Understand the words used in websites and digital marketing.

Comparing two versions of a page, message or feature with similar audiences to learn which performs better against an agreed measure. A valid test needs enough suitable data and should not be stopped simply when one version briefly moves ahead.

The part of a webpage visible before a visitor scrolls. It should establish context and a useful next step, but important information should not be forced into a crowded first screen.

Making a website usable by people with different abilities and assistive technologies. It includes keyboard operation, readable contrast, clear labels and understandable content. An overlay alone does not establish accessibility.

Google’s calculation used to determine whether an advertisement can show and where it may appear. Bid, ad quality, expected effect of assets, competition and search context can all contribute.

A controlled entry point through which AI agents reach tools, data or other agents. A gateway can apply identity, permissions, rate limits and logging consistently, but it does not decide whether the underlying action is appropriate.

A distinct technical identity assigned to an AI agent so its permissions and actions can be attributed and revoked. It should not share a broad human administrator account.

An inventory of approved AI agents, their owners, versions, permissions and deployment status. It helps limit unrecorded agent sprawl only when entries are reviewed and obsolete agents are removed.

AI designed to pursue a goal through several steps, often by choosing tools, checking results and adapting its next action. The term describes a capability pattern rather than guaranteed autonomy; permissions, limits, monitoring and human approval still determine what the system can safely do.

The European Union regulation that applies different obligations according to an AI system’s role and risk. Some UK organisations can be affected when they provide systems or use outputs in the EU. It is separate from UK data-protection, consumer, equality and sector-specific rules.

Software that uses an AI model to work through a goal and may call other tools or take several steps. Unlike a simple chatbot reply, an agent can look up information or trigger an action, so its permissions, checks and stop conditions need to be tightly defined.

Software that uses an AI model to help with tasks such as answering questions, drafting, searching approved information or operating connected tools. Its reliability depends on the model, instructions, source data, permissions and the way uncertain answers are handled.

A standardised set of tasks and scoring rules used to compare model performance. Benchmark results can help shortlist models, but may not predict performance on an organisation’s own data, users, costs or safety requirements.

A repeatable test of an AI system against representative tasks and agreed scoring rules. Useful evals examine accuracy, unsupported claims, safety, cost and failure handling on the actual use case rather than relying only on a provider benchmark.

The policies, responsibilities and evidence used to decide which AI systems may be used, what data they may access, who approves important actions and how risks, incidents and changes are reviewed. Governance should match the real consequences of the use case.

The practical knowledge needed to use and oversee AI responsibly, including what a system can and cannot do, how to question its output, when human review is required and what privacy, security or fairness risks may arise. The appropriate level differs by role and use case.

A Google Search experience that uses generative AI to answer and explore queries conversationally. Its interface, availability and reporting can change. Website eligibility continues to rely on accessible, indexable and useful content rather than a special guaranteed-inclusion technique.

The software and rules that coordinate models, retrieval, tools, memory and review steps. Good orchestration makes state, failures, permissions and costs observable instead of hiding the whole process behind one chat response.

A generated summary that Google may show within search results, supported by links selected for that query. An appearance is not guaranteed, and a cited page is not necessarily the highest conventional organic result.

Structured adversarial testing intended to uncover unsafe, unreliable or exploitable behaviour before and after deployment. Findings need prioritisation and remediation; a single red-team exercise is not permanent assurance.

An informal, critical term for low-value AI-generated material produced at scale with little checking, originality or care. It is not a technical classification, but it highlights why publication volume should not replace editorial judgement and accountability.

Software that infers how to produce outputs such as predictions, content, recommendations or decisions from the inputs it receives. In practice, the surrounding data, interfaces, rules, integrations and people are part of the operational system, not only the model.

A repeatable process that combines model calls with business rules, data sources, software actions and review steps. A workflow can be useful without being fully autonomous; deterministic checks should handle decisions that must always follow fixed rules.

AI optimisation: making useful, accurate information easier for AI-powered search and answer systems to retrieve and interpret. It builds on sound SEO and does not guarantee citations.

A text alternative describing the purpose or meaningful content of an image for people who cannot see it. Decorative images should usually have an empty alt attribute rather than a repeated keyword.

A search or AI service that tries to provide a direct response rather than only a list of links. Clear, accurate and well-structured source content can make information easier to understand, but no organisation can guarantee that an answer engine will quote or recommend it.

Application programming interface: an agreed way for software systems to exchange information or request actions. An integration depends on the permissions, limits and reliability of the provider’s API.

The method used to assign credit for an enquiry or sale to marketing activity. Attribution reports are models, not complete records of every influence on a customer’s decision.

A sequence of rules that moves information or starts an action without somebody completing every step manually. For example, a valid website enquiry could create a CRM contact, notify the correct team and schedule a follow-up task while keeping an audit trail.

Recoverable copies of website files or data. Their value depends on what they include, how recently they were created and whether the recovery process works.

A measure describing sessions that did not meet an engagement condition. Its exact calculation depends on the analytics product and configuration; a high or low number is not automatically good or bad without page context.

A Google Ads keyword setting that can match searches related to the keyword’s meaning, not only the same words. It can expand reach, but it relies heavily on conversion quality, bidding signals, exclusions and search-term review.

Copies of website resources stored on a visitor’s device so they do not need to be downloaded again immediately. Appropriate cache rules improve speed, while changed files need versioning so visitors receive the new copy.

The preferred address for a page when similar or duplicate versions exist. A canonical tag signals that preference to search engines; it does not redirect visitors or guarantee which version will be indexed.

Content delivery network: servers that help deliver assets closer to visitors. A CDN can also provide caching or security features, depending on the service and configuration.

Content management system: software for editing and publishing website content without changing every page’s code. WordPress is one example. Editing permissions and updates still need management.

An AI capability that interprets a screen and operates an interface through actions such as clicking, typing or scrolling. It can work with software that has no suitable API, but it is more fragile than a direct integration and needs strict approval boundaries for consequential actions.

A freely given, specific, informed and unambiguous agreement to a particular use of information. Optional marketing consent should not be bundled into an unrelated enquiry or service.

A Google tagging framework that changes how supported tags behave according to a visitor’s consent choices. It does not create valid consent by itself; the website still needs an appropriate banner, default state, accurate categories and a way to withdraw consent.

Designing the information supplied to an AI system at each step, including instructions, retrieved records, tool results, examples and conversation history. It extends beyond writing a prompt by managing what enters the model’s limited working context and when.

The maximum amount of input and generated material a model can consider in one interaction, measured in tokens. A large window does not guarantee that every detail will be recalled or weighted correctly, so important information still needs structure and testing.

An action chosen as a useful business outcome, such as a completed enquiry or booking. A tracked action is not automatically a qualified lead or a completed sale.

The proportion of eligible visits or users completing a defined action. The denominator, attribution window and action definition must be consistent before periods or channels can be compared.

The technical setup used to record selected actions after marketing or website activity. It needs testing, consent controls and clear definitions; the recorded event may still differ from a qualified sale.

A small value a website can ask a browser to store and return. Cookies can support essential sessions, preferences, measurement or marketing; the purpose and legal basis matter more than the technology’s name.

Google’s user-experience metrics covering main-content loading, responsiveness to interactions and visual stability. Real-user measurements can differ from a single laboratory speed test.

Cost per acquisition or action: advertising cost divided by the number of attributed results. The report should say exactly which action is being counted and whether it represents a lead, customer or another step.

Cost per click: advertising spend divided by the number of clicks. A lower CPC is not necessarily better if those visitors are unlikely to become customers.

Cost per lead: marketing cost divided by recorded leads. Lead quality, duplicate contacts and offline outcomes need to be considered before comparing campaigns.

Whether search-engine crawlers can discover and request a page. Crawlable content can still be excluded from search results for other reasons.

Customer relationship management software: a system for organising contacts, enquiries and customer activity. Useful CRM integration reduces repeated entry while preserving clear ownership and permissions.

Conversion rate optimisation: researching and improving the journey towards a useful action. It can involve content, usability, performance and testing; it is broader than changing button colours.

Content Security Policy: response rules telling a browser which sources may load scripts, images and other resources. It reduces some attack opportunities but does not replace secure development.

Cascading Style Sheets: the language that controls presentation, including colours, spacing, typography and responsive layouts.

Call to action: a prompt helping someone take an appropriate next step, such as requesting a consultation or reading supporting evidence. The wording should describe what actually happens.

Click-through rate: clicks divided by impressions, usually expressed as a percentage. Its meaning depends on the channel, placement and people who saw the result.

The organisation or person that decides why and how personal data is processed. A controller remains accountable for those decisions even when suppliers carry out parts of the processing.

A structured store used to pass website and business events to a tag-management or analytics setup. A clear event contract reduces duplicated tags and inconsistent naming; personal information should not be added without a defined, lawful need.

The principle of collecting and using only the personal data that is adequate, relevant and necessary for a stated purpose. Keeping extra fields “in case they become useful” is not the same as demonstrating a need.

A supplier that processes personal data on a controller’s documented instructions. The relationship, security duties, subprocessors, international transfers and deletion arrangements normally need a suitable contract.

Synthetic or manipulated audio, imagery or video that realistically depicts a person saying or doing something they did not do. Legitimate creative uses still require attention to consent, rights, disclosure and the risk of misleading an audience.

A digital representation of a physical object, process or environment that is updated with relevant data. It may support monitoring or simulation, but its usefulness depends on how accurately the model and incoming data reflect reality.

Domain Name System: records that connect a domain name to services such as website hosting and email. Changing the wrong record can interrupt those services.

The memorable website address people type, such as londondigitalworks.co.uk. Registration, DNS and website hosting are connected but separate services, so access and renewal responsibility should be recorded clearly.

Data protection impact assessment: a documented process for identifying and reducing privacy risks before high-risk personal-data processing begins. It is a decision-making record, not a one-off form or an automatic approval.

Experience, Expertise, Authoritativeness and Trustworthiness: qualities discussed in Google’s search-quality guidance. They are not a single score or a piece of markup that guarantees rankings.

A numerical representation that places related meanings near one another in a mathematical space. Embeddings help semantic search and retrieval find relevant material even when the query and source use different words; they can still expose or reproduce sensitive information.

A Google Ads measurement feature that sends selected first-party contact data in hashed form to help match conversions with advertising interactions. Hashing is not anonymisation, so transparency, consent or another valid basis, minimisation and configuration still matter.

A recorded interaction or occurrence in an analytics system, such as a submitted form or viewed page. Event names, parameters and triggering conditions need consistent definitions.

A paid-search keyword setting focused on searches with the same meaning or intent as the keyword. It is narrower than phrase or broad match, although Google may still match close variants rather than only an identical character-for-character query.

Additional training that adjusts a pre-trained model using task-specific examples. It can improve consistent behaviour or specialised performance, but it is not the usual way to provide frequently changing facts; retrieval or tools may be more suitable for that.

Information an organisation collects through its own direct relationships and services. It still requires an appropriate purpose, lawful handling, security and retention policy.

A model trained on broad data that can be adapted to many downstream tasks through prompting, retrieval or further training. An application built on a foundation model still needs use-case testing because general capability does not prove suitability for a particular business process.

An informal term for a highly capable model near the current edge of general-purpose AI performance. There is no single permanent threshold, and greater capability can bring higher cost, latency or risk as well as better results.

A way for an AI model to request a defined software function using structured arguments. The application—not the model—must validate those arguments, enforce authorisation, run the function and decide how its result is returned.

A simplified model of stages between awareness and a business outcome. Real customer journeys are rarely linear, so a funnel is a planning aid rather than a complete description of behaviour.

Google Analytics 4: Google’s event-based analytics product. Its reports depend on correct implementation, user choices, browser restrictions and the definitions used for events.

General Data Protection Regulation: a framework governing personal-information processing. UK GDPR applies in the UK alongside other relevant rules. Compliance involves organisational decisions, not just a cookie banner.

AI that produces new material such as text, code, images, audio or video from learned patterns and current inputs. Generated output can be useful and fluent while still being inaccurate, unoriginal or unsuitable for publication without review.

Generative engine optimisation: improving the clarity, evidence and machine-readable context of information so generative search systems can understand it. The term is still evolving and does not imply guaranteed AI citations.

A Google service for understanding search visibility, indexing and some technical issues affecting a website. It reports search activity rather than all website traffic.

A tool for managing website tags and their triggers. A published container can change data collection, so access, testing and consent controls matter.

General-purpose AI: a term used in the EU AI Act for models able to perform a wide range of tasks and integrate into many systems. The regulation distinguishes model providers from organisations that deploy a resulting AI system.

A control pattern in which an AI agent begins with restricted permissions and may receive narrowly expanded authority after task-specific evidence. High-impact actions can remain subject to human approval regardless of past performance.

Connecting an AI response to supplied evidence, such as approved documents, search results or database records. Grounding can improve relevance and traceability, but it does not prove that the evidence is current, complete or interpreted correctly.

A technical or organisational control intended to prevent, detect or limit unwanted AI behaviour. Examples include access checks, content filters, spending limits and human approval. No single guardrail should be treated as a complete safety boundary.

A plausible-sounding but unsupported or factually wrong output presented by a generative AI system. Retrieval, citations and checking can reduce the risk, but important facts should still be verified against authoritative sources.

A visual summary of where sampled visitors click, move or scroll on a page. It can suggest where people hesitate or miss information, but it does not explain motivation on its own and must be configured with appropriate privacy controls.

An annotation identifying equivalent pages for different languages or regional audiences. It is for genuine alternatives, not for inventing location versions of an English-only page.

HTTP delivered over an encrypted TLS connection. It protects information in transit but does not prove that a website’s content, owner or application is trustworthy.

A workflow in which a person reviews, approves, corrects or takes over at a defined point. The review is meaningful only when the person has enough information, time and authority to challenge the system rather than merely confirming its recommendation.

A count recorded when content or an advertisement is eligible to be shown under a platform’s rules. An impression does not prove that a person noticed or understood it.

The process by which a search engine evaluates and stores information about pages for possible search results. Inclusion and rankings are not guaranteed by submitting a sitemap.

The stage when a trained model processes new input to produce an output. Inference cost and speed depend on factors such as model size, input length, output length, hardware, caching and the number of tool calls.

Interaction to Next Paint: a Core Web Vital measuring responsiveness to user interactions throughout a page visit. Lower times generally indicate that the interface responds more promptly.

A link from one page to another page on the same website. Useful internal links help visitors continue a task and help search engines understand how pages and topics relate.

A programming language used to add behaviour to websites, such as menus, filters and application interactions. Essential content and links should not depend unnecessarily on it.

In GA4, an event marked as especially important to the business. Choosing the right event and testing when it fires is essential to meaningful reporting.

A third-party estimate of how competitive it may be to rank for a search term. Different tools calculate it differently, so it should support judgement rather than decide a strategy alone.

The likely purpose behind a search, often described as informational, navigational, commercial or transactional. One phrase can still reflect several needs, so the actual results page should be reviewed.

A page designed around a visitor’s arrival context and a specific next step. It should match the promise in the advert, link or search result that brought the visitor there.

A model trained to predict and generate sequences of language tokens. Large language models can perform many text and reasoning tasks, but fluent output is not proof of factual knowledge, permission or understanding in the human sense.

The time between a request and a useful response. For an AI workflow, total latency can include retrieval, model processing, tool calls, network delay and validation rather than only the model’s generation speed.

The legal justification relied on for processing personal data under UK GDPR, such as contract, legal obligation, legitimate interests or consent. Consent is one option, not the default answer for every activity.

Delaying non-essential resources until they are needed. Below-the-fold images can benefit; delaying the main visible image can make the page feel slower.

Largest Contentful Paint: a Core Web Vital estimating when the largest main content element becomes visible. Image delivery, server response, fonts and rendering can affect it.

A person or organisation that has shown potential interest, commonly by making contact. A submission can be invalid, duplicated or unsuitable, so lead quantity should not be confused with revenue.

An assessment of whether an enquiry is relevant, genuine and realistically likely to become useful business. Measuring quality helps avoid optimising campaigns merely for a larger number of unsuitable form submissions.

An optional text file that points AI tools towards important website information. It is not a universal standard, a crawler access rule or a guarantee of inclusion in AI answers.

The map-led group of local business results that Google may show for searches with local intent. Relevance, distance and prominence influence visibility; no provider can guarantee inclusion.

A relatively specific search phrase that often has lower individual search volume but clearer context. A useful content plan groups closely related needs instead of making a thin page for every variation.

A field of AI in which systems learn patterns from data to make predictions or generate outputs instead of relying only on hand-written rules. The data, objective and evaluation method strongly shape the behaviour produced.

Model Context Protocol: an open standard for connecting AI applications with external tools, prompts and data resources. MCP standardises the connection; each server still needs trustworthy code, appropriate permissions, clear descriptions and review of the data it receives.

Information describing a page or asset, such as its title, search description or image dimensions. Good metadata supports interpretation but cannot compensate for weak or inaccurate visible content.

Multi-factor authentication: requiring more than one type of evidence when signing in. Two-factor authentication is a common form. Recovery procedures also need to be protected.

Training a smaller model to reproduce useful behaviour from a larger model or set of outputs. The result may be cheaper and faster for a focused task, while losing capability or reproducing errors present in the teacher data.

A numerical value learned during model training. Parameter count can indicate scale, but it does not by itself measure accuracy, reasoning quality, efficiency or suitability for a business task.

Software that chooses which AI model or workflow should handle a request based on factors such as task type, risk, latency, capability or cost. Routing rules need evaluation because an apparently simple task can still contain sensitive or difficult content.

Marketing-qualified lead: a contact judged to meet agreed marketing criteria before sales review. The criteria should be documented and tested against later outcomes.

AI that can work across more than one type of input or output, such as text, images, audio or video. Support for a modality does not mean equal accuracy across every format, language, device or accessibility need.

A paid-search control that blocks ads from searches matching an excluded word or phrase. It can reduce irrelevant spend, but an overly broad list can also prevent valuable searches from reaching the correct advert.

A directive asking search engines not to include a page in results. The crawler generally needs access to the page to read the directive. It is not a privacy or access-control mechanism.

A business outcome recorded after the initial website or advertising interaction, such as a qualified enquiry or completed sale. Importing it can improve optimisation, provided records are deduplicated, minimised and matched lawfully.

A model whose trained parameter weights are available under stated licence terms. Open weights can support local deployment and inspection, but do not necessarily include the training data, training code or unrestricted permission to use the model.

Visits attributed to unpaid search results by an analytics platform. Consent choices, attribution rules and browser behaviour mean analytics totals may not match search-platform clicks.

How quickly a webpage loads and becomes usable. Laboratory tests help diagnose problems, while field data reflects real devices, networks and visitors.

An additional value passed with a URL, analytics event or software request. Parameters can carry useful context, but sensitive personal information should not be placed in public URLs or analytics data.

An agreed limit for page weight, loading time or the amount of JavaScript and imagery a page may use. It helps a team make design and technology choices before attractive additions quietly make the experience slow.

Information relating to an identified or identifiable person. Names and email addresses are obvious examples, but online identifiers, location, behavioural data and combinations of records can also be personal data.

A paid-search keyword setting that can match searches containing the keyword’s meaning, including additional words before or after it. Search-term review remains necessary because meaning-based matching is broader than a literal phrase.

Pay-per-click advertising: a model in which advertisers pay for clicks. Campaign management should connect targeting, budgets, landing pages and the quality of resulting enquiries.

The instruction or context supplied to an AI model. A good prompt can improve relevance, but reliable business use also depends on approved source information, permissions, testing and how the system handles uncertainty.

Reusing previously processed portions of an AI input to reduce repeated cost and latency. Cache behaviour, retention and eligibility differ by provider, so sensitive context and changing instructions still need careful handling.

Designing instructions, examples and output requirements so an AI model can perform a task more consistently. It is useful, but it cannot replace access control, authoritative data, evaluation or conventional software validation.

An attack or unintended conflict in which untrusted text tries to override an AI system’s instructions. Indirect injection can arrive through a webpage, document or email the model reads. Impact is reduced through restricted tools, external authorisation and human confirmation—not prompt wording alone.

A Google Ads diagnostic based on expected click-through rate, ad relevance and landing-page experience for a keyword. It helps identify weaknesses, but it is not a direct auction input, business-profit score or guarantee of performance.

Representing a model’s numerical weights with lower precision so it can use less memory and often run faster. The trade-off can include some loss of output quality, so the actual task should be evaluated after conversion.

Retrieval-augmented generation: a method that finds relevant information from an approved source and gives it to an AI model when forming an answer. It can reduce unsupported responses, but the source quality, retrieval settings and final output still need testing.

Restricting how many requests a user, device or system may make within a period. It can reduce abuse and unexpected cost, but legitimate traffic patterns and distributed attacks require more than one simple limit.

A model designed to spend additional computation working through multi-step problems before answering. It may improve difficult analysis, but longer processing does not guarantee correctness and can increase latency and cost.

A response that sends a visitor or crawler to another URL. Permanent redirects help consolidate replaced addresses; avoid unnecessary chains and redirects to unrelated pages.

Advertising aimed at people based on earlier interaction with a website or business. It requires careful audience rules, consent handling and platform configuration.

Layouts that adapt to different viewport sizes and input methods. A responsive site still needs checks for readable text, usable controls and content that does not overflow.

Return on ad spend: attributed revenue divided by advertising cost. It normally excludes management fees, fulfilment costs and untracked sales unless the calculation explicitly includes them.

A public file giving cooperating crawlers instructions about what they may request. It does not secure private files, force indexing or reliably remove a page already in search results.

Handling or transforming input to reduce unsafe content. It is not a substitute for validation, context-appropriate output escaping or parameterised database queries.

A shared vocabulary for structured descriptions of content, often published as JSON-LD. Markup should reflect visible, accurate information; unsupported reviews or ratings should not be added.

The task or answer a searcher is likely trying to complete. Understanding intent helps decide whether a service page, guide, comparison, tool or local result is the appropriate format.

An estimate of how often people search for a term during a period. It is useful for comparing demand, but tools use sampled data and a high-volume phrase may be less commercially relevant than a smaller, more specific one.

Search engine optimisation: improving a website’s technical foundations, content and discoverability for relevant searches. It cannot guarantee rankings or enquiries.

Search engine results page: the page showing organic results, advertisements, maps, answers and other search features. Its composition can differ by location, device and query.

Producing a page’s main HTML on the server before it reaches the browser. It can improve resilience and discovery, although performance and accessibility still depend on the complete implementation.

A measurement setup that routes selected events through a controlled server endpoint before sending them to analytics or advertising providers. It can improve governance and resilience but does not bypass consent, transparency or data-minimisation duties.

A group of interactions that an analytics system associates with one visit over a defined period. Sessions are measurement constructs rather than a perfect count of people, especially when consent, devices and browser restrictions differ.

A list of important website URLs, often supplied as XML for search engines. It helps discovery but does not replace internal links or guarantee indexing.

A comparatively compact language model designed to use fewer computing resources. It may be faster, cheaper or suitable for local use, while offering less capability than a larger model on complex or unfamiliar tasks.

Google Ads auction-time bidding strategies that use machine learning to optimise for configured conversions or conversion value. The strategy can only pursue the signals supplied, so duplicate, low-quality or wrongly primary conversions can distort it.

Single-page application: an interface that changes views without loading an entirely new document each time. Public sites still need meaningful URLs, accessible navigation and content that search engines can retrieve.

Machine-readable information added to a page to describe visible content such as an organisation, service or breadcrumb trail. Valid markup can help systems interpret a page, but it does not guarantee a special search result.

Artificially generated data designed to resemble useful patterns in real data. It can support testing or model development, but may reproduce bias, omit rare cases or remain linkable to source information depending on how it was created.

Images, audio, video or text produced or substantially altered by generative systems. Publication decisions should consider accuracy, copyright, consent, provenance, disclosure and the overall impression given to the audience.

Work that helps search engines request, interpret and consolidate website pages, including status codes, internal links, canonicals, sitemaps and rendering. It supports good content rather than replacing it.

A model setting that influences how varied or predictable generated output is. Lower temperature can make responses more consistent, but it does not make facts correct or remove the need for validation.

A statistical pattern a model provider may place in generated word choices so authorised tools can estimate whether that model contributed to text. It is not a statement of truth, ownership or responsible authorship and can weaken after substantial editing.

Transport Layer Security: the protocol commonly used to encrypt HTTPS connections. Certificates must cover the requested domain before an HTTPS redirect can be received.

A unit into which an AI model divides input and output, often a word fragment, punctuation mark or symbol. Providers commonly use token counts for context limits, usage reporting and pricing; one token is not the same as one word.

The pattern in which an AI model selects from defined external capabilities, such as searching a database or creating a calendar event. The surrounding application must control permissions, validate requests, handle errors and log consequential actions.

User interface: the controls and visual elements people interact with. Clear labels, predictable behaviour and readable presentation matter more than decoration alone.

A labelled URL parameter commonly used to describe campaign source, medium and name in analytics. Consistent naming prevents fragmented reports; UTMs should not contain personal information.

User experience: how successfully and comfortably someone can accomplish a task. It includes the content, navigation, speed, accessibility and response when something goes wrong.

Checking that input meets expected rules, such as an acceptable length or email format. Server-side checks remain necessary even when the browser validates the form.

A system designed to store and search embeddings by similarity. It can support semantic retrieval for RAG, but access controls must be enforced on the underlying records so similarity search does not reveal information a user should not see.

An informal term for creating software mainly through natural-language instructions to an AI coding tool. It can speed up exploration, but production work still requires understanding, testing, security review, licensing checks and maintainable ownership of the resulting code.

Web application firewall: a service that filters or challenges web requests using security rules. It can reduce common automated attacks and abusive traffic, but it does not repair vulnerable application code or replace monitoring and updates.

Web Content Accessibility Guidelines: a set of testable criteria for accessible digital content. An accessibility claim should identify the version, level, scope and testing performed.

Software used through a web browser to complete interactive tasks, such as managing bookings, accounts or workflows. Unlike a simple content page, it usually needs roles, data rules and ongoing operational support.

The service that stores and serves a website or application. A suitable plan depends on traffic, software, storage, security and recovery needs; buying hosting alone does not maintain the website.

A request sent automatically from one system when an event occurs. Receivers should verify authenticity, validate the message and safely handle retries or duplicates.

A simplified representation of page structure and priority used before visual styling. It helps teams discuss content, navigation and actions without mistaking decoration for a settled experience.

A security approach that avoids automatically trusting a request because it comes from a particular network or device. Access is continually limited and checked according to identity, context and the minimum permissions needed.

A search in which the person does not visit an external result, often because the results page provides an answer or enables an action directly. Visibility can therefore change without a matching website visit.

Information a person intentionally gives an organisation about their needs or preferences, such as selecting a service in an enquiry form. It should be requested only when useful and handled according to the stated purpose.