The GEO Handbook, Part 1: Opportunities, Risks and the Trust Problem

GEO can make your company visible in AI answers. It cannot guarantee a mention or a sale. A serious GEO programme therefore does three things: it improves verifiable sources, measures observable signals and states clearly what the platforms keep hidden.

This is Part 1 of the GlasBox GEO Handbook. The series is written for Swiss SMEs, marketing leads and management teams. It explains GEO without magic words. Each part ends with controls you can verify yourself.

01 — What GEO is, and what it is not#

Who are you? What do you offer, to whom, in which region and on what terms? What evidence supports your claims? Consistent answers on your website and in reliable external sources make interpretation easier. They do not replace technical accessibility or a platform’s source-selection decision.

GEO does not replace SEO. In 2026, Google states explicitly that SEO fundamentals remain relevant to generative Search features. Google also rejects supposed shortcuts: Google Search does not need a special llms.txt file, artificially fragmented copy or pages rewritten just for AI. Clear technology, distinctive content and reliable information still matter. [1]

Technology, evidence and contact paths belong together. For each project, we establish which signals are available and how to check them. Missing analytics or CRM data remains an explicit measurement gap. A mention alone proves neither an enquiry nor a sale.

Planned chapters in the series

Parts 02 to 08 are the planned continuation. This first part focuses on opportunities, limits and how to assess an engagement. The introductory guide and local examples are linked at the end.

02 — Why GEO is a management task#

A conventional search result shows a page. A generative answer can combine several sources, omit claims, form a recommendation or show no link at all. The question shifts from ‘What position do we hold?’ to ‘Which claim appears in which decision context, on what basis and with what consequence?’

Marketing cannot answer that alone. Technology controls access. Subject experts own the truth. Editorial and design make it understandable. Sales and service know the questions that precede a purchase. Legal and privacy set limits. Management decides which risk and cost are acceptable.

03 — Five challenges every GEO strategy must withstand#

1. The system remains partly a black box

OpenAI, Anthropic and Google document access controls, product features and selected measurement signals. They do not expose the complete selection and weighting logic behind every answer. Even when your page is accessible, you cannot see every competing source, internal subquery or personalised variant.

The practical consequence is simple: you are not optimising a fixed formula. You are improving the quality and availability of your evidence. Then you test samples. Anyone who turns ten prompts into a general visibility guarantee is claiming more certainty than the data supports.

2. Measurement remains incomplete

Part of the journey is visible. Since 31 August 2026, Google has made its separate report for generative AI features available worldwide. It shows organic impressions in AI Overviews and AI Mode, including pages, countries, devices and changes over time. An impression means a link to the site appeared in an included generative feature. [2] The report may be absent when impressions are insufficient.

Website visits from AI systems can be identifiable in analytics. GA4 documents an “AI Assistant” channel; Google AI Overviews and AI Mode fall under Organic Search. Check source and medium as well as actual collection. Missing consent or referral information can leave gaps. Visits are not a complete measure of activity inside a platform. [5]

3. Prompts do not form a stable keyword list

People do not only ask ‘fiduciary Lucerne’. They describe a situation, budget, language, deadline and concern. Then they ask follow-up questions. Another person frames the same problem differently. One keyword cannot represent that variety.

The answer is not an endless prompt inventory. Build a small, versioned panel around decision situations: understand the problem, compare providers, test risk and trigger an action. Vary only features that matter to the business. For a Lucerne SME, these are often region, language, audience and scope.

4. Zero-click can be both useful and costly

An answer without a click is not automatically worthless. It may communicate the correct company name, specialism or location. It may also satisfy the entire information need without sending a visit. Then there is no referral, and you can barely see whether the brand was noticed.

Separate four cases: answer without brand, answer with source, answer with brand and answer with action. Only the last case contains a clear next step. That step still has to work: phone number, form, route, appointment or booking.

5. GEO requires coordination

GEO involves coordination: which service is actually offered, which figure is current, and who approves claims and corrects conflicting profiles? Editorial staff, subject experts and IT need clear responsibilities.

04 — Five opportunities that go beyond reach#

1. You can appear while the decision is being framed

A strong GEO source does more than answer ‘Who offers this?’ It explains differences, limits and selection criteria. Your company can act as a guide early in the research process. That role is valuable when the claim is correct and the brand remains recognisable.

2. Owned evidence is harder to copy than prose

Original evidence can distinguish an offer: documented methods, first-party data, traceable examples and local experience. The creative work turns that material into clear explanations, useful comparisons and meaningful visuals. Specialisation helps identify the questions and limitations that matter to customers. None of this guarantees a defensible position in AI answers.

3. Answer tests expose gaps in the offer

When a system describes your service incorrectly, the system is not always the only problem. The service page may be vague. Two locations may conflict. A price range or exclusion may be missing. GEO tests then become a stress test for your own information system.

4. Early teams build a learning lead

Early tests can build operational knowledge: a source inventory, test panel, clear approvals and a team that handles errors. Whether this becomes a commercial advantage depends on the market and execution. Starting early is not protection from competition.

5. Good GEO work improves the purchase decision

Clear answers also help outside AI systems. They improve service pages, sales conversations, onboarding and support. This is the durable part of the business case. Even if a platform changes, better information and shorter internal approval paths remain.

05 — The dangerous claim: ‘People trust AI’#

Do not present that sentence as a universal fact. Trust depends on the person, topic, interface, source display and level of risk. A restaurant idea is checked differently from a tax question. Without a defined sample and context, ‘high trust’ is not a defensible business metric.

The providers themselves call for verification. OpenAI explains that ChatGPT can produce incorrect or misleading answers and can even fabricate sources. Important information should be checked against reliable sources. That warning is not a footnote. It is a core principle for responsible GEO. [4]

The goal is not to ‘look objective’. The goal is to be verifiable. Show the date, author, scope, method, source and next step. Remove superlatives that nobody can support.

06 — GEO as a credence service: a useful model#

Economics distinguishes qualities that can be inspected before purchase from qualities revealed through use. Phillip Nelson described search and experience qualities in 1970. Michael Darby and Edi Karni extended the model in 1973 to qualities buyers may struggle to assess even after purchase. [7][8]

An offer can combine search, experience and credence qualities. A smartphone has specifications you can inspect beforehand and usability you discover in daily use. GEO consulting has credence qualities: clients cannot see every platform interaction, and the cause of a change is often difficult to isolate. This is a conceptual model, not a measured ranking of products.

That does not make every part unverifiable. A serious provider turns part of the service into inspectable goods: a source register, a documented robots.txt review, a list of changed pages, version history, screenshots, raw data, test rules and acceptance criteria. The uncertain outcome stays uncertain. The work performed becomes visible.

07 — Trust is built, not claimed#

A good proposal separates four layers. First, it defines the scope. Second, it names deliverable outputs. Third, it describes observable signals. Fourth, it frames business outcomes. More factors influence the result as you move right. The wording must become more careful at the same time.

Acceptance must not depend on one live answer. Answers change. Fair acceptance checks the agreed artefacts, technical correctness, documented tests and implemented changes. Visibility is then observed as an outcome, not owed by the provider as a ranking.

08 — A measurement model without false precision#

Our working model separates directly available data, controlled test observations and platform processes that are not fully visible. The collection intervals below are suggestions; scope, access and frequency are agreed for each project.

Illustrative calculation, not a GlasBox measurement: “18 citations in 96 documented answers, of which 14 were factually correct” is more inspectable than “18 citations”. Record system, language, location, account state, date and repeats. Log failed requests separately. Repeats are not independent people, and the rate is not market share.

Avoid a single GEO total if nobody understands the formula. Management often needs four separate figures: accurate source rate, accurate brand mention rate, actionable answer rate and confirmed qualified leads. Each figure should include raw data and a short limitation.

09 — Worked example: a GEO audit for a Lucerne fiduciary firm#

The fiduciary firm and programme below are fictional. Test volume, page count and timing illustrate a possible engagement; they are neither client results nor the standard SOURCE/01 scope. The example company serves small Swiss limited companies in Central Switzerland in German and English. Old team pages and inconsistent service names make interpretation harder.

Step 1: define the business question

The business question is not ‘How often does ChatGPT mention us?’ It is: ‘When decision-makers in Lucerne ask concrete fiduciary questions, do they receive an accurate account of our offer and a working route to an initial consultation?’ That question creates measurement fields connected to the business.

Step 2: inspect sources and access

• Inventory all service, team, contact and location pages.

• Assign every concrete claim to a responsible subject expert.

• Flag stale names, services, opening hours and language versions.

• Check robots.txt, status codes, canonicals, redirects and server-side blocks.

• Document OAI-SearchBot, Claude-SearchBot and Google access separately from training choices.

• Check Google Business Profile and important company listings for consistency.

• Test the mobile contact path in both languages.

OpenAI separates OAI-SearchBot for search visibility from GPTBot for potential model training. Anthropic distinguishes Claude-SearchBot, Claude-User and ClaudeBot. These controls serve different purposes. A blanket instruction to ‘allow all AI bots’ is too crude. [3][6]

Step 3: build a small, repeatable test panel

The example uses 24 questions: six each for orientation, comparison, risk checking and action. Each question runs in three defined systems and is repeated four times. That produces 288 documented observations per measurement wave. The number is not market share. It is a controlled sample.

Step 4: code outcomes separately

Every observation receives the same fields: answer present, own domain cited, brand mentioned, explicitly recommended, claim accurate, region accurate, language appropriate and action possible. The team also records an error class and supporting evidence. A screenshot becomes a dataset.

Every error receives an owner. A wrong location goes to operations. A conflicting service goes to the service owner. A blocked retrieval goes to IT. A vague page goes to editorial. GEO becomes operational only when an observation creates an assigned task.

Step 5: turn the audit into a clear proposal

The example agrees deliverables and verification rules, not a platform recommendation. Changes to the test panel are versioned and analysed separately. If no improvement appears, review data quality and assumptions before deciding whether to continue or stop.

The annual rhythm

10 — How to review a GEO proposal before signing#

• Does it name the domain, markets, languages, platform interfaces and period?

• Are prompts, repeats and test conditions available as an appendix?

• Are source, mention, recommendation, accuracy and action measured separately?

• Does the report include raw data rather than only a proprietary score?

• Are technical checks and editorial work described as concrete outputs?

• Does it say who approves factual claims?

• Does it separate training, search and user-directed retrieval in bot controls?

• Is a qualified CRM lead defined?

• Does it state known measurement gaps and platform dependencies?

• Can acceptance happen without relying on one random live answer?

• Do you retain ownership of the site, raw data and created content?

• Is there a stop, learn or reprioritise decision after the pilot?

11 — Seven warning signs#

An eighth warning sign is excessive certainty. GEO is young, but that does not excuse weak method. The less transparent the channel, the more precise the scope, method and limitation must be.

12 — Frequently asked questions#

Does GEO replace conventional SEO?

No. Crawling, indexability, clear pages, internal structure, useful content and reputation remain foundational. GEO adds answer contexts, source review, platform tests and more granular measurement of representation and action.

Can an agency guarantee visibility in ChatGPT?

No. It can inspect access, improve sources, run tests and reduce errors. It does not control the system output. A fixed placement promise is therefore not a responsible acceptance criterion.

Is a citation the same as a recommendation?

No. A source may support one claim. A brand mention may be neutral. A recommendation makes a positive selection. Measure all three events separately.

Is zero-click always bad?

No. Correct brand or service information can be useful without a visit. Zero-click remains hard to attribute to sales. Add action signals and CRM data when commercial value is the goal.

Does Google require an llms.txt file?

No. Google explicitly says llms.txt is not required for Google Search and does not improve Search visibility or rankings. Other services may have their own rules. Decide per system and maintain only files with a clear purpose. [1]

How often should a company test prompts?

As often as the decision requires. A pilot may measure monthly. A stable area may only need quarterly testing. Run a focused test after major page, offer or platform changes. Keep the core panel identical and document variants.

Why is GEO consulting a credence service?

Because the client cannot see every system interaction and effects are rarely attributable to one action with certainty. That uncertainty demands more documentation. It does not justify vague reports.

What should the first GEO audit deliver?

At minimum: a source inventory, technical access review, versioned prompt panel, raw observations, error classes, prioritised actions, owners, acceptance criteria and an explicit limitations list.

13 — Conclusion: the advantage is not a trick#

The biggest GEO opportunity is not another channel. It is a better information system: clear claims, strong evidence, reachable sources, visible ownership and a working next step. That work helps people and machines.

The biggest danger is also clear: false precision. A dashboard can make uncertainty look tidy. An agency can sell tests as market share. One answer can look like proof. Protect the budget with defined scope, raw data, repeatable checks and honest limits.

Sources and editorial basis#

Product and technical claims use official platform documentation only. The economic model refers to the original papers. Platform details can change, which is why this article carries a review date.