MarTech Glossary

Marketing technology (MarTech)

The software a team uses to plan, run, measure, and optimize marketing, together with the operating practices — data flows, naming, ownership — that keep those tools working as one system.

The 2026 marketing technology landscape counts 15,505 products, and for the first time in fifteen years the number has essentially stopped growing. The practical consequence: almost every capability now exists in several tools you already pay for, so stack decisions are less about finding software and more about deciding which overlapping tool owns which job.

In depth: How to Plan a MarTech Stack Before Buying Tools

MarTech stack

The specific set of tools a company runs and the integrations between them, usually described in layers: collection, storage, decisioning, activation, and measurement.

A stack is defined by its data flows more than by its logos: two companies can own identical tools and have completely different stacks because the connections differ. Mapping capabilities and flows before buying anything is the argument of the stack planning guide.

Composable architecture

An approach that assembles interoperable best-of-breed components around shared data — usually a warehouse — instead of adopting one vendor’s integrated suite.

Composable trades licence cost for engineering ownership: every integration a suite would have shipped becomes something your team maintains. You need it when data volume and team maturity justify it; you do not when "composable" is a way to avoid choosing an owner for the glue.

Not the same as Composable CDP: Composable architecture is an approach to the whole stack; a composable CDP applies it to the customer-data layer only.

iPaaS (integration platform as a service)

Middleware such as Zapier, Make, or Workato that moves data between tools through prebuilt connectors, without custom code.

Good for edge cases and internal workflows; risky as the core pipe between systems of record. Per-task pricing grows faster than volume, and retries and error handling are opaque exactly where reliability matters most.

Not the same as Reverse ETL: iPaaS moves records between apps through connectors, usually one event at a time; reverse ETL syncs modeled tables from the warehouse into apps.

API (application programming interface)

The documented surface through which one system reads or writes another’s data — the raw material of every integration.

"Has an API" is where vendor diligence starts, not where it ends. Check that the objects and fields you need are actually exposed, what the rate limits are, and whether the authentication model fits how you deploy — those three points are where martech integrations die.

Not the same as Model Context Protocol (MCP): An API is the vendor’s documented interface for any program; an MCP server packages selected capabilities as tools an AI model can discover and call, usually on top of that API.

Not the same as Webhook: With an API, your system asks for data when it wants it; with a webhook, the other system calls you when something happens.

Webhook

A push notification between systems: instead of being polled, a tool calls a URL you configure the moment an event happens — a form submit, an unsubscribe, a purchase.

Webhooks are usually delivered at-least-once, so the receiving side must handle duplicates, and a webhook that fails silently loses data with no one watching. If a workflow depends on one, monitor it.

Not the same as API (application programming interface): With an API, your system asks for data when it wants it; with a webhook, the other system calls you when something happens.

Authoritative system (system of record)

The system agreed to win when two tools disagree about a field — the source of truth for that specific piece of data, such as consent, lifecycle stage, or email address.

Authority is per field, not per tool: the CRM can own lifecycle stage while the consent platform owns opt-in status. Most sync problems trace back to two systems that each believe they are authoritative for the same field — the ownership-table exercise in the CDP versus CRM comparison exists for exactly this.

Not the same as Data warehouse: The system of record is where a field is created and corrected; the warehouse holds copies of it for analysis. Fix a wrong value at its source, not in the copy.

Model Context Protocol (MCP)

An open protocol for connecting AI applications to tools and data: a vendor runs an MCP server that exposes actions and records, and any compatible assistant or agent can call it. Anthropic introduced it on November 25, 2024, and donated it to the Linux Foundation’s Agentic AI Foundation on December 9, 2025; the current specification revision is dated July 28, 2026.

“We have an MCP server” is now a line in every martech pitch, and it says less than it sounds. The questions are which specification revision the server implements, how clients register and authenticate, and what each OAuth scope lets an agent do. The three-command vendor check asks them before the demo.

Not the same as API (application programming interface): An API is the vendor’s documented interface for any program; an MCP server packages selected capabilities as tools an AI model can discover and call, usually on top of that API.

Not the same as WebMCP: An MCP server is a service an assistant connects to over the network; WebMCP registers tools from inside a web page, for an agent already working in the browser.

In depth: What Can AI Agents Do in Your Ad Accounts and CRM?

Consumption pricing (credits)

Billing by metered units of use — actions, messages, records, AI responses — usually through prepaid credits converted at a published rate, instead of or on top of a fee per user seat. Salesforce’s Flex Credits and HubSpot Credits are the current martech examples.

The price per credit is the least important number. Read what one business action consumes, whether unused credits expire (Salesforce’s do, with no rollover), what stops when they run out — on HubSpot’s flexible pricing, email can pause — and whether the rate card can change mid-term. What an AI agent action costs works the conversion on three platforms, and the HubSpot credits and seats analysis covers overage defaults.

Web analytics

Software that records visitor behavior on sites and apps as events — page views, clicks, conversions — and aggregates them into reports; Google Analytics 4 is the dominant implementation.

Analytics describes behavior on properties you own. It is not a system of record for revenue, not a CDP, and its counts will never match the CRM exactly — reconciliation gaps within roughly five percent are normal attrition from blockers and consent, not bugs to fix.

Not the same as Attribution: Web analytics records what happened; attribution is one analysis on top of it, deciding which touchpoints get credit for a conversion.

In depth: Why Analytics Implementations Fail Before Reporting

Event (and event parameter)

The unit of measurement in modern analytics: a named action (the event) with attached details (parameters) — for example a sign_up event carrying a plan parameter.

Naming decides usability: GA4 treats form_submit and Form_Submit as two different events, and a property without a naming convention accumulates parallel events nobody can report on. The conventions that survive contact with real teams are in the GA4 and GTM notes.

Tag management system (TMS)

A tool such as Google Tag Manager that deploys and governs measurement and marketing tags through a managed container, without a code release for every change.

The value is governance, not convenience: versioned changes, preview before publish, and one place to answer what fires where. A TMS without a named owner and a review step drifts into tag soup nobody dares delete.

Not the same as Data layer: The data layer holds the facts: page type, user state, transaction values. The TMS reads them and decides which tags fire, with which values.

In depth: Why Analytics Implementations Fail Before Reporting

Data layer

A structured JavaScript object on a page that exposes agreed, named values — page type, user state, transaction data — for tag management and analytics to read.

It is a contract between the site and measurement. The alternative — scraping values out of the DOM — breaks silently on the next redesign. If a tracked value matters, it belongs in the data layer, put there deliberately by the site code.

Not the same as Tag management system (TMS): The data layer holds the facts: page type, user state, transaction values. The TMS reads them and decides which tags fire, with which values.

In depth: Why Analytics Implementations Fail Before Reporting

Server-side tagging

Running tags in a server container you control: the browser sends one stream to your endpoint, and the server forwards events to analytics and advertising platforms.

It buys durability (fewer blocked requests), control over what leaves the browser, and faster pages. It does not create consent — forwarding data a user declined is the same violation with an extra hop — and it means owning infrastructure: a misconfigured server container fails invisibly.

Not the same as Conversions API (CAPI): Server-side tagging is where your tags run, in a server container you control; a conversions API is one destination that container, or your backend, sends events to.

Measurement plan

A document that maps business questions to the events, parameters, and reports that answer them — written before implementation, and used as the acceptance test after it.

The cheapest artifact in analytics and the most skipped. Without one, tracking accumulates by request until hundreds of events serve nobody. The analytics implementation checklist starts here for a reason.

In depth: Why Analytics Implementations Fail Before Reporting

Attribution

Assigning credit for a conversion across the touchpoints that preceded it — last click, first click, or model-based approaches such as GA4’s data-driven attribution.

Every model redistributes credit; none measures whether the marketing caused anything — that is incrementality testing, a different discipline. Attribution answers "compare channels under consistent rules", so read relative movement, not absolute numbers.

Not the same as Web analytics: Web analytics records what happened; attribution is one analysis on top of it, deciding which touchpoints get credit for a conversion.

Not the same as UTM parameters: UTM parameters label where a visit came from; attribution decides how much credit that source gets. Inconsistent UTMs make every attribution model wrong in the same way.

UTM parameters

Query parameters (utm_source, utm_medium, utm_campaign and companions) appended to inbound links so analytics can classify the traffic they bring.

They only work as a taxonomy: lowercase everything, document allowed values, and generate links from a shared sheet or tool. Two gotchas cost real data — UTMs on internal links restart sessions, and inconsistent casing splits one campaign into several reporting rows.

Not the same as Attribution: UTM parameters label where a visit came from; attribution decides how much credit that source gets. Inconsistent UTMs make every attribution model wrong in the same way.

Measurement Protocol

GA4’s HTTP endpoint for sending events directly from a server — used for offline conversions, subscription renewals, and other events that never happen in a browser.

Its status is widely misreported. Google has placed the Measurement Protocol in maintenance mode with no plans for deprecation, while recommending the newer Data Manager API for new server-to-server integrations — so existing integrations are safe, and new builds should start on the newer path.

Not the same as Conversions API (CAPI): Both are server-to-server endpoints. Measurement Protocol sends events to GA4; a conversions API sends conversions to an advertising platform.

Conversions API (CAPI)

The generic name for advertising platforms’ server-side conversion endpoints — Meta’s CAPI is the best known — which receive conversion events from your server instead of the browser pixel.

Platforms promote CAPI as recovering signal lost to blockers and privacy features, and it does — but it also moves matching from an anonymous pixel to identifiers you send deliberately. Decide what you send with a privacy review, not from the ad platform’s setup wizard.

Not the same as Server-side tagging: Server-side tagging is where your tags run, in a server container you control; a conversions API is one destination that container, or your backend, sends events to.

Not the same as Measurement Protocol: Both are server-to-server endpoints. Measurement Protocol sends events to GA4; a conversions API sends conversions to an advertising platform.

Core Web Vitals

Google’s three field metrics for page experience: Largest Contentful Paint for loading (good at 2.5 seconds or less), Interaction to Next Paint for responsiveness (200 milliseconds or less; it replaced First Input Delay on March 12, 2024), and Cumulative Layout Shift for visual stability (0.1 or less), each judged at the 75th percentile of real visits.

They come from real users’ devices through the Chrome UX Report, so one lab run proves little, and they are assessed per group of similar URLs, not per site. Martech is a frequent culprit: every tag, chat widget, and personalization script competes for the main thread that INP measures. The website launch checklist has the test.

In depth: WordPress vs. Headless CMS: When to Make the Move

First-party data

Data a company collects directly from its own audience with consent — site behavior, purchases, form fills, email engagement. Second-party data is someone else’s first-party data shared under agreement; third-party data is aggregated and resold.

"Zero-party data" — information customers volunteer, such as stated preferences — is a Forrester coinage that is legally still first-party data. Treat the label as marketing; treat the consent under which it was collected as the thing that matters.

Not the same as Third-party cookies: First-party data is collected by you from your own audience; third-party cookies are set by other domains to follow people across sites. Losing the second is why the first became a strategy.

Not the same as Personally identifiable information (PII): First-party describes where data came from; PII describes whether it identifies someone. Much first-party data is PII, and collecting it yourself does not exempt it from consent rules.

Third-party cookies

Cookies set by a domain other than the site being visited — historically the mechanism behind cross-site ad tracking and retargeting.

The category is a lesson in not building plans on announcements: Safari and Firefox have blocked them by default for years, Chrome spent six years announcing deprecation — and then kept them in April 2025, dropping even the planned choice prompt. Signal loss is real and growing anyway; the reversal changed the deadline, not the direction.

Not the same as First-party data: First-party data is collected by you from your own audience; third-party cookies are set by other domains to follow people across sites. Losing the second is why the first became a strategy.

Customer Relationship Management (CRM)

The system of record for known contacts, accounts, and deals, usually owned by sales or service — HubSpot and Salesforce are the reference points.

A CRM tracks relationships, not behavioral event streams, and its data decays as fast as people change jobs. The distinction that matters in stack conversations: a CRM records what your team knows; a CDP assembles what your systems observed.

Not the same as Customer Data Platform (CDP): A CRM is the system of record for the contacts, accounts, and deals that sales and service work in; a CDP assembles profiles from many sources, behavior included, for other tools to use.

Not the same as Marketing automation platform (MAP): A CRM records who the contacts, accounts, and deals are; a MAP decides which message each contact gets next.

In depth: CDP vs. CRM vs. Marketing Automation: What Each Does

Customer Data Platform (CDP)

Packaged software that builds persistent, unified customer profiles from many sources and makes them available to other systems — the CDP Institute’s definition, and still the test worth applying.

The label is the most rebranded in martech. The tells that a “CDP” is something else wearing the badge: it cannot ingest raw behavioral events in near real time, marketers cannot build audiences without engineering tickets, or profiles cannot leave for other tools. Most teams asking about a CDP need ownership decisions first — the argument is in the CDP versus CRM comparison.

Not the same as Customer Relationship Management (CRM): A CRM is the system of record for the contacts, accounts, and deals that sales and service work in; a CDP assembles profiles from many sources, behavior included, for other tools to use.

Not the same as Data Management Platform (DMP): A DMP builds anonymous audience segments for ad targeting, mostly from third-party cookies and device identifiers; a CDP builds persistent profiles of known people from first-party data.

Not the same as Data warehouse: A data warehouse stores and models any company data for analysis; a CDP exists to activate customer profiles in marketing tools. A composable CDP is the second built on top of the first.

Not the same as Composable CDP: A packaged CDP keeps its own copy of the data and does identity, audiences, and syncs inside one product; a composable CDP leaves the data in your warehouse and assembles those jobs from parts.

Composable CDP

CDP capabilities assembled from parts — event collection, warehouse storage, identity modeling in SQL, and reverse ETL activation — instead of bought as one packaged product.

Both camps argue with pricing decks: packaged vendors call composable a science project, warehouse vendors call packaged a data silo. The real fork is staffing — composable means your team owns identity logic in SQL. If nobody can name that owner, packaged is the honest choice.

Not the same as Customer Data Platform (CDP): A packaged CDP keeps its own copy of the data and does identity, audiences, and syncs inside one product; a composable CDP leaves the data in your warehouse and assembles those jobs from parts.

Not the same as Composable architecture: Composable architecture is an approach to the whole stack; a composable CDP applies it to the customer-data layer only.

Not the same as Reverse ETL: Reverse ETL is one part, syncing warehouse tables into tools; a composable CDP also needs event collection and identity modeling around it.

In depth: CDP vs. CRM vs. Marketing Automation: What Each Does

Data Management Platform (DMP)

An advertising-era system for building anonymous, mostly third-party audience segments from cookies and device identifiers.

The category has effectively ended: Oracle shut down its advertising business — including the best-known DMP — in 2024, and the surviving functions moved into CDPs, retail media, and clean rooms. A vendor pitching a DMP in 2026 should be asked which identifiers it still has.

Not the same as Customer Data Platform (CDP): A DMP builds anonymous audience segments for ad targeting, mostly from third-party cookies and device identifiers; a CDP builds persistent profiles of known people from first-party data.

Not the same as Data clean room: A DMP pools audience data so one party can target it; a clean room lets two parties measure their overlap without either seeing the other’s raw records.

Data warehouse

A queryable analytical store — BigQuery, Snowflake, Databricks — where raw data from every system lands and where modeling happens in SQL.

In marketing architecture the warehouse is the candidate single source of truth and the foundation of composable stacks. The honest prerequisite is engineering capacity: a warehouse without owned, maintained models is a swamp with a monthly bill.

Not the same as Customer Data Platform (CDP): A data warehouse stores and models any company data for analysis; a CDP exists to activate customer profiles in marketing tools. A composable CDP is the second built on top of the first.

Not the same as Authoritative system (system of record): The system of record is where a field is created and corrected; the warehouse holds copies of it for analysis. Fix a wrong value at its source, not in the copy.

In depth: CDP vs. CRM vs. Marketing Automation: What Each Does

Reverse ETL

Tooling that syncs modeled data out of the warehouse into operational tools — CRM fields, advertising audiences, email lists; Hightouch and Census defined the category.

What it actually buys: audience logic lives in version-controlled SQL instead of a vendor interface, and every tool receives the same answer. What it does not buy: real time — most syncs are scheduled batches, and the interval belongs in the design conversation.

Not the same as Composable CDP: Reverse ETL is one part, syncing warehouse tables into tools; a composable CDP also needs event collection and identity modeling around it.

Not the same as iPaaS (integration platform as a service): iPaaS moves records between apps through connectors, usually one event at a time; reverse ETL syncs modeled tables from the warehouse into apps.

In depth: CDP vs. CRM vs. Marketing Automation: What Each Does

Identity resolution

Stitching events and records from different devices, browsers, and channels into one customer profile — deterministically (shared identifiers such as email or login) or probabilistically (inference from signals).

Deterministic matching is the only kind worth building decisions on; probabilistic fills reports, not audiences. Vendor match rates are measured on the vendor’s best data — ask for the rate on records that look like yours.

In depth: CDP vs. CRM vs. Marketing Automation: What Each Does

Data clean room

A controlled environment where two parties analyze overlapping customer data — brand and retailer, brand and platform — without either side seeing the other’s raw records.

Genuinely useful at scale for measurement partnerships; mostly irrelevant below it. The prerequisites are meaningful audience overlap, legal review on both sides, and analysts to do the work — a mid-market team is rarely missing insights a clean room would unlock.

Not the same as Data Management Platform (DMP): A DMP pools audience data so one party can target it; a clean room lets two parties measure their overlap without either seeing the other’s raw records.

Marketing automation platform (MAP)

Software that executes multi-step programs across channels — email, SMS, push, in-product — from triggers, audience rules, and branching logic.

The distinction from an ESP is the trigger-and-branch engine, and it is exactly what most teams underuse: a MAP sending two undifferentiated newsletters a month is an expensive ESP. The automation planning template exists to force the design questions first.

Not the same as Email service provider (ESP): An ESP’s core job is sending email well; a MAP runs multi-step programs across channels from triggers and rules, and email is only one of those channels.

Not the same as Journey orchestration: A MAP executes programs; journey orchestration coordinates programs and channels around one customer’s state, so they stop running in parallel. Many MAPs claim both: check whether one program can see what another just sent.

Not the same as Customer Relationship Management (CRM): A CRM records who the contacts, accounts, and deals are; a MAP decides which message each contact gets next.

In depth: CDP vs. CRM vs. Marketing Automation: What Each Does

Email service provider (ESP)

Software whose core job is sending email at scale: list management, templates, scheduled sends, and delivery infrastructure.

An ESP is enough when your programs are calendar-driven sends to segments. Upgrade when programs genuinely need behavior triggers and branching — not because a vendor tier renamed itself "automation".

Not the same as Marketing automation platform (MAP): An ESP’s core job is sending email well; a MAP runs multi-step programs across channels from triggers and rules, and email is only one of those channels.

Journey orchestration

Coordinating messages across programs and channels around one customer’s state, so the sequence behaves as one conversation rather than parallel campaigns.

The unglamorous parts are the product: cross-program frequency caps, quiet hours in the recipient’s timezone, and exit rules for people who convert. A journey map without those three is a wish, not orchestration.

Not the same as Marketing automation platform (MAP): A MAP executes programs; journey orchestration coordinates programs and channels around one customer’s state, so they stop running in parallel. Many MAPs claim both: check whether one program can see what another just sent.

Segmentation

Dividing an audience into groups by attributes or behavior — lifecycle stage, engagement, purchase history — so programs and messages can differ by group.

Segments rot: the rule that defined one two years ago rarely matches how the business talks today. Audit membership logic on a schedule, and delete segments nobody can explain — an audience no one understands is a targeting bug waiting to send.

Not the same as Lead scoring: Segmentation sorts contacts into groups that get different treatment; lead scoring ranks individuals by fit and behavior to decide whom sales contacts first.

Lead scoring

Ranking leads by fit and behavior — points for firmographics, engagement, and actions — to decide which ones sales touches first.

Scores inflate: every added rule adds points somewhere, and thresholds set once are rarely recalibrated. The only defensible calibration is backward-looking — check what scores the leads that actually closed had, then set thresholds from that.

Not the same as Segmentation: Segmentation sorts contacts into groups that get different treatment; lead scoring ranks individuals by fit and behavior to decide whom sales contacts first.

Not the same as MQL and SQL: A score is a number; MQL and SQL are the stages a lead moves through when the score crosses an agreed threshold and someone accepts the handoff.

MQL and SQL

Marketing qualified lead and sales qualified lead: the two stage labels marking marketing’s handoff to sales, defined by criteria both sides agree to.

The letters are trivial; the definition is the entire game. If marketing and sales have not written down the criteria and the exit rules — what returns a lead, and when — the MQL count is a vanity metric that funds arguments instead of pipeline.

Not the same as Lead scoring: A score is a number; MQL and SQL are the stages a lead moves through when the score crosses an agreed threshold and someone accepts the handoff.

Suppression list

The list of addresses a program must never contact: unsubscribes, hard bounces, complaints, and legally excluded contacts.

This list is legally load-bearing. Unsubscribes must survive every migration and reach every sending tool — consent does not transfer between channels, and a suppression dropped in an export is a compliance incident, not a data cleanup item.

Email deliverability

Whether sent email reaches inboxes: a function of authentication (SPF, DKIM, DMARC), sender reputation, complaint rates, and list quality.

Since February 2024, Gmail and Yahoo require bulk senders — 5,000 or more messages a day — to authenticate with SPF, DKIM, and DMARC, offer one-click unsubscribe, and keep user-reported spam under 0.3%. Those are floors, not targets; reputation damage starts well below the enforcement line.

Personally identifiable information (PII)

Data that identifies a person directly (name, email, phone) or in combination (an IP address plus context), with exact definitions varying by law.

The classic martech leak is the URL: form pages that put an email address in query parameters feed PII into every analytics and advertising tag on the page. Google Analytics prohibits collecting it, and the fix — checking what your forms append to URLs — takes an afternoon.

Not the same as First-party data: First-party describes where data came from; PII describes whether it identifies someone. Much first-party data is PII, and collecting it yourself does not exempt it from consent rules.

Consent management platform (CMP)

The tool that presents privacy choices, records them, and exposes the resulting consent state for other systems to enforce.

Buying a CMP is not compliance — wiring it is. The project is the enforcement path: tags that actually wait for consent, a documented denied state, and a mid-session change test. The banner is the visible tenth of the work.

Not the same as Google Consent Mode: The CMP asks for the choice and records it; Consent Mode only passes that choice to Google’s tags. Without a CMP behind it, Consent Mode has nothing true to pass.

Google Consent Mode

Google’s API for passing consent state to its tags, which then adjust behavior — from full measurement down to cookieless pings that feed modeled conversions.

Version 2 — two added advertising signals — has been required for ad personalization features in the EEA since March 2024. Read modeled numbers as modeled: when consent rates drop, Consent Mode fills reports, not audiences, and the model’s input is the behavior of people who did consent.

Not the same as Consent management platform (CMP): The CMP asks for the choice and records it; Consent Mode only passes that choice to Google’s tags. Without a CMP behind it, Consent Mode has nothing true to pass.

Not the same as IAB TCF (consent string): Both carry a consent choice to tags. The TCF string is the advertising industry’s standard format, read by programmatic vendors; Consent Mode is Google’s own API for its tags.

IAB TCF (consent string)

The IAB Transparency and Consent Framework: the advertising industry’s standard for encoding consent choices in a machine-readable string that programmatic vendors read; version 2.2 is current.

It matters in proportion to your programmatic footprint in the EEA and UK. If you run no programmatic advertising, your CMP does not need TCF registration — and a simpler consent implementation is easier to make honest.

Not the same as Google Consent Mode: Both carry a consent choice to tags. The TCF string is the advertising industry’s standard format, read by programmatic vendors; Consent Mode is Google’s own API for its tags.

Data governance

The rules and ownership around data: who may create fields, events, and audiences; what names they use; which systems are authoritative; who reviews changes.

In martech, governance is mostly naming and permission discipline — an agreed event taxonomy, one owner per system, a review step before new tags ship. It sounds bureaucratic and costs hours a month; its absence is why analytics audits find hundreds of events serving six questions.

Law 25 and CASL (Canada)

Québec’s Law 25 modernized private-sector privacy law in phases from 2022 to 2024 — privacy officer, breach reporting, consent standards, impact assessments. CASL, Canada’s federal anti-spam law, has required consent for commercial electronic messages since 2014.

For Canadian teams these bind daily martech practice more than the GDPR does. The CASL detail programs forget: implied consent expires — roughly two years after a purchase, six months after an inquiry — so a database that never re-earns express consent is quietly shrinking its lawful audience.

AI crawler

A bot that fetches web content for AI systems. The name covers three different jobs: training crawlers building datasets, search indexers feeding AI answers, and user-triggered fetchers retrieving a page during a conversation.

The three deserve separate policy — OpenAI alone runs GPTBot (training), OAI-SearchBot (search), and ChatGPT-User (on-demand fetches). Blocking “AI” as one category usually means blocking the answers channel you wanted while the training you objected to happens elsewhere. Telling them apart in server logs is covered in the AI search traffic guide.

Not the same as AI agent: An AI crawler fetches content for training, for an index, or for one answer; an AI agent acts, filling forms, calling tools, and editing records toward a goal.

In depth: Can AI Engines Read Your Site? Check Three Layers

AI visibility

How often, and how favorably, a brand appears in AI-generated answers — the AI-era analogue of rankings, measured by querying assistants with prompt sets and scoring mentions and citations.

Treat any single score as one run of a probabilistic system: the same prompt re-run produces different answers, so a score without sample size and variance is a screenshot, not a measurement. The full argument is in the AI visibility scores analysis.

Not the same as Generative engine optimization (GEO): GEO is the work; AI visibility is the measurement of whether it worked.

Answer engine optimization (AEO)

Optimizing content to be selected and cited by answer engines — AI assistants and AI-augmented search — typically through extractable structure: direct answers, question-form headings, and machine-readable metadata.

Most AEO work is SEO work under a new invoice line: crawlability, clarity, and authority still decide selection. The genuinely new part is answer extractability — the first paragraph under each heading has to survive being quoted alone.

Not the same as Generative engine optimization (GEO): AEO is about being selected and cited for a specific answer; GEO is the wider umbrella for presence in generative results, technical surface included. Vendors use the two almost interchangeably.

Generative engine optimization (GEO)

The umbrella term for earning presence in generative AI results — content structure, entity clarity, citations, and the technical surface (feeds, markdown renditions, metadata) that makes content easy for AI systems to ground on.

GEO and AEO circulate as competing labels for overlapping work; no vendor taxonomy is standard, so judge deliverables, not vocabulary. Ask what would change on the site — if the answer is structure, metadata, and citations, the work is real; if it is a dashboard of unstable scores, see AI visibility.

Not the same as Answer engine optimization (AEO): AEO is about being selected and cited for a specific answer; GEO is the wider umbrella for presence in generative results, technical surface included. Vendors use the two almost interchangeably.

Not the same as AI visibility: GEO is the work; AI visibility is the measurement of whether it worked.

In depth: Can You Trust AI Visibility Scores?

llms.txt

A proposed convention (llmstxt.org, 2024): a Markdown file at /llms.txt giving language models a curated map of a site’s most useful content.

No major AI provider has committed to reading llms.txt, so treat it as a cheap map, not a lever. This site publishes one and serves every page as Markdown alongside it; the honest framing is hospitality for agents, with adoption as the open question.

Not the same as Content Signals: llms.txt suggests what to read; Content Signals state how fetched content may be used. Neither decides access: robots.txt rules and the CDN or firewall in front of the site do.

In depth: Can You Trust AI Visibility Scores?

Content Signals

A robots.txt extension (Cloudflare’s Content Signals Policy, 2025) declaring how fetched content may be used: search, ai-input — grounding AI answers — and ai-train, each set to yes or no. Cloudflare added a fourth field, use, on July 1, 2026: it takes a level rather than yes or no — immediate (interact, store and reuse nothing), reference (index, excerpt, and link back — the default), or full (summarize and reproduce).

Two honest caveats: compliance is voluntary — Google has said the directive does not change its crawlers’ behavior — and Cloudflare’s own defaults moved under it. New-domain defaults were scheduled for September 15, 2026; on August 21, Cloudflare revised them so that only a site saying it monetizes with ads starts with Training set to Disallow, and other new domains start with no blocks. As of October 10, 2026, Cloudflare’s documentation still describes the July 1 plan in the future tense, so read your own dashboard, not the docs — the three layers of AI crawler access has the detail. This site declares search=yes, ai-input=yes, ai-train=no, use=reference: being usable in AI answers — and linked back to — is the point of the exercise.

Not the same as llms.txt: llms.txt suggests what to read; Content Signals state how fetched content may be used. Neither decides access: robots.txt rules and the CDN or firewall in front of the site do.

AI agent

Software that pursues a goal by taking actions — browsing, calling tools, editing systems — rather than returning one answer; in marketing operations, agents draft, verify, and operate under human review.

The delegation test that holds up: give agents work that is structured, reversible, and cheap to verify, and keep judgment and accountability human. The reasoning, and where the line sits in content work, is in the AI agents guide.

Not the same as AI crawler: An AI crawler fetches content for training, for an index, or for one answer; an AI agent acts, filling forms, calling tools, and editing records toward a goal.

In depth: What Can AI Agents Do in Your Ad Accounts and CRM?

AI Overviews and AI Mode

Google’s generative answers inside Search. AI Overviews summarize an answer above the results and rolled out to everyone in the U.S. on May 14, 2024; AI Mode is a separate conversational search experience that rolled out in the U.S. on May 20, 2025. Both cite and link to web pages.

Google says neither needs special optimization or schema to appear in. The trap is measurement: their clicks arrive with an ordinary google.com referrer, so GA4 counts them in Organic Search rather than its AI Assistant channel, and no standard report isolates them. How to measure AI search traffic sets out what each system can and cannot see.

Not the same as Zero-click search: Zero-click is an outcome, a search that ends without a visit; AI Overviews are one feature that produces it, alongside featured snippets and knowledge panels. AI Overviews also link out, so not every one ends without a click.

Zero-click search

A search that ends on the results page, without a click to any website, because the answer was shown there: a featured snippet, a knowledge panel, or an AI-generated summary.

Published zero-click percentages come from clickstream panels with their own biases, so treat any single figure loosely; the direction is not in doubt. The practical consequence is in reporting: when impressions rise and clicks do not, a page may still be doing its job as a cited source. Read the two separately, with the same caution as AI visibility scores.

Not the same as AI Overviews and AI Mode: Zero-click is an outcome, a search that ends without a visit; AI Overviews are one feature that produces it, alongside featured snippets and knowledge panels. AI Overviews also link out, so not every one ends without a click.

Query fan-out

The technique Google says AI Overviews and AI Mode may use to build an answer: issuing multiple related searches across subtopics and data sources, then combining what comes back.

It is why a page can be cited for a question it never targeted: it ranked for one of the sub-queries. Tools that claim to reveal the exact fan-out queries for a prompt are inferring them, since Google does not report them to site owners. Covering a topic’s natural sub-questions in self-contained sections is the durable response — the first item in the SEO architecture checklist’s AI section.

Grounding (retrieval-augmented generation)

Supplying a language model with retrieved documents at answer time so it responds from them rather than from its training alone; retrieval-augmented generation (RAG) is the common architecture. AI search answers are grounded in the pages they fetch.

Grounding is what the ai-input signal in Content Signals governs, separately from training. For martech, ask any vendor selling an AI assistant which sources it grounds in, how fresh they are, and whether its answers cite them: an ungrounded answer about your own customer data is a guess in a confident tone.

In depth: How to Measure AI Search Traffic: Four Systems, Four Blind Spots

Structured data (schema.org)

Machine-readable statements about a page — what it is, who wrote it, what it describes — written in the schema.org vocabulary, usually as JSON-LD. Search engines use it to understand pages and to decide eligibility for rich results.

Markup buys eligibility, not rank, and it must describe what the page visibly shows. Google limited FAQ rich results to authoritative government and health sites on August 8, 2023, and stopped showing HowTo results as of September 13, 2023, so a vendor still selling those as a growth tactic is years behind; Google also says no special schema is needed for AI Overviews. Generate markup from the same data as the page, as the breadcrumb item describes.

IndexNow

A protocol that lets a site notify search engines the moment a URL is added, updated, or deleted, instead of waiting to be recrawled. Microsoft Bing and Yandex launched it in October 2021; indexnow.org lists Amazon, Bing, Naver, Seznam.cz, Yandex, and Yep as participating.

Google is not among them — it said in 2021 that it would test the protocol and has not adopted it — so the sitemap and Search Console remain the route there. IndexNow speeds discovery; it is not a ranking signal, and resubmitting unchanged URLs spends crawl budget and earns rate limits. Submit what changed, after it is live.

WebMCP

A proposed browser API that lets a web page register tools that an AI agent in the browser can call, instead of the agent driving the page by clicking through it. It is a draft of the W3C Web Machine Learning Community Group — not a W3C standard — with editors from Microsoft and Google, and Chrome opened an origin trial in Chrome 149, announced on June 9, 2026.

The API is document.modelContext.registerTool(), with an AbortSignal to unregister; many write-ups still quote an older navigator.modelContext shape. This site registers read-only tools that way, including on the stack assessment. Treat tool arguments as untrusted input and validate them as you would a form.

Not the same as Model Context Protocol (MCP): An MCP server is a service an assistant connects to over the network; WebMCP registers tools from inside a web page, for an agent already working in the browser.

In depth: What Can AI Agents Do in Your Ad Accounts and CRM?