Being found is not the same as being recommended. Visibility in Google, ChatGPT and other AI answers takes more than publishing additional blog posts. It requires a clear offering, accessible information, credible evidence and measurement that distinguishes a source citation from an actual customer enquiry.
By Kacper Ruta · GlasBox — SEO & GEO Agency · Editorial reference date: 6 September 2026.
I am Kacper Ruta. My approach at GlasBox is straightforward: collecting as many AI mentions as possible is not the objective. I want to establish whether a business is represented accurately in relevant answers, and whether that visibility can contribute to qualified demand. This guide sets out how I approach that task for Swiss businesses.
The starting point is Double Digital’s AI SEO research review, which lists 75 materials. The factual sections below reference selected original studies and current platform documentation. The practical examples and recommendations are my own interpretation. [1]
01 — The decision for business leaders
My recommendation is to extend established SEO with a controlled assessment of AI answers, not replace SEO with GEO. The management question is not “Which new tool should we buy?” It is “Where are we losing a relevant decision: discovery, understanding, trust or the next step?”
Technical problems, unclear service pages and missing local information remain work to be done. A new label does not remove these responsibilities.
Record a source citation, a brand mention and an explicit recommendation separately. Then assess their commercial significance.
Published studies inform priorities. Budget decisions also need your starting position, a defined scope and explicit acceptance criteria.
Google continues to describe optimization for its generative search features as SEO. GEO and AEO are commonly used terms, not separate Google admission procedures. My service commitment is therefore to controlled execution, not control over what an answer system selects. [2]
02 — SEO, GEO and AI search: related work, different outcomes
I use the following working definitions in this guide. SEO concerns discovery and presentation in search engines. GEO — Generative Engine Optimization — focuses on representation within generated answers. AEO — Answer Engine Optimization — emphasises direct answers. The acronym matters less to me than a precise definition of the intended outcome.
The interfaces must also be distinguished. An AI Overview is an AI summary within Google Search. AI Mode is a search experience for deeper questions and comparisons. Both may use related subqueries. Seeing a summary in search does not establish how many people subsequently use a separate AI mode. [29]
RAG — Retrieval-Augmented Generation — describes a technical method: information is retrieved and used to support a generated answer. Google describes this approach for its search features too. However, building a company-owned RAG assistant is a different project from improving public visibility. It needs a defined knowledge base, access rules and its own quality controls. [2]
I would not bundle these projects into a single “AI package”. A company with weak service pages may first need better public information. A team that searches internal documents every day may have a knowledge-management problem. These are two different investment decisions.
03 — What the studies show, and what their numbers do not prove
A percentage becomes useful only when we know what was counted. Is it all searches, selected purchasing decisions, retrieved pages or answers containing citations? For each figure, I keep the measured outcome and the sample limitation together.
Seer examined 3,119 informational search terms across 42 clients from June 2024 to September 2025. For queries with AI Overviews, organic click-through rate (CTR) fell from 1.76% to 0.61%. This is not a forecast for every Swiss website or an isolated causal finding. [3]
AirOps analysed 548,534 pages retrieved by ChatGPT. The study began with 15,000 prompts; including follow-up searches, the dataset contained 43,233 queries. Only 15% of retrieved pages were cited. This measures selection after retrieval, not a general loss of customers. [5]
Semrush reports higher AI-search visit value based on conversion rates. The article particularly addresses digital-marketing and SEO topics. I do not translate this into a universal conversion rate or guaranteed revenue growth for other industries. [4]
In an Ahrefs experiment, 43% of answers citing a conference-promoting page did not mention the promoted conference. For the product pages studied, the figure was 11%. The 43% figure therefore does not apply to every cited promotional page. [9]
My conclusion is not that traffic has become worthless. Visitors remain relevant when they take meaningful action. What I reject is reporting that counts sessions without explaining search intent, enquiry quality, source use or the accuracy of how the offering is described.
Frequent citation does not prove strong methodology either. For decisions, I distinguish official platform rules, published observations, experimental findings and my own working hypotheses. Several articles discussing one measurement do not create several independent measurements.
04 — From question to answer: five points where visibility can be lost
Imagine a managing director asking: “Which SEO and GEO agency suits a Swiss B2B company with German and English website versions?” The following example uses the starting report’s five steps as an explanatory model, not a recorded AI test. Its platform details come from different original studies; they do not describe one identical architecture used everywhere. [1]
A system may research services, regional suitability, references, costs or language capabilities. Google documents related search queries of this kind. My planning response is to consider the information a purchasing decision needs, not just the phrase “GEO agency”. The specific subqueries in this example are hypothetical. [29]
Relevant search results may enter a working set of sources. Inclusion there is not publication in the answer. I therefore ask separately: can the page be found, was it actually retrieved in the observable process, and was it selected as a source? Intermediate steps that cannot be observed remain unknown.
The system needs relevant information, not necessarily the entire page. Dejan reports a median of 1,929 grounding words per query in an analysis of 7,060 queries involving Google grounding outputs. This is a measurement of the studied process, not an official limit for every Google product or AI system. [7]
Dejan’s analysis describes sentence-level selection and a preference for opening passages. My editorial response is to make important statements understandable without distant context. Scope, limitations and validity belong together, rather than placing an attractive number at the top and its decisive condition at the bottom. [8]
Now I assess the visible result: is GlasBox named, in what context, and which source supports the statement? An answer could use our guide while recommending a different agency. That is not a contradiction: information source and recommended provider are different roles.
An important clarification to the starting report: the frequently quoted AirOps figure of 58.4% concerns the first result returned in ChatGPT’s web-search results. The original study does not confirm the underlying search provider. I therefore do not describe it as “58.4% for Google position one”. [6]
My diagnostic rule is simple: without visibility into retrieval, I do not claim to know why a company is absent. “No mention observed” is a finding. “The model does not trust your brand” is an explanation, and needs additional evidence.
05 — Access, selection and prior: three different workstreams
The starting report distinguishes access, selection and prior — the model’s pre-existing tendencies. I use this distinction as a diagnostic aid, not as a readable ranking score. [1]
My workstream would review important URLs, status codes, crawling rules, indexability and the text actually delivered. The output is an evidenced list of obstacles, not yet a promise of more mentions.
Here I would improve offer explanations, comparison criteria and decision questions. Every change needs a rationale: which ambiguity are we removing, which statement becomes more reliable, and how will we recognise improvement?
My work here would strengthen the publicly documented identity and subject expertise. I can resolve contradictory information and publish substantiated contributions. I cannot promise when an external model will change its internal representations.
These workstreams partly run in parallel. A technical defect may be fixed quickly, but its observed effect still needs measurement. Equally, I would not prescribe a universal number of years for brand development. The useful horizon depends on the starting position, competition, platform and quality of the information actually published.
06 — Technical foundations: accessible, readable and consistent
The first review is of the most important public pages, not an AI plugin. Google’s technical requirements include an unblocked Googlebot, working pages returning HTTP status 200 and indexable content. Meeting those requirements does not guarantee indexing. [13]
Review a representative selection of homepage, service, comparison, contact and relevant language pages.
Document redirects, preferred URL versions and accidental noindex instructions.
Compare delivered HTML text with the visible browser experience.
Link important detail pages from relevant editorial context.
Record the URL, date, method, result and owner for each finding.
Search access and model training are separate decisions. OpenAI documents OAI-SearchBot for search, GPTBot for potential training use, and ChatGPT-User for certain user-initiated requests. Search can be allowed while training use is declined. robots.txt rules may be handled differently for ChatGPT-User. [11]
Google-Extended is a control token for certain Gemini-related training and grounding uses. Google states that it affects neither inclusion nor ranking in Google Search. I would therefore not simply “allow all AI bots”, but document a decision for each purpose. [12]
Where a block occurs also matters. A file on the web server and an upstream firewall are not the same control. A request with a changed user-agent alone does not prove that the genuine provider can access the site. I would examine authorised logs, published provider guidance and actual responses together, without bypassing protections or changing another organisation’s production system. [11][28]
JavaScript requires precise language. Google can render JavaScript. The Deep Research sessions examined by Peec instead used a text browser without equivalent page interaction. This does not mean every AI system can never click. An FAQ accordion is not automatically invisible either: the question is whether its text is already present in the delivered document or loaded only after an action. [14][15]
My requirement for new service pages would therefore be to include the essential answer as real text in the document. A graphic can support it, but should not be its only carrier. A video needs meaningful context. Navigation should remain useful without displacing the actual service explanation in the document. Human accessibility and robust machine processing should be planned together.
Structured data complements this work. Its claims must match the visible content, and correct markup does not guarantee a special search appearance. I would rather review a smaller set of accurate statements than add as many schema types as possible. [16]
Two common misunderstandings deserve attention. Google documents the removal of FAQ rich results from 7 May 2026. And while llms.txt is a proposal for machine-readable guidance, Google states that it provides no visibility benefit in Search. Useful FAQs remain useful; an extra file should be prioritised only when it serves a specific, testable purpose. [17][18][2]
Confidential information does not belong in a public file, even with a polite request for crawlers to ignore it. robots.txt is not access control. Internal documents require genuine access restrictions. [24]
07 — Write content that helps someone make a decision
I start content planning with a decision, not a word count. What does a prospective customer need to know before taking the next sensible step? For an agency, this might include responsibilities, audit scope, implementation, languages, dependencies or measurement. A page promising only “comprehensive solutions” does not answer these questions.
My editorial pattern is a direct answer followed by prerequisites, process, evidence and limitations. This is a recommendation for clear communication, not a claimed ranking law. A self-contained passage names its subject instead of relying on “this”, “here” or “as mentioned above” to point to distant context.
“Our innovative AI strategy takes your business to the next level.” This leaves the reader without an explanation of what is reviewed, what they receive or who implements a change.
“The agreed audit examines selected pages, their technical accessibility and answers to defined customer questions. You receive documented findings and prioritised changes. Implementation is included only when explicitly commissioned.” This is a writing template, not an additional GlasBox service commitment.
For every important service, I would collect five customer questions: who is it suitable for, when is it unsuitable, what is delivered, what determines the effort, and how is quality assessed? The answers belong on the relevant service page. A comprehensive guide can explain the broader context and link to those details.
My architecture rule is that a separate article should serve a separate information need. I would not publish near-identical pages for slightly different wording. Conversely, not every detail belongs in one enormous article. Overview and detail work together when their roles and internal links are clear.
08 — Information gain: what does the reader genuinely learn?
I use information gain here as a practical quality criterion: which important uncertainty does a page resolve that other available explanations leave unanswered? Google recommends original information, analysis and demonstrable expertise. I do not turn that guidance into an official scoring formula. [19]
On-Page.ai examined 150 pages from the first three results for 50 selected US queries. The vendor found a median of four unique numerical claims per page. This establishes neither that “four numbers are enough” nor that originality has no effect on rankings: the sample primarily examines already well-ranked pages and uses the vendor’s own assessment method. [20]
For a Swiss service business, I would first make existing expertise accessible. A documented decision process can be more useful than an inflated statistic. A dramatic number is unnecessary when explaining when a solution is unsuitable or why two similarly named offers contain different work.
Explain the object of the review, its sequence, responsibilities and acceptance criteria. A process description becomes a case reference only when it points to work actually completed and approved for publication.
Publish a measurement with its method, period, sample and limitations. Prefer aggregated or anonymised information to confidential customer data.
Explain alternatives, exclusion criteria and dependencies. A fair comparison names the properties assessed and acknowledges the limits of your own perspective.
My editorial boundary is clear: no invented client cases, fabricated expert quotations or decorative statistics. AI can help structure drafts. Subject responsibility and approval remain with people. Google’s spam policies address, among other practices, scaled content created to manipulate rankings without adding value; they do not categorically prohibit all AI-assisted writing. [27]
09 — Brand and external sources: do not manufacture credibility
An Ahrefs analysis of 75,000 brands found strong associations between external brand mentions, particularly on YouTube, and AI visibility. The study measures correlations. It does not prove that an additional video or a purchased mention causes a specified number of new recommendations. [21]
Seer also documents cases where ChatGPT included brand names in its follow-up searches. This is a useful observation point. It is not a direct view into model weights, nor proof of which individual earlier exposure caused a particular brand to be selected. [22]
My response is not “more mentions at any cost”. I would first clarify the publicly visible identity: a consistent name, clear services, accessible contact routes and identifiable responsibility. Relevant external contributions could follow, such as an original analysis that provides genuinely new information to an industry publication.
For GlasBox, this article makes the distinction explicit: Kacper Ruta is the author and GlasBox provides the agency perspective. The figures come from the cited studies, not from supposed GlasBox client results. This separation is part of the trust I want the article to establish.
A self-published list placing the publisher’s own company first is not independent evidence of quality. I prefer transparent criteria, disclosed interests and sources with identifiable responsibility. Success would mean an accurate, relevant representation of the actual offering, not merely one more mention.
10 — What Swiss businesses need to assess separately
A Polish market report cannot be localised simply by replacing the country name. The Swiss application below is my recommendation, not an effect measured in Switzerland by the starting report. Language, service territory and commercial conditions need to be assessed for the individual business.
Peec analysed 64.77 million Reddit citations. Within the studied Google AI Overviews subset, machine-translated Reddit pages accounted for 71.7% of Reddit citations in Poland and 52.2% in Germany. The denominator is Reddit citations, not all AI answers. I do not derive a Swiss percentage from this finding. [23]
My working hypothesis is that publishing in a local language alone is insufficient differentiation. A local provider should make the information for which local context genuinely matters particularly clear. This can include the actual service area, supported languages, binding service boundaries and applicable offer conditions.
Check DE-CH and English for equivalent meaning, not merely translated headings.
Keep every quoted price with its scope, currency and genuinely applicable conditions.
Do not suggest a local office or nationwide service coverage that does not exist.
Make contact routes and next steps clear in every language version.
Measure Swiss queries separately; do not present German, Polish or US data as your own market.
Technically, hreflang annotations can tell Google how actual language and regional versions relate to one another. The references must point to the correct pages and be reciprocal. An English page is not automatically intended for the United States, nor is a German-language page automatically intended for Germany. [33]
11 — Three shortcuts I would not trust
A short page can be incomplete; a long page can be precisely organised. I remove repetition, not necessary answers. My editorial test is whether readers can find and interpret relevant information without contradictions.
I update when facts, services, conditions or reliable findings change. I would not sell a new timestamp without a substantive reason as an improvement.
A screenshot records one answer under particular conditions. Assessment requires repeated observations, relevant questions and a consistent counting method.
Measurement also matters when assessing the relationship with rankings. An updated Ahrefs analysis distinguishes the first ten SERP elements from the first ten standard organic results. For the latter, it reports 37.1% overlap with AIO citations. I treat this neither as a universal citation probability nor as evidence that SEO is unimportant. [10]
An Ahrefs freshness analysis covering 17 million citations found differences between AI assistants. That does not establish a requirement to redate every page every three months regardless of its substance. My rule remains a factual review with a documented reason for each change. [30]
One further distinction matters: a benchmark is not a revenue report. The E-GEO research paper examines product-text optimization in a constructed evaluation environment containing more than 7,000 questions. Such experiments inform methods. They do not establish that a change in your own shop creates additional profitable orders. [31]
12 — Measuring GEO: a useful system, not a decorative score
I would build measurement in two layers. The first describes observed AI answers. The second describes real usage and qualified demand. Results from the first layer must not automatically be sold as results from the second.
The brand appears in the answer text. Record the context and spelling, keeping an incidental mention separate from an answer addressing a purchasing decision.
A specific URL is used as a source. Separate your own domain, an external source about the brand, and other sources. A brand name without a source link is not an owned-site citation.
Is the company explicitly presented as a suitable option? Are the stated properties correct? A frequent but incorrect recommendation is not a sound success.
Which traceable visits, suitable enquiries and completed sales result? Define quality from your own business rather than a generic GEO score.
A practical pilot is small enough for someone to read the answers. My example uses twelve real decision questions, two language versions, two platforms and three repetitions. This produces 144 answers, with 36 per language-platform combination. I would assess branded questions separately from questions without the brand name.
Illustrative calculation, not measured GlasBox data: in one group, nine of 36 answers mention the brand and six cite the company’s own website. That produces a 25% mention rate and an approximately 16.7% owned-source rate within this sample. It is neither market share nor a probability for every future user.
Save the question, question ID, language, platform, interface and available model information.
Document the date, time, assumed market and conversation context.
Retain answer text and sources, removing confidential information before sharing.
Count technically failed attempts separately rather than treating them as “brand absent”.
Define mentions, recommendations and incorrect statements before repeating the test.
I would use new conversations and avoid telling the system beforehand that the brand is the desired answer. OpenAI documents that search can use rewritten queries and available context. I therefore record test conditions rather than equate a personal chat with extensive history to a fresh search. [28]
Current platform update: Google’s newer documentation describes a separate “Generative AI performance report” in Search Console, citing 31 August 2026 for its broader rollout. It documents impressions from AI Overviews and AI Mode, including page and country breakdowns. These data are also part of the general Web report and must not be double-counted. Access and data availability for your own property require a separate check. [25]
Bing’s AI Performance provides additional citation information across supported Microsoft and partner surfaces, including cited pages and a sample of grounding queries. These data are neither complete coverage of all AI platforms nor a ranking of each source’s importance within every answer. [26]
For management reporting, I would connect these observations with website analytics and the enquiry process. An optional question about how someone found the company can supplement uncertain technical attribution. I would not count every direct visit following a GEO change as an AI success. Attribution assigns credit; incremental business impact requires further assessment.
My reporting template would keep the same fields for every language and platform: usable answers, mentions, owned citations, accurate recommendations, factual errors and observed qualified enquiries. Alongside them, record changes to the offering, website or measurement method. That context makes comparison meaningful.
13 — My 90-day plan: understand first, then change with control
The following is my planning proposal, not a guaranteed service scope or time to impact. The actual scope needs to fit the website, access, subject expertise and budget. A small business does not automatically need the same programme as an international shop.
Define the business objective, priority services and customer questions. Document a representative page sample and initial AI answers. Marketing owns the questions, subject owners verify the offering and technical staff assess accessibility. Acceptance: a reproducible baseline and separate lists of confirmed problems and unresolved hypotheses.
Improve a small set of priority pages. Address access obstacles and contradictions first. Every change gets an owner, a before-state record, a test criterion and a rollback route. Acceptance: the agreed information is delivered correctly, while contact routes and other important functions remain intact.
Add missing explanations, fair comparisons and approved original findings. Review language versions together. Pursue relevant external publication only with genuine subject substance. Acceptance: accountable, supportable content, without invented references or purchased credibility signals.
Repeat the question panel under documented conditions. Compare factual errors, source use and suitable demand against the baseline. Acceptance: a decision report explaining observations, uncertainty and the next priority, rather than a collection of positive screenshots.
For a technical or editorial handover, I use five questions: what is the finding, what evidence supports it, which business mechanism is plausible, who should change what, and how will we demonstrate the correction? This turns “We need more GEO” into a testable work item.
A hypothetical example: a service page does not define whether implementation is included. The saved page version would be the evidence. The possible business effect is unsuitable expectations in enquiries, not a calculated revenue shortfall. The fix would be an approved paragraph defining scope and exclusions. Verification would combine a checked publication with subsequent enquiry review. This is not a claimed audit finding about GlasBox.
My stop rule is to reassess a programme when its questions do not reflect genuine demand, the data cannot be collected reliably or necessary subject approvals are missing. More text does not solve these problems. I would reduce the scope or prioritise another action with a clearer business benefit.
14 — Frequently asked business questions
My answer would be: not without a diagnosis. The important information can often be improved within existing pages. A rebuild is a separate investment that should be justified by specific technical or business requirements.
I would create a separate page when it serves a distinct information need and fits the website’s structure. Small wording differences alone are not sufficient reason to publish another page.
For a manageable pilot, I would first document questions, answers and sources carefully. Additional software becomes useful when it reliably reduces recurring work and its measurement method is understandable. A tool does not replace subject judgement.
As many as the task needs, with as little repetition as possible. I would assess whether it fully supports the intended decision rather than use a fixed word count as evidence of quality.
In my working model, yes: for structure, language variants and critical review. Not as a substitute for sources, actual experience or approval. Confidential customer information would not be sent to an external service without an appropriate review.
I would not promise a universal deadline. Deliverables, responsibilities and measurement checkpoints can be agreed. Whether observed answers and qualified enquiries subsequently increase is an outcome to assess.
I would assess it as a separate use case. An internal or customer-facing assistant needs approved knowledge, appropriate access and a defined operational purpose. It does not replace a clear public website.
In my view, when internal ownership, implementation capacity or measurement discipline is missing. The assignment should close that gap. A large report without an implementation owner is not an outcome for which I would recommend an ongoing budget.
15 — My conclusion: make better-supported claims, not more claims
I do not see SEO and GEO as a contest to make the loudest AI promise. The useful task is more specific: identify relevant questions, explain the actual offering clearly, make accurate information accessible and connect observed representation with the real business.
My sequence at GlasBox is therefore baseline first, prioritised change second and verification afterwards. A citation is interesting. An accurate recommendation is better. A suitable enquiry matters commercially. I would never collapse these stages into one success story when the evidence is missing.
As an initial step, GlasBox offers a free AI visibility short check covering one domain, three purchase-related questions and three initial observations. This is an orientation exercise, not a full crawl or a 90-day plan. Any deeper work is defined around the scope actually required. [32]
GlasBox scope and starting options
Kacper Ruta · GlasBox — SEO & GEO Agency · glasbox.ch
Sources, interpretation and limitations
The Double Digital report is the starting point, not a dataset collected by GlasBox. This article is not an independent meta-analysis of all 75 materials. The original sources below support the identified factual claims; my examples, priorities and workflows are separate editorial recommendations. Platform documentation is considered as of 6 September 2026 and may subsequently change.
Evidence status: VERIFIED OBSERVATION refers here to the documented content of a source, not a study replicated by GlasBox. Confidence is high for source reporting and limited by the original study design. WORKING HYPOTHESIS describes a possible application to a Swiss business; confidence is medium and requires project-level testing. Effects, rankings, enquiries and revenue from a specific GlasBox client project are NOT VERIFIED: no such outcomes were measured for this article.
Starting report and research overview; not a GlasBox meta-analysis.
SEO, generative search, RAG and limits of special AI optimization.
Observed click-through rate; note the period and client sample.
Value and conversion comparison within the studied topics.
Source retrieval and citation are different events.
Position within returned search results; provider not confirmed.
Measured grounding volumes, not a universal word limit.
Analysis of passage and sentence selection.
Citation without a brand mention; distinguish page-type results.
Standard organic top ten and the first ten SERP elements are different.
OAI-SearchBot, GPTBot and ChatGPT-User.
Purpose and limitations of Google-Extended.
Technical search requirements and indexing limitations.
Google’s processing of JavaScript.
Observed Deep Research sessions, not every AI system.
Consistency with visible content and no guaranteed appearance.
8 May 2026 entry: FAQ rich results removed from 7 May.
Proposal for structured guidance, not a ranking standard.
Originality, demonstrable expertise and usefulness.
150 pages and a proprietary assessment method; limited generalisability.
Brand visibility and external mentions; correlation, not causation.
Documented brand names within follow-up searches.
Reddit subset; not a Swiss whole-market rate.
Crawling rules are not access control.
Newer reporting documentation; rollout note dated 31 August 2026.
Citation data across supported Microsoft and partner surfaces.
Scaled content for ranking manipulation without additional value.
Search context, sources and guidance for website owners.
AI Overviews, AI Mode and technical requirements; see [25] for reporting.
Platform differences; no universal update interval.
Research benchmark, not an observed GlasBox revenue effect.
Public entry options and the short check’s limited scope.
Language and regional versions, including hreflang.
How do AI answers represent your business?
Start with your domain and three purchase-related questions. The free short check provides three initial observations, not a complete audit or a ranking promise.
Request the free short check