The data product lexicon

Semantic layer#

A shared definition layer between raw tables and the tools that read them, where a business metric like revenue or active user is defined once so every dashboard, query, and agent computes it the same way.

Assumed: a BI feature, the place your dashboards keep their numbers.

Actually: the contract that makes "revenue" mean one thing across every tool, query, and agent. It is the layer most teams gesture at and never finish, which is exactly why two dashboards disagree.

Data contract#

A formal, versioned agreement between the team that produces a dataset and the teams that consume it, specifying the schema, the meaning of each field, and the quality and freshness guarantees, checked automatically so a breaking change fails a test before it ships.

Assumed: documentation, a schema written down somewhere.

Actually: an enforced agreement with consequences. If no test fails and nobody gets paged when the shape changes, it is a wish, not a contract.

Data product#

A dataset, model, or metric that is managed like a product rather than a one-off export: it has a named owner, known consumers, a service-level agreement, and a lifecycle of versions and deprecations.

Assumed: a dashboard, or a dataset someone published.

Actually: something with an owner, a consumer, an SLA, and a lifecycle. If nobody is on the hook when it breaks, it is exhaust, not a product.

Single source of truth#

One agreed place, and one agreed definition, that every team refers to for a given fact, so that the same question returns the same number no matter who asks.

Assumed: a database, usually the warehouse.

Actually: an agreement about which definition wins. No database enforces truth by itself; people and governance do. A second "truth" is one rogue spreadsheet away.

Reverse ETL#

The practice of copying modeled data out of the warehouse back into the operational tools people work in, such as a CRM, an ad platform, or a support desk, so that trusted numbers are usable at the point of action.

Assumed: a niche sync tool for moving rows around.

Actually: the step that operationalizes the warehouse, pushing modeled data back into the tools people actually work in. Analytics nobody acts on is a cost center; this is where it becomes an action.

Data governance#

The policies, roles, and standards that decide who owns each data asset, who may access it, what quality and definitions it must meet, and how it stays compliant with regulation.

Assumed: compliance, the department that says no.

Actually: the layer that lets you move fast without breaking trust. Done right it is an accelerator, the thing that makes self-serve safe, not a brake.

North Star metric#

The single metric a company chooses to organize around because, better than any other, it captures the core value customers get from the product.

Assumed: the one number the company should grow.

Actually: the metric that best predicts durable value for the customer. Pick the wrong one and you optimize the company straight into a wall, efficiently.

Resolution#

The grain at which a dataset captures reality: what one row represents, and therefore the smallest question the data can still answer. A daily sales rollup is lower resolution than an order-line event stream; you can always aggregate resolution away, but you can never add it back after collection.

Assumed: image sharpness, or a synonym for having lots of data.

Actually: a product decision disguised as an engineering one. The resolution you capture sets the ceiling on every future question, model, and agent. Storage is cheap; a question the grain cannot answer is gone forever. Choose resolution for the questions you will want in two years, not the dashboard due this quarter.

Altitude →

Altitude#

The level in a dimensional hierarchy at which a question, metric, or report is flying. Revenue across all channels sits at one altitude; channel, source, campaign, and keyword are each a step down the same ladder, and every hierarchy has one: region to store, portfolio to SKU, year to day. Every number has an altitude whether or not anyone chose it.

Assumed: summary versus detail: executives fly high, analysts fly low.

Actually: a property of the question, not the audience. The skill is naming the altitude a question is actually asking at, answering at that altitude, and knowing how far you can descend before the data's resolution runs out. Most bad data conversations are a question asked at one altitude and answered at another.

Resolution →

Back margin#

The money a retailer earns from suppliers rather than shoppers: trade spend, slotting fees, display allowances, and volume rebates, paid for shelf access and promotion before a single unit sells.

Assumed: a discount the retailer negotiates on wholesale prices.

Actually: a second P&L running behind the shelf. For grocers it can rival front margin, and it quietly reprices the shelf as advertising inventory: the brand pays for position, the shopper pays for product.

Retail media#

Advertising sold by a retailer on its own properties and data: sponsored placements on its site and app, in-store screens, and offsite ads targeted with its shopper data.

Assumed: banner ads on a grocery website.

Actually: the retailer monetizing its closed-loop data: it knows what you bought, so it can prove an ad worked. Software margins on top of a two-percent business, which is why every large retailer suddenly runs an ad platform.

Planogram#

The diagram that dictates exactly where every product sits on a shelf: which aisle, which height, how many facings, next to what, typically generated from sales data and reset on a schedule.

Assumed: a stocking chart for shelf workers.

Actually: a ranking algorithm rendered in steel: eye-level is position one, the end-cap is the featured slot, and a reset is an algorithm update shipped to physical space. Whoever controls the planogram controls the market share.

TAM (total addressable market)#

The total annual revenue on offer if you sold your category of product to every possible customer who could ever buy it, with no competition and no reach limits.

Assumed: the big number at the top of the deck.

Actually: a boundary you choose to draw, and drawing it is the strategy. Uber's TAM was never the taxi fleet. The number follows the boundary, not the other way around.

SAM (serviceable addressable market)#

The portion of the TAM that your current product, business model, and sales channel can actually serve, before you subtract competitors.

Assumed: TAM minus a bit, some vague fraction of the big number.

Actually: the slice of the total market your product, model, and channel can actually serve today, before you even subtract competitors. It is the honest ceiling on the business as it is now, not as you imagine it at scale.

SOM (serviceable obtainable market)#

The share of the SAM you can realistically win in a given period, given your team, budget, and the competition already in the field.

Assumed: TAM times some confident-looking percentage.

Actually: what you can actually operate to, built from reps you can hire, deals they can close, and contract value that pays for the motion. If the math needs 400 deals and you can fund eight reps, it is fan fiction.

Profit pool#

The total profit earned at a given stage of an industry's value chain, which can sit very differently from where the revenue is concentrated.

Assumed: where the revenue is.

Actually: where the margin is, which is where the power is. Revenue shows you the activity; margin shows you who captures it. The PC industry ran on hundreds of billions in revenue while Intel and Microsoft kept nearly all the profit and everyone else fought for scraps.

Buyer journey#

The path a customer travels from first recognizing a need to evaluating options and committing to a purchase.

Assumed: a tidy funnel: awareness, consideration, purchase.

Actually: a trigger makes the status quo suddenly unacceptable, one worried person spends months turning that into a funded priority, and by the time an RFP exists the winner has usually already been chosen by whoever framed the problem. Nobody buys on a calm Tuesday.

Economic buyer#

The person in a purchase with the budget authority to approve the spend, as distinct from the end users and evaluators who influence the choice.

Assumed: the person who uses the product.

Actually: the smallest coalition whose incentives you can align to get a purchase that survives deployment. Often it is not the user. Employees hated Concur for two decades; finance held the budget, so it won.

Jobs to be done#

A framework that defines a product by the progress a customer is trying to make in a given situation, rather than by the customer's demographics or the product's features.

Assumed: a fancier word for personas.

Actually: the progress a customer is trying to make, and the trigger that starts them looking. People do not buy products, they hire them for a job and fire them when something does it better.

Product-market fit#

The point at which a product satisfies a strong market need well enough that customers adopt it, keep using it, and pull for more.

Assumed: a good score on a "how would you feel if you could no longer use this" survey.

Actually: retention and pull, measured by what customers do rather than what they say. People promising they would use it is the cheapest, weakest evidence there is.

Moat#

A durable, structural advantage that protects a company's profits from competitors over time, named for the water around a castle.

Assumed: a feature your competitors do not have yet.

Actually: a benefit paired with a barrier, a reason rivals cannot or will not copy you. Great UX is not a moat; anyone with more engineers copies it by Friday.

Switching costs#

The costs, in money, time, data migration, and retraining, that a customer must absorb to move from your product to a competitor's.

Assumed: lock-in, the contract that traps a customer.

Actually: the cost a customer would eat to leave, built up over time through data, workflow, and integrations. It pays off late, so it is a power you invest in before you need it.

Closed-loop economy#

An ecosystem in which a company's currency can be earned and spent only inside its own walls, so the issuer sets both sides of every exchange.

Assumed: a loyalty program with extra steps.

Actually: a structural moat: whoever owns the loop owns the exchange rate, the float, and the switching costs, and a competitor cannot enter without building an entire economy first.

Flywheel#

A self-reinforcing loop in which each part of the business feeds the next, so that momentum compounds and growth gets cheaper over time.

Assumed: a growth-loop buzzword for a slide.

Actually: a loop where each turn lowers the cost of the next one. If turn ten costs as much as turn one, you have a treadmill, not a flywheel.

Right to win#

A company's specific, defensible reason it can succeed in a particular market, separate from how attractive that market is to everyone.

Assumed: the same thing as an attractive market.

Actually: a separate axis from market attractiveness: whether you specifically can win here. Google had unlimited capital and distribution and still could not will Google+ into existence. A boring market you can dominate beats a glamorous one where you are a tourist.

Blitzscaling#

Deliberately prioritizing speed over efficiency to seize a market before rivals do, accepting waste and risk to capture network or scale effects while they are still up for grabs.

Assumed: grow as fast as humanly possible, always.

Actually: rational only inside the narrow window where network or scale effects are still up for grabs. Outside that window it is arson with a pitch deck.

Conservation of attractive profits#

Clayton Christensen's observation that when one stage of a value chain commoditizes, the ability to earn attractive profits does not disappear but shifts to the adjacent stage that has become the new bottleneck.

Assumed: when a layer commoditizes, the profit is gone.

Actually: the profit does not vanish, it migrates to the adjacent layer that has become the bottleneck. When the build layer commoditizes, the margin moves to governance and scale. Reposition to own where it went, do not defend where it left. (Christensen's law.)

Commoditize your complement#

A strategy of deliberately driving down the price of products used alongside yours, because demand for what you sell rises as its complements get cheaper.

Assumed: give things away free as a growth gimmick.

Actually: demand for what you sell rises as the price of its complement falls, so you drive the complement to zero on purpose. Give away the layer that is commoditizing anyway to pull demand toward the layer you monetize. (Joel Spolsky, Strategy Letter V.)

Positioning#

The deliberate choice of the market context in which a product is understood: what kind of thing it is, who it is for, and what alternative it beats.

Assumed: the tagline, the words on your homepage.

Actually: the frame that makes a buyer instantly grasp what you are, who it is for, and why it beats the alternative. When a product that works is still not selling, the problem is usually here, not in the product. (April Dunford, and Ries and Trout.)

Counter-positioning#

A newcomer adopting a superior business model that the incumbent cannot copy without undermining its own existing business.

Assumed: simply being different from the incumbent.

Actually: a move the incumbent could copy but will not, because copying you torches their own P&L. The barrier lives inside their income statement, which also means it expires.

Reverse positioning#

A positioning move that strips away attributes the category treats as mandatory while adding a few unexpected ones, resetting what the product is.

Assumed: a synonym for counter-positioning, or just differentiating by adding more.

Actually: the opposite of adding more. You strip features the category trained people to expect, then add a few they never asked for. IKEA drops delivery and assembly and adds a showroom and a restaurant. It resets what the product even is. (Not counter-positioning, which is about a rival who cannot afford to copy you.)

Breakaway positioning#

Deliberately associating a product with a category other than the obvious one, so buyers judge it by a different set of rules.

Assumed: just being different inside your category.

Actually: deliberately attaching your product to a different category than the obvious one, so buyers judge it by new rules. Swatch framed a watch as a fashion accessory, not horology. (Youngme Moon.)

Stealth positioning#

Introducing a product that people distrust or cannot place by framing it inside a familiar, friendly category so they give it a chance.

Assumed: hiding what your product does.

Actually: smuggling a product people distrust or cannot place into a familiar, friendly category so they give it a chance. Sony framed the AIBO robot as a pet, not a machine. (Youngme Moon.)

Category design#

Creating and naming a new market category and establishing your product as its defining example, so the buyer's frame of reference is one you set.

Assumed: marketing your product really well.

Actually: defining and naming a new category that you then lead, so the buyer's first question becomes yours to frame. You do not win an existing game, you invent the game. (Play Bigger.)

Context layer#

The structured schema, semantic, and business information an AI system needs to answer correctly about a specific organization, assembled and supplied to the model alongside the question.

Assumed: prompt engineering, the words you type at the model.

Actually: the schema, semantic, and business context an agent needs to be right. The model is rented; the context is owned.

Read: The context layer is the product →

RAG (retrieval-augmented generation)#

A technique that first retrieves relevant documents or data and places them in the model's prompt, so the answer is grounded in that material instead of the model's memory.

Assumed: the fix for hallucination.

Actually: a retrieval problem wearing a generation costume. Retrieval quality is the whole game; feed it the wrong context and you get confident, well-cited nonsense.

Text-to-SQL#

A system that translates a plain-language question into a SQL query and runs it against a database, so non-technical users can ask for data in words.

Assumed: natural-language analytics, solved.

Actually: only as good as the semantic layer beneath it. Without defined metrics it invents columns and hallucinates joins, then hands a plausible wrong number to someone who trusts it.

Evals#

Automated tests that score an AI system's outputs against a set of known-good answers, so quality can be measured and tracked over time rather than judged by eye.

Assumed: a benchmark score you cite once.

Actually: your golden set of question-and-answer pairs running in CI. They are the difference between "the demo seemed right" and "we know it is right," and almost nobody builds them early enough.

Hallucination#

A confident, fluent AI output that is factually wrong or unsupported by the data the system was given.

Assumed: the model lying or making things up.

Actually: usually missing context, not a model defect. Most hallucinations in data work are the system answering a question it was never handed the ground truth for. Fix it upstream.