roasters.
What a British shop's tracking code may do before and after a visitor presses accept, what Google's consent mode really sends on a no, and how to estimate the sales your ad platform cannot see.
How to choose between options when you cannot know how they will turn out: the payoff table, expected value, utility for when the bad year could sink you, regret, decision trees, and what a pilot is worth before you pay for it.
How a data warehouse such as BigQuery or Snowflake joins your shop, analytics, ads and email data on the customer, turns segments into queries that run every morning, and sends them back to the tools that act.
Data science delivers better decisions, measured, not dashboards. The five step loop, a way to price a better decision, and three small models at an invented Brighton garden centre chain.
The same campaigns in two languages across four countries, and three systems that each count the orders differently. How I keep country and language apart, map everything to a market, correct for consent per market and reconcile the platforms against the orders.
Selling the same product as everyone else and selling something only you make are different games with different rules. How the two markets differ, the four numbers that tell you which one each product line is in, and what each one rewards, worked through for two invented Bristol shops.
Why the customers who got your campaign always look better, and how potential outcomes, propensity score matching and instrumental variables find the effect the campaign actually caused.
When a bigger rival cuts the price of a product you both sell, matching feels safe and usually loses money. The economics of price wars, elasticity, contribution, suppliers, switching costs and bundles, worked through for an invented Leeds tyre retailer.
How to feed margin, lead quality, stock and consent state into Google Smart Bidding and Meta's Conversions API, with the formulas, the weekly loop and a worked example from an invented Brighton supplier.
