Terms that appear in my writing, from A to Z. Anyone who argues that brands should make their rules explicit should start with their own.
Accessibility Tree Accountable Decision Address Agent Authority Agent Payments Protocol / AP2 Agent Skills Agentic Brand Readiness Agentic Browsing Agentic Commerce AI Act AI Agent AI Slop Answer Engine / AEO ARIA Landmark Bot Defence Brand as Software Brand Infrastructure Brand Kit Brand Specification Browser Agent Canonical Link Citation Commitment Connector / Plugin Consensus Taste Consistency / Coherence Crawler Design System Design Token Digital Omnibus Domain Authority Execution Layer / Layer of Meaning Failure Signal GEO llms.txt Machine Customers Machine-Readable Machine Readable Brands Machine View Markdown Model Context Protocol / MCP Open Brand Repo Preprint Repository robots.txt Schema.org / JSON-LD Special Markup Stateful Structured Data Synthetic Respondents TDM Opt-out Tool Poisoning Universal Commerce Protocol / UCP User Agent WebMCP Wikidata
The structure the browser computes from every page: all elements with role and name, no layout.
Built for screen readers, and according to Google's developer documentation now the primary interface for browser agents. An element without an accessible name does not exist in this view, however prominently it is designed. Accessibility is no longer a niche; it is the ticket into the machine view.
The judgment about which option is right: justified, dated, versioned, with a named sender who stands behind it.
What remains of brand work when machines take over execution. An opinion is stated and left behind. An accountable decision carries consequences. Opinions are free, including from machines.
The person who stands behind a decision, and whom everybody knows to go to.
Taste has no address: it lives in people and leaves the building with them. Responsibility is bound to an address and therefore cannot be copied. For agents this means there is someone who stands behind their actions. The address does not have to do everything itself. It is the one everybody knows to go to: I noticed something, I report it there, and it gets written down. Delivering and checking can be automated, running it can be outsourced, the address cannot.
The deliberately granted, bounded, and accounted-for mandate a brand gives its agents.
Granted means: decided rather than drifted into. Bounded means: defined what the agent may promise and where a human takes over. Accounted for means: there is an address that stands behind it. The expression comes from the law of agency and is carried into the brand context here.
An open protocol from Google for payments an agent triggers on a person’s behalf.
Presented on 16 September 2025 with more than 60 payments and technology partners. At its core are signed mandates: proof that a person authorized a purchase in this form before the agent pays. Since April 2026 the FIDO Alliance has been building on it to develop standards that let users empower agents only for approved actions. For brands this is the technical half of Agent Authority; what the agent may pay for, the protocol does not decide.
An open format in which an AI agent receives expertise and workflows as a folder it loads when needed.
At its core is a text file named SKILL.md with a name, a description, and instructions, plus scripts, templates, and material as needed. The agent first reads only the name and description and loads the rest when a task fits. Anthropic introduced the format in October 2025 and released it as an open standard in December 2025; tools from Google, Microsoft, OpenAI, and many others now read it. For brands it is an obvious place for rules an agent is meant to apply. Of 21 open-source design systems examined, 18 ship official Agent Skills; among the most visited domains, only 0.47 percent offer them.
The measurable maturity level of a brand toward AI systems: being found, described correctly, and recommended.
Stage one of the ladder, and the only one you can test yourself in ten minutes: three questions to ChatGPT or Perplexity, four points of evaluation.
An audit category in Google's Lighthouse measuring how well agents can read and operate a website.
Since Lighthouse 13.3.0 of 7 May 2026 an experimental category in the default configuration, usable from Chrome 150. The audit tool itself is built into Chrome. Agent readability turns from claim into score. For context: a good score proves readability, not visibility, and Google itself states that AI search needs no special files.
Commerce in which an agent on at least one side compares, negotiates, or buys.
An actual purchase by an agent is currently the rarest case. In May 2026 Forrester described true autonomy as rare: very few consumers let an agent complete a task or a purchase without direct oversight, and fewer still do so regularly. More common is the step before it: agents that compare, filter, and recommend. None of the sources checked puts a revenue figure on it. The recommendation decides earlier than the transaction.
The European Union regulation on artificial intelligence.
For brands what counts is above all Article 50: chatbots must identify themselves, and AI-generated content must carry machine-readable marking. The duty does not always fall on whom it appears to. The marking is owed by the provider of the generating system, not by the brand using a third-party model. The deadlines have been moved twice since 2024.
Software that pursues a goal on its own over several steps: it plans, uses tools, and acts.
A chatbot answers, an agent gets things done: it searches, compares, fills in, and books, often without a person seeing every step. The browser agent is one form of it. How well that works depends on the task. On real websites the success rate of the best web agents rose from 61 to 97.7 percent in a little over 16 months; asked to turn what they found into a document, the best completes fewer than 3 percent of the tasks in full. For brands, what an agent can do matters less than what it may do in their name: its Agent Authority.
AI content of low quality, produced in bulk, that nobody asked for.
The programmer Simon Willison popularized the word in May 2024: the way spam became the term for unwanted email, slop would become the term for unwanted AI content. Merriam-Webster chose “slop” as its word of the year for 2025. What it describes is less the error than the middle, smooth, predictable, interchangeable: consensus taste from the machine. For strategy there has been a name of its own since March 2026, trendslop: seven language models took the same side across all vendors on six of seven strategic either-or questions.
A system that answers a question directly instead of linking to results. AEO is the practice of optimizing for it.
The difference from a search engine is the missing list: there is one answer, not ten options. If you are not in it, you are not there. In practice AEO overlaps heavily with GEO; the split is more the market's vocabulary than a real distinction.
Markup telling a system which part of a page is main content, navigation or footer.
It answers where the content begins and the navigation ends. Without it a program has to infer that from the design, and design is what it does not see. In our own survey, 52 of 135 home pages of German listed companies carry no main-content landmark.
Technology that detects and refuses automated access.
In our own survey, 12 of 160 home pages refuse an automated request, mostly through Akamai or Cloudflare. The point of it: the defence recognises the tool, not the intent. An agent sent by a customer to check availability is turned away just like a scraper.
Brand as something that runs rather than a document you read: with an owner, versions, tests, and maintenance.
The word changes the part of speech: machine-readable describes a property of files, software something that runs at the client. Paul Jun sums up the weakness of the old model: “The old system distributed rules without distributing judgment.” The software only becomes binding once it is set out what applies, how serious a violation is, and what is deliberately left unautomated. The term describes production with the brand, not the question of how machines find and cite it.
The machine-readable layer of a brand: explicit, parametrizable, for systems.
It complements brand identity, which addresses people, rather than replacing it. A brand needs both at once: identity for people, infrastructure for systems. Older uses of the expression exist in a different sense, for example for organizational brand processes.
The version of a brand that a tool stores for itself: logo, colors, typefaces, often guidance on brand voice as well.
Canva, Microsoft, and others call their feature by that name. Each tool thus keeps its own copy of the brand, and each copy ages on its own. Microsoft’s Copilot learns a brand from exactly one PDF; a new one overwrites the existing values, brand voice included. A brand kit ensures consistency within one tool. Which copy applies when two contradict each other, it does not decide.
The brand in a form that can be applied rather than interpreted.
Positioning, language rules, facts, and boundaries, written for people and systems at once. A brand book describes a brand and leaves the reading to the reader. A specification states what holds, and can therefore be executed. That difference decides whether a brand appears in automated processes at all.
Software operating websites on a user's behalf: comparing, booking, buying instead of just reading.
Distinct from two neighbors: a crawler collects content for an index, a chatbot answers from what it has learned. A browser agent acts, in real time and with consequences. It reads pages through their structure, the accessibility tree, and takes literally what it finds there. For brands it is the new audience: the reader that fills in nothing.
The machine-readable commitment declaring which version of a piece of content is the original.
A reference in a page's source code (rel=canonical) stating where its original lives. Search engines and AI systems then credit the original address; duplicates do no harm. This site applies it itself: the guest essay "The Gap That Doesn't Respond" lives here in the archive and points via canonical to its first publication at The Business of Brand Management.
The named mention of a source in an AI system’s answer, usually with a link.
The currency of AI search: whoever is cited appears. A visit it is not. When an AI summary appears, users click a classic result on 8 percent of visits, and on 15 percent without one. And it is volatile: Reddit’s share of the sources ChatGPT cites fell from 3.83 to 0.52 percent within a few days.
The decision that ends interpretation: what applies, how you would notice it does not hold, and who stands behind it.
A brand book explains, a commitment decides. “We communicate at eye level” lets three readers understand three things, “No more than one exclamation mark per text” only one, and a system can apply it without guessing. What gets committed is the rule, not the meaning: a secret stays a secret, it just no longer depends on who happens to be in the room. In strategy this means a thesis, a named failure signal, and a sender who bears the consequences. A language model delivers the most probable consensus instead.
A ready-made link through which an AI assistant uses an outside service; a plugin bundles it with skills into a package you install.
Technically, a connector usually runs over the Model Context Protocol. For brands, connectors and plugins are the way into the assistants’ directories: Monotype is testing a connector for typefaces and licenses in Claude, and Meta reviews submitted connectors for its agent Muse for function, security, and legal requirements. Anyone who wants a plugin in ChatGPT’s directory has to set out what it should not do, with exactly five positive and three negative test cases.
Taste that can be learned from many examples, and therefore reproduces the middle.
What many people consider good can be trained. When that middle comes out of a machine, we call it slop. The objection is not aimed at the machine but at the assumption that taste is an average. A decision that risks something is by definition not the middle.
Consistency means every application keeps to the rules. Coherence means all applications together still add up to one brand.
Consistency can be looked up: color values, logo variants, type weights, product names, all questions with one right answer that a machine sticks to. Coherence arises one level up, in the layer of meaning, and each time it is a judgment about a single case. Consistency tolerates many rules, because they are cheap to check. Coherence needs few, and those have to really hold. Choosing them is brand management: fix hard the few things a brand is recognized by, and you can release the rest.
A program that retrieves websites automatically and collects their content.
What matters is the purpose, and there are three: a search index, training AI models, and retrieval on behalf of a user who is asking right now. OpenAI separates them into three tokens: GPTBot for training, OAI-SearchBot for search, ChatGPT-User for retrieval. Not every vendor separates them. Crawlers that collect for search and training at once make up 36.6 percent of verified crawler traffic, according to Cloudflare; blocking training with them means blocking search as well. This website’s robots.txt explicitly allows all three purposes.
The collection of reusable building blocks a brand uses to build its digital products: components, colors, typography, and rules for their use.
A design system structures the execution layer, that is, how the brand looks, mostly as code and as design tokens. At IBM, Porsche, and Deutsche Bahn it is public and versioned. For machines it is the most developed part of a brand: of 21 open-source design systems examined, 18 ship their own MCP server and 18 official Agent Skills. What the brand means, and what may be said in its name, is not in it.
The smallest machine-readable design decision: a color, a spacing, a typeface as a named value.
Instead of "blue as in the PDF", a token holds a named, versionable value tools can apply directly. Since 28 October 2025 a stable specification exists, published by the Design Tokens Community Group at the W3C. It is explicitly not a W3C Standard and not on the W3C Standards Track. Tokens prove that a brand's execution layer can become machine-capable. For the layer of meaning, the counterpart is still missing.
Umbrella name for the packages with which the European Commission wants to simplify its digital rules, presented on 19 November 2025.
There are two proposals. One amends the AI Act and has been in force since 27 July 2026: high-risk obligations are postponed to December 2027 and August 2028, and generative systems already on the market must retrofit machine-readable marking by 2 December 2026. The other concerns the GDPR, the ePrivacy Directive, and the Data Act, among others, and was still a proposal in September 2026. Anyone planning with AI Act deadlines is planning with the state after the omnibus.
A metric from SEO tools estimating the strength of a domain from its link profile.
It comes from tool vendors, not from search engines. For visibility in AI answers it carries little: across a survey of 75,000 brands, third-party mentions correlate at 0.66 to 0.74, domain authority at 0.27 to 0.33. What others say about a brand weighs more than what the brand writes about itself.
A brand's two layers: how it looks and sounds (execution), and who it is and what holds for it (meaning).
The execution layer corresponds to classic corporate design and has become partly machine-capable through design systems and component libraries. The layer of meaning corresponds to corporate identity and almost everywhere lives in sentences a human has to interpret. Toward machines it needs a sentence that would be unnecessary for a human audience: what is meant to apply must also exist as text.
The named event at which a recommendation visibly fails.
Without a failure signal, a strategy paper is not finished; it is plausibility in a nice format. A failure signal is a sentence, not a model with two decimal places: how will we know in twelve months that this recommendation does not hold? The RAND Corporation calls such a signal a signpost.
Optimizing content for the answer a generative system gives, rather than for a ranking in a list of results.
Generative Engine Optimization. SEO gets you found; GEO gets you into the answer. Both remain work on discoverability. Machine Readable Brands start one step earlier: with the question of whether anything applicable exists that a system can work with.
A text file summarizing a website's most important content for language models.
Proposed in 2024 and heavily promoted since. For coding and browser agents that are handed the file directly, it is useful. As a visibility lever it is, by measurement, ineffective: a server-log analysis across 137,210 domains found that 97 percent of all llms.txt files received no request at all in May 2026. This site has one, for the first reason, not the second.
Systems that buy on their own or prepare purchasing decisions.
Coined by Gartner. In a survey by the firm, chief executives estimate that by 2030, 15 to 20 percent of their own company's revenue will come from machines as customers. That is a self-assessment by those surveyed, not a house forecast from Gartner. What matters is not the number but the shift in addressee: brand communication then speaks to a customer that sees no images.
Prepared so that software can process content without guessing: with fixed fields, unambiguous values, and a declared structure.
Digital is not yet machine-readable. A brand book as a PDF can be opened and summarized by a machine, but not applied without guessing; a rule like “No more than one exclamation mark per text” can be applied. The word now has legal consequences: online, the reservation against text and data mining only counts in machine-readable form, and the AI Act requires AI-generated content to carry machine-readable marking. Machine Readable Brands describes how far a brand has come along that path.
Brands whose identity, language, and design logic exist as a specification that people and systems can work with equally.
Third stage of a maturity model: being read, producing with AI, becoming applicable. A brand book describes a brand; a specification can be applied.
A brand as systems process it: source code, structure and markup instead of design.
What AI search, crawlers and browser agents receive from a page is not the designed but the marked-up: text, roles, names, structure. The demonstration of the same name sets both views side by side and shows what gets lost on the way. Since August 2026 also the title of the Maschinensicht column at The Business of Brand Management.
A simple text format that marks structure with a few characters: a # for a heading, a * for a list item.
John Gruber created it in 2004. People read it like ordinary text, and language models usually write their answers in it themselves; that is why instructions for agents almost always come as Markdown, from llms.txt to the SKILL.md file. GitLab even keeps its brand voice that way. Every article on this website is also available as Markdown: append .md to its address. The evidence is available as one Markdown file per topic.
An open standard through which AI systems reach external tools and data sources.
Introduced by Anthropic in November 2024. OpenAI adopted it in March 2025, other providers followed the same year. Since 9 December 2025 stewardship has rested with the Agentic AI Foundation, a directed fund under the Linux Foundation. Alongside it stand Agent2Agent (A2A), through which agents are meant to talk to each other, and Microsoft’s NLWeb, which turns a website into an answer field. For brands the technology matters less than the consequence: there is now a standard route by which an agent reaches brand data. Anyone with nothing available along that route does not exist for agents.
The brand specification as a public, versioned repository.
Identity, language, and facts readable for agents and verifiable for people, while the boundaries of what an agent may promise stay internal. Trust through verifiability instead of assertion. Every rule a documented decision, with date and sender.
A scientific paper that is published but not yet peer reviewed.
That makes it citable and unchecked at once. On the evidence page this stands as a limit on every entry resting on a preprint. The difference is not a formality: peer review is where specialists get to attack the method before a figure goes into circulation.
A store for files that records every change with date, author, and rationale.
The term comes from software development; a repository is usually kept with the tool Git and hosted on services such as GitHub or GitLab. Each change is a commit: you see what changed, who made the change and why, and can return to any earlier state. GitLab keeps its own brand this way, brand voice, naming rules, and values as public files, every change with a real name and a rationale. For a brand it is the difference between a PDF nobody knows still applies and a state with a date.
Text file in the root directory telling crawlers which areas they may retrieve.
Older and more effective than llms.txt, and a request rather than a barrier. For AI crawlers it is today the only widely used means of control, and it steers through tokens that each vendor assigns and can change at any time. No common standard for this exists.
A shared vocabulary for marking up who or what a website is. JSON-LD is the usual format for it.
The least exciting and most effective part of brand infrastructure. A person can see that a page holds a person, a product, or a price. A system has to be told. Without that markup, a statement stays prose a system has to interpret instead of adopt.
Additional markup intended specifically for AI systems.
Google declares it expressly unnecessary in its own guidance, together with llms.txt and custom chunking. That is the inconvenient answer for a whole market of tools. What helps is less spectacular: reachable pages, and clean markup of what is there anyway.
A brand carries its entire history with it. Every statement lands against everything said before.
A language model is stateless: each request starts from zero, with no memory of the last. Brands work the other way around. What you promise today is measured against what you promised three years ago. So it is not enough to tell a system how a brand sounds. It has to know what the brand has already said.
Statements on a page that do not merely sit there but are marked up as what they are.
The umbrella term for Schema.org and JSON-LD. Google records in its own guidance that AI search needs no special markup; for rich results in classic search the markup remains useful. That both hold at once is why the question of whether markup helps is so often answered wrongly.
AI-generated participants who answer a survey in place of people, also called digital twins or silicon samples.
As a replacement they do not work so far: the Pew Research Center re-ran three of its own surveys from 2026 with AI respondents. Across nearly 300 questions the results were off by 12 percentage points on average, by more than 15 on around 28 percent of the questions. Uncertainty hardly occurred: people chose “not sure” around four times as often as the AI. Carried over to brand research, not tested there: synthetic personas answer more confidently and more uniformly than the people they are meant to stand for.
A rights holder’s objection to text and data mining, the automated analysis of their content.
Text and data mining of lawfully accessible content is permitted, commercially too, as long as the rights holder has not reserved it. For content online, the reservation under Section 44b of the German Copyright Act is effective only if it is made “in a machine-readable format”. The law demands machine readability, then, without saying which format suffices. In circulation are comment lines in robots.txt, Cloudflare’s Content Signals, and the drafts of the IETF working group AIPREF; there is no common standard. Of 77 German-language news and trade media, 20 declare their reservation only as a comment line, which no crawler evaluates. US law knows no such reservation; there the question is fair use.
An attack in which instructions are hidden in a tool’s description, and an agent follows them without a person seeing them.
Invariant Labs described it in April 2025: a manipulated tool got an agent to send the private SSH key to an outside server. Measured on 45 real MCP servers with 353 tools, the attack succeeded in 36.5 percent of cases averaged over 20 model settings, at most in 72.8. Anyone who offers a brand as an MCP server or connects outside servers therefore has to check the tool descriptions too, not just the content.
Open standard for purchasing through agents, carried by a body of several vendors.
The counterpart to the Model Context Protocol on the commerce side. The body is staffed and the announcements are numerous; to this day no source gives a figure for adoption, Google included. A seat on a standards body does not establish adoption.
The string with which a program identifies itself when retrieving a page.
It is self-declared and can be set freely. Bot defence nonetheless rests largely on it. For brands that means two things: steering crawlers means steering through a statement anyone can forge, and locking out an agent may mean locking out a customer.
A proposed web standard through which a website offers agents in the browser tools: functions with defined inputs, instead of leaving them to guess which button does what.
Where the Model Context Protocol needs a server, WebMCP runs in the page itself. Chrome has been testing the standard since June 2026 in an origin trial, a test phase websites sign up for. The regular launch is planned for Chrome 157 in November 2026, as a target, not a commitment. It is not yet in use: among the most visited domains, not a single one offers tools via WebMCP. For brands the question shifts from whether an agent understands a page to what the page explicitly offers it.
The free knowledge base of Wikimedia, in which every person, company, and thing has a fixed identifier.
The identifier consists of Q and a number and stays the same even when a name changes. Websites point to it in their structured data to make clear who is meant, especially among namesakes. For brands a well-kept entry is one of the few places where they help decide how machines identify them. This website points to its author’s entry, Q141223781, and its own surveys take the companies’ addresses from Wikidata.