Prompt and target
The human set the problem, audience and ambition: intensive research, a German Herrnhut focus, bilingual content, ALT text, real visitor conditions and a demonstrator that stakeholders could explore.
Inside the demonstrator
Why this site looks the way it does
This website is designed not only to tell an attractive story, but to organise knowledge so that people and machines can understand Herrnhut as unambiguously as possible. This guide explains the idea in plain language.
The model's five layers
They work together and describe the same places from different perspectives.
Visible explanationunambiguous entitiesrelationships & geographystructured datasources & maintenance
The visitor story layer
Storytelling and the knowledge graph belong together
A person remembers the fitting door. A machine additionally needs to know which places it connects and how strongly the claim is evidenced.
The page “Seven stories connecting Herrnhut with the world” therefore combines memorable people, spaces and objects with place, time, entity relationship, source and evidence note. Stories have stable anchors, matching German and English versions and a dedicated story dataset.
The door demonstrates the method: the familiar Bethlehem association remains understandable, while a fact-check explains that the official source more broadly names Moravian buildings in the USA or Gracehill. Narrative strength and factual reliability support one another.
The idea in one sentence
A shared truth layer
GEO - Generative Engine Optimization - creates better conditions for AI systems to identify a place correctly, connect it with reliable sources and consider it in a suitable answer or travel recommendation.
It does not guarantee a mention by ChatGPT, Gemini, Copilot or any other system. It is quality work on the information from which such systems learn, search or assemble answers.
How the site was made
From a human prompt to a tested GEO model project
It began with a detailed brief from the project initiator to Codex: research the German World Heritage component at Herrnhut far beyond the obvious results, examine its digital representation and build an idealised GEO website from scratch.
Starting pointA human briefwith a purpose, stakeholder perspective and quality ambition
Research base115 sources & data servicesidentified, assessed and classified by source role
ImplementationVibe coding with Codexfrom the data model to site, planner and editor
Result18 pages · DE & ENplus JSON-LD, story data, entity register, GeoJSON and documentation
Research and source critique
Codex reviewed the project knowledge base, researched official and tourism sources, and distinguished primary evidence, operator information, contextual sources and changeable visitor data.
Knowledge model before layout
Entities, identifiers, part-whole relationships, spatial sectors, operator roles, sources and freshness rules were modelled before designing the pages.
Vibe coding and testing
Natural-language instructions became HTML, CSS, JavaScript, PHP and data files. Browser, mobile, link, JSON and content checks then drove further revisions.
Collaboration, not one-click magic
The division of labour between human and AI
Human direction
The project initiator set purpose and limits, contributed the stakeholder perspective, made taste and priority decisions, tested deployment and supplied concrete feedback - including the comment that the criteria box was hard to understand.
AI contribution
Codex provided research breadth, comparison and synthesis, developed the entity and data model, drafted the content, generated working code, documented decisions and ran reproducible technical checks.
Shared iteration
The result did not emerge from one uncontrolled prompt. Drafts were inspected, questioned, tested on the server and improved. Feedback flowed back into content, design, data and project memory.
What “vibe coding” means here
A human describes in everyday language what an application should do and how it should feel; the AI translates that intent into working code. This project went beyond rapid prototyping: research, source critique, a data model, visual checks, validators and human approval gave the process its professional frame.
Why deliberately idealised?
The demonstrator was designed to show as many relevant GEO criteria as possible in one coherent model: clear answers, entities, relationships, geography, multilingual content, structured data, sources, image context, planning conditions, freshness and governance. A real operator website need not copy the design or every file; it can adopt the components that match its role and connect them to a shared knowledge network.
The site is therefore a reference architecture and learning object - not an official World Heritage website and not proof of a particular AI ranking. Factual approval, real photography, operator maintenance and institutional responsibility remain human tasks.
What to notice as you click
Six visible traces of the GEO logic
The demonstrator deliberately makes its method visible. These details help people and machines identify, relate and plan.
Direct answers
Headings and answer boxes address real questions: What is part of the property? What is merely nearby? What must be checked before visiting?
Clear roles
The site distinguishes the town, the World Heritage component, the Moravian Church, individual places, operators and related visitor services.
Explicit relationships
“Component of”, “contained in”, “operated by” and “thematically related” communicate more than a vague association with Herrnhut.
Planning conditions
Indoor or outdoor, slopes, duration, weather, mobility and seasonal access make places usable in context-sensitive itineraries.
Sources and review date
Stable World Heritage facts point to primary sources. Opening and access information should be checked with the responsible operator.
Images with context
ALT text describes the motif. Current images are clearly marked as AI placeholders so they cannot be mistaken for documentary photographs.
Officially precise, plainly explained
Why are there strange-looking codes?
Codes remain where they make an entity unambiguous. Their plain-language meaning appears alongside them.
1468bis
The serial property
This is the official identifier used for the whole Moravian Church Settlements property. “bis” is part of the identifier, not a value judgement.
1468bis-002
The Herrnhut component
The suffix identifies the German component within the serial property - not the whole municipality and not all four settlements together.
(iii) & (iv)
Cultural testimony and settlement model
In simple terms, criterion (iii) recognises exceptional testimony to Moravian culture; criterion (iv) an outstanding type of planned Protestant settlement.
45.56 ha + 185.08 ha
Property and buffer zone
The first number is the property area and the second its buffer zone. A representative coordinate is not an official boundary.
3 sectors
One spatially distributed component
The historic centre, God's Acre/Hutberg and Berthelsdorf are connected parts of the Herrnhut component, not an arbitrary attraction list.
reviewed …
Knowledge needs maintenance
A review date records editorial checking. Timetables, openings and access still require current operator information.
The design principle:
The official code creates precision in the data space. Its plain-language translation keeps people from being stranded in internal terminology.
Entities and relationships
AI needs more than place names
It is not enough for “Herrnhut” to occur in a text. A system needs to know which thing is meant and how it relates to other things.
serial World Heritage property · ID 1468bisMoravian Church Settlements
contains
German component · ID 1468bis-002Herrnhut
connects
Historic settlement centre
God's Acre & Hutberg
Berthelsdorf
not identical with · but related to
One body of knowledge, several representations
Five layers, one content
Core statements are deliberately repeated across aligned layers. These layers should reinforce one another, not contradict one another.
01Visible textanswers, advice, descriptions
02Semantic HTMLheadings, lists, navigation
03JSON-LDtypes, IDs, relationships
04Entity register & GeoJSONshared reference data
05Primary sourcesevidence and correction
Inspect the machine layer
The demonstrator makes central files public. They are structured views of the same places and relationships, not a separate story.
From data to a travel answer
Why weather, time and mobility belong here
Generative travel advice is often expected to say not only what exists, but what makes sense under particular conditions.
A place therefore needs more than a name and coordinate. Indoor or outdoor setting, slope, duration, access, interests and operator responsibility support a reasoned route.
2 daysrainfew slopeshistory
A different route from two sunny hoursIndoor places receive more weight. Hutberg is not silently added despite the mobility constraint. Opening and access still need a live check.
Planning relationships with evidence status
Accessibility, parking and paths are knowledge-graph data
They determine whether an AI can responsibly recommend a route for a particular person.
Accessibility as features
Step-free entrance, lift, toilet, door width, surface, gradient and sensory/cognitive support must be kept separate. “Accessible” is not a universal yes/no field.
Parking as an entity
A parking facility has its own coordinate, operator, capacity, fees, opening and designated spaces. Nearby does not mean owned or currently available.
Path as a directed edge
A connection has a source, destination, mode, distance, duration, surface, gradient, steps, source and review status. Return travel may differ.
01Operator statementfor a clearly named scope
02Map signalexistence, not a promise
03Route modeldistance/duration, not accessibility
04Field checkfeatures, date, responsibility
05Maintenancechange and live status
Herrnhut example
The exhibition fit-out is described as accessible, but that does not establish the complete building route. A car park is mapped near Zinzendorf Square, but that does not establish capacity or free spaces. The path to God's Acre is modelled, but that does not establish surface, segment gradient or wheelchair suitability.
The potential network effect
The network effect with more organisations
One good site is a start. The benefit grows when municipalities, attractions, hosts, restaurants, mobility providers and regional portals describe and connect the same entities consistently.
Organisations maintain data
Official name, type, location, operator, access, booking, review date and reliable sources.
Entities are connected
Shared IDs and precise relationships turn isolated pages into a regionally connectable knowledge network.
Regional data becomes more useful
Portals, maps, entity registers and Open Data structures can receive more consistent signals.
Answers may fit better
AI systems gain better conditions for correct identification, source use and contextual recommendations.
This is a quality and interoperability proposition, not a ranking promise.
Structured data, an entity register or llms.txt cannot guarantee inclusion in an AI answer. They can reduce avoidable ambiguity and improve machine usability.
A mini guide for operators
Nine steps for your own website
Not every website needs a large knowledge graph. It should publish its most important entities clearly, currently and in a connectable form.
- Name: official name and clear short description.
- Type: museum, hotel, restaurant, route, event or another precise role.
- Locate: full address and verified coordinate.
- Relate: part of, operated by, near, booked through or thematically related.
- State conditions: opening, indoor/outdoor, duration, individual accessibility features and booking.
- Connect infrastructure: parking, stops and routes as entities or directed edges with evidence status.
- Evidence: primary sources, responsible contact and last review date.
- Structure: suitable schema.org/JSON-LD types, stable URLs and consistent IDs.
- Synchronise: visible text, structured data and regional feeds must not disagree.
This guide follows the principles it describes.
It has a stable URL, German/English language relationship, semantic headings, visible statements, structured TechArticle markup, sources, a review date and internal links to the data. Reviewed 14 July 2026.