# Glossary 50 terms across MCP, advertising, analytics and measurement, written for the place where the two vocabularies meet. Each entry says what the term means, and where it exists, what makes the reported value differ from what the word implies. ## MCP and agents ### Account scoping Restricting which advertising accounts a server may reach, enforced on the server rather than left to the model to respect. A platform token usually sees every account the person can see, which for an agency is every client. Scoping narrows that to an explicit list, checked on every call, so an account that was not ticked is refused even when the model asks for it by identifier. The distinction that matters is where the check runs. A scope described in a tool description is a request. A scope enforced server side is a boundary. **Through an agent:** Models sometimes guess identifiers, especially when a conversation has mentioned several accounts. Server-side scoping turns a guess into a refusal instead of a cross-client data leak. Read more: https://www.getmcpads.com/glossary/account-scoping ### Context window The total amount of text a model can hold at once, shared between the conversation, every tool description, and every result returned. Tool descriptions are loaded before the first question and stay for the whole session. Results accumulate on top. A window that looked generous at connection can be mostly spent on a catalogue the conversation never uses. This is why the size of a tool catalogue is a design decision rather than a feature list. Every tool announced is rent paid on every question, including the ones that will never call it. **What makes it wrong:** Large result sets are the other half of the problem. A report pulled at daily granularity across a year returns hundreds of rows the model must carry for the rest of the conversation, when the question could have been answered with a weekly aggregate. Read more: https://www.getmcpads.com/glossary/context-window ### Dynamic client registration An OAuth mechanism that lets an MCP client register itself with a server automatically, so no one has to create credentials by hand before connecting. Without it, connecting a new client means someone creating an application record, copying an identifier and a secret, and pasting them somewhere. With it, the client receives a 401 carrying a pointer to the server's metadata, registers itself, and walks the authorisation flow. The user sees a consent screen and nothing else. It is what makes a hosted MCP server feel like pasting a URL rather than like configuring an integration. Read more: https://www.getmcpads.com/glossary/dynamic-client-registration ### MCP (Model Context Protocol) An open protocol that lets an AI assistant call external tools and read external data through a uniform interface, instead of each integration being built by hand. MCP standardises the shape of the connection between a model and everything outside it. A server exposes tools, resources and prompts; a client, which is the assistant application, discovers them at connection time and can then call them during a conversation. The model itself never talks to the advertising API: it decides which tool to call and with what arguments, and the client performs the call. The practical consequence is that an integration written once works in every client that speaks the protocol. A Google Ads server built for Claude works in Cursor, in ChatGPT and in an n8n workflow, without a second implementation. **What makes it wrong:** MCP standardises the plumbing, not the quality of what flows through it. Two servers for the same advertising API can expose wildly different surfaces, one with four hundred generated tools and one with thirty written by hand, and the protocol has nothing to say about which will actually answer a question. The compatibility promise is about clients, never about capability. Read more: https://www.getmcpads.com/glossary/mcp ### MCP client The application that hosts the model and performs the tool calls it decides to make, such as Claude, Cursor or ChatGPT. The client is the half of the protocol people actually use. It connects to servers, presents their tools to the model, executes the calls, and hands the results back into the conversation. It also owns the parts a model cannot do safely on its own: storing credentials, asking a person for permission, and refusing a call. This is the distinction that gets reversed most often. A website that exposes an MCP endpoint is a server, not a client, even if it has a user interface. A client is what consumes servers. **What makes it wrong:** Clients differ in what they will actually do with a server. Some cap how many tools they will load, some do not support resources at all, some require confirmation before every write and some do not. A server that behaves well in one client can be unusable in another, and the protocol does not guarantee otherwise. Read more: https://www.getmcpads.com/glossary/mcp-client ### MCP gateway A single MCP endpoint that fronts several underlying platforms, so one connection reaches all of them under one authentication and one set of conventions. A gateway exists because the questions people ask cross platforms and the servers do not. Answering whether search spend moved organic rankings means holding Google Ads and Search Console in the same reasoning context, with periods that line up and metric names that mean the same thing on both sides. The alternative is several servers wired into the same client, which works, and costs a reconciliation the model performs on every question: different authentication, different naming, different date semantics, different write behaviour. **What makes it wrong:** A gateway is also a single point of scope. What it can reach is what every tool can reach, so the account allowlist matters more here than on a single-platform server: it is the only boundary between one client's data and another's. Read more: https://www.getmcpads.com/glossary/mcp-gateway ### MCP registry The official index of MCP servers, where a publisher declares a server under a verified namespace so clients and directories can discover it. A server is published with a manifest declaring its name, its package and its environment variables. Ownership of the namespace is proved either through a code host account or through a DNS record on the publisher's own domain, which is what a reverse-DNS name such as com.example is asserting. Several third-party directories read from the registry, so an entry there tends to propagate without further submissions. **What makes it wrong:** The registry verifies the published package, not the repository. A manifest can be correct while the package on the package manager carries no matching ownership marker, and publication fails for a reason that names neither file. Read more: https://www.getmcpads.com/glossary/mcp-registry ### MCP server A program that exposes a set of tools and data to an AI assistant over the Model Context Protocol. A server declares what it can do at connection time, then waits. It holds the credentials for whatever it fronts, translates a tool call into an API request, and returns a result the model can read. In advertising, one server usually fronts one platform: its API version, its authentication, its rate limits, its error vocabulary. Servers come in two shapes. A local server runs on the user's own machine and speaks over standard input and output. A remote server runs on someone else's infrastructure and is reached over HTTP, which is what makes hosted products possible. **Through an agent:** The number of tools a server announces is the number of tool descriptions the model carries in its context for the whole conversation. A server announcing four hundred tools spends a large slice of the window before the first question is asked, which is why some servers announce a small core and expose the rest through a search tool. Read more: https://www.getmcpads.com/glossary/mcp-server ### MCP tool A single named function a server exposes, with a written description and typed parameters, that a model can choose to call. A tool is the unit of capability. It carries a name, a description in prose, and a schema for its arguments. The model reads the description to decide whether the tool answers the question in front of it, then fills the schema. The description is not documentation, it is the interface. A tool whose description says what it returns, over what period, and what it costs will be chosen correctly. A tool named after an API endpoint, with the endpoint's own wording, will be chosen wrongly or not at all. **Through an agent:** Tool descriptions are the only thing the model sees before calling. Everything a person would learn from trying, the shape of the response, the units, the fact that money arrives in micros, has to be in the description or it is not known. Read more: https://www.getmcpads.com/glossary/mcp-tool ### Preview-then-confirm writes A write pattern where the first call describes the change it would make and applies nothing, and only an identical second call carrying an explicit confirmation performs it. The first call returns what would change: which entity, from what value to what value, in which currency. Nothing has happened yet. The same call repeated with a confirmation flag performs the change. The model cannot skip the first step because the second one refuses without it. The related convention is that newly created campaigns are created paused, so activating spend stays a human action taken in the platform. **Through an agent:** The pattern exists because a model that misreads a budget by a factor of a million spends real money before anyone reads the transcript. Two calls make the intended change visible in the conversation while it is still reversible. Read more: https://www.getmcpads.com/glossary/safe-writes ### Remote MCP server An MCP server reached over HTTP at a URL, rather than installed and run on the user's own machine. A remote server is pasted into a client as a URL and works immediately: nothing installed, no runtime, no developer token to obtain. Authentication happens through OAuth in the browser, so no credential is typed into a chat window. The trade is where the credentials live. A local server keeps the platform token on the machine that runs it. A remote server holds it on the operator's infrastructure, which is a real decision to make deliberately rather than by default. **Through an agent:** Remote servers can update their tool surface without the user reinstalling anything, which is convenient and also means the set of tools a model saw yesterday is not guaranteed to be the set it sees today. Read more: https://www.getmcpads.com/glossary/remote-mcp-server ### stdio transport The connection mode where a client launches the server as a local process and talks to it over standard input and output. This is how a self-hosted server runs. The client starts the process, usually through a package runner, passes credentials as environment variables, and exchanges protocol messages over the process pipes. Nothing listens on a port and nothing leaves the machine except the calls the server itself makes to the platform API. It is the mode that gives the strongest answer to the credentials question, because the token is only ever on the machine that uses it. **What makes it wrong:** Everything the server writes to standard output is protocol traffic. A stray log line, a debug print, a dependency that greets on startup, and the connection breaks with an error that names none of them. Servers that log must log to standard error. Read more: https://www.getmcpads.com/glossary/stdio-transport ### Structured output A tool result returned as typed data with a declared schema, rather than as prose the model has to parse. A tool that returns a table of numbers with named columns lets the model compute on them. A tool that returns the same numbers inside a sentence forces it to extract them first, which it usually does correctly and occasionally does not. The gain is largest exactly where the cost of an error is highest: currency amounts, date ranges, and identifiers that will be passed into a later write call. Read more: https://www.getmcpads.com/glossary/structured-output ### Tool call One request from the model to a server: a tool name plus arguments, answered with a result that returns into the conversation. A tool call is the atom of billing, of latency and of failure. The model emits a name and arguments; the client sends them; the server answers; the result becomes part of the context. A question that needs three platforms and two periods is not one call, it is six, and each one spends context on the way back. Most quotas in this space are counted in tool calls rather than in questions asked, which is why one badly shaped question can cost more than a whole afternoon of well shaped ones. **What makes it wrong:** A failed call still costs a call, still costs context, and on some platforms still counts against a quota. Worse, some APIs answer HTTP 200 on failure and put the error in the body, so a server that only checks the status code reports a failure to the model as a success, and the model reasons on an empty result as if it were a real zero. Read more: https://www.getmcpads.com/glossary/tool-call ### Tool poisoning An attack where instructions hidden in a tool description, or in data a tool returns, are read by the model as commands to follow. The model cannot tell the difference between content and instruction by position alone. Text arriving in a tool result carries no marker saying it is data, so a campaign name, an ad headline or a support ticket can contain a sentence addressed to the model, and it may be acted on. In advertising this is not hypothetical: campaign names, creative copy and audience labels are all free text, often written by people outside the account owner's organisation. **What makes it wrong:** The defence is not sanitising strings, which fails on the first encoding anyone tries. It is that side effects require an explicit confirmation from a person, so an instruction found in returned data cannot spend money on its own. Read more: https://www.getmcpads.com/glossary/tool-poisoning ## Advertising ### Audience saturation The point at which additional budget mostly buys more impressions to people already reached, rather than new people. Saturation is visible before it is expensive: reach flattens while impressions keep rising, so frequency climbs and response falls. It is a property of the audience size relative to the budget, not of the creative. It is the reason a campaign that scaled well suddenly stops scaling at a particular daily budget. **What makes it wrong:** Saturation is often misread as creative fatigue, and the two have opposite fixes. Fatigue is solved by new creative into the same audience; saturation is solved by a larger audience for the same creative. Frequency against reach separates them. Read more: https://www.getmcpads.com/glossary/audience-saturation ### Bid strategy The rule the platform follows when deciding how much to bid in each auction, ranging from manual amounts to fully automated targets. Automated strategies optimise toward a stated goal, a target cost per action or return, and need conversion volume to learn. Below that volume they behave erratically, which is usually blamed on the strategy rather than on the signal feeding it. Changing strategy resets learning, which is why comparing the week before and the week after a change measures the transition rather than the strategy. **What makes it wrong:** A target that is far from what the account currently achieves does not produce that result, it produces low delivery. The platform declines auctions it cannot win at the stated price, and the campaign quietly stops spending. Read more: https://www.getmcpads.com/glossary/bid-strategy ### Budget pacing How a platform spreads a budget across the period, rather than spending it as fast as the auction allows. Pacing smooths delivery so a daily budget is not exhausted in the first hours. It also means a budget raised mid-day does not take effect at the new rate immediately, and that a campaign can underspend while appearing budget-limited. On a monthly view, the interaction between pacing and daily caps explains most of the gap between the budget set and the amount actually charged. **What makes it wrong:** Daily budgets are averages on several platforms, not caps: a single day can exceed the stated amount as long as the period averages out. A day-level alert built on the daily budget as a ceiling fires on normal behaviour. Read more: https://www.getmcpads.com/glossary/budget-pacing ### Creative fatigue The decline in response that comes from an audience having seen the same creative too many times. Fatigue shows as a falling click-through rate at stable delivery and stable cost per thousand. The auction has not changed and the audience has not been exhausted: the same people are simply responding less to the same thing. It is the most common cause of a slow performance decline in an otherwise unchanged account. **What makes it wrong:** Refreshing a creative that was never fatigued costs production time and resets learning for nothing. Before rotating, check whether the cost per thousand rose, which points at competition, or whether frequency rose, which points at saturation. Read more: https://www.getmcpads.com/glossary/creative-fatigue ### Custom audience An audience built from the advertiser's own data: site visitors, a customer list, app activity or engagement with the brand's content. Custom audiences are the input to both retargeting and lookalike expansion. Their quality decides both: an audience built on all site visitors behaves very differently from one built on purchasers. They carry personal data by construction, which is why the matching happens on hashed identifiers and why the platform reports a match rate rather than a list. **What makes it wrong:** Audience membership expires. A thirty-day site visitor audience shrinks continuously unless traffic replenishes it, so a campaign that performed well at launch can be delivering to a fraction of the intended pool weeks later without any setting having changed. Read more: https://www.getmcpads.com/glossary/custom-audience ### Frequency The average number of times each reached person saw the ad over a period. `Frequency = impressions / reach` Frequency is the first place to look when performance decays without anything having changed. A rising frequency on a fixed audience means the same people are being served more often, which reliably lowers response. It is period-dependent by construction: the same campaign has a low weekly frequency and a high quarterly one. **What makes it wrong:** An average hides the distribution. A frequency of three can mean everyone saw the ad three times, or that most saw it once and a small group saw it fifteen times. The second is where budget is actually being wasted, and it needs a distribution, not an average. Read more: https://www.getmcpads.com/glossary/frequency ### Impression One instance of an ad being served, which is not the same as an ad being seen. An impression is recorded when the ad is delivered. Whether it entered the viewport, for how long, and whether a person was looking are separate questions that separate metrics answer. Platforms that also report a viewable or qualified impression are reporting a subset, sometimes a much smaller one. **What makes it wrong:** Mixing served and viewable impressions in the same calculation silently changes every rate built on them. A CTR computed against viewable impressions is not comparable with one computed against served impressions, and neither is labelled in most exports. Read more: https://www.getmcpads.com/glossary/impression ### Impression share The share of the impressions an ad was eligible for that it actually received, with the remainder split between budget and rank limitations. `Impression share = impressions received / impressions eligible` The value on its own says how much of the available demand was captured. The two lost-share figures beside it say why the rest was not: lost to budget means the money ran out, lost to rank means the auction was not won. The distinction decides the action. Budget loss is solved with money, rank loss is not. **What makes it wrong:** Eligibility is estimated by the platform, not counted, so impression share is a modelled figure. It is reliable for direction and unreliable as a precise denominator. Read more: https://www.getmcpads.com/glossary/impression-share ### Lookalike audience An audience the platform builds by finding people who resemble a source audience the advertiser provided. The platform models the source and returns a population ranked by similarity, usually cut at a percentage of the country. A one percent lookalike is tighter and smaller, a ten percent one is broader and closer to general targeting. Everything depends on the source. A lookalike of all visitors models visiting; a lookalike of repeat purchasers models buying twice. **What makes it wrong:** A source audience below the platform's minimum produces a lookalike anyway, built on too little signal, and nothing in the reporting says the model was thin. Small sources give expensive audiences that look like general targeting with extra steps. Read more: https://www.getmcpads.com/glossary/lookalike-audience ### Micros A unit used by several advertising APIs where a currency amount is expressed as one millionth of the currency unit. `amount in currency = micros / 1,000,000` Money crosses these APIs as integers to avoid floating point drift, so ten euros is ten million. Google Ads and the X Ads API both use micros; others use minor units, where ten euros is one thousand. The conversion is trivial and the consequence of getting it wrong is not, because it lands on a budget field. **What makes it wrong:** Zero-decimal currencies break the minor-unit assumption. The Japanese yen has no subunit, so multiplying by a hundred to get minor units inflates a yen budget by a factor of a hundred. Any conversion has to be per currency, never a constant. **Through an agent:** A budget written into a write call is where this becomes expensive. A preview that shows both the raw value and the converted amount makes a factor-of-a-million error visible before it applies. Read more: https://www.getmcpads.com/glossary/micros ### Quality Score A diagnostic rating of expected click-through rate, ad relevance and landing page experience, reported per keyword. It is a diagnostic rather than a live input: the auction uses its own real-time signals, and the reported score is a summarised view intended to point at what to fix. Read as three components rather than one number, it is genuinely useful, because each component has a different remedy. **What makes it wrong:** Averaging Quality Score across keywords produces a number that means nothing, since it is unweighted by volume. A keyword with three impressions counts as much as one with thirty thousand. Read more: https://www.getmcpads.com/glossary/quality-score ### Reach The number of distinct people who saw an ad at least once, as opposed to the number of times it was served. Reach is deduplicated and impressions are not, which is why impressions divided by reach gives frequency. It is the denominator that matters for saturation questions: whether an audience has been exhausted, and whether spending more will find new people or the same ones again. Deduplication only holds inside one platform. The same person reached on two platforms is two reached people in the sum. **What makes it wrong:** Reach cannot be added across date ranges. Two weeks of a hundred thousand reach each is not two hundred thousand people, because the overlap is unknown, and pulling the two weeks as one range gives a different and correct number. Read more: https://www.getmcpads.com/glossary/reach ### Retargeting Showing ads to people who have already interacted with the business, identified from a pixel, a customer list or an app event. Retargeting is cheap on every attributed metric, because it addresses an audience already close to converting. That is also its weakness as an investment: proximity to the purchase is not the same as causing it. It is bounded by the size of the pool feeding it, which makes audience saturation the constraint that eventually decides its ceiling. **What makes it wrong:** Retargeting budgets tend to grow because the reported return is high, and the reported return is high partly because the conversions would have happened anyway. Incrementality is the only way to see the difference. Read more: https://www.getmcpads.com/glossary/retargeting ### Search term What a person actually typed, as opposed to the keyword the advertiser bid on that matched it. Keywords are bids, search terms are reality. Broad and phrase matching mean one keyword collects many terms, some of which have nothing to do with the intended intent. The search terms report is where wasted spend is visible, and it is the input to negative keywords. **What makes it wrong:** The report does not show every term. Terms below a volume threshold are withheld for privacy, so the terms shown never sum to the campaign total, and the withheld tail can hold a meaningful share of the spend. Read more: https://www.getmcpads.com/glossary/search-term ## Measurement ### Attribution window The period after seeing or clicking an ad during which a conversion is still credited to it. A seven-day click and one-day view window credits a purchase made six days after a click, and a purchase made the same day as a view. Widening the window increases attributed conversions without changing behaviour; narrowing it decreases them. Because platforms default to different windows, comparing attributed results across platforms compares different definitions before it compares anything else. **What makes it wrong:** Windows also make recent data provisional. A report pulled today for yesterday is missing every conversion that will land inside the window over the coming days, so the last few days of any series are systematically understated and look like a decline. **Through an agent:** Ask for the last seven days and a model reports a downward trend that is an artefact of the window. Comparable series need the tail excluded, or the same lag applied to both periods. Read more: https://www.getmcpads.com/glossary/attribution-window ### Conversion An action the advertiser decided to count as a result, recorded and attributed by the platform. A conversion is a definition before it is a number. Which events count, how they are deduplicated, and which touchpoint gets credit are all configuration, and all of them change the reported total without changing what people did. The same underlying purchase can arrive through a browser pixel, a server-side event, or both, and whether that is one conversion or two depends on the deduplication key being sent correctly. **What makes it wrong:** Duplicate counting between browser and server events is the most common inflation. It shows up as a conversion count that exceeds what the order system recorded, and it is invisible in the platform's own interface, which reports both paths as one figure. Read more: https://www.getmcpads.com/glossary/conversion ### Conversions API A server-to-server path for sending conversion events to an advertising platform, alongside or instead of the browser pixel. Events sent from the advertiser's own systems are not subject to browser tracking prevention, so the server path recovers conversions the pixel loses. It also allows sending events the browser never sees, such as an order confirmed later in a back office. Running both paths is the normal configuration, with a shared event identifier so the platform can recognise the two reports as one event. **What makes it wrong:** Without a correctly propagated deduplication key, running both paths double-counts every conversion that both saw. The symptom is a platform conversion count noticeably above the order system's, with no obvious source. Read more: https://www.getmcpads.com/glossary/conversions-api ### CPA (Cost per acquisition) The advertising spend divided by the number of conversions attributed to it. `CPA = ad spend / attributed conversions` CPA answers what one result cost. It is the inverse of the conversion side of ROAS and inherits every one of its attribution problems, with one addition: what counts as a conversion is a setting, not a fact. Changing which events are marked as conversions changes every historical CPA the platform reports, without changing anything that happened. **What makes it wrong:** CPA falls when a cheaper conversion event is added to the account, which looks like improvement and is arithmetic. A CPA compared across a period where the conversion configuration changed is comparing two different questions. Read more: https://www.getmcpads.com/glossary/cpa ### CPC (Cost per click) The advertising spend divided by the number of clicks it produced. `CPC = ad spend / clicks` CPC sits between what the auction charges and what the creative earns: it is the product of the cost of reaching people and the rate at which they respond. A CPC can therefore worsen either because the auction got more expensive or because the creative stopped working, and the number alone does not say which. Splitting it into its two parts is usually the first useful diagnostic step, because the two have different remedies. **What makes it wrong:** Platforms count clicks differently. Some count any click on the ad unit, including profile and hashtag clicks, and some count only clicks through to the destination. A CPC compared across platforms without checking which click is being counted compares two different denominators. Read more: https://www.getmcpads.com/glossary/cpc ### CPM (Cost per mille) The cost of one thousand impressions, which is the price of reaching an audience rather than of any result. `CPM = (ad spend / impressions) x 1000` CPM is the auction's own price signal. It moves with demand for the audience being bought, and it moves for reasons that have nothing to do with the advertiser: a competitor's launch, a seasonal peak, an election. Because it is upstream of everything, a CPM change explains many performance changes that are otherwise blamed on creative or targeting. **What makes it wrong:** A rising CPM with flat performance is often read as fatigue when it is competition. The distinction is visible: fatigue shows a falling click-through rate on stable delivery, competition shows a rising CPM on a stable click-through rate. Read more: https://www.getmcpads.com/glossary/cpm ### CTR (Click-through rate) The share of impressions that produced a click. `CTR = clicks / impressions` CTR is the cleanest available signal of whether a creative is working, because both of its terms come from the same system and the same moment. It is the metric to watch when diagnosing fatigue, and the one to distrust when comparing across placements. A feed placement and a story placement produce different rates for the same creative, so a blended CTR across placements mostly measures the placement mix. **What makes it wrong:** The denominator is not the same everywhere. Some platforms report a rate against served impressions and some against viewable impressions, and the second is always the higher number for the same creative. Read more: https://www.getmcpads.com/glossary/ctr ### Incrementality The share of results that would not have happened without the advertising, measured by holding out a group rather than by attribution. Attribution asks which ad touched the customer. Incrementality asks whether the customer needed touching. The two answer different questions and the second is the one that decides whether a budget should exist at all. It is measured with a holdout: a comparable group that is not shown the advertising, and whose behaviour becomes the baseline. **What makes it wrong:** Retargeting scores best on attribution and often worst on incrementality, because it addresses people who were already going to buy. A channel can be the top performer in every attributed report and add nothing. Read more: https://www.getmcpads.com/glossary/incrementality ### Lift study A platform-run experiment that splits an audience into exposed and held-out groups to measure the difference advertising made. A lift study is incrementality implemented inside the platform, with the platform managing the split. It produces a measured difference rather than an attributed credit, which makes it the strongest evidence available without leaving the platform. It requires enough volume to separate the groups, and it costs the holdout: a share of the audience is deliberately not advertised to for the duration. Read more: https://www.getmcpads.com/glossary/lift-study ### Pixel A snippet on the advertiser's site that reports visitor actions back to the advertising platform from the browser. The pixel is the browser half of conversion measurement. It builds audiences, records events, and feeds the optimisation models that automated bidding depends on. It is also the half most affected by tracking prevention, ad blocking and consent choices, which is why the server-side path exists alongside it. **What makes it wrong:** A pixel that fires on a page rather than on an action counts page views as conversions. This inflates every downstream number and is invisible in the platform, which reports what it was sent. Read more: https://www.getmcpads.com/glossary/pixel ### ROAS (Return on ad spend) The revenue attributed to advertising divided by the amount spent on it, over the same period. `ROAS = attributed revenue / ad spend` ROAS is the ratio everything else is judged against, and it is a ratio of two numbers that come from different systems. The spend is what the platform charged, which is close to exact. The revenue is what the platform decided to credit itself with, which depends entirely on its attribution rules. Two platforms reporting on the same week will both claim revenue for the same order if both touched the customer, so the sum of platform ROAS across a business is generally higher than the business's real return. **What makes it wrong:** A ROAS pulled per platform and then added up double-counts. The overlap is not an error in either platform: each is answering the question it was asked, which is what happened inside its own attribution window. Only a shared source of truth on the revenue side, or an incrementality test, settles it. **Through an agent:** Ask a model for ROAS across four platforms and it will happily sum them unless the tool descriptions say the values are not additive. This is the single most common wrong answer in cross-platform reporting. Read more: https://www.getmcpads.com/glossary/roas ## Analytics and search ### Average position In Search Console, the mean ranking of a site's highest-ranking result for a query, averaged over impressions. The value averages positions across every impression, so it mixes results shown at the top of the first page with results shown far down the third. A page ranking first for half its impressions and thirtieth for the other half reports an average of about fifteen, a position it never actually held. It is a trend signal, not a rank tracker. **What makes it wrong:** Average position improves when a page stops appearing for its worst queries, which looks like a ranking gain and is a loss of impressions. Position always has to be read next to impressions. Read more: https://www.getmcpads.com/glossary/search-console-position ### Cannibalisation Two or more pages of the same site competing for the same query, so neither accumulates the signals a single page would. Search engines pick one result per site for most queries. When several pages target the same intent, the chosen one varies over time, splitting clicks and making performance look unstable for reasons that have nothing to do with the pages themselves. It is found by grouping query and page data and looking for queries served by more than one URL over a period. **What makes it wrong:** It is invisible in any report aggregated by query alone, which is the default view. The two competing pages appear as one line with mediocre performance rather than as two lines that should be merged. Read more: https://www.getmcpads.com/glossary/cannibalisation ### Cardinality The number of distinct values a dimension can take, which decides whether its values are reported individually or collapsed. High-cardinality dimensions, page paths or search terms for instance, exceed the number of rows a report can hold. Values beyond the limit are grouped into a single catch-all row rather than dropped. The grouped row is real data, but it is unusable for anything except knowing how much was grouped. **What makes it wrong:** The catch-all row is easy to filter out as noise, which removes a share of the total without any warning. On a high-cardinality dimension it can hold more volume than the named rows above it. Read more: https://www.getmcpads.com/glossary/cardinality ### Data stream One source of events feeding a GA4 property: a website, an Android app or an iOS app. Each stream has its own measurement identifier and its own configuration, including which enhanced measurements are enabled. Events from every stream land in the same property and are reported together unless a dimension separates them. Adding a stream therefore changes the property's totals from that day forward, with no change in behaviour. **What makes it wrong:** A test or staging site pointed at the production stream inflates every metric with traffic that is not customers. It is common, it is invisible in aggregate reporting, and it is only findable by looking at hostnames. Read more: https://www.getmcpads.com/glossary/data-stream ### Dimension An attribute used to break a metric down, such as country, device, campaign or landing page. Metrics are the numbers, dimensions are the ways of cutting them. A report is a choice of both, and not every pair is valid: platforms maintain compatibility rules, and an incompatible request is refused rather than approximated. The more dimensions applied at once, the smaller each resulting group, which is what eventually triggers sampling or thresholding. **What makes it wrong:** Adding a dimension changes totals, and not only their breakdown. Rows too small to report are withheld under thresholding, so a broken-down report can sum to less than the same report without the breakdown, with nothing indicating the difference. **Through an agent:** Compatibility matrices are exactly the kind of knowledge a model cannot infer from an error message. Servers that carry the matrix can refuse an impossible combination before spending the call. Read more: https://www.getmcpads.com/glossary/dimension ### GA4 property The container in Google Analytics 4 that holds a set of data streams and the configuration applied to them. A property is the unit of reporting, of retention settings and of access. Several sites and apps can feed one property through separate data streams, which is what makes cross-platform reporting possible and also what makes a property's numbers depend on which streams are attached. The property identifier, not the account, is what every reporting call takes. **What makes it wrong:** Data retention is a property setting with a short default, and it applies to the event-level data behind custom explorations, not to the standard aggregated reports. A question that needs event-level history often cannot be answered at all beyond the retention horizon, and the interface answers with an empty result rather than an explanation. Read more: https://www.getmcpads.com/glossary/ga4-property ### Key event An event marked in GA4 as one that matters, which is the analytics counterpart of a conversion. Marking an event as key changes how it is reported, not how it is collected. The event was already being recorded; the mark decides whether it appears as a result in reporting and can be imported into advertising platforms for optimisation. The renaming from conversion to key event was made precisely to separate the analytics definition from the advertising one, because the two count differently. **What makes it wrong:** The mark is not retroactive in the advertising sense: a newly marked event will be reported historically in analytics, but platforms that imported the old definition keep optimising toward what they were given. The two systems disagree until both are updated. Read more: https://www.getmcpads.com/glossary/key-event ### Metric A quantitative measure in a report, such as sessions, users, revenue or event count. Metrics carry definitions that differ between systems using the same word. Users in analytics and reach in an advertising platform both mean distinct people and are counted by entirely different mechanisms, over different populations. Comparing them directly is the most common cross-system reporting error, and it is one a shared vocabulary makes worse rather than better. **What makes it wrong:** Some metrics are not additive across rows or dates. Users is the standard example: the users of two weeks is not the sum of each week's users, because the same person appears in both. Summing a deduplicated metric always overstates. Read more: https://www.getmcpads.com/glossary/metric ### Sampling Computing a report from a subset of the data and extrapolating, rather than from every event. Sampling keeps large queries fast. The result is an estimate with a confidence interval rather than a count, and the wider the query, the more aggressive the sampling. It matters most for the small numbers inside a large report: a segment representing a fraction of a percent of traffic can be estimated from very few actual events. **What makes it wrong:** A sampled report is labelled in the interface and easy to miss in an export or an API response. Two runs of the same sampled query can return different numbers, which is correct behaviour and reads as a bug. Read more: https://www.getmcpads.com/glossary/sampling ### Top N truncation The Search Console API returning only the highest-volume rows for a query, so the returned rows never sum to the site total. A request returns rows ordered by clicks up to a row limit. Everything below the cut is absent from the response, and the response says nothing about how much was left out. For a site with a long tail, which is most sites, the withheld rows can be a large share of total traffic. **What makes it wrong:** Summing the returned rows and presenting the result as the site total is the single most common Search Console error, and it always understates. The site total has to come from an unsegmented query, not from adding up a breakdown. **Through an agent:** A model handed a list of rows will add them up, because that is what a list of numbers invites. Only the tool description saying the rows are a truncated top N prevents it. Read more: https://www.getmcpads.com/glossary/top-n-truncation