Method — how the figures and the briefs are computed
One calculation contract for the whole page: the weekly group rows above are the only input; country, channel, funding and loyalty splits are modelled on top and rescaled so every table reconciles to the group totals in all directions. The Decision Brief follows the same contract — it is generated from the computed aggregates only, so it can never quote a figure that is not on the page, and anything the data cannot support is flagged as not derivable rather than guessed.
Conventions: growth attribution in pp on a common group base (additive in both directions); the commercial equation runs a chained four-factor bridge, traffic → conversion → items per basket → average price, whose steps sum exactly to the gap (the reference-to-actual waterfall on the Overview tab runs the same bridge in its three-factor form, traffic → conversion → basket, since basket depth is not modelled per channel); comparisons only on comparable periods (same week vs Plan or vs LY), and a return rate compared with last year is stated on one announced deterioration in points — the same one on the channel cut, the category cut and the group row, so both sides of the comparison come from one book; contribution = gross margin, itself net sales less cost of goods plus the supplier funding credited less the marketplace payout, less the variable cost of serving the order.
The book and the warehouse node. The S&OP model is the whole merchandise book, shop floor and warehouse together, and the cover bands, the presentation floors of the policy matrix and the opening stock of the open-to-buy are all read against it. Outbound is every unit the group sells in the week across the five channels; the weekly row carries digital orders only, so the level comes from the five channel equations and the shape from the row. The warehouse is a node inside that book, not a second book: it holds what the shop floor does not, and it ships store replenishment for every unit that left a shop floor — sold across the counter, collected, or picked for a ship-from-store order — plus everything that goes direct to a customer or a wholesale account. Those two lanes add to the units the group sells, which is why one outbound figure now serves S&OP and Logistics. Its cover falls out of the book rather than being stated, and it reads as what it is: a flow-through building, not where the stock sits. And the book has no channel. Stock in units and at cost, cover, the age of the layers, the provision, the carrying and the open money are the NETWORK’s under every channel and every store-cluster filter: the same carton serves the shop floor, the site and the marketplace, and a flagship and a local store draw on one book. Only the sales and the outflows follow those two filters. Cover is therefore network stock over network outflow — a “retail cover” is not a quantity this book has — and rotation, days on hand and the provision share are numerator and denominator on one population. The country and the category stay on the book, because a market has its stock and a category has its stock.
The provision ledger. One function, provisionOf(category, cause, location), read by the cost layers, the old-stock table, the markdown exposure, the inventory scorecard and the P&L. Three rules. A provision is taken at the book's own average unit cost, because splitting the book into an old and a current layer changes what a unit costs and never what it is provisioned at. A category has one cause: Beauty is provisioned on expiry, and the near-expiry units leave the over-twenty-six-week block first, so its age provision is nil rather than charged as well. And a provision has a location, so the shop-floor and warehouse lines add to the book. The P&L takes the movement of that ledger and never its level, which is why a provision released by a markdown appears once, as a smaller charge, and never again as extra margin on a cost-layer panel or as a saving on the action board.
S&OP identities: outbound gross units = the units the group sells, split by category share; closing stock = opening + inbound + returns to stock − outbound; forward cover = closing ÷ trailing 4-week outbound, on a trailing window that walks out of the actual weeks into the projected ones, computed by one function, projectCover(category, week), on the committed PO book and read by the stock flow, the planned-productivity table, the forward-cover chart and the scenario; sell-through = outbound ÷ (opening + inbound). The ordered → invoiced bridge applies stated erosion factors (shipping fees, gift cards, cancellations by origin, fraud, backlog in/out, returns by reason) to ordered value excluding Click & Collect, and reconciles demand-side sales to invoiceable revenue. Brands, media levers and care reasons are modelled splits that rescale to the group weekly totals, like every other dimension on the page.
Financials & CRM: the P&L expands the steering margin into stated cost ratios (customer care computed from the care model) and reconciles to the same contribution figure as the Sales Overview; the GM bridge and operations bridge are weight-based decompositions whose steps sum exactly to the movement they explain; recruitment, multi-category baskets, funnel steps and top products rescale to the week's new customers, transactions, sessions and AOV respectively. The S&OP forward projection chains the stock identity six weeks ahead at the committed intake, and solves the intake ratio that lands cover mid-band — a deterministic calculation, not a forecast model.
Onsite, merchandising and the initiative portfolio: search conversion is solved, not typed — given the share of sessions that use search and the observed ratio between searchers and non-searchers, the two conversions are the pair that reconciles to the conversion of the surface, so the panel can never contradict the funnel. Recommendation surfaces carry two figures on purpose: attributed revenue, which counts orders containing a clicked product, and the part of it proved incremental against a held-out control, which is the only one used in a plan. Push and in-app send volumes are built from an app base derived from app sessions and visits per user, so they move with the same traffic as every other panel; the orders and the revenue those sends produce are levelled to the owned-channel table on CRM & Media, which is the one place owned channels are sized, so this tab and that one cannot state two figures for the same notification. The copilot paths model three effects separately — selection (advisors receive the conversations with the strongest intent, so their order rate rises as their share falls), handler skill (solved from today's two order rates once selection is removed), and escalation (unresolved conversations return as a contact-centre contact and a lost sale, at a rate that rises faster than the volume automated). The initiative portfolio sizes each line against the run-rate of the business it touches, separates the contribution of incremental sales from the effect on the rate of the business that already exists, and states a confidence weight rather than implying one.
Channel taxonomy. A sale belongs to the channel where it is created; how it is served is a route. Endless aisle is a store sale served from the warehouse, so it is retail. Ship-from-store is an online sale served from a shop floor, so it is web & app. Omnichannel is Click & Collect and nothing else. Digital stays web & app + omnichannel + marketplace, which is why digital penetration reads 11.1% of total business rather than the 11.5% it read when endless aisle was counted online. The three routes keep their own panels on Logistics with their cost to serve and their promise, because the route is what the operation is planned on even when the sale is booked elsewhere.
How the cost ratios followed the sales. The P&L carries a cost ratio per channel, not per route, so the old omnichannel column had to be split before it could move. Each line of it is shared across the three routes by the store cost the route actually incurs — AED 6.2 an order for Click & Collect, 21.8 for ship-from-store, 9.4 for the endless aisle — times the volume of the route, and each part is then attached to the channel the sale is created in. Two rules are stated rather than weighted: a Click & Collect order is never delivered, so it takes none of the last-mile line, which goes to ship-from-store and endless aisle only; and gross margin follows the sales rather than the cost, so it is split on volume. Orders are taken proportional to sales inside the omnichannel layer, the only order split the model supports. The identity the split preserves, verified at load: for every line of the P&L, the AED charged across all channels is the same before and after, to the dirham.
Attach at Click & Collect pickup. A customer collecting an order buys something else on the shop floor often enough to change the economics of the service. The attach rate and the size of that second basket are stated per store cluster; a country reading weights them by the clusters that country operates, on the same store matrix every other store panel uses. The basket is applied to the average value of a collected Click & Collect order, which is the announced base. That second sale is created in store, so it is booked in Retail and counted once: it is shown beside the omnichannel line as a consequence of the service, never added to it. The attach is earned at the counter, so it is stated on the orders that reach it. An order cancelled before pickup produces no second basket, so the rate applies to the orders collected and the basket to what a collected order is worth. The order counter itself does not move — it is the created book, the same 12,980 the route counters and the channel equation read — and both are printed, because a panel that shows one of the two invites the reader to divide by the wrong one.
Click & Collect cancelled before pickup. This event was carried four times and no two agreed: the country table stated 1.9% to 4.8% of orders, the store-cluster table 1.2% to 4.1%, the order-to-pickup distribution a flat 2% “not collected — cancelled”, and the ordered-to-invoiced bridge on Financials −2.3% of ordered value. One event, four rates, three tabs. ccCancel(country) is the one counter: the country table is the anchor because it is the operational figure a market owns, the cluster rates keep their own shape and are rescaled onto it exactly as the fulfilment mix is rescaled onto the route counters, and the bridge step is derived from it on the stated assumption that a cancelled order carries the basket of a collected one. Not collected is a different event and keeps its own line: the customer never came, the order was picked, held and restocked. The cause split is not stated per market. One announced input gives the group share caused by preparation — a line short at the moment of picking, or a promise out of time — and it is distributed across markets on the availability the inventory policy computes, which is itself built from that market's lead time, review period and service target. Egypt is served on a five-week lead time at a 93% target and holds the lowest availability of the five, so it cancels most in preparation; the UAE, one week from the hub at 97%, cancels least. The spread is a consequence of the stock policy rather than a second opinion about it, and the deviations are scaled rather than clipped so the weighted mean stays exactly the input. What it costs is three kinds of money and they are named apart: the contribution on turnover that never invoiced — the turnover itself is not a value, it was never recognised; the pick already done, at the same 9.5 minutes a collect pick takes on the labour panel and the market's own loaded hourly rate; and the attach that never happens. The second announced input is the share of customer cancellations that arrive after the pick has started — a preparation cancellation is picked by definition.
Price position by category. The page carried a price index twice before this panel: by country on the market-share panel, with its own online and store columns, and by range family on the price-architecture panel. Neither is by category, which is the grain a trading meeting decides in, and a third free-standing index would have put three price positions on one tab. So the category index is an announced deviation and it is recentred: the sales-weighted blend of the six categories is the country panel's index on whatever scope is selected — the blend of the five markets, one market under a country filter, the online column under a digital channel and the store column under retail. It is the same recentring the inventory policy uses to spread availability across markets without moving the level, and it is a check. Brand deviations are recentred the same way onto the index of the category each brand sells in, so the cloud describes the dispersion inside a shelf rather than restating its level; the brand table labels its categories differently from the stock model, so nothing maps onto Menswear and that is stated rather than papered over. The horizontal axis of the cloud is the share of sales of the latest week, which is what the model carries: it holds no twelve-month brand history and inventing one from a growth rate would be a figure nobody could reconcile. The SKU matching rate is part of the reading: an index on 47% of an assortment is a weaker statement than the same index on 76%. Announced input: the full-price elasticity, stated as a share of the aged-block elasticity the Markdown Engine already carries, because CAT_EL is the response of a stuck unit — one that only moves on price — and using it on full-price demand would price every price move at roughly twice what it earns. One ratio rather than six new numbers. What a permanent move is worth is a margin figure and not a volume one: closing the gap costs the price on every unit that was going to sell anyway, which is pure margin, and buys elasticity times the gap in extra units at the category's own contribution. It is positive only where the elasticity clears one over the contribution, which is why most price moves are a share decision taken with the margin cost in front of you — and why the two halves are printed apart.
The loyalty programme is a cost, and it is not the discount. The offer table already carries a Loyalty rewards mechanic, 7% of the face value granted: that is a discount off a ticket and it is already inside the markdown, the gross margin and discountFacts(). The line added to the P&L is a different object — the value a member earns on spend and later collects as goods. It never touches a ticket, it leaves the building as merchandise and it costs what merchandise costs: the cost of goods of the channels a member shops on, at the cost layer the accounts recognise. The two are disjoint by construction. Loyalty rewards — collected, at cost sits under CRM & owned channels in the variable block and is read by the columns on the member orders each one takes, which is why the store column carries most of it: members shop mostly across a counter. It is 2% of member spend collected at a stated rate, so it moves with the sales that earn it; putting it below the contribution line so the media payback floors would not move was a convenience rather than a treatment. The floors move, and the payback panel states the movement. A point is a material right, and the statement says so (IFRS 15). The same standard is applied two paragraphs away to state the marketplace as an agent, and it applies here too. Part of the transaction price is allocated to the point and deferred until it is redeemed or expires, and the three movements are stated in the memo under the P&L: the value granted and deferred out of net sales, the value redeemed and recognised as the goods go out, and breakage recognised at the rate the redemptions run — the part expected never to be collected, which is a product of the measurement and not "information". In a steady state the three offset, so the operating result does not move; what changes is that the balance sheet stops being silent and carries the contract liability outstanding, a week's grant over the weeks a point stays alive, which is the one announced input the treatment needs. The other-costs panel keeps gift-card breakage separately, because a gift card is cash received and a point is revenue deferred. The member base is derived rather than stated: member turnover divided by what a member of each status spends, summed. A base that cannot place the turnover its own members are credited with is a base that is wrong, not a base that needs a scaling factor. The frequency and the basket cancel in that derivation, so the tier mix comes out exactly as the tier table states it, and the card bridge keeps its movements' shape levelled onto the derived base. Announced inputs: the earn rate on member spend and the collection rate by status. The threshold to the next status is not announced: a member reaches it by spending what that status spends, which the tier table already carries. The programme has no budget version anywhere in the model, so its rates are held across the scenarios and only the sales they apply to move — the same treatment the return rate and the shipping rate get.
One device split for the page. The sales equation by device is the source of truth: the onsite surface table and the funnel step multipliers are both derived from it at load, so the funnel, the search panels and the device table cannot state three different conversions. The step multipliers keep their stated shape — which step each device wins or loses on — and their level is calibrated so the product of the five steps equals that device’s conversion index. The app and web funnels are levelled the same way and on the same source: the typed pass rates are the shape, one factor per branch is solved so the product of the branch’s steps is that surface’s conversion in the device table — the app against its own row, the single web branch against mobile web and desktop blended on their sessions — and the factor multiplies each step’s loss rather than its rate, because a rate rescaled upward can pass 100% and a loss cannot. The install and the first open keep their announced level: they are per-impression acquisition steps above the session a conversion is measured on. The account screen is a step of the app funnel and a share of the web’s checkout step, printed beside it and never multiplied into the chain, because the app asks at the first open and the web asks at checkout. Tablet and other surfaces, 2% of sessions, are folded into desktop on a session-weighted blend rather than shown as a fourth line nobody acts on. Search behaviour is measured on two surfaces only, the app and the web, so mobile web and desktop share the web search profile; that is stated rather than split into a third row nobody measured. Frequency by device is an announced input, not a derived one: the model carries frequency by loyalty tier and by banner and never by device, and the case for an app cannot be argued without it.
Failing search queries are priced by intent. A product question that finds nothing costs a sale, and is priced on the demand behind it: queries × exit rate × searcher conversion × searcher basket. A service question typed into a search box costs a contact, and is priced on the contacts it creates at the blended cost per contact of the care model. The two are never added: one is revenue at risk, the other is cost avoided, and putting them in one column is how a search backlog gets mis-ranked.
The window a release decision is allowed to read. A dynamic tranche is released day by day on a ROAS that is not finished: conversions land after the click and the platforms backfill them, so the same spend reads 62% of its eventual value at D+1 and 94% at D+7. The curve and its day grid are an announced input and there is one of them on the page — the stabilisation chart, the maturation panel and the release rule all read it. Two rules follow and both are written down rather than assumed: the decision window closes at D+7, because earlier than that the reading is noise dressed as a number and later than that the tranche has already been spent; and the floor is compared to the mature figure, the observed one divided by the share of the return the curve says has arrived. Reading yesterday's number against a floor is not a stricter policy, it is a different and wrong one, and it is biased against precisely the levers with the longest consideration window. The maturation and the incrementality are two different corrections applied in that order and never blended into one fudge.
The contact belongs to a lane, and the refund to a reason. The care panel says how many contacts there are and what they cost; the carrier table says who delivers. Neither said which carrier the contacts come from, which is the number that decides a lane change — and a contact is what a broken promise costs after the delivery has already been paid for, charged to customer care rather than to the lane that caused it, which is how a cheap lane keeps winning a cost-per-order comparison. The rate here is derived and not stated: a contact for “order not received” scales with the lane's own late rate, one for “improper order received” with its damage and loss rate, and the rest of the care mix is carrier-neutral — three figures the care model and the carrier table already carry. The volume is the care panel's: the lanes are normalised onto the digital contact volume, so the table redistributes contacts and never creates any, and the lines that travel on no carrier of ours — collect and the marketplace — are on the table rather than dropped, which is what makes the total the whole digital book. A ship-from-store parcel is handed to the same carriers and travels in the same mix. The refunds are the ordered-to-invoiced classification the Financials tab already carries, applied to the returns the statement credits on the two channels we deliver ourselves — a marketplace return goes back to the seller and never touches our cash — and the refund float below the other-costs table is that same base carried for 12 days. A marketplace refund never touches our cash, so it is not in that book.
Who a brand actually recruits. The brand table says what a brand sells and at what margin, and the recruitment split says how many customers the week brought; neither said which brands did the recruiting, which is the question a delisting conversation turns on. Nothing new is stated: a brand's share of the week's new customers is its share of sales times the recruitment power the brand-power panel already plots, renormalised so the twelve brands and the long tail add to the week exactly — so the index printed on the brand table is that panel's, and its sales-weighted mean is one by construction. The split between a customer new to the brand and a store customer new to the channel follows the brand's own web share of its sales, the column two places to the left: a brand sold across a counter converts the customers it already has, one sold online has to win them. It is renormalised onto the group's new-to-brand share from the recruitment split table, scaled rather than clipped, so the two tabs share one recruitment book. The long tail carries the group index because no recruitment is measured on it, and it sits in the total row rather than in a row of its own.
Generative AI referrals. An assistant that answers a shopping question and links a product sends a visit that lands in the analytics as organic search: no campaign, and a referrer nobody has a bucket for. It is small and it grows at a rate nothing else on the acquisition table grows at, which is exactly the combination that stays invisible inside a big line until it is no longer small. It is carved out of organic search and never added beside it — the shares of sessions and of revenue come off the organic line, so the table still adds to one hundred and the digital week is untouched. Its share, its growth and its incrementality are announced: the model carries no referrer-level split of its own, and a server-log or analytics export would replace them. Its incrementality is the earned figure of the line it comes from, because it is the same kind of visit.
Next iterations: v2 fits channel and country elasticities from transaction history instead of assumed parameters; v3 automates the weekly refresh and posts the brief to the trading channel.
Decision rules
The seven rules behind This week’s decisions on the Sales Overview tab are deliberately simple and readable: each one is a test on figures another panel already computes, and a value formula built from the same figures. The threshold of every test is a stated parameter rather than a hidden one, so a rule can be argued with, moved, or switched off without touching the rest of the page.
Two columns of value, and one of them is not a weekly result. Every rule states its unit, and the board carries two: P&L / wk, a result the week actually carries, and One-off, money recognised once. Stock at cost is neither and is never a value on the board — it is a balance, it belongs in the signal, and pricing an action at stock times margin values the same dirham again every week it stays unsold. The provision a cleared block releases is a balance-sheet movement and belongs in the one-off column, not divided across the weeks of a clearance window. The release is now in the one-off column, capped at what the provision ledger has actually taken on the category, and the weekly column carries the holding cost avoided and the trading contribution of the clearance route and nothing else. The total of the weekly column is stated against the operating result of the same week and the same selection, so the board can be read as a share of what the business earns rather than as a number with no denominator.
Grain is stated per rule because it is not always the grain the question is asked at. Assortment ratios are a group discipline; stock ratios are set by market. Sell-through and buying depth answer how well a range was chosen and bought, and that judgement does not change at a border, so they stay group. Cover and availability answer how much stock a market has to hold to serve its customers, and lead time, demand variability and the service level the business is willing to pay for all differ by country, so they are computed per market by the inventory policy and read through coverOf(category, country), targetCoverOf(category, country) and availOf(category, country). The discount rate is carried by channel and country and not by category, so the discount rule is stated where the figure lives. The out-of-stock panel has no country dimension, so its rule names no country to expedite from.
The P&L, and where the markdown is deducted. The statement is structured the way an omnichannel retailer closes one: a gross-to-net bridge, a gross margin, a contribution after the cost of serving the order, an operating result after the costs that do not move with it, in six columns because a store, a website, a click-and-collect counter, a marketplace and a wholesale account do not share a cost structure. The turnover every other tab carries is ordered demand, net of VAT and after the discount the customer received at the till, so gross sales at full ticket is derived from it: turnover plus the face value of the markdown granted. The markdown is deducted once, in that bridge, at face value; the part the supplier funds comes back below as a credit to cost of goods, because a supplier-funded markdown is a reduction of the purchase price and not a sale. Charging it a second time as a cost line under a gross margin labelled "at full ticket price", applied in AED to a turnover already net of it, would deduct AED 14.5M a week twice. Returns are the second correction. Ordered demand is gross of returns and the statement deducted them nowhere, so the credited sale appeared only as a processing cost; they are now a deduction from gross sales at one rate per channel, read from the channel-economics table that already states it, and the goods that come back release their cost of goods with them. Cost of goods is the full-ticket cost of what shipped at the layer the cost method recognises, so the FIFO / weighted-average / lot toggle on the cost-layer panel moves the statement, which is the point of having it. The marketplace is stated as the agent position it is (IFRS 15): net sales there IS the commission, the merchandise value and the seller payout are the seller’s and sit in a memo under the statement rather than inside our revenue. Stating the gross merchandise value as net sales consolidated a seller’s turnover into ours and made the group operating margin read on a base that included it. A line a channel does not have prints an em dash and never a zero, because zero is a measurement and an em dash is a structure. Two gross margins exist on purpose and both are labelled: the full ticket rates, which are the rate a ticket price is derived from and which the cost layers, the Markdown Engine and the bundle panel read, and the P&L's own gross margin, after markdown, returns, supplier funding and payout, which the bridge and every panel valuing a revenue at a margin read.
The store network is generated once, and everything about it is an aggregation. Four hundred and thirty-five stores, each with a country, a cluster, a turnover, a selling area, a week of paid hours and an opening date. The cluster scorecard, the store journey, the labour panel, the store S&OP table, the Top 20, the like-for-like, the route counters and the two store keys of the P&L all read that one list. The turnover spread inside a cell is a symmetric ladder ordered by a hash of the store index, so its mean is exactly one and the cell sums to the dirham, no two stores share a turnover, and the modulo of period eleven that made three pairs of stores identical is gone. Openings are dated, ten of them, and the like-for-like falls out of the dates rather than being stated: a store that opened this year has no same week last year, so its whole turnover is the network effect and the comparable base is what is left. A new store trades on a ramp, an announced 55% of the cluster mean at opening rising to the mean at one year.
One growth against last year, at country × format. growthOf(cluster, country) is the only derivation now, and it is stated at the grain the network is run at. For every country × format cell the opening effect is the turnover of the stores that have no last year over the turnover of the ones that have one, read from the dated opening list; the like-for-like is the announced shape of the format plus one adjustment per market, solved so the cells of that market close exactly on that market’s own retail line; the total is their sum, and the cell’s last year is its turnover divided by it, so nothing is renormalised afterwards. Cluster, country and network rows are aggregations of those cells: the total is turnover over last year, the opening effect is new over comparable, the like-for-like is the difference. That is why like-for-like + openings = total on every row that is printed, why the country columns of the Deep Dive matrix are the retail lines of those markets, and why the Overview text now states the effect its own cells produce instead of a share of the store count. The like-for-like rule on the action board reads the same cells, at country × format, because a format average hides a spread three times wider than itself.
One session book, and there is no such thing as an omnichannel session. There is one session total and one marketplace base.sessionFacts() starts from one base, the digital sessions of the input file on the taxonomy the page uses, and splits it in two: the marketplace, a different catalogue and a different visit, carries its own sessions solved from its own conversion index; what is left is the own-site base, and Web & App and Click & Collect share it. A customer opens one session and chooses at the checkout whether the parcel is delivered or collected, so the collect route is stated as a share of the digital checkouts and as a conversion on the shared base — never as traffic of its own. The omnichannel equation is flagged shared, so the group visit base counts that session once, the same discipline the marketplace already had. The order counters are the route counters, 73,894 / 12,980 / 14,011, so the rows of the digital-layer table add to its total instead of being sessions times an index under a total read from the raw input column. Every share “of online sales” — media, CRM, the Journey segment table — is taken on the same digital week, and the app base is derived once from the app sessions and read by both the Onsite cards and the owned-channel table.
One attach book, on one grid. The second basket a collection triggers on the shop floor is built on a country × cluster grid and every panel is an aggregation of it: the created orders are the country’s, split on the same catchment key the route model uses; the cancellations carry the cluster’s own shape rescaled onto that market’s rate; the attach rate applies to the orders collected, because a cancelled order never reaches a counter; and the second basket is the cluster’s index on orderValue("cc"), the one order value the Overview, the channel-economics table and this panel read. The consequence is that the NETWORK row is the collected-weighted average of the four rows above it, on the rate and on the basket, and sits inside their range. The intraday table states the execution and reads the money, instead of building a second total 32% larger three panels below.
One parcel count, one contact base, one book of returns. The parcels that travel are the route counters, home delivery plus ship-from-store. The last-mile line of the statement is two lanes — the parcels a customer receives and the freight that moves a wholesale order — and only the first is what the carrier panel prices, which is what the reconciliation page now tests rather than comparing two perimeters that happened to add up. Contacts are stated per 100 digital orders on a named base, the same order book on the Financials cost-to-serve panel, the D+7 table and the lane table. The returns are one book: the units come from the rates the statement charges each channel times the units those orders carry, valued at each channel’s own realised net price, so the S&OP panel returns the units the accounts pay for. The marketplace is absent from both sides, because the goods are the seller’s and the credit never reaches our statement. The planning model keeps the category shape and the reverse-logistics lag; only the level is the statement’s — and since a return is credited to the stock book, correcting it also removes the carrying cost of units that were being credited to stock and to no line of the accounts.
The routes are counted once, per store. Click & Collect, ship-from-store and the endless aisle are carried by each of the four hundred and thirty-five stores, on one key: the cell’s own digital demand times the country’s route rate, so a store serves the collect orders of its own catchment. The store-logistics table, the attach table and the intraday table read that one split. One order value per route, its own sales over its own orders, and the Omnichannel channel and the Click & Collect route state the same basket because they are the same object. Home delivery is Web & App less ship-from-store and nothing else. A marketplace order is shipped by the seller, so it is not a fulfilment route of ours at all: it carries its own counter, it has no lane in the warehouse allocation, and that is the same reason the P&L gives it no cost of goods and no warehouse.
One session split, and it is derived. The share of sessions is derived, not stated.The inputs are now what a session of each kind is worth: a conversion index — App 1.18, Desktop 1.14, Mobile web 1.02, in that order, because an app session is opened by a customer who is already signed in, a desktop session is deliberate, and a mobile-web session is the unauthenticated end of a search click — and a basket index, App 1.06, Mobile web 0.86, Desktop 1.26, a desktop basket being the largest and a mobile-web one the smallest. The share of sessions is solved from those indices and from the app turnover the route model carries: there is exactly one app share that produces it. The relative split of the non-app sessions, mobile web against desktop, is the announced input that closes the system. The result is that the app is under half the sessions and not seven tenths of them, which is the difference between counting app opens and counting sessions that reach the site. The sales equation by device, the device table beside it, the onsite surfaces and the traffic-source panel all read that one split; the App line of the route table on Deep Dive is the route alone and is labelled that way. The ordering of the three conversions is an identity on the reconciliation page, so the solve cannot invert what the inputs state.
Replenishment is solved from a rate. The share is now whatever ninety units an hour requires, an announced input, and the hours that frees go to selling, where the labour panel already shows the network short. The payroll does not move: it is hours times the country rate, and neither changes. The warehouse has one stock. The inventory control scorecard reads the orchestration’s own units for stock speed, turnover and days on hand, so a flow-through node the same tab describes as 2.2M units and 1.4 weeks can no longer be stated at 14.7% speed and 53 days. Days on hand decompose rather than average: the days a unit spends in the building plus the days it spends on the shop floor are the days it spends in the book.
Fixed costs are allocated on what drives them. A fixed cost spread pro rata of sales says the marketplace consumes a warehouse it never touches and that a click-and-collect counter is free because its turnover is small. Each family is allocated on its own driver, and the family total is unchanged by the allocation, which is one check per family on the reconciliation page. Stores — Retail and Omnichannel only. The shop floor is retail’s except for the standing capacity the collect service occupies: the counter hours the rota reserves, taken from the picking hours the Logistics tab already computes as orders times minutes per route, against the paid store hours of the same store model; and the collect desk and holding shelf, an announced input of 14 m² a store, against the sales area of the network. The payroll and occupancy halves of the block are split from the store model’s own rota and hourly rates, and each half takes its own driver. The pick of each order is a variable cost and sits in fulfilment above; this line is the space and the rota, which is a different cost. Warehouse fixed and inventory carrying — the channels with a stock flow, on units through the building from the warehouse flow: store replenishment and the endless aisle to Retail, home delivery and ship-from-store to Web & App, click-and-collect to Omnichannel, the wholesale lane to Wholesale. The marketplace holds no stock and moves none of ours, so it carries none of this and prints an em dash, the same as its cost of goods. Technology and platform — the digital channels, on orders, because a till transaction does not call the order-management system. CRM and owned channels — Retail, Web & App and Omnichannel, on the orders the owned channels are credited with, with an announced input of 35% of them redeemed in a store rather than online. Brand and commercial animation and overhead — every channel, on net sales, because neither has a causal driver in this model and saying so is better than inventing one; with one exception that is not a convenience: on the marketplace, net sales is the seller’s merchandise value and not our revenue, so the key there is the commission we keep. Allocating group overhead on somebody else’s turnover charged the marketplace eight times what it earns.
Media payback floors. One definition of contribution, and it is a line of the P&L. A dirham of media does not pay for itself when it covers the gross margin. It pays for itself when it covers what the order leaves once it has been served and once the discount that won it has been granted. That line is the Web & App column of the P&L on Financials, printed as a memo under the statement: gross margin — already after the markdown granted, the goods that came back and the supplier funding credited — less the variable cost of serving the order: fulfilment and packaging, the last mile including its cash-on-delivery and failed-delivery share, reverse logistics, payment fees including cash handling, and customer care. Media is not deducted, because media is what the floor is judging, and nothing else is deducted, because nothing else moves when one more order is served. It is stated on the column's ordered turnover and not on its net sales, because a ROAS is measured on the order value the platform attributes and the same dirham has to be on both sides of the division. The transactional floor is one over that contribution. It reads 2.72, against 3.43 when the markdown was still being charged twice, 2.86 when the markdown was left out of the base entirely and 2.19 when the floor was one over the budget gross margin; and it is the same number on the payback panel, on the dynamic budget and on the year secured in advance.
The acquisition floor, and the input it rests on. A source that recruits is not judged on the order it brings. R = 26% × 3.01 × 0.982 = 0.769: the eight-week repurchase rate of a new customer, read from the loyalty-state table on Journey, times the orders a repeating customer places over the following twelve months less the first, times their basket index. Announced input: the model carries no order frequency of its own for the new-customer cohort, so the frequency and the basket index stay the loyalty tiers’ figures, which describe a customer who has come back; cohort frequency and cohort basket from transaction history would replace them, and they are the first two numbers to fit against real data. The acquisition floor is one over the contribution times (1 + R), 1.94, and the effective floor of a source is the two weighted by the share of its attributed orders that come from a customer new to the brand. That share is an announced input too — Paid search 34%, Paid social 62%, Affiliation 41%, Retargeting 11%, Display 48% — because the model carries new customers by commercial mechanic and never by media source. The acquisition floor is a twelve-month contribution floor, not a cash floor: the first order alone returns the transactional floor and the rest arrives over the year, so acquisition spend is capped by working capital as well as by this floor. CAC is stated beside it, media over the new customers media brought, so the two are read together. The point of the two floors is that an acquisition source is judged on the value of the customer and not on the value of the order, which is why retargeting carries the highest floor of the five: it re-sells to customers the business already owns.
Owned channels are sized once. Email, app push, in-app messages, SMS, the loyalty statement and web push are sized on the owned-channel table of CRM & Media and nowhere else. The push and in-app panels on Onsite keep their own shape — which notification carries the volume, which one earns its sends, which one burns the file — and their orders and revenue are levelled to that table, scaled to the share of the digital week the Onsite selection covers. One owned-channel incrementality, the revenue-weighted holdout result of the owned lines on the acquisition table, 65.1%: the growth bridge opens its slider on it rather than on a typed 28%, and both pages now say the same thing about how much of a push is demand created rather than demand moved forward.
Discount by category. The rate is carried by channel and country. The offer calendar carries, by category, the mechanic run each week, the participation of the category and a depth index against the group rate, and that is enough to derive a rate by category, channel and country. It is only worth having if it reconciles, so one scale factor ties the discount it implies in AED to the discount the channel-and-country table already states, and the factor and the identity are printed on the offer panel. The plan side carries no discount by category anywhere, so the same offer structure is assumed on the plan as on the actual; a plan built by category would replace that assumption. The rule that reads it keeps the channel in its grain, because retail is four fifths of the base and runs a small gap to plan: a rate blended across channels never reaches the threshold, and blending would hide exactly the cells that matter.
Promoted stock velocity. The ratio between a promoted and a non-promoted unit is not a stated constant per category. It is now read from the price response the Markdown Engine already uses to clear a block, (1 − depth) to the power of minus the category’s elasticity, evaluated at the depth the category actually runs — its reconciled discount rate divided by the share of it sold on a mechanic, because a rate on all sales understates the depth granted on the slice that carries it. The consequence is the one a trading meeting needs: a deep markdown on a slow category still turns slowly. Home & Lifestyle at 36% depth moves 9.7% of its stock a week against Beauty at 15% depth moving 30.2%, so the response factor is higher on Home and the speed is three times lower.
Bundles. Incremental units = units × (1 − substitution rate). Margin after substitution = incremental units × the bundle’s cash margin, less substituted units × the margin given up when a full-price hero becomes a discounted bundle. Substitution rates are an announced input in the 45 to 75 per cent band; measured rates from a held-out control would replace them. Kit availability is the minimum of the components’ category availability, read from the market inventory policy, because a bundle is only sellable when every part of it is. Bundle revenue is a reallocation of the categories’ existing sales and not an addition: the turnover of the six is already inside the sales of their hero categories, nothing on this page is grossed up by them, and the incremental units are attach units financed by the hero’s margin rather than new demand. The fifth clearance route on the Markdown Engine is priced the same way, on the margin after substitution and never on the units.
Cost layers. The same SKU enters the book at different costs depending on when it was received, and the model carries one cost per category, which is the blend. It is split into the two layers the ageing data distinguishes, the block over twenty-six weeks old and the rest, solved so their weighted average is exactly that blend: ccurrent = cbook ÷ (1 − share over 26 w × inflation), cold = ccurrent × (1 − inflation). Intake inflation between the oldest and the current layer is an announced input per category — Womenswear 9%, Menswear 8%, Kidswear 7%, Footwear 11%, Beauty 5%, Home & Lifestyle 14% — and the old layer carries the whole gap. Break-even discount d* = (ccurrent − cold) ÷ ticket, the depth at which a discounted old unit stops matching a new one at full price on cash. Economic margin = ticket margin on the units, minus the holding cost over the remaining cover at the scorecard’s own rate. The provision released is not added here — it is a balance-sheet movement recognised once, it is stated on its own line under the ledger and in the one-off column of the action board, and adding it to a trading result is how the same money came to be counted twice. The ticket is derived exactly as the Markdown Engine derives it, cost over one minus the renormalised category margin, so the two panels price the same unit at the same shelf price.
The accounting distinction, which is the point of the panel. Under strict FIFO on one SKU the accounts deplete the oldest layer whatever unit physically ships, so choosing to sell the old carton does not change the cost of goods recognised: what the decision is worth is the provision not taken or released, the holding cost avoided, and the pace, because selling faster consumes the cheap layers sooner. The cost-of-goods lever is real only where the layer is a different accounting object — lot costing or specific identification, where the season code changes the SKU — or under weighted average, where the blend itself moves. Intake inflation between layers is a pricing question and not a rotation one: at the same ticket the new layer earns less, and the answer is the price. The method toggle recomputes the columns on all three bases, and the three answers are deliberately different.
Layer mix on the margin bridge. The layer effect is not a step added to the bridge, it is a sub-line of the supplier-cost step. That step is split into what the current intake costs against plan and what consuming the old, cheaper layers gave back, and the two sum to the step: the eight stated steps keep their own values and their signs, nothing is rescaled to make room, and the bridge still closes on its own sum. When the layer line is larger than the step the intake line turns negative, and that is the true reading rather than an artefact — the old layers helped the rate, the current intake costs more. The effect itself is computed on the cost of goods, so it does not depend on which sales base is read: the book sits older than the excess-and-obsolete target the plan was built on, so it draws on cheap layers the budget did not price. Under FIFO this is a pace effect and not a choice of unit: the accounts deplete the oldest layer whatever ships, so selling faster consumes the cheap layers sooner and the gain runs out earlier; it is not earned by picking the old carton.
Beauty is provisioned on expiry, not on age. The share of its book inside six months of expiry is an announced input, 9%, provisioned in full. Those units are taken out of the over-twenty-six-week block first, so the age provision applies only to whatever is left of that block, which on the current book is nothing. Age is secondary for a dated product: a unit is not written down because it is old, it is written down because it will not be sellable.
The inventory policy. Weekly demand d is the category’s trailing four-week outbound, split across markets on that category’s own country mix, so the group cover of every category is exactly the one the stock flow panel states and the policy only distributes it. σd = d × the category’s coefficient of variation × the variability factor of the market. Cycle stock = d × R ÷ 2. Safety stock = z × √((LT + R) × σd² + d² × σLT²), z being the normal quantile of the service target, computed from a rational approximation rather than a table: z(93%) = 1.48, z(95%) = 1.65, z(97%) = 1.88. Target cover = assortment cover + (cycle + safety) ÷ d. Actual cover = stock ÷ d, the stock being the target times an execution factor and renormalised, one factor per category, so each category ties to its own line of the group book. Expected availability = Φ(z of the actual position), taken on one option of the assortment plan rather than on the category aggregate, because availability is a shelf measure and a category holding eighteen weeks of cover can still be out of the middle of its size curve; the spread is then recentred on the availability the control table states, so the policy sets the difference between markets and not the level.
Seasonal and replenishable. A seasonal category is bought once for a season, so the middle of its cover band is the assortment it has to present and the replenishment component sits on top. A replenishable category is reordered continuously, so its target is a presentation floor — the bottom of its band, the least a shelf can hold and still be shoppable — plus the replenishment cycle and its safety stock, capped by shelf life rather than by a season, Beauty at 26 weeks. The two kinds therefore answer the same question with different weights: for Beauty the floor is 6 weeks and the market component adds between 1.5 and 3.6 depending on the lead time it is served on, so a market with a five-week lead time has to hold materially more of the same category than one replenished weekly from the hub.
Execution factors e(c,k) above: what the book actually carries against the policy it should carry, stated and relative within a category. The unit costs of the categories are calibrated once at load by a single factor, currently 1.57, so that units × unit cost equals the cost of goods implied by the plan at plan margin; cover, sell-through and every ratio are unaffected by it, only the valuation is.
Product affinity — what the transactions identify, and what they do not.
Integrated business planning — the four reviews, and what forecast value added would need.
The risk & opportunity register — the amount is a function, the probability is the input.
Working capital — three legs, two bases, and what is deliberately not recomputed.
Four rules read panels that carry no category dimension at all: the out-of-stock lines, media, onsite search and the delivery promise. Under a category filter their value is shared out on that category’s weight in the turnover of the selection, and the line itself says so. That weight is an allocation key and not a measurement — it is there so that the six categories add back to the group rather than each carrying the whole of it. The out-of-stock lines do carry a product category, but named on a different taxonomy than the six categories of this page, so they are shared out with the others until the two are reconciled. Note also that a category filter can produce more lines than the group rather than fewer, and that is the point of it: a category can be behind plan in a market whose overall achievement looks perfectly healthy, and a rule reads the cell rather than the average above it.
