AI For E-Commerce

The six knowledge files every later lesson reads: the store profile, the catalogue, the customer file, the policies, the brand voice and the raw customer voice. Build them once here, and nothing asks you to go and find them again.

Download: the set-up zip holds Northmill Coffee, a complete practice store. Use it if your own data is not ready yet. Northmill is a UK shop, so its money figures are in £; yours will be in your own currency.

Creating the Project, loading the six files and running the setup prompts on camera, so you can see it working before you are asked to use it.

Name it once, refer to it forever. This is the convention that lets the twenty-ninth prompt read what the fourth one made, without you pasting anything between them.

Running the convention through a short chain, so you can see a file written by one prompt picked up by another.

Four types of AI, three things they can move, and one rule about numbers that the rest of the course runs on. You finish with a map of where AI pays in your store, not a list of tools.

You finish with: ai-opportunity-map.md

Nine columns of your customers' own words, built from material you already have, and read by eleven later lessons.

The rule that makes it worth anything is verbatim or it does not go in. A summary is your words about their words, and the phrasing is the entire value.

You finish with: customer-voice.md

Three or four personas built from what customers did and said, each carrying the number of real customers it covers and the phrases that support it. If a persona covers nobody in your file, it is a character you invented.

Fifteen cells: three stages of intent against five motivations. You have already built it, because it is two columns of 1.2's file, pivoted. What this lesson adds is what to do with each cell, and what an empty cell is telling you.

You finish with: intent-matrix.md

Three audits and five numbers: which fields are really populated, what your stack already does, and the baselines everything after this is measured against. This is the lesson that stops Section 2 being built on a column that is half empty.

The voice dataset, personas, intent matrix, stack audit and baselines in one session. It loads a deliberately broken catalogue, so the reconciliation genuinely fails on camera instead of being described.

A trend is not a product, and a rising line is not a reason to buy stock. Four stages of signal that turn a trend into three to five clusters worth shortlisting from, and the four questions that kill a candidate at each stage.

Ten proven core, ten trend-emerging, five experimental. One number, one mix and four filters, including the margin filter that does the real work, and the two gates that used to throw away a third of this store's revenue.

Most competitor research produces a feeling. This produces a figure out of ten that the next lesson's matrix reads straight off the page: five gaps scored nought, one or two, gated on five competitors and a collection date.

You finish with: competitor-grid.md

Five factors, five weights, five defined scales and one set of thresholds. Two of this store's three candidates flip verdict if you average instead of weight, in opposite directions, which is the point.

You finish with: scored-matrix.md

Marketplaces hold thousands of people describing the product you are considering, in the hour they bought it. This lesson reads four things off those pages, files them in the voice file you already have, and sends one capped number back to the matrix.

Five packs and fifteen prompts, and it cannot be split: 2.4 reads 2.3's output, and 2.5 writes back into 2.4's. Nine unscored candidates ship with the course, so the scoring happens live rather than being read out.

A tool can bring you four hundred rows of your market talking about your category overnight. The whole skill is choosing where to point it, and knowing which of those rows you are allowed to put in front of a customer. The source check is per dataset, not per platform.

You finish with: rows in customer-voice.md

An answer engine returns an answer rather than a page, which changes the unit of work: not a string with volume attached, but a question you can answer better than the three competitors a buyer is also going to read. Four of this store's six gaps already had an answer sitting unpublished.

You finish with: keyword-clusters.md

Fifteen angles generated cold are fifteen sentences about a category. Fifteen with a source row each are fifteen things your customers said, shortened, and somebody else can check them.

You finish with: ad-angles.md

Ten hooks across three formats is a week of work and a real budget, so the first thing this produces is a channel decision rather than a script. Every hook carries a physical object or a real number, because that is what a model cannot invent for you.

You finish with: creative-matrix.md

The obvious prompt calls this predictive and asks a model to multiply five numbers you guessed. The useful version is one formula, three measured inputs, and a break-even this store turns out to be sitting just below — 2.1 times against a break-even of 2.15.

You finish with: forecast.md

This store's margin bands have break-even returns of 3.49, 2.06 and 1.68 times, and its feed bids on all three identically, which starves one end and overspends the other at the same time, invisibly.

You finish with: feed-rewrite.csv

A creator list is out of date in a quarter. The brief — what is in frame, what cannot be claimed, how it is disclosed, who may run it as a paid ad — is still true next year, and it decides whether you can use what arrives.

You finish with: creator-brief.md

All seven packs in one sitting: community rows and the theme list, the question harvest and the answer pages, fifteen angles, thirty scripts, the break-evens, the feed rewrite and the creator brief. Three props are available and none needs staging, because they are real defects and real limits in the store's own files.

This store holds 239 phone numbers and may lawfully market to 22 of them. That gap is the whole lesson, and the reason it comes before anything in the section gets built rather than after. One of its sixty-eight best customers has SMS consent.

You finish with: consent-audit.md

VIP has at least four defensible definitions. Run all four against the same six hundred customers and one gives you seventeen people and another two hundred and sixty-seven. The difference is one word.

You finish with: customer-score.md

The abandoned-cart delay gets quoted as an hour, as three hours and as thirty minutes. Two of those are round numbers; one arrives as the first step of a sequence with a reason attached, and that is the one to use. Thirty minutes, three hours, a day, two days, and no discount until the fourth.

You finish with: lifecycle-flows.md

Most customer exports hold no purchase history — no order lines, no join to the catalogue — so recommendations come from four signals you do have, each stated as a rule somebody could read and disagree with. Coverage goes from 304 to 573 of 600 on one fallback rule and no new data.

You finish with: recommendation-rules.md

Average order times frequency times retention is a revenue number wearing a value number's name. Put gross margin into it and this store is worth $96.54 a customer rather than $207.16, and three of the four levers change rank.

You finish with: ltv-levers.md

Northmill is a UK shop, so the video and SOP show these figures in £. Your own will be in your currency.

A campaign to every consented address on this list is worth $12.75 in gross profit, derived from two numbers the store already measured. It changes what is worth building, and it makes split testing arithmetically unavailable rather than merely slow.

You finish with: email-measurement.md

Northmill is a UK shop, so the video and SOP show these figures in £. Your own will be in your currency.

Note: the prompts in this lesson’s SOP name consent-register.md. That is the file you built in 4.1 as consent-audit.md, so use that one.

Consent register, score and segments, the six flows, the recommendation rules, the bump and bundle set and the measurement file. Six packs and eighteen prompts.

A discount is quoted as a share of the price and paid out of the profit, and the two numbers are not close. Ten per cent off an average order here costs 21.5% of the gross profit on it, and the multiplier that does it is one divided by your margin — the same division that gives break-even ROAS.

You finish with: discount-cost.md

A discount lowers the margin your break-even is calculated from, so it moves the break-even too. Split your range into three bands and the answer differs in each: the fattest can carry about twenty-two per cent, the middle under two, and the thinnest none at all.

You finish with: price-architecture.md

Free carriage is an offer with a price on it, and the price is the postage you stop charging. Here that is $3.95, or 12.5% of the gross profit on an average order, given away against a threshold two thirds of the customer base already clears.

You finish with: shipping-economics.md

Northmill is a UK shop, so the video and SOP show these figures in £. Your own will be in your currency.

An offer is five decisions and only one of them is a percentage. Four forms of near-identical perceived value cost between $2.30 and $6.80 on the same order, because a free item costs what you paid for it rather than what you charge for it.

You finish with: discount-offer.md

Northmill is a UK shop, so the video and SOP show these figures in £. Your own will be in your currency.

A calendar with no cost column is a list of themes. Put the cost in gross profit beside every slot and the year gets shorter: this store's profit funds nine promotions at twenty per cent off, not twelve — and every one of them is also a send that counts against the cadence ceiling.

You finish with: promo-calendar.md

A promotion ends on a date; a price change does not, so it has to be right. This store can read a ten per cent change on either of its volume bands inside a quarter and cannot read a change on any single product at all — the busiest one would take thirty-five weeks.

You finish with: price-changes.md

Six packs: the cost of a discount, the band ceilings, the carriage waiver, the offer, the calendar and the price-change log.

Put the feed you already architected into three Shopping campaigns bid at their own break-even, then build search — the one channel the course has not taught — out of questions you already collected. Clears the eleven disapproved products first, because a rewritten title on a disapproved product shows nowhere.

Spend against a ceiling you measured rather than a target somebody else published. The acquisition ceiling is $31.67 against a measured $31.00, which is sixty-seven cents of headroom, and that figure governs every decision in the session. Audiences are limited to what your consent record actually allows.

Northmill is a UK shop, so the video and SOP show these figures in £. Your own will be in your currency.

Make the thirty cells you already built survive a vertical frame, a first second and a voice file that bans eleven words — and reshoot the ones that cannot. Placement work only; the campaign structure stays in Workshop 2.

Find out what a community is saying without breaking a term you would have had to agree to, then publish from your own reviews, because those are the only words you own. The store owns 498 quotable reviews against fifteen community posts it may not quote.

Decide whether your own numbers point at this channel before you build for it — this store's did not — and if they do, build against a break-even you worked out rather than one borrowed from the channel that already works.

Build the first of your six flows for the audience you may lawfully reach, with a job for every step and every claim traceable to a review somebody actually left you.

Build the four emails on timings that are already settled, and answer the only open question in the flow — what the fourth email offers — by pricing it before writing it.

Build the flow the rest of Section 4 hangs off, a sequence with two named slots in it, and fill the recommendation slot with rules you could defend to the customer who asked why.

Build the flow for your largest segment and your weakest reach, and find the handful of customers inside it worth reaching first — because a win-back list of nine is not the same job as a win-back list of 378.

Build the one flow that is not trying to sell anything: the one that catches the addresses the other five cannot see, and keeps the other five arriving at all. The governing figure is a complaint count, not a rate.

Build the branch that sits inside Workshop 8's flow — ranked on gross profit rather than margin percentage, under a ceiling you set first, behind a floor that is allowed to say no.