Glossary.

The address of a site typed into a browser, for example velvetum.com. Rented per year with renewal.

A domain is the name of your site by which people find it — for example, velvetum.com. Behind the server's numeric address sits a convenient, human-readable name. A domain is not bought outright but rented, usually for a year with annual renewal. If the lease is not renewed, the site at that address stops opening.

The place where a site's files are stored so it stays available on the internet around the clock.

Hosting is the place where a site's files live and from which they are served to visitors: a server in a data center running around the clock. Without hosting there is nowhere to put a site. It is paid for monthly or annually. Modern sites are often placed on cloud platforms that take on speed and uninterrupted uptime.

A technology that encrypts data between a site and its visitor. The padlock and the 'https' in the address bar.

An SSL certificate turns on a secure connection: the data a visitor enters on the site is transmitted in encrypted form. Visually it is the padlock in the address bar and the 'https' prefix instead of 'http'. Without a certificate the browser flags the site as 'not secure', and search engines demote it in results. Today it is a mandatory baseline for any site.

Work on a site so it ranks higher in the classical results of Google and Yandex for customer queries.

SEO (search engine optimization) is the set of work that lifts a site in the classical results of Google and Yandex and brings visitors without paying per click. It covers clear structure, speed, correct titles and descriptions, and useful copy. SEO does not pay off immediately but over weeks and months — after which the traffic flows in steadily.

A site that opens equally well on both a computer and a phone.

A responsive site automatically adjusts to the device: it shows several columns on a computer, a single column on a phone, and enlarges buttons and text. This matters because more than half of visitors arrive from a phone, and an awkward mobile site loses customers.

A single-page site with one goal: capture a lead, registration, sale, or subscription.

A classical landing follows a familiar arc: a first-screen offer, value blocks, social proof, lead form. In the higher price segment a landing is not a template but a bespoke project with original typography and a written-out scenario. In AI-First, a landing gains Schema.org markup (Service, Offer, FAQPage) and starts to appear in AI search answers for the precise service query.

A site built from several pages with their own navigation: home, services, portfolio, about, contacts, blog.

A multi-page format suits companies with multiple offerings or audiences. For AI-First architecture a multi-page site multiplies entry points into neural search: each service page answers its own narrow user question and can land in its own AI answer.

A platform through which the owner adds and edits pages, posts, and products without a developer.

Common CMS options: WordPress, Tilda, Webflow, Sanity, Strapi, plus regional incumbents like 1C-Bitrix. In the higher price segment headless CMS dominates: content lives separately from layout, which buys full design freedom and fast page rendering. For AI-First the headless approach is preferred — content is structured machine-readably and is easy to wrap in Schema.org.

A CMS whose admin layer is decoupled from the site's front-end and delivers content via API to React, Next.js, or Vue. Examples: Sanity, Strapi, Contentful, Directus.

In a classical CMS, design and content are bound by a theme. A headless CMS serves content via API, and the front-end is built separately on a modern stack such as Next.js, React, or Vue. The result is fast rendering, design freedom, and easy AI-First markup. Examples: Sanity, Strapi, Contentful, Directus.

Google's quality metrics for sites: load speed (LCP), interaction responsiveness (INP), visual stability (CLS).

Core Web Vitals directly influence Google rankings and user experience. Healthy values: LCP under 2.5 s, INP under 200 ms, CLS under 0.1. In AI-First the importance doubles — slow pages are indexed worse by neural search engines and land in AI answers less often. Speed is not a PageSpeed Insights score in a vacuum; it is the real time to first meaningful content for the end user.

Related reading:SEO future 2025–2030: 8 shifts from keywords to AI citability · Web performance in 2026: why speed isn't PageSpeed

Design approach where the mobile version is built first and desktop is scaled up from it.

More than 70 percent of traffic in many markets arrives from mobile devices. Google and Yandex index sites primarily by their mobile version. The studio first designs the phone layout and then extends to desktop. AI-First is fully compatible with mobile-first: clean structure reads equally well to neural search and on a small screen.

Related reading:Why an e-commerce store doesn't sell: 7 pitfalls and the fixes · SEO future 2025–2030: 8 shifts from keywords to AI citability · Generations and product 2026: how habits of different ages shape UX

UX is the user experience — the path from first click to goal. UI is the visual interface — color, typography, buttons, icons.

UX and UI solve different problems. UX design answers 'is this site easy to use'. UI design answers 'how does this site look and feel'. In the higher price segment, weak UI undermines trust; even a beautiful site without solid UX fails to convert. AI-First adds a third layer: does the model understand what one can do on this site and how.

Meta's UI library, the de facto front-end standard for modern websites and web applications.

React expresses interfaces as functions of state and reusable components. It powers some of the largest products on the web — Facebook, Instagram, Airbnb, Vercel, the OpenAI Console. In AI-First architecture React is particularly handy because Server Components, in combination with Next.js, render pages on the server and ship clean HTML to crawlers and AI search engines without requiring them to execute JavaScript.

Related reading:Next.js vs React in 2026: when a library is enough · 10 UX/UI design tools in 2026: Figma, Sketch, Framer, Penpot · Ad creative rotation 2026: scheduled replacement cuts CPL by 30–47%

A framework on top of React, made by Vercel, with built-in routing, server rendering, and optimization for fonts and images.

Next.js is the backbone of the studio stack. App Router, Server Components, caching at the Vercel CDN edge, and built-in Schema.org support through generateMetadata form an ideal AI-First base: crawlers get fast HTML with markup, users get an interactive SPA. The studio uses Next.js 16 and Tailwind 4 as the standard for client sites since 2025.

A utility-first CSS framework: layouts assemble from short class names like flex, gap-4, or text-lg directly in HTML.

Tailwind speeds up development and tightens design-system consistency — paddings, colors, and fonts come from config tokens. Version 4 supports CSS variables, layered cascade, and runs faster than its predecessors. In the studio stack Tailwind 4 covers 100 percent of styling for client sites; bespoke CSS is reserved for the heavy effects — parallax, scroll-driven animations, 3D scenes.

Hosting and CDN platform from the creators of Next.js — deploy via git push, edge network across 70+ regions, built-in analytics and WAF.

Vercel is the studio's primary platform. Preview deployment on every push, instant rollback, edge functions, image optimization, defensive WAF, and Core Web Vitals analytics out of the box. For a brand it means a new journal piece or a homepage fix ships to production in minutes, while the CDN caches pages globally — directly improving TTFB and LCP, the two key metrics for landing in AI search.

Related reading:Next.js vs React in 2026: when a library is enough

React components that run only on the server: they ship HTML to the browser and do not send their code to the client.

Server Components dramatically reduce the JavaScript a user has to download and simplify access to databases and APIs straight from a component. In AI-First design they are especially valuable: the page arrives in browsers and crawlers as clean HTML with Schema.org markup already inlined, and neural search engines read it without running scripts. The studio renders most content pages as Server Components.

Related reading:Next.js vs React in 2026: when a library is enough

React components with interactivity — event handlers, state, effects — executed in the user's browser.

In App Router, Client Components are marked with the 'use client' directive and reserved for places that genuinely need interactivity: forms, players, 3D scenes, scroll animations. Over-using them is a common mistake that breaks SEO and LCP. The studio's practice: Server by default, switch to Client only with a proven need.

Rendering a page on the server on every request: users and crawlers receive ready HTML.

SSR guarantees the freshest content and suits dynamic surfaces — catalogs, search, account areas. The cost is higher server load and a worse TTFB compared to pure static. In Next.js, SSR is one of several rendering modes alongside SSG, ISR, and PPR; picks the mode based on load and update cadence.

Related reading:Next.js vs React in 2026: when a library is enough · How to land in Google AI Overviews and ChatGPT Search answers · GEO optimization for service pages

Building every page to HTML at deploy time and serving it from the CDN without hitting the origin.

SSG delivers maximum speed and is ideal for content sites, journals, and landings. Most of velvetum.com is static, cached at Vercel's edge: TTFB under 200 ms from anywhere on the planet, which is critical for AI-First indexing. When content changes in Sanity, on-demand revalidation fires and the page rebuilds without a full redeploy.

Related reading:Next.js vs React in 2026: when a library is enough

A hybrid of static and SSR: a page is cached but rebuilt on a schedule or by signal.

ISR resolves the trade-off between freshness and speed. The page is generated once, served via CDN to every visitor, and rebuilt in the background by a revalidate timer or a revalidateTag call. The studio uses ISR on sections with regular but not constant updates — journal, case studies, pricing — to avoid paying for SSR on every request while staying free from manual redeploys.

Related reading:Next.js vs React in 2026: when a library is enough

Rendering a page in the user's browser via JavaScript after the empty HTML shell loads.

CSR was dominant in the SPA era (React, Angular, Vue) and remains useful for admin panels, dashboards, and account areas. For public pages it loses to SSR and SSG on speed and indexability: a crawler must execute JavaScript, and AI search engines often refuse to. The studio's standard forbids pure CSR on the marketing surface of client sites.

Related reading:Next.js vs React in 2026: when a library is enough

A hybrid where the static shell of a page ships instantly and dynamic blocks stream in.

PPR landed in Next.js 14-15 and closes a vulnerable spot in pure static rendering: a page can carry a personalized block (cart, recommendations, login state) without sacrificing the speed advantages of static caching for the bulk of the content. The studio applies PPR on client home pages with a dynamic hero or personalized recommendations.

A network of edge servers caching a site closer to its users, usually across dozens of regions.

A CDN dramatically lowers TTFB and LCP, especially for users far from the origin data center. Vercel Edge Network has 70+ regions; Cloudflare has 300+. In AI-First the CDN brings a second benefit: AI search crawlers come in from many regions, and without a CDN half of those requests run into a slow origin.

Running code not in a single data center but on edge nodes of the CDN closest to the user.

Edge Functions in Next.js and Vercel handle personalization, A/B tests, geo redirects, and middleware right at the edge — without round-tripping to the origin. That cuts latency for users worldwide and enables scenarios like 'switch the page language by cookie or IP in 10-20 ms'. The studio uses the edge for cookie-based localization on velvetum.com and for its own AI agents.

Time from request to the first byte of the response. Target: under 200 ms; affects both user perception and crawler indexability for Google, Bing, and Yandex.

TTFB depends on distance to the server, rendering complexity, and caching. On velvetum.com the global average TTFB is roughly 170 ms thanks to the Next.js plus Vercel CDN pairing. In AI-First, TTFB directly affects indexability: crawlers with strict timeouts abandon slow pages, and the site falls out of AI search visibility.

Core Web Vitals metric: time to the appearance of the largest visible element — usually the hero image or headline. Target: under 2.5 s.

LCP reflects 'when the user first sees that the page contains content'. Poor LCP is almost always tied to heavy images, missing next/image, a slow server, or blocking scripts in the head. The pre-deploy protocol checklist enforces an LCP measurement on three devices and a 4G profile — production releases are not allowed with LCP above 3.5 s on a mid-tier phone.

Core Web Vitals metric introduced in 2024 to replace FID: the slowest interface response time across a session. Target: under 200 ms.

INP measures not the first but the worst delayed response on the page during the user session. A poor INP is visible on heavy React pages with many states and animations. The studio's practice: defer heavy hero animations to IntersectionObserver, lazy-load chunky components via dynamic import, and debounce form input. After these fixes the INP on velvetum.com sits steadily below 100 ms.

Core Web Vitals metric: total visual shift of a page during load. Target: under 0.1.

CLS deteriorates when images lack explicit width and height, when font swaps cause FOIT/FOUT, or when ad blocks load below the fold. The studio standard: width and height on every image, fonts via next/font with display swap and tuned fallback metrics, ads in pre-reserved slots.

Performance metric: time to the first visible content on a page — text, icon, or image.

FCP reflects rendering start and typically precedes LCP. A healthy FCP sits below 1.8 s at the 75th percentile. FCP suffers from the same culprits as LCP: slow server, render-blocking CSS, heavy JavaScript in the head. An FCP measurement compares mobile and desktop — on premium projects we expect FCP under 1.2 s on 4G.

Next.js built-in component for optimized image delivery: automatic WebP/AVIF, responsive sizes, lazy load.

Next/image closes 80 percent of the visual half of Core Web Vitals: format, size, lazy loading, placeholder, blur-up. The studio standard: on every client site, images go through next/image with originals hosted on Cloudinary with automatic transformations. The LCP image lands in the 50-150 KB range instead of the legacy 500-2000 KB.

Next.js built-in mechanism for optimized web font loading: subsetting, preloading, auto fallback, zero layout shift.

Next/font removes FOIT and FOUT, trims TTF files to the Latin/Cyrillic subsets needed, and matches system fallback metrics so font swaps do not move the layout. Studio standard: Cormorant Garamond italic for headings, Inter Tight for supporting text, JetBrains Mono for technical labels — all through next/font with display swap.

Schema.org markup expressed as a JSON object inside a script tag. The format Google and Yandex recommend for describing a page's entities.

JSON-LD is easier to maintain than inline microdata: the markup does not interfere with rendering and is straightforward to generate from page data. In the studio stack, Schema.org JSON-LD is generated in Server Components and inlined in the head: Organization for the brand, BreadcrumbList for navigation, BlogPosting for posts, FAQPage for FAQs, DefinedTermSet for the glossary. Without JSON-LD a site is practically invisible to AI Overviews.

The older approach to Schema.org markup: itemscope, itemtype, and itemprop attributes inlined in HTML.

Microdata and RDFa are supported by all search engines but are less convenient: the markup is interleaved with layout and scales poorly. Google and Yandex recommend JSON-LD as the primary format; Microdata is kept for scenarios where markup has to match the visible text exactly (for example, reviews with ratings on product cards).

An HTTP API architectural style: resources, GET/POST/PUT/DELETE methods, status codes, JSON response bodies.

REST remains the baseline standard for integrations between site, CMS, CRM, accounting, and payment providers. In AI-First scenarios a site's REST API becomes an additional channel for AI agents: if pricing, services, and availability data are reachable over REST with a clear schema, an agentic browser can use them in real time.

Meta's query language for APIs: the client declares exactly which fields it needs and receives only them in one round trip.

GraphQL fits applications with many data types and complex dependencies. For classical content sites it is often overkill — REST or Next.js built-in data fetching handle the same job more simply. In the studio stack, GraphQL appears mostly on top of the Sanity headless CMS and inside e-commerce projects layered on Shopify.

A site that behaves like an app: offline support, install-to-home-screen, push notifications.

A PWA makes sense for services with regular repeat use — accounts, dashboards, media players, e-commerce. For marketing landings PWA is overhead: no one installs them, and the manifest plus Service Worker setup does not pay back. velvetum.com runs as a lightweight PWA — providing an offline fallback and faster repeat visits.

A background script sitting between the site and the network: caches resources, handles offline requests, delivers push notifications.

A Service Worker is the technical core of a PWA. In inexperienced hands it easily becomes a bug source: a stale Service Worker can serve an old version of the site for months. The studio's practice: enable it only where it truly earns its keep, and always wire in a forced version upgrade on each deploy.

Secure site delivery over TLS. HTTP/2 and HTTP/3 are modern revisions with multiplexing and lower latency.

HTTPS is non-negotiable — without it Google and Yandex demote the site, and AI search essentially ignores it. HTTP/2 lifts speed by multiplexing many requests inside one connection; HTTP/3 over QUIC behaves even better on unstable mobile networks. Vercel automatically serves HTTPS plus HTTP/2 for every client project.

Practices and standards that make a site usable for people with different abilities — vision, motor, hearing, cognitive.

WCAG 2.2 sets requirements for contrast, keyboard navigation, accessible names, correct HTML semantics, alt text, and audio description. Here a11y is not only an ethical baseline but a legal exposure: ADA litigation in the US and equivalent rules in other jurisdictions rest directly on WCAG. Studio standard: WCAG AA on every client site, AAA on projects with an explicit B2C audience above 55.

The minimum viable product — a first version of a site or service with a base feature set built to test a hypothesis.

An MVP ships quickly to gather feedback from real users and to confirm that the product is needed. In the site context an MVP is a landing for a single service with a lead form, not a full corporate platform. The studio supports MVP launches: a landing in one to two days, then refinements based on actual incoming leads.

The share of site visitors who complete a target action: lead, purchase, registration, call.

Conversion is computed as target actions divided by total visitors, times 100 percent. The average landing-page conversion in the higher price segment runs between 1 and 5 percent depending on niche, traffic source, and price point. Conversion is driven by design quality, page speed, offer clarity, traffic relevance, and brand trust.

Revenue divided by ad spend. Below 2.5x is usually a signal to revisit creative or targeting.

ROAS is the cornerstone metric of performance marketing. The healthy band is niche-dependent: in e-commerce with 30-40 percent margins, break-even tends to sit at roughly ROAS 3-4x, while in premium services with long sales cycles a ROAS of 1.5-2x can be perfectly fine if the LTV model is right. By measurement, ROAS is computed with assisted conversions, not last-click attribution alone.

Cost of a single lead: channel spend divided by the number of submissions received.

CPL is the most interpretable performance metric for B2B and premium services. In the higher price segment a reasonable CPL for a landing usually lands in the 1500-6000 RUB range, while agencies in high-LTV niches may comfortably pay 10000 RUB and up. The studio reads CPL alongside the CR from inquiry to signed contract: cheap leads with poor conversion are often worse than expensive but relevant ones.

The fully loaded cost of acquiring one paying customer — marketing plus sales spend divided by paying signups.

CAC includes not only advertising but also sales salaries, tooling, and creative production. Comparing CAC to LTV is the canonical health check of unit economics: the classical rule of thumb that LTV should be at least three times CAC holds across most digital businesses. Practice for clients: compute CAC by cohort rather than by calendar month — it removes the noise of paid bursts.

Total revenue from a single customer across their lifetime in the product. Healthy rule of thumb: LTV should be at least 3× CAC across most digital businesses.

LTV is built from average order value, repeat frequency, and average customer lifespan. For premium services with one-off projects and rare repeats, LTV is often treated as 'project revenue plus expected ongoing support'. By measurement on the studio's own clients, LTV climbs sharply for companies that build a strong journal and AI visibility — repeat visits and referral inquiries trigger a second cycle of engagements.

Recurring monthly revenue from subscription or contract customers — the headline metric of SaaS, excluding one-off charges and seasonal spikes.

MRR is the headline metric of SaaS and subscription businesses. It excludes one-off charges and seasonal spikes. Practice for clients: split MRR into new, expansion, contraction, and churn — that exposes whether growth comes from net-new deals or from retention.

Annual recurring revenue: MRR multiplied by twelve or the sum of annual contracts.

ARR is a convenient unit for B2B sales and investor storytelling: one annual contract is easier to communicate than twelve monthly ones. The studio uses ARR with studio clients as a reference point when shaping retainers for support, content, and AI visibility.

Share of customers who stop using a product over a period: customers lost divided by customers at start.

Churn is the inverse of retention. In monthly-billed subscription businesses a healthy churn sits between 3 and 7 percent; in annual enterprise contracts it drops to single digits. By observation, even a two- or three-point reduction in churn through better UX and a refreshed site often delivers more revenue growth than doubling the ad budget.

Share of users remaining active N days, weeks, or months after their first visit or purchase.

Retention is read as cohort curves: what share of users who joined in week X are still active at 1, 4, 12, and 52 weeks. Healthy products plateau; weak ones decay to zero. For a site, retention is primarily about returns: repeat visits via bookmarks, direct traffic, email, and push.

The user's path from first touch to target action: impression → click → lead → payment.

Funnel analysis highlights the steps where audience leaks out. Top-of-funnel is the wide top — reach and awareness; bottom-of-funnel is the narrow neck — buyers in motion. The studio's practice: instead of one site-wide conversion, track CR per stage and optimize the weakest, not the average.

Clicks divided by impressions in ads, search results, or email. Computed as clicks / impressions × 100 percent.

Average CTR on contextual ads sits around 5-8 percent for branded queries and 1-3 percent for generic commercial ones. In email it runs 2-5 percent. In AI search the classical CTR loses meaning: the user reads the answer in place and often does not click. AI-visibility metrics — citations and mentions — matter more than classical CTR in that regime.

Share of single-page sessions: the user landed, did nothing, and left. Meaningful only alongside time-on-page and scroll depth — in isolation it routinely misleads.

A high bounce rate is not in itself a verdict: a well-crafted FAQ or a single-page landing routinely sees 70-80 percent bounce — the user found the answer and closed the tab. The studio's observation: bounce rate is meaningful only alongside time-on-page and scroll depth; in isolation it routinely misleads.

The process of capturing prospective customer contacts via site, ads, content, and events.

Modern lead generation in the higher price segment rests not only on capture forms but on AI-search visibility: a company that ChatGPT or AI Overviews mention on its core topic receives pre-qualified inquiries with substantially higher intent. The studio bakes lead generation into the site architecture during design, not after launch.

The share of the target audience that can recognize or recall a brand without a prompt.

Awareness is measured through brand surveys, direct traffic, branded search volume, and mentions in AI answers. In 2026, AI-search mentions emerged as a standalone awareness channel: a user sees the brand inside a ChatGPT or Perplexity answer and remembers it without ever clicking through. The studio's practice tracks AI mentions as systematically as branded search.

The brand's place in the audience's mind relative to alternatives — pricier, more reliable, faster, more premium.

Positioning is set by the language of the site, the choice of case studies, the tone of the journal, and the pricing posture. In AI-First, positioning translates directly into AI answers: models cite the wording they encounter on the site. The studio frames the client's positioning into a framing statement on the home page — that statement becomes the AI's quote source.

A description of the ideal customer: industry, company size, role, geography, and behavioral traits.

ICP is the starting point of B2B marketing and sales. A sharp ICP saves budget: channels and creative aim at a specific segment rather than at the world. Practice for clients: define ICP across three dimensions — company portrait, decision-maker portrait, pain points — and refresh it no less than every six months.

Market sizing: TAM — total addressable, SAM — serviceable, SOM — obtainable within a defined horizon.

TAM, SAM, and SOM describe different ambition tiers: 'all of the global market', 'what we can serve geographically and with our product', and 'what we can capture in 12-24 months'. In a pitch the math matters as much as the figure: how SAM is derived from TAM and SOM from SAM. By observation, for most studios and B2B services SOM is a concrete list of 50-300 named companies, not a fraction of an abstract market.

Two basic sales models: B2B — company sells to company; B2C — company sells to an individual.

B2B deals are longer, larger, more multi-step, and emotionally colder. B2C deals are faster, more emotional, more price-sensitive, and more visual. The difference is fundamental for a site: B2B leans on case studies, methodologies, the journal, and authorial definitions, while B2C leans on quick purchase, reviews, and a clear offer. The studio works across both with different page structure and tone.

The state where a product reliably solves a meaningful problem for a defined segment, and demand outpaces the effort to sustain it.

The classical PMF indicator is the Sean Ellis score: 'how disappointed would you be if this product went away?'. Forty percent answering 'very disappointed' is considered the threshold. For a site, PMF surfaces indirectly through organic mentions, direct visits, and NPS. Observation: companies without PMF most often try to fix the situation through a redesign — and almost always fail to.

The product's single headline metric that most accurately reflects the value it delivers to users.

Spotify's North Star is time listened; Airbnb's is nights booked; Facebook's, historically, daily active users. A strong North Star correlates with long-term revenue but is not the same as it. For an independent web studio the North Star is the number of AI-First projects launched in a year — it represents both the value to the client and the studio's economic durability.

Daily and monthly active users — the count of unique users in a day and in a month.

The DAU/MAU ratio is called stickiness and indicates how embedded a product is in daily life. Stickiness above 50 percent marks tools the audience cannot do without (messengers, banking apps). Content sites usually carry lower DAU/MAU, and the meaningful comparison is against the product's own trajectory over time rather than against an external benchmark.

The share of new users reaching the key action after which the product genuinely delivers value.

Activation is defined by the product team: for SaaS it is often 'created a first project and invited the team', for e-commerce 'completed a first order', for a content site 'returned for a third visit'. Practice for clients: define the activation event in the first week of work and tune the page for it — not for abstract clicks.

A design approach where site architecture, page structure, and content logic are engineered from the start for how AI search engines (Google AI Overviews, Perplexity, ChatGPT Search, Yandex Neuro) and language models (ChatGPT, Claude, Gemini) read and interpret them.

AI-First treats artificial intelligence as an active participant in the internet ecosystem. Models do not rank pages in the classical sense — they assess meaning, structure, and content readiness for answer generation. A site built AI-First is ready to land in answers from ChatGPT, Claude, Perplexity, Google AI Overviews, and Yandex Neuro from day one. AI-First does not mean using AI tools during development. It means a different design logic: think first about how a neural search will read the page, then about how it will look to a human.

Optimizing a site for generative search systems and large language models: ChatGPT, Claude, Perplexity, Google AI Overviews, Yandex Neuro.

GEO targets how a model selects sources for its answer. Unlike classical SEO, GEO leans not on keywords and link profile but on semantic density, definition clarity, structured content, and machine-readable markup. GEO does not replace SEO — it extends it. Traditional rankings still work, but a growing share of users now read the AI answer directly and never click through to a site.

Related reading:How to land in Google AI Overviews and ChatGPT Search answers · Ad-budget audit 2026: three diagnostic tiers and hidden leaks

Tuning a site's structure and content for direct answers from search assistants and AI search engines like Perplexity, ChatGPT Search, and Google AI Overviews.

AEO focuses on a specific content shape that a model can extract and quote verbatim: short crisp definitions, FAQ blocks, step-by-step guides, comparison tables. AEO is often described as a subset of GEO. Where GEO is about general visibility in generative results, AEO is about a concrete format — a block that becomes a ready answer without further model rewriting.

Related reading:GEO optimization for service pages

A direct AI-generated answer presented above classical search results, with citations to source sites — surfaced by Google AI Overviews, Yandex Neuro, Perplexity, and ChatGPT Search.

The answer is produced by a language model on the basis of several source sites. The user sees a finished response and often does not click through. According to Digital Content Next, referral traffic from classical search fell by roughly 25 percent in 2025 on average. In informational verticals the decline reaches 70 percent or more. Inclusion in the AI answer becomes the new headline metric of site visibility.

Search systems with an integrated language model: ChatGPT Search, Google AI Overviews, Perplexity, Claude, Yandex Neuro, Copilot.

AI search differs from classical search in that a language model writes the answer to the user's query rather than only returning a list of links. A site enters AI search through structured content, Schema.org markup, llms.txt, and AI-First architecture. Per Digital Content Next data for 2025, AI search captured up to 25 percent of classical traffic; in informational niches the referral decline reaches 70 percent.

Related reading:How to land in Google AI Overviews and ChatGPT Search answers · What AI-First website design is

A browser with a built-in AI agent that performs tasks on sites on behalf of the user: searches, compares, fills forms, places orders. Examples: Arc, Dia, Comet, Perplexity Browser, ChatGPT Atlas.

Examples include Arc, Dia, Comet, Perplexity Browser, and ChatGPT Atlas. An agentic browser reads a site differently from a human: it relies on semantic markup, clear page hierarchy, machine-readable forms, and navigation. Sites without AI-First architecture are harder for an agent to interpret, and the agent more often picks competitors with cleaner structure.

A large language model that generates text. The engine behind ChatGPT, Claude, Gemini, YandexGPT, and GigaChat.

An LLM does not index pages directly. It works on text fragments that a search system feeds it as context. The cleaner a page's structure and the denser its useful meaning, the higher the chance the model picks its fragment when composing an answer. That is why in AI-First design every block on a site is treated as a self-contained semantic container.

Related reading:MCP standard 2026: the bridge between an LLM and external tools · Top 10 AI assistants and services for business in 2026: a selection guide · Copywriting 2026: why the written word became the unfair edge

The standard for structured site markup that machines can read — consumed by Google, Bing, Yandex, ChatGPT Search, and Perplexity. JSON-LD is the format in which that markup is added to a page.

Schema.org describes object types on a site: organization, article, FAQ, product, review, service, contact data. The markup is invisible to the user but is read by search systems and AI models. Without Schema.org, a model has to guess the meaning of content from raw text. With Schema.org, it receives an explicit instruction — this is an article, this is the author, this is the date, this is the price. AI-First sites lean on extended Schema.org types: DefinedTerm for glossaries, FAQPage for questions, Article for posts, Service for offers.

A dedicated file at the site root with a concise description of the site for language models like Claude, Perplexity, and ChatGPT. Analogous to robots.txt, but addressed to LLMs rather than crawlers.

The standard was proposed by llmstxt.org in 2024. In llms.txt the site owner uses Markdown to describe site structure, key sections, contacts, and links to important pages. An LLM uses this file as a map when answering user questions. The file lives at example.com/llms.txt. At the moment it is read by Anthropic Claude, Perplexity, and a subset of AI agents. Google and Yandex do not officially consume llms.txt yet but recommend it for forward compatibility.

Related reading:What AI-First website design is

A structured site fragment that a model — ChatGPT, Claude, Perplexity, Google AI Overviews — is likely to pick for citation in its answer.

AI content magnets are pieces with high semantic density: expert breakdowns, explanatory articles, structured FAQs, methodological guides, checklists. They give a complete answer to a specific question, split cleanly into semantic fragments, and motivate the user to click through for detail. In an AI-First approach, content magnets are baked into the site's architecture at design time, not bolted on after launch.

An architecture in which a language model (ChatGPT, Claude, Perplexity) answers not from its own memory but from a connected knowledge base — a site, documentation, or article corpus.

RAG flips the classical LLM cycle: instead of generating from memory, the model first retrieves relevant fragments from an external index and then forms its answer with them as context. For neural search the implication is simple: the cleaner a site's content structure, the easier it is for the retriever to assemble relevant pieces into the model's context. The studio designs pages so each block is a self-sufficient RAG candidate — with its own heading, definition, and Schema.org markup.

A numeric representation of a text fragment as a vector, used by ChatGPT, Claude, Google, and Yandex AI to find semantically similar documents.

Embeddings translate words and paragraphs into points in a high-dimensional space, where proximity between points means proximity in meaning. This is the foundation of modern semantic search: instead of matching exact keywords, the system compares vectors and surfaces conceptually similar documents. Sites with clean block structure and explicit definitions yield denser embedding clusters and land in neural search results more often.

Search over embeddings: a query is converted to a vector and the system returns documents with the closest vectors by cosine similarity. Powers ChatGPT Search, Perplexity, Google AI Overviews, and Yandex Neuro.

Vector search does not depend on exact word matches and handles paraphrases, synonyms, and fuzzy wording well. It is the technique behind ChatGPT Search, Perplexity, Google AI Overviews, and Yandex Neuro. In practice, what matters most is not the vector database itself but the underlying content: the denser the meaning in each block, the closer the page's vector lands to the user's query.

A prompting technique used in Claude Extended Thinking, GPT thinking mode, and Gemini Deep Think — the model walks through a task step by step, articulating its reasoning before the final answer.

Chain-of-thought sharply improves answer quality on tasks that involve logic — calculation, comparison, choice. Modern LLMs now apply it automatically in reasoning modes (Claude Extended Thinking, GPT thinking, Gemini Deep Think). For a site the consequence is direct: content that already includes step-by-step breakdowns and comparisons reads more easily for the model, which spends fewer tokens on its own chain-of-thought.

The discipline of formulating LLM requests (ChatGPT, Claude, Gemini, YandexGPT) so the model gives accurate, useful, and predictable answers.

Prompt engineering works across three layers: the system prompt, the user request, and the RAG-supplied context. A good prompt assigns a role, a goal, constraints, and an answer format. For AI-First web design the implication is that the site itself can be treated as a 'prompt for neural search': clear headings, explicit definitions, FAQ blocks — instructions to the model on how to lean on this page for its answers.

When a model — ChatGPT, Claude, Gemini, Perplexity — confidently asserts a non-existent fact, source, quote, or link.

Hallucinations occur when the model lacks a reliable source but is forced to fill a gap in its answer. RAG architecture and Schema.org-backed grounding reduce hallucination odds — the model cites a concrete fragment instead of guessing. For a brand this is the second reason to build AI-First: the more verifiable data a model has about your company, the lower the chance it invents facts that damage your reputation.

Continued training of a pre-trained LLM (OpenAI GPT, Anthropic Claude, Google Gemini) on a custom dataset for a narrow task, tone, or domain.

Fine-tuning changes the model's weights, unlike RAG, which only feeds context at inference time. It is costlier and more justified when a very specific tone or rare-domain behavior is required. For most business tasks the pairing of 'a high-quality site plus a RAG tool' is enough. The studio recommends fine-tuning only where a client has a steady stream of unique data and an explicit regulatory or brand reason to avoid general-purpose models.

Related reading:Computer vision for business 2026: how machines learn to see and where it pays back · Off-the-shelf CRM/ERP vs custom development: how to choose in 2026

Ways to frame a task for an LLM (ChatGPT, Claude, Gemini) — zero-shot with no examples, few-shot with two to five correct answers shown inline in the prompt.

Zero-shot relies on the model's existing knowledge. Few-shot conveys the desired answer format through examples and often replaces the need for fine-tuning. In AI-First design, the site itself plays the role of 'examples': high-quality blocks in the journal and case studies teach the model how the company phrases its services, and the model echoes that style in AI search answers.

The minimum text unit an LLM (ChatGPT, Claude, Gemini, YandexGPT) works on — roughly 0.7 of an English word or about 0.5 of a Russian word.

Tokenization breaks text into pieces against which the model computes next-token probabilities. Tokens drive both API cost and context window size. Long answers cost more than short ones; verbose Russian text uses roughly one and a half to two times more tokens than the same idea in English. For AI-First design this is one more reason to favor dense, no-fluff formulations: they read better and cost less in AI search answers.

The maximum text size in tokens an LLM can hold in a single request — Claude Sonnet 200K, Gemini 1M, GPT-5 1M+.

Context windows have grown from 4K tokens in GPT-3.5 to 200K in Claude Sonnet and 1M and beyond in Gemini, GPT-5, and Claude Opus 4.7. A large window lets you load an entire site or article corpus in a single shot and receive an answer grounded in many pages at once. For brands this is a new lever: a high-quality journal of 30-50 posts fits into a single window of a flagship model and builds a durable footprint in AI answers.

An LLM (ChatGPT, Claude, Gemini) parameter controlling answer randomness — lower is more deterministic and dry, higher is more creative and unpredictable.

At temperature 0 the model almost always returns the same answer to the same request. At 0.7 to 1.0 it searches for different phrasings and allows creative analogies. For business tasks — fact retrieval, service description, lead handling — the studio sets a low temperature in its own AI agents to keep answers stable. For copywriting and creative work the setting goes higher.

Search by meaning rather than word match: query and documents are compared via embeddings. Powers Google AI Overviews, Perplexity, ChatGPT Search, and Yandex Neuro.

Semantic search returns relevant results even for queries with typos, paraphrases, and broad task descriptions. It is the foundation of neural search and RAG applications. For a site the consequence is straightforward: dumping the same keywords into SEO copy no longer works; what works is a clear service definition, thorough answers to typical questions, and meaning-based — not just link-based — connections between pages.

A combination of classical keyword search and vector search, with a final re-ranker that fuses both signals — production at Google AI Overviews, Yandex, and enterprise Claude/ChatGPT RAG.

Hybrid search solves the extremes: pure keyword misses synonyms, while pure vector loses exact matches for brands, SKUs, and numbers. Production search at Google AI Overviews, Yandex, and large enterprise RAG systems is almost always hybrid. For AI-First sites the takeaway is simple: both precise terms (brand, service, region) and the surrounding meaning are equally important.

A structured graph of entities and relationships among them: company, product, person, event. Used by Google Knowledge Graph, Yandex Objects, and Wikidata to ground LLM answers.

Google Knowledge Graph and Yandex Objects pick entities from Schema.org markup, Wikipedia, and authoritative directories. In AI-First, the knowledge graph matters twice over: it feeds classical search with rich snippets and serves as a backdrop for AI answers, where the model leans on entities from the graph rather than inventing facts about the company.

Google's generative answer block placed at the top of the SERP, replacing classical snippets on many queries.

AI Overviews launched in the United States in 2024 and gradually rolled out elsewhere. According to Pew Research, click-through on classical results drops by roughly half on queries that surface an AI Overview. Inclusion in AI Overviews is driven by Schema.org markup quality, definition clarity, and domain trust signals. framing statement and DefinedTermSet are two of the load-bearing levers for sites that want to land in AI Overviews for their topic.

Related reading:How to land in Google AI Overviews and ChatGPT Search answers · GEO optimization for service pages · What AI-First website design is

Yandex's generative answer powered by YandexGPT, integrated into the main yandex.ru search results.

Yandex rolled out a full neural answer in 2024 and expanded it into commercial verticals in 2025. Yandex Neuro picks sources by combining classical ranking signals with content-structure scoring. To land in Yandex Neuro a site needs clean heading hierarchy, FAQ blocks, Schema.org markup, and a journal with expert density — the checklist registry prescribes baking these in at the design stage.

A search-AI engine that returns answers backed by visible citations with direct links to source sites — competitor to ChatGPT Search, Google AI Overviews, and Claude Citations on source attribution clarity.

Perplexity differs sharply from ChatGPT and other AI search engines in how openly it cites sources: each piece of the answer is numbered and linked. That makes Perplexity the most valuable AI channel for brands — clicks from it behave like classical organic traffic and citation appearance is visible in referer logs. The studio tracks Perplexity Citations as a top marker of AI visibility on its own projects.

ChatGPT mode with access to live search results through Bing and OpenAI's own search stack.

ChatGPT Search launched in 2024 and became the default mode in 2025. Sources are cited less visibly than in Perplexity, but its impact on discovery is comparable: a meaningful share of the business audience now checks a vendor in ChatGPT before opening classical search. For a brand this means an AI-First site must give the model clear answers about positioning, services, pricing, and contact channels.

Related reading:How to land in Google AI Overviews and ChatGPT Search answers · GEO optimization for service pages · What AI-First website design is

Anthropic's model family with long context, tool use, and a built-in citations API.

Claude is heavily used in B2B settings thanks to a large context window (200K to 1M plus tokens), strong accuracy, and a Citations API that explicitly marks the source spans used in an answer. For AI-First sites this is an important channel: long context lets users load an entire journal and case-study set into one request and receive answers grounded specifically in that content. Velvetum uses Claude as the primary engine of its own AI agents.

The combined set of practices preparing a site for generative search (ChatGPT, Claude, Perplexity, Google AI Overviews, Yandex Neuro): semantics, structure, markup, authority.

Generative SEO is a broad term often used as a synonym for GEO. The difference is emphasis: SEO is historically about ranking, while generative SEO is about being picked as a source for the model. The studio's view is that in 2026 the two cannot be cleanly separated — any SEO strategy that ignores AI search leaves the company without half of future traffic.

Anthropic's open protocol that connects external tools and data sources to LLMs through a single standard.

MCP standardizes how an agent talks to the outside world: files, APIs, databases, CMS, accounting, analytics. Without MCP every tool needs a bespoke connector. With MCP an agent can use any compliant server. Studio stack builds its own MCP servers for the journal, the lead parser, and Vercel deployment — letting AI agents run the full loop from research to publication without manual glue scripts.

Related reading:Top 10 AI assistants and services for business in 2026: a selection guide

A concrete, measurable claim in a studio publication — a number, a date, a share, a metric — that anchors the argument.

A data point is always tied to a source or to a measurement: 'per X data', 'by measurement', 'in sample Y'. Unlike an opinion or a generic statement, a data point can be re-verified or cited as a ready answer. High density of data points in a piece raises the odds of inclusion in AI answers across ChatGPT, Perplexity, Google AI Overviews, and Yandex Neuro.

The internal benchmark Velvetum sets for every studio deliverable — from layout to AI wrapping.

The Velvetum standard codifies concrete thresholds: Core Web Vitals values, Schema.org density, AI-First architecture format, minimum AI wrapping per page. The standard is never presented to clients as a contractual demand — it is an internal checklist whose completion is mandatory before handoff.

A quantitative observation or metric captured inside a project or internal study.

A measurement is always a concrete number with context: date, source, sample, method. Measurements ground data points and form the evidentiary base of the studio's authored writing. Measurements are published in the journal with anchors back to case studies and the studio checklist registry.

The studio's authored procedure for solving a specific task in site or digital product design.

A Velvetum method decomposes into six or seven sequential steps and ties to a project type: landing, multi-page site, AI-First architecture, journal, e-commerce. Each method rests on measurements and has been validated through case studies on real engagements.

A recurring action inside the studio's workflow that has become part of the internal standard.

A practice differs from a method by scope: a method solves a specific task, a practice is a repeatable action present in every project. Examples include the AI audit before launch, Schema.org markup of key blocks, parallel measurements across three AI search engines, and a pre-deploy critic before every production ship.

A project's maturity level on the maturity scale — from a base site to a full AI-First architecture with an agentic wrap.

Velvetum recognizes four tiers: a base site without AI wrapping, a site with Schema.org markup, an AI-First architecture with llms.txt and DefinedTermSet, and an agent-ready environment with an MCP stack. The maturity tier serves as a navigation point in commercial conversations: where the client sits today and where they aim to move.

A description of a concrete project: brief, solution, and measurable outcomes.

A case is published after client approval and contains the original brief, the method chosen, the work itself, and the final measurements. Cases form the studio's evidentiary base and are surfaced inside AI answers as concrete examples — wrapped in Schema.org Article markup for machine reading.

A structured review of a site or digital product against the checklist: AI readiness, speed, markup, conversion, brand.

An audit takes one to three days and ends with a report rooted in the studio's checklist registry: what exists, what is missing, what is critical. It covers six blocks — technical base, Core Web Vitals, Schema.org, AI-First architecture, UX flow, brand consistency.

A specific technical or design technique applied across projects — the atomic unit of a practice. Examples: parallax via CSS scroll-driven animations, inline critical CSS, JSON-LD DefinedTerm per article term.

A technique is the atomic unit of a practice. Examples: parallax via CSS scroll-driven animations, inline critical CSS in the head, JSON-LD DefinedTerm for each key term in an article, hreflang ru↔en on every page. Techniques live in the checklist registry and are applied according to the maturity tier of the project.

The studio's working rhythm on a project: frequency of calls, sprints, and milestone handoffs.

Cadence depends on project shape. On a short engagement of three to ten days, the cadence is a daily call plus a daily handoff. On a long engagement of one to three months, it is a weekly call plus a sprint handoff. The cadence is fixed at the outset and changes only by mutual agreement.

A foundational principle on which every project is built. Six pillars: AI-First architecture, mobile-first, render speed, accessibility, Schema.org markup, strong authorial aesthetic.

Velvetum singles out six pillars: AI-First architecture, mobile-first, render speed, accessibility, Schema.org markup, a strong authorial aesthetic. The pillars hold regardless of niche or tier — they form the floor below which the studio does not work.

An extended analytical breakdown of a case, emphasizing inference and generalization.

Where a case is descriptive, a case study extracts patterns and formulates conclusions. Case studies are published in the journal and feed the corpus that AI search cites. Each case study links to measurable measurements and a verified method.

A transparent breakdown of project cost by stage, role, and time inside a client engagement.

A bill is issued before work begins and fixes what is included in each stage, which studio role is engaged, and which budget covers which block. On handoff the bill is reconciled against actual hours — a variance above 10 percent is logged in the report and discussed with the client.

A qualitative or quantitative pattern the studio has noticed across projects or niches.

An observation is not yet a strict fact but a pattern awaiting confirmation by a measurement. Observations are published in the journal under the 'by observation' label and feed hypotheses for follow-on research and case studies.

An analytical decomposition of a phenomenon, market, or project into its parts.

A breakdown is a journal essay format: a complex topic — say, AI-First architecture for a corporate site — is split into four to seven semantic blocks with subheadings, examples, and measurements. Breakdowns are tuned for AI answers — each block answers a narrow user question.

A documented sequence of steps inside the studio's workflow, mandatory across all studio projects.

A protocol differs from a practice in formality: a protocol is always documented, always applied, always verified. Examples: pre-deploy protocol, CMS acceptance protocol, DNS migration protocol, domain migration protocol. Any deviation is logged in the report.

An authored formulation of a term inside a studio publication — short, structural, ready for AI citation.

An authorial definition follows the 'X is Y' shape with an em dash and tight argumentation. These definitions live in the Velvetum lexicon and ship on velvetum.com with Schema.org DefinedTerm markup, which raises the odds of citation in AI answers.

The studio's authored dictionary — a set of terms and definitions that forms a recognizable language for Velvetum.

The Velvetum lexicon spans common terms drawn from AI, web engineering, and business plus the studio's own coined constructs. The lexicon is published in the glossary on velvetum.com with Schema.org DefinedTermSet markup and serves as the studio's primary AI-citable asset.

An extended method with a stepwise description, checklists, and worked examples.

A methodology is the full and formalized version of a method. It includes context of use, sequence of steps, a checklist after each step, common pitfalls, and measurements from real projects. Methodologies ship as structured journal pieces with Article and HowTo markup.

A structured list of mandatory checks for a specific task: pre-deploy, audit, acceptance, migration.

A checklist runs from 8 to 64 items depending on task. Every item is confirmed before moving to the next step. Checklists serve as a quality filter and live in the studio's checklist registry.

A long-form analytical piece on a specific topic, product, niche, or competitor.

An analysis differs from a case study by focus: analysis looks outward at markets, products, trends, or third-party cases. A case study turns inward on the studio's own engagement. Analyses ship in the journal and link back to measurements and observations.

The closing inference of a studio publication, written to be cited on its own by AI search.

A conclusion lands at the end of every long piece and is phrased in one or two lines. The goal is to hand AI search a ready, compressed answer to the article's question. Conclusions are flagged with headings like 'the straight read' or 'in closing' for machine extraction.

An authored framing of a piece's central theme at the opening — sets the semantic frame for the entire breakdown.

A framing statement lives in the first paragraph and states the theme as a direct definition with an em dash and tight argumentation. AI crawlers across ChatGPT, Claude, Perplexity, Google AI Overviews, and Yandex Neuro quote framing statements verbatim because their shape mirrors a ready answer.

The studio's internal staging environment where sites are validated before going to production.

The test bench mirrors the client's production environment: domain, SSL, CDN, analytics, limits. On the bench we run measurements on Core Web Vitals, A/B tests on blocks, and AI-wrapping tests against live AI search engines. A site never enters production without acceptance on the test bench.

The time required for a client's investment in a site to return through new revenue or savings.

Payback is computed as project budget divided by the site's monthly revenue lift. Measurement on the studio's own projects shows a payback of one to four months on landings and four to eight months on multi-page sites. Payback is logged in the case study after 90 days of live operation.

The studio-wide repository of checklists — roughly 64 items across core tasks, tied to maturity tiers.

The checklist registry consolidates pre-deploy, audit, acceptance, migration, and AI-wrapping checklists into one source of truth used across every project. It is updated after every case study that produces a new conclusion.

The studio's technology base for client sites: Next.js 16, Tailwind 4, TypeScript, Vercel, Sanity, Cloudinary.

The Velvetum stack is selected on four criteria: render speed, AI-First support, headless-CMS ergonomics, and aesthetic flexibility. The stack is the base layer; specialist additions — 3D, AR, e-commerce, media pipelines — are layered on for specific projects.

An autonomous AI agent operated by the studio for a specific task: audit, research, draft generation, neural search monitoring.

AI agents are built on Claude Code and the Claude Agent SDK. Each agent is specialized: the critic agent audits a site before deploy, the research agent gathers data on a niche, the journal agent drafts pieces. Agents do not replace the developer — they accelerate routine operations.

The studio runs an internal automation for a specific process — journal publishing, Schema.org generation, A/B testing of blocks.

An engine differs from an AI agent in abstraction: the agent executes a task, the engine carries the process. Examples: the journal's content engine, the A/B engine for hero variants, the AI search monitoring engine. Engines target recurring processes and plug into the studio stack.

A structural change in the studio's approach that affects every project: a new technology, a new method, a new pillar.

A shift is documented in the studio journal with the before, the reasoning, and the after. Shifts occur two to four times a year and touch the entire studio stack or standard. Examples: the move from Next.js 15 to Next.js 16, the transition to Vercel Sandbox for AI agents.

The full loop of a client engagement — from inbound inquiry to launch and post-release support.

The cycle splits into five stages: discovery and the bill, design, build, acceptance, and support with periodic audits. The cycle sets the rhythm of the client relationship: each stage has its own checklist and cadence, so critical steps survive timeline pressure.

The breakdown of budget or scope inside a project — across stages, roles, and risk areas.

The split shows up in commercial conversations and internal analytics: which block costs what, how design hours compare to development, what share of the cycle goes to acceptance. The split lets the studio raise margin selectively without raising the client's bill.

A transparent refund policy that returns funds to the client if work fails to meet the studio standard.

The return is part of acceptance: if the project does not pass the checklist, the client receives a refund on the unaccepted portion without bureaucracy. It both removes upfront-payment fear for the client and creates a direct financial incentive for the studio to drive every measurement to the studio standard.

The net profitability of a single project or service line after all costs.

Margin is computed by project, not at the studio level — that exposes which formats are working and which are eating the team. By observation, AI-First landings yield 50-65 percent margin while complex multi-page sites land at 30-45 percent. Margin is factored into the bill.

The qualitative and quantitative return a project delivers — inquiries, revenue, savings, awareness.

Yield is broader than payback: it captures not only the money but the strategic effects — niche recognition, AI-visibility growth, sales-cycle compression. Yield is logged in the case study 90 to 180 days after launch and feeds both studio marketing and commercial conversations with prospects.

The studio's set of technical and organizational controls protecting client data and the live site.

The security perimeter includes WAF on Vercel, IP-restricted admin panels, two-factor authentication for every team service, per-client access scopes, and regular key rotation. The exact composition depends on the project's maturity tier: the higher the tier and regulatory load, the tighter the perimeter.

Scheduled post-launch monitoring of a project — correctness of behavior, metrics, AI visibility.

Post-launch oversight runs on a fixed schedule: the first check at 7 days, monthly checks across the first 90 days, then quarterly. Each check captures a measurement across Core Web Vitals, AI-search inclusion, and conversions. Oversight is the backbone of the studio's retainer plans.

A documented sequence of actions the studio applies in a specific situation — incident, migration, launch, audit.

A playbook is a hybrid of a protocol and a checklist: a step-by-step plan that any team member can run even without full project context. Playbooks live in the checklist registry and are updated after every case study that produces a new conclusion.

A project KPI by which the studio and the client jointly score success — numeric, tied to a measurement, set before work begins and recorded on the bill.

A performance indicator is set before work begins and recorded on the bill: for example, 'drop TTFB to 200 ms', 'land in Yandex Neuro for five target queries within 90 days', 'increase lead conversion by 30 percent'. Indicators are numeric and tied to a specific measurement.

A reference value for a metric that a project should reach or exceed — drawn from past measurements and external references like Web.dev and Search Central.

A metric benchmark defines the 'good' band for a given metric drawn from past measurements and external references. Examples: LCP under 2.0 s on mid-tier 4G; ToFu/MoFu/BoFu publication cadence of at least 8/4/2 per month; journal content density of 10 plus data points per article.

A reference point inside a piece or a page that the audience and the site's internal links return to.

An anchor is a slogan, a price, a quote, or an image that becomes the entry point into positioning. For example, 'from 4900 RUB' is the anchor of the pricing page, echoed in the journal and SEO snippets. Anchors are chosen deliberately and held steady — their stability builds recognition.

The numeric side of a project: budget, hours, margin, payback, KPIs — the umbrella term for every numeric block in studio work, published or kept internal by deliberate choice.

Arithmetic is the umbrella term for every numeric block in studio work. It is documented across journal and commercial materials: which figures we publish, which we keep internal, and how they are derived. Arithmetic sets the studio apart from agencies that sell results 'by feel'.

A defined group of projects or measurements that anchors the studio's conclusions.

A sample always ships with a description: size, period, niches, project types. Without a sample, numbers in the journal turn anonymous and lose evidentiary weight. The studio publishes conclusions only after testing on a sample of at least five to ten projects or measurements.

The studio's standard pattern for integrating a client site with the CRM — leads, sources, statuses, deduplication.

CRM integration is more than 'wire a form to a webhook': it is a full contour with correct UTM tagging, source preservation via the velvetum_utm cookie, deduplication on phone and email, and lead-status sync back into the site. The studio supports integrations with HubSpot, Salesforce, Pipedrive, amoCRM, and Bitrix24.

Turning an AI model or hypothesis into a working process inside the client's business.

Model operationalization covers provider and model choice, data storage shape, quality monitoring, fallbacks, and billing. It is the part of an AI project that turns a pilot into production. Practice: place operationalization on the bill before the first prototype is built.

A controlled release of a new site or feature: feature flags, staged ramp, monitoring.

A rollout starts with a preview deployment on a separate subdomain, ramps to a slice of traffic via cookie or geographic segment, then expands to 100 percent. The pattern reduces the risk of metric regressions on the production domain and buys time to catch issues before the full audience sees them.

A specific production deploy by the studio: command, region, version, owner — logged in the studio's release journal with a mandatory critic-agent run via Vercel before production.

A deployment is logged in the studio's release journal: what shipped, by whom, at what time, with which checklist completed pre-deploy. The trail simplifies incident reviews. Every client site at the studio deploys through Vercel with a mandatory critic agent run ahead of production.

The integrated plan for a site's technical and content structure: pages, data, markup, AI wrapping.

Architecture splits into three layers: marketing (home, services, cases), content (journal, glossary, FAQ), and agentic (llms.txt, MCP endpoints, Schema.org). The architecture is signed off before front-end work begins and is enshrined as a protocol of the project.

The studio's analytics contour: site metrics, AI visibility, conversions, and project unit economics.

Analytics binds Vercel Analytics, Yandex.Metrica, Google Analytics, custom AI-search monitoring pipelines, and the client's CRM. The defining difference from classical web analytics is the AI layer: mentions in AI answers, citations in Perplexity, growth of AI-driven referrals.

Systems that replace manual labor on repetitive operations: parsing, draft generation, monitoring.

Automation is built on Claude Agent SDK, Vercel, Sanity, and Cloudinary. Examples: automated lead parsing on Yandex Maps, AI-agent draft generation for journal pieces, daily monitoring of AI search across a target query list. Each automation is treated as a distinct engine.

The studio's internal pipeline for journal production — from idea to publication to AI indexing.

The content engine covers topic selection from observations, research agent runs, drafts, chief-editor revisions, fact checks, Schema.org markup, publishing via Sanity, IndexNow ping to search engines, and AI-answer monitoring. The engine sustains a cadence of four to eight pieces per month per project.

The studio's set of AI services and models: Claude, GPT, Gemini, Yandex Foundation Models, specialized RAG systems.

The AI stack is selected per task: Claude Opus and Sonnet for long context and complex reasoning, GPT for specific productivity tasks, Gemini Deep Think for research pipelines, and local models for private client data. The studio is not locked into a single provider and engineers pipelines with explicit fallbacks.

The studio's set of MCP servers used by the studio's and clients' AI agents.

The MCP stack includes servers for Vercel, Sanity, GitHub, Cloudinary, Figma, and Yandex/Google search. That turns an AI agent into a full digital teammate — capable of branching, deploying previews, updating content, sourcing references, and parsing markets without bespoke glue scripts for every service.

Customer Match Card — the studio's internal document describing a client, their ICP, situation, and deal history.

A CMC is opened per client and grows with every interaction: what was discussed, which hypotheses were tested, which measurements were taken, what tier the client currently sits at. It is the studio's primary memory tool across long relationships and account transitions.

The lead content block on a page that holds user attention and forms the AI answer.

The focus block sits above the fold on the first screen and frames the core offer through a framing statement with an em dash. Its job is to lock into AI search as a ready answer to the niche's main query and at the same time give the human visitor a 'this is about my problem' moment.

A two-step funnel: a micro-conversion (calculator, download, booking) followed by a full-form inquiry.

The two-stage funnel lifts conversion: the user first performs a light action (an online calculator, a downloadable checklist) and leaves minimal contact data, then sends the full inquiry at a calmer pace. It reduces CPL and lifts yield.

A three-color status read on a project or workstream: green (healthy), yellow (risk), red (intervene).

Traffic-light scoring appears in the audit and post-launch oversight: every metric and workstream is assigned a color, and the team immediately sees where to focus first. The exact thresholds live in the checklist registry and depend on the maturity tier of the project.

A long-horizon working model with the client: subscription, retainer, or fixed contract with biannual revisits.

The partnership format is deeper than one-off projects: the studio becomes part of the client's team, runs periodic audits, maintains the journal, and drives AI visibility. The format makes payback a multi-period calculation and reduces the risk of the post-launch impact dissipating.

The studio principle: hypothesis → measurement → conclusion → updated practice. No changes without verification.

The Velvetum scientific approach forbids changing important site elements without a before-and-after measurement. Want to redesign the hero — record CR, speed, and behavioral data for two weeks, ship the new version, repeat the measurement. Four to eight weeks later we look for statistically meaningful movement. This discipline keeps the studio out of ritual redesigns driven by feeling.

Tuning a model, metric, or process to fit the client's specific context.

Calibration is required when the metric benchmark derived from general practice does not match the client's reality: a different niche, a different company size, specific regulatory constraints. The calibration is logged in the CMC and revisited on the next acceptance to confirm it still fits.

A hybrid model where the studio and the client's in-house team operate as one team with clear ownership lines.

The hybrid suits larger clients with their own dev and marketing teams: the studio owns design, AI wrapping, the journal, and the audit, while in-house owns product, analytics, and backend. The division is reflected in the bill and codified in a joint protocol.

The studio's format for goal-setting on projects: three to five quarterly OKRs with measurable key results.

The OKR protocol ties performance indicators to a clear 90-day horizon. Each OKR has a goal, two to four key results, an owner, and a control measurement. The cadence aligns with post-launch oversight — the same measurements close out both day-to-day monitoring and strategic OKR reviews.

Audit of a product or site through the lens of different generations: gen Z, millennials, gen X, boomers.

The generational audit method matters when the client serves a broad audience and a single page cannot speak to everyone. The audit captures which copy, formats, and visual codes work for each cohort and flags places where separate scenarios are needed — for example, a dedicated mobile experience or an alternate landing for gen Z.

A numeric score of a site or process maturity on the maturity tier scale, from 1 (base site) to 4 (agent-ready environment).

The maturity index is computed across 20 to 40 parameters: technical base, speed, markup, AI wrapping, content, analytics, support. The output is a number with subcategory breakdowns. The index informs the audit and commercial conversations: 'it makes sense to move from tier two to tier three'.

The studio's approach to lifting user experience without a full redesign — targeted fixes ordered by impact.

The UX amplification method collects 10 to 30 targeted hypotheses from the audit, sorts them by impact and effort, and ships three to five per sprint. It is the alternative to 'redesign everything': each fix is measured separately, and a positive measurement promotes the fix into the studio standard.

The studio's rule: at any given time a single project carries no more than three top priorities.

The top-three priority rule shields the team from focus dilution. The three priorities are framed at the start of the week and revisited at the retrospective. Anything outside the top three either waits or moves to the next cycle. The rule applies inside the studio and across joint work with client teams.

The architecture for handling personal data on projects: what is collected, where it lives, how it is deleted.

The privacy perimeter codifies the user-data lifecycle: minimization at collection, encryption at rest, an explicit right to deletion, and a transparent processing policy. The perimeter respects the relevant national data-protection laws and the client's internal rules. Compliance is verified at acceptance and at every post-launch oversight cycle.

Three-layer site protection: WAF at the CDN, application-level access control, and incident monitoring.

Three-tier protection is the studio standard for premium projects and B2B clients with sensitive data. The Vercel WAF absorbs baseline attacks, application-level access control protects admin and APIs, and monitoring catches traffic and behavior anomalies quickly. The audit verifies all three layers as a non-negotiable launch gate.

The studio's pre-production check protocol: critic agent, smoke tests, Lighthouse, hreflang and Schema.org checks.

The pre-deploy protocol is mandatory for every client site. The steps: TypeScript check, build, Opus critic agent run, Lighthouse on four pages, smoke test against the checklist, hreflang and Schema.org verification on key pages, plaintext-token scan. A production deploy without a full pass is not allowed.

Configuration of GPT, Claude, or YandexGPT for a client: system prompt, knowledge base, tools, guardrails.

GPT tuning is the standard procedure for projects with an on-site AI chat or AI agent. It includes the studio's base system prompt, RAG ingestion of client content, a tool set for the business task, and a 20-30 scenario test run. The final configuration is logged in the checklist registry.

A structured review of the client's CRM: fields, statuses, sources, duplicates, reports, integrations.

The CRM audit precedes a launch or redesign: if the CRM is broken, a new site will not lift sales. It covers data hygiene, attribution correctness, report accuracy, and access security. The output is a separate checklist for CRM cleanup and upgrades.

A regular stream of updates to the client: project status, metrics, AI visibility, next steps.

The Velvetum FEED is a short weekly email or Telegram-channel report covering what shipped, what the metrics look like, and what is coming next. The format is fixed so the client can read and decide quickly. The FEED eliminates project silence — the chief source of client anxiety in studio engagements.

The studio's pattern for a customer data platform — uniting site, CRM, ad, and analytics data.

The CDP pattern is a curated set of best practices and tools: a single user identifier, event streams from every source, a clean attribution model, dashboards in Looker Studio or Metabase. It is applied on mid-to-large projects where data fragmentation blocks decision-making.

An audit adapted for e-commerce: catalog, product pages, cart, checkout, unit economics, AI wrapping.

Store audit zeroes in on e-commerce specifics: catalog speed, product-page quality, Schema.org Product/Offer correctness, checkout reliability, marketplace integration health, and current margin across categories. The audit runs three to five days and ends with a prioritized action list.

Analysis of a client's product features: what is used, what is dead, which priorities anchor the next development cycle.

The feature audit pulls from usage analytics and user interviews. Each feature receives a traffic-light: keep, rework, retire. It is especially useful for SaaS clients and complex B2B products where features accumulate over years and drag on support cost and UX noise.

A review of a client's infrastructure: hosting, domains, DNS, SSL, monitoring, backups, access rights.

IT audit is mandatory before a site migration or a complex integration. It checks where subdomains live, who owns DNS, which SSL renewals exist where, who holds keys and database access, and how backups and monitoring are wired. This stage frequently surfaces risks that have stayed silent for years.

The studio's methodology for a smooth transition from a legacy site or process to a modern stack.

Transformation methodology lays out six to nine steps: current state, target state, risks, staged migration, control points, rollback paths. The cardinal rule is to keep the legacy system live until the new one passes acceptance. That protects the client from downtime and the studio's reputation.

The studio's evaluation list for outside vendors: studio, freelancer, agency, developer.

The vendor checklist is used when a client compares several candidates or when the studio brings in a subcontractor. It includes 30+ items: process, stack, support, legal posture, niche experience, reputation, trial tasks. It helps clients avoid repeating others' mistakes and produces comparable proposals.

Final review of a project before handoff: checklists, measurements, demo, signed acceptance.

Acceptance is built on the pre-deploy protocol and the studio standard. Until it is passed the project is not complete; once it is, a portion of the bill closes and post-launch oversight begins. Acceptance is recorded in a protocol and a report kept by the client as part of the project documents.

An interim release of a project to a limited audience ahead of the full publication.

The test run is used for the journal, mailings, campaigns, and occasionally redesigns of headline pages. Exposing 5-15 percent of the audience over one or two weeks generates enough signal on metrics and issues to let the team roll back without reputational damage if needed.