Representation Changed the Rules

Imagine looking for information about a company. You might type its name into a search engine, open its website and read its services and About pages. Customer reviews, social media accounts and other sources also enter the picture. In this process, you are the one putting the pieces together.

When you delegate the same research to a generative AI system, you do not merely ask for the company’s web address. You expect an assessment:

‘Is this company trustworthy?’ ‘Which services does it genuinely specialise in?’ ‘What distinguishes it from its competitors?’ ‘Is it suitable for my project?’ ‘What is known about its prices?’ ‘Can it work internationally?’ ‘What evidence does it have?’

A system equipped with search and source-inspection tools can base its answer on several sources. It can bring together accessible web pages, service descriptions, social profiles, publications, structured data, independent references and older records. These sources may conflict. What the user receives is not just a list of links, but a narrative constructed from sources: a representation. We must also remember that not every model has web access or a tool capable of reading all these sources. This is no small change.

A company describes itself on its own pages. A generative system can build a new account of the company from the information it accesses. How the brand is perceived is then shaped not only by its own words, but by how those words are connected to other sources. Representation changed the rules.

Representation is not a summary

When an AI system describes a company in three paragraphs, it is doing more than shortening longer texts.

It is making many decisions at once:

• which source matters more,
• which information is current,
• whether records with the same name refer to the same entity,
• which services are central and which are secondary,
• which claims are supported,
• which contradictions to disregard,
• which details matter to the user,
• and which uncertainties to reflect in the answer.

Representation is therefore not mechanical compression. It is the construction of a meaningful model from selected parts of reality. A company may have published hundreds of true sentences online. Yet if those sentences do not fit together, the representation an AI system constructs may be wrong. If a service page says the company works in six languages while another profile says it serves only a local market, which will the system use?

If the homepage mentions AI automation but the service catalogue does not include it, will the system treat it as a genuine capability or a marketing phrase?

If one pricing page lists an inexpensive annual hosting package and another a monthly managed-operations service, will the system interpret them as the same product?

If a company has moved into a new specialism but its old profiles still emphasise past activities, which identity will the machine represent?

A person who notices these contradictions can ask a question. If the system builds a plausible narrative instead of pointing out the conflict or seeking clarification, the error can become invisible. The most plausible account is not always the true one.

A company has more than one face online

A brand may believe its identity resides on its homepage. The brand a machine encounters is much broader than that.

It exists simultaneously across:

• the main website,
• service pages,
• price and scope records,
• social media profiles,
• business directories,
• customer reviews,
• press coverage,
• academic publications,
• downloadable files,
• old web pages,
• partners’ websites,
• employee profiles,
• domain and company records,
• structured data,
• accessible web archives,
• and texts produced by other AI systems.

People see only a few of these surfaces. Machines may encounter a much wider footprint. Digital identity can therefore no longer be managed on a single page. A company can prepare a flawless homepage, yet outdated, contradictory or incomplete records elsewhere can still distort the machine’s representation. Think of an orchestra. Every instrument may be playing the right note. But if they play in different keys, at different tempos and from different pieces, the result is confusion, not music. Digital representation works the same way. Each sentence being true on its own is not enough.

Together, the sentences must describe the same reality.

Six fundamental questions about representation

In this book, we use six questions to assess whether a representation is coherent:

Who is this? What do they do? What evidence supports that? What do they not do? When is this information valid? Which source is more authoritative?

If any of these remains unanswered, the representation becomes weaker.

Who is this?

Identity is the most basic layer. The same company may appear under different names on different platforms. Its brand name may differ from its registered business name. A person’s author name, company title and social media name may differ. A brand may operate under another company. If these relationships are unclear, a machine may merge different entities into one or treat a single entity as several disconnected records. Identity is more than a name.

It brings together:

• the official or trading name,
• the brand name,
• the domain,
• ownership relationships,
• the founder or authorised representative,
• the field of activity,
• geographic scope,
• language coverage,
• and current status.

If identity is uncertain, every other layer of representation becomes uncertain too.

What do they do?

Many companies answer this question too broadly. ‘We provide digital solutions.’ ‘We design innovative experiences.’ ‘We prepare businesses for the future.’ These phrases may impress people, but they give a machine no sound basis for action.

A record suitable for assessment must distinguish concrete capabilities:

• Does the company develop websites?
• Does it build client portals?
• Does it provide technical SEO?
• Does it develop AI agents?
• Does it provide hosting?
• Does it offer managed site operations?
• Does it prepare content only, or implement the system too?
• Is the service advisory only, or does it include delivery?

When a company tries to say it does everything, it often fails to make any one capability sufficiently clear. Strong representation comes from precise capabilities, not sweeping claims.

What evidence supports that?

Saying that a company provides a service is not the same as supplying sufficient evidence of its ability to provide it.

Evidence can take different forms:

• published work,
• a sample implementation,
• technical documentation,
• a version record,
• verifiable customer feedback,
• an independent citation,
• a live product,
• a measurement report,
• a test result,
• a canonical publication,
• or a work process with acceptance criteria.

Evidence is not limited to success stories. Explaining how a company works, the limits it accepts and its verification methods can provide grounds for understanding its process. Describing a process, however, does not by itself prove that it has been implemented successfully. For a machine, trust does not arise simply from the number of positive sentences. It arises from the connection between a claim and its evidence.

What do they not do?

This question is absent from most marketing copy. Companies want to describe what they offer, not what they do not. Yet exclusions are extremely valuable to machines. When a service’s exclusions are not explained, a system may fill the gap by guessing.

For example:

• Does the hosting package include managed operations?
• Does logo design mean a complete brand system?
• Does AI avatar production include unrestricted use of a person’s face and voice?
• Does the SEO service guarantee rankings?
• Does consultancy include implementation and software development?
• Does ‘support’ mean continuous, round-the-clock intervention?
• Is data migration included in the main project price?

A company clearly stating what it does not do is not showing weakness. It filters out unsuitable requests and gives the right customers more reason to trust it. Representation is not made up solely of positive capabilities.

Limits are part of identity too.

When is this information valid?

Undated information leaves its currency uncertain. A pricing page contains a figure, but we do not know which year it belongs to. A company profile shows an old team structure. A service page refers to technology that is no longer used. A press article presents a field of activity from years ago as if it were current. When freshness signals are weak, a machine may give old and new information equal weight. Representation therefore needs a time dimension. It should make the following clear:

• publication date,
• last-updated date,
• version number,
• period of validity,
• observation date,
• measurement period,
• and archive status.

For information that can change, saying it is correct is not enough. We must also know when and under what conditions it holds true.

Which source is more authoritative?

Different sources may say different things about a company. Its own website may publish one price, a business directory may show an old one, and a social media post may describe a temporary promotion. A blog may misinterpret a service. One AI answer may repeat an older text produced by another AI. It may then be unclear which source the system should rely on. This is where the idea of a canonical source matters.

Canonical source: in this book, a canonical source is the reference point used for a particular piece of information, whose currency and authority can be verified. This is broader than the choice of a canonical URL in search engines. A company designating its own statement as the primary source does not make the claim independently verified. Nor is it enough to say, ‘This page is canonical.’ Other surfaces must not contradict it. If the canonical record and the visible web page give different prices, the system gains no reliability. Canonical status is not a label; it is a responsibility to maintain consistency.

How does missing information become misrepresentation?

When people notice missing information, they often describe it as uncertainty. ‘I don’t know.’ ‘There isn’t enough information.’ ‘We need to ask the company.’ Under pressure to produce an answer, machines may instead construct plausible connections between the missing pieces. If a company is known to build client portals, the system may assume data migration is included. If a brand publishes pages in six languages, it may conclude that live support is available in all six. If a company designs 3D experiences, it may assume WebXR support on every device.

If a consultant works with AI automation, the system may imply that they can autonomously manage all a client’s systems. These inferences may sound reasonable without being true. A representation gap is therefore a source of risk. An error can arise not from explicitly false information, but from a plausible, unverified connection between incomplete facts. Good representation is consequently more than adding correct information. It also closes the gaps in which a machine might draw the wrong inference.

This is why statements such as these matter: ‘Content is available in six languages; the language of live support is specified separately in the contract.’ ‘Data migration is scoped separately after the source system has been assessed.’ ‘The annual hosting package does not include monitoring or incident response.’ ‘Permission to create an avatar does not automatically authorise publication of every piece of content.’ ‘Acceptance of an IndexNow notification is not evidence of indexing or ranking.’ These statements clarify scope and reduce the risk of mistaken inferences. An explicit limit also keeps the sales promise aligned with the service actually provided.

Representation debt

Over the years, companies produce content with different people, on different platforms and for different purposes. A marketing agency writes the homepage. Someone else creates the social profiles. Sales prepares a pricing file. A developer adds structured data. Human resources uses another version of the company description. A directory retains old information. An AI tool generates new copy. Each piece may be reasonable in its own context. Small differences nevertheless emerge over time. ‘Global’ in one place, ‘Europe-focused’ in another. ‘24/7 support’ here, ‘replies during business hours’ there.

One surface advertises ‘operations from $500’ without defining the scope; another lists ‘$200 hosting’, which gets confused with it. ‘AI consultancy’ appears in one place, ‘autonomous agent development’ in another. ‘Free maintenance’ sits alongside a ‘monthly maintenance package’. As these small differences accumulate, the company’s digital representation fragments.

We can call this accumulation representation debt. Representation debt is the risk that outdated, incomplete, contradictory or context-free statements accumulated across an organisation’s digital surfaces will distort future machine interpretations. Like technical debt, it may initially be invisible. The website works. The profiles are active. The content is published. But as the system grows, each new piece of information is layered over old contradictions.

Eventually, nobody can easily answer:

• Which price is current?
• Which service is actually provided?
• Which description is canonical?
• When this page changes, which other records must change with it?
• Which language version reflects the underlying facts?
• Which claim has evidence?
• Which content is merely marketing language?

Human teams can live with this debt for a long time. It becomes visible when AI systems begin reading these surfaces together. The machine tries to resolve the inconsistency. Sometimes it chooses the right source; sometimes the most repeated, the one that looks newest, or the one it considers most trustworthy. Sometimes it misses the contradictions and produces a coherent but incorrect account. Reducing representation debt is one of the responsibilities of the GEO approach advocated in this book. GBO must prevent that debt from turning into inappropriate action.

Repetition is not truth

Information repeated in many places online can gain the appearance of truth. But repetition alone is not evidence. A false statement about a company may be published in a directory. Other sites may cite it. AI-generated content may reproduce the statement, and other models may then read those new texts. The misinformation spreads without independent verification. Eventually, the same claim appears in several sources, although all of them trace back to one error. We can call this representation laundering.

A false or weak claim starts to look independent and reliable because it has been repeated across many surfaces. A company might first call itself ‘the world leader’ on its own site. The claim is then repeated in automatically generated company profiles, followed by AI-written lists. Other systems use those lists as sources. The claim grows. The evidence does not. GEO should not aim to produce the internet’s most repeated narrative. It should build an accurate representation with traceable sources and explicit limits. This matters even more for GBO.

Repeated falsehood can influence not only answers, but choices and actions. If a supplier’s capacity has been exaggerated, an agent may select it for a real project. If a product certification has been misunderstood, a system may recommend the product for a regulated use. If old records show a low service price, an agent may assess budget fit incorrectly. GBO therefore asks about the origin of evidence before counting repetitions.

Where did this claim originate? Is it independent? Is it current? Is it direct? Within what scope is it valid? Do other sources genuinely verify it, or merely repeat it?

In the machine era, authority should be measured by the strength of the source, not the volume of its echo.

From source to sentence, from sentence to identity

When a generative system answers a user, it often does not repeat source statements directly. It combines information and constructs a new sentence. That sentence may not appear in the same form in any of the sources.

Different sources might, for example, contain these facts:

• The company provides content in six languages.
• It develops AI agents.
• It works with international companies.
• It provides SEO and GEO services.
• It has technical infrastructure for multilingual publishing.

Combining these, the model might construct this representation: ‘The company is a specialist agency developing multilingual SEO, GEO and AI automation systems for international brands.’ The sentence may draw on the sources’ shared meaning. Yet every word carries a new responsibility. What supports ‘international’?

What evidence justifies ‘specialist’?

Does ‘developing’ imply an implementation capability distinct from consultancy?

Does ‘agency’ fit the company’s legal or commercial structure?

Are ‘brands’ its actual customers?

Representation is therefore more than collecting information. It creates relationships between facts. New relationships create new meaning, and new meaning creates new risks. A system can combine true pieces through a false connection. A company may publish content in six languages without offering customer support in all six. It may have developed an AI agent without selling it as a product to every client. It may seek international visibility without yet operating in particular countries. An audit of representation must therefore check more than whether individual sentences are true.

It must also check whether the connections between them are true.

A machine sees the same reality in different forms

A person reads a service page.

On that same page, a machine may encounter several layers of representation:

• the title,
• the meta description,
• the page copy,
• link text,
• the menu structure,
• structured data,
• the price record,
• the FAQ section,
• image descriptions,
• the canonical tag,
• language relationships,
• the sitemap entry,
• and date and version information.

When these layers agree, representation becomes stronger. When they conflict, a machine may reach different conclusions. The visible page may say ‘pricing upon request’, while its structured data contains a fixed price. A person sees one; a machine can read the other as well. The visible copy may limit a service to existing clients, while the service catalogue omits that restriction. The homepage may claim ‘global operations’, while language or country records cover only one market. A page may say ‘we do not provide legal advice’, while its FAQ implies a guarantee of legal compliance. The company is then publishing two different realities at once.

One for people. One for machines.

A basic principle of this approach is that the reality people read must not differ from the reality machines read.

We can call this representation parity. Representation parity means that visible content and machine-readable records convey the same identity, capabilities, prices, scope, limits and evidence. It does not require identical wording on both surfaces. An explanation for people can be natural and detailed; a machine record can be more structured. But the underlying semantic contract must remain the same. A significant limit on one surface must not disappear from the other. A guarantee rejected in one must not be implied in the other. A price available ‘upon request’ must not be invented elsewhere. A service restricted to certain clients must not appear open to everyone in another record.

Representation parity is an important part of the connection this book draws between GEO and GBO. Its importance becomes apparent when an agent bases its action on a machine-readable record. If that record differs from what the person sees, an apparently correct action may rest on the wrong contract.

Multilingual publishing does not simply multiply the same reality

A company publishing in six languages may believe it is creating six copies of the same reality. In fact, it is creating six separate surfaces of representation.

Each language has its own:

• vocabulary,
• commercial conventions,
• legal connotations,
• search intentions,
• technical terminology,
• and reading conventions.

A sentence may be clear in English and become ambiguous when translated word for word into Turkish. In Arabic, changing the text direction alone may not suffice. A German service term may carry a scope that its approximate Turkish equivalent does not. An English technical term borrowed directly into Russian may sound artificial or meaningless. Professional terminology in Spanish may vary between countries. Multilingual work is therefore not simply translation.

It is the reconstruction of reality within different language systems.

The difficulty is this:

Every language must sound natural, but no language may invent a new reality. A service’s scope must not expand in one version. A price limit must not disappear in another. An accessibility commitment in one language must not become a vague promise of quality in another. ‘Support’ in one version must not turn into ‘continuous operations’ elsewhere.

Accurate multilingual representation must satisfy two conditions at once:

The language must be natural. The facts must not change.

Without this balance, a company may be visible in six languages yet be represented through six different identities. For GBO, this is a serious problem. If an agent encounters different scope, pricing or authority information depending on the user’s language, it may behave differently towards the same company. Language versions of a behavioural contract must therefore be equivalent not only as translations, but in the outcomes of the actions based on them.

Representation does not have to be entirely positive

Companies want to present themselves in the best light. That is understandable. But representation is not solely about producing a favourable image.

Accurate representation sometimes includes information such as:

• The service is not currently available.
• Capacity is limited.
• Results are not guaranteed.
• The price is determined after the project has been assessed.
• The service is unavailable in certain countries.
• Third-party costs are excluded.
• Without sufficient data, no experimental conclusion can be drawn.
• Nothing is published without human approval.
• Certain uses are not accepted for ethical or legal reasons.

Such statements may appear to reduce sales opportunities in the short term. Making these limits explicit is intended to reduce mismatches, disputes and loss of trust. This information is particularly important for an AI agent. It needs information not only to choose suitable options, but also to exclude unsuitable ones. If a brand presents itself as suitable for every need, an agent may choose incorrectly. If the brand makes its limits clear, a system may choose it in fewer situations, but it has firmer grounds for identifying the right match.

That is exactly what GBO seeks.

Not more choices, but better choices.

Negative information is therefore not a weakness in representation.

It is an input to behavioural quality.

Citation visibility is not accurate representation

An AI system may cite a brand as a source. That is an important visibility signal, but being cited does not mean being represented correctly. A system can cite the right page and misinterpret its contents. It can present a starting price as the price of every package, describe a company’s methodology as a certification or official standard, or turn a framework proposed in a book into a universal rule. It can mistake a historical record for current service scope. Asking only ‘How often were we cited?’ is therefore insufficient.

We must also ask:

• Which sentence were we cited to support?
• Does the source actually support that sentence?
• Was the company identified by its correct name?
• Were the service boundaries preserved?
• Were uncertainties removed?
• Was old information presented as current?
• Were claims from other sources attributed to us?
• Does the citation take the user to the right page?

I argue that GEO assessment must extend beyond citation counts to the quality of representation.

GBO goes a step further:

What inappropriate behaviour could this misrepresentation cause?

If a service price is wrong, an agent may compare budgets incorrectly. If a company’s geographic scope is wrong, it may request an unsuitable quotation on the user’s behalf. If a product feature is exaggerated, the purchasing decision may be flawed. If the consent boundaries of an AI avatar service are incompletely represented, the agent may treat unauthorised content generation as normal. Auditing representation is therefore more than reputation management. It is a safety check on future actions.

The representation state is an input to decisions

Before performing an action, an agent forms an internal state representing the world.

That state may include:

• the user’s purpose,
• the available options,
• their capabilities,
• prices,
• risks,
• evidence,
• limitations,
• required permissions,
• and possible outcomes.

The agent’s decision arises from this representation state. If it is wrong, the result may be wrong even when the decision logic is flawless. If a navigation system uses an incorrect map, the best routing algorithm cannot take you to the right place. If a doctor’s patient information is wrong, sound medical methods can be applied to the wrong premise. If a purchasing agent has incorrect product specifications, an excellent comparison system can still select the wrong product. GBO therefore cannot begin at the behavioural layer alone. It must first establish a reliable representation state.

Put simply, the quality of an action cannot be independent of the quality of representation. Yet representation quality is not enough on its own. An agent may know everything correctly and still act without authority, exceed a risk limit, skip human approval or take an irreversible step.

The relationship is therefore:

Accurate representation is necessary for appropriate behaviour. It is not sufficient on its own.

This is where GEO and GBO connect. GEO helps the agent see the world more accurately. GBO sets limits on how it behaves within the world it sees.

Five representation states

For this assessment, we can distinguish five states:

Unknown

The system does not know the entity well enough. The correct response is not to guess, but to investigate further or explain the uncertainty.

Partially known

The identity may be known, while services, prices or limits are missing. The system may present the entity as a candidate, but must seek clarification before transacting.

Known through conflicting information

Different sources describe different realities. The agent must make the contradiction visible rather than choose whichever information is easiest to use.

Sufficiently known

Identity, capability, scope, evidence and currency are sufficient for a particular decision. This state may permit assessment.

Known and ready for action

Alongside an adequate representation, authority, entry conditions, the transaction path, risk limits and a reversal or remedy mechanism are defined. Remedy does not mean that every effect can be fully undone. In this state, the agent may act within the specified boundaries.

This classification makes an important distinction: being well represented does not automatically make an entity ready for action. Everything about a company may be known, yet there may be no API or transaction channel. Its price may be current, yet its capacity unknown. Its service may be suitable, yet the user may not have authorised contact. Where GEO’s questions have been answered, GBO’s further questions begin.

Representation governance

An organisation’s digital representation cannot be left to chance. If teams update content across platforms and language versions without a shared verification process, representation debt can accumulate. Organisations therefore need representation governance, not just a content strategy.

Representation governance answers questions such as:

• What is the canonical source for identity information?
• Who may update service records?
• When a price changes, which surfaces must change together?
• Which claims require evidence?
• Which languages are bound to the same factual contract?
• How are older versions archived?
• How are machine-readable records compared with visible copy?
• How are third-party profiles audited?
• Who corrects a misrepresentation when it is found?
• How are AI answers sampled and measured?

This is not one editor’s responsibility. Marketing, legal, technical teams, sales, operations and senior management must work together, because representation consists of more than words. A price is a commercial decision. An authority boundary is a legal and operational decision. A structured-data record is a technical decision. A service exclusion is a sales decision. An update date is a governance decision. GEO is therefore inadequate if treated merely as copywriting. The GEO approach I propose here governs the information and limits through which an organisation is represented in the digital world.

GBO extends that governance to action.

Three quiet qualities of good representation

Good representation is often unshowy.

It has three quiet qualities:

It is consistent

Different surfaces convey the same underlying facts. Wording, language and format may change, but identity, services, scope, prices and limits do not contradict one another.

It has boundaries

It does not claim to be capable of everything, conceal uncertainty or hide where it does not apply. Its statements do not reach further than its evidence can support.

It can be updated

Representation is not a one-off publication. When prices, capacity, services or technology change, all relevant surfaces can be brought back into alignment. Old records are retained but not confused with current ones. These qualities do more than help machines understand. They strengthen human trust, which often grows not from great promises, but from the absence of small contradictions.

How should an organisation describe itself?

The answer is not more adjectives. ‘Best.’ ‘Leading.’ ‘Revolutionary.’ ‘Unique.’ ‘World-class.’ Without evidence, these words do not improve the value of a representation.

An organisation should describe itself through this structure:

Who we are What we do Who we do it for Under what conditions What we do not do How we substantiate it When the information was updated

This applies not only to companies, but to products, specialists, universities, public bodies, datasets and AI agents. In a machine-mediated world, reliability comes from explicit relationships, not ornate storytelling.

A view of GEO beyond visibility

GEO is sometimes described merely as a way to appear more often in AI answers. That view is incomplete. If GEO’s only purpose is to make a brand more frequently mentioned, it will soon become another visibility race. More content. More repetition. More artificial sources. More requests for citations. More machine signals. Yet more visibility does not mean more accurate representation.

In this book, I propose a broader role for GEO: to ensure that an entity is represented to people and machines in a way that is verifiable, current, appropriate to the context and explicit about its limits.

This role should produce three outcomes:

Correct identity: the system should know whom it is discussing. Correct meaning: it should understand what the entity does and does not do. Correct basis: it should be able to trace the evidence supporting that meaning. To act safely, an agent needs sufficient information about the relevant identity, capabilities and evidence. This need not come from a separate GEO programme. A reliable organisational record or verified transaction data can also provide that basis.

The chain from representation to behaviour

An AI agent’s behaviour often follows this chain: Source → Information → Representation → Assessment → Selection → Authority check → Action → Outcome. Errors can occur at every link. A source can be wrong. Information can be old. Representation can be incomplete. Assessment can skip suitability conditions. Selection can rely on popularity. Authority can be assumed. An action can be irreversible. An outcome can be measured incorrectly. GBO does not look only at the last link; it examines the whole chain. Representation sits at its centre, because it turns sources into a world that a decision-maker can use.

An agent first settles on what it thinks the world is like. Then it acts within that world.

This makes the following statement one of GBO’s foundations: every machine action is preceded by a representation, visible or invisible. Trying to correct inappropriate behaviour only at the moment of action may be too late. Much of the error has already arisen in the representation formed beforehand.

Not what we can tell a machine, but what we can verify

Organisations will increasingly ask: ‘How should we describe ourselves to AI systems?’ It is a useful question.

A better one is: ‘Which facts about us should AI systems be able to verify reliably?’ The first focuses on communication; the second on evidence. The first may strengthen the narrative; the second strengthens trust. What a company wants to tell machines may differ from what it can actually substantiate. GBO does not accept that gap. If an agent is to act, its decision should not rest solely on the company’s own claim. The connection between claim and evidence must be visible, and the evidence’s scope must not be exceeded. A performance test can show a particular page’s result on a particular date. It does not prove the entire website will always perform the same way.

A customer review can describe a particular experience; it does not prove every customer will obtain the same outcome. A publication explains a methodology; it does not prove the whole industry has accepted it. Using a standard does not mean being certified by the organisation that publishes it. Representation quality should be measured by its ability to preserve these boundaries.

Accurate representation makes correct rejection possible

An agent must not be designed only to make positive selections. It must recognise unsuitable situations too. If a company accepts only large corporate projects, declining a small-budget request may be the right behaviour. If a service is unavailable in certain countries, the agent must suggest alternatives. If an AI avatar project lacks explicit permission to use the person’s face and voice, the system must stop. If a conversion test lacks enough traffic, the agent must not manufacture statistical certainty. If microservices are unnecessary for a DevOps project, it must not recommend them simply because they are expensive.

For this behaviour to be possible, representation must include not just capabilities, but conditions of unsuitability. Good representation does not make an entity attractive in every situation. It makes it understandable in the right context. Its success therefore cannot be measured solely by how many positive sentences it produces.

It must also be measured by this question: can the system recognise when this entity is not the right choice?

This is one of the most important thresholds connecting GEO to GBO.

From representation to responsibility for action

An AI system has represented a company correctly. It has understood the services, recognised pricing boundaries, assessed the evidence and concluded that the company matches the user’s needs. What happens next?

Will it request a quotation?

Add a meeting to the calendar?

Share the user’s information?

Create a draft contract?

Start the payment process?

Responsibility for accurate representation is now joined by responsibility for the authority and consequences of action. Accurate representation gives an agent a strong foundation. It does not automatically confer the right to act. Knowing is not being authorised. Finding an option suitable is not permission to transact. Correctly identifying an entity does not legitimise every interaction with it.

At the end of this second chapter, we can therefore draw a clear distinction:

GEO governs how reality is represented within the machine. GBO governs the conditions of behaviour based on that reality.

At the next stage, an answer will no longer be enough. The answer will become an action.

When that happens, accuracy is no longer the only concern. Others enter the picture:

• authority,
• consent,
• safety,
• explainability,
• reversibility,
• and human responsibility.

All of these become part of the decision.

Why did representation change the rules?

Because companies are no longer known only through what they say about themselves. They are also known through the narratives machines construct about them. Because users no longer have to read every source individually; they can rely on a combined answer. Because an entity’s visibility has become distinct from the meaning attributed to it. Because being cited does not mean being understood correctly. Because many true pieces, connected wrongly, can produce a false representation. Because contradictions people previously noticed can be turned into automated decisions by agent systems.

Most importantly, representation now informs not only perception, but action. Misrepresentation can also cause harm by influencing human decisions. Tool-using agents add another route: the system itself can directly make an incorrect selection, send a message or complete a purchase. Representation is therefore no longer a narrow concern of the communications department.

It is now infrastructure with several dimensions:

• technical,
• commercial,
• legal,
• ethical,
• and operational.

Representation underpins all of them.

The chapter’s conclusion

Before an AI system acts, it constructs a world. That world contains people, companies, products, prices, capabilities, risks and possibilities. The system makes decisions within it. If the world has been constructed incorrectly, stronger decision-making can magnify the danger. If it is incomplete, the agent may fill the gaps with guesses. If it is contradictory, the system may trust the wrong source. If it contains only positive claims, it cannot filter out unsuitable choices. Behavioural optimisation therefore cannot bypass representation.

The following conditions must first be established:

Identity must be clear. Capability must be real. Evidence must be traceable. Limits must be visible. Information must be current. People and machines must see the same facts.

Only then can we ask about behaviour. Representation changed the rules because machines no longer merely describe the world. Agents connected to tools can also transact on the basis of the representations they construct. In the next chapter, we cross that threshold. We will examine when an answer becomes an action, how a seemingly small recommendation produces real consequences, and why responsibility changes fundamentally when agents begin not to speak, but to act.

The way a machine sees the world shapes how it behaves within it.

GBO’s first task is therefore not to tell the agent what to do. It is first to show it accurately what is real.