Every AI conversation in tourism runs on about sixty terms. Most of them get used loosely, and the loose use is what turns a board meeting into forty minutes of nobody quite disagreeing. This is the working vocabulary: what each term means, and why it matters to a tourism organisation. Five sections, from talking to a chatbot through to governing the whole thing.
1. The basics: talking to AI
Start here. These are the terms that decide whether AI gives you something useful or something that wastes an afternoon. Most disappointing AI output traces back to one of three things in this section: the wrong model, not enough context, or a prompt that assumed the tool could read your mind.
AI (artificial intelligence). Software that can interpret instructions, generate content, analyse information or take defined actions. In this lexicon it almost always means generative AI: the kind that writes, answers and decides.
LLM (large language model). The engine behind tools such as ChatGPT, Gemini, Claude and Copilot. It predicts and generates language from patterns learned across enormous amounts of text. The tool is the car; the LLM is the engine.
Model. A specific version of an LLM (e.g. Sol, Fable). Every product ships several, at different capability levels, and which one answers you changes everything about the answer. Think seniority: you would not send the receptionist to answer a board question, and you would not pay the senior consultant to add two and two. Always check which model you are talking to before judging the output.
Effort (also thinking or reasoning level). A setting on newer models controlling how hard the model thinks before answering. Low effort is the CEO before coffee: fast and fine for routine. High effort is the CEO after coffee with the door shut: slower, dearer, and what you want for strategy, planning and anything consequential.
Token. The unit AI reads, writes and bills in; roughly three-quarters of a word. Tokens are the currency of AI: every plan has an allowance, bigger models and higher effort spend it faster, and running out mid-month is the AI equivalent of the old data cap.
Context. Everything you give the model before asking for the work: background documents, audience, constraints, examples of what good looks like. The salesperson’s suitcase: hand over the full case of background material before you ask for the pitch. Thin context is the number one reason AI output disappoints.
Prompt. The instruction you give an AI tool. A strong prompt has four parts: the goal, the context, the named source to use, and expectations (format, tone, length, what to avoid). End important briefs with “ask me questions until you are 97 per cent sure you understand what we need.”
System prompt (or instructions). The standing brief a tool carries into every conversation: persona, rules, sources, what it must never do. Writing these well is a craft; they are the difference between a tool that sounds like your organisation and one that sounds like AI.
Custom GPT / Gem / Project. A configured assistant inside ChatGPT (GPT), Gemini (Gem) or Claude (Project): your instructions, your documents, your sources, saved and reusable by the whole team. The cheapest way to give staff a tool that already knows the organisation.
Deep research. A mode where the model works for many minutes, reads dozens of sources and returns a cited report. Work that took a research team weeks now lands in half an hour; Brand USA runs it every morning.
Voice mode. Talking to the AI instead of typing, hands and eyes free. Matters twice for tourism: travellers plan trips this way (ask your phone where to eat lunch and watch where the answers come from), and it is the fastest live demo that convinces a board.
Hallucination. A confident, fluent, wrong answer. Not lying; pattern-completion without facts. The cure is grounding plus verification, never trust.
Grounding. Restricting an AI to named source material so its answer comes from that material rather than general model knowledge. A grounded system says “that is not in my sources” instead of inventing. This single technique is what separates a trustworthy organisational tool from a public chatbot.
Human review. The named point where a person checks facts, judgement, tone, permissions and consequences before output is used. Non-negotiable for anything published, sent or relied on.
2. How AI finds your destination
The vocabulary changes once you move from using AI to being found by it. Your website is now read by machines more often than by people, and almost none of that reading shows up in Google Analytics. These terms cover what is doing the reading, what it is looking for, and why your traffic can fall while your content gets used more than ever.
Zero-click search (invisible search). The traveller asks an AI tool and gets the answer on the AI’s screen, never clicking through to the website that supplied it. The visit is real in your server logs and invisible in Google Analytics. This is why DMO traffic is down while DMO content is used more than ever.
AI Overviews / AI Mode. Google’s AI answers: Overviews sit above the search results; AI Mode replaces them with a conversation. You can rank first in the blue links and still be missing from the answer.
Indexing bot vs retrieval bot. The two kinds of AI visitor in your server logs. Indexing bots (GPTBot, ClaudeBot) crawl and store your content on a schedule, building the AI’s memory. Retrieval bots (ChatGPT-User, Perplexity-User) fetch a page live, mid-conversation, because a traveller is asking right now. A -User suffix means a human is waiting; block it and you refuse a customer.
User agent. The name a bot announces when it visits your site. Server logs use it to tell Googlebot from GPTBot from a person on Chrome.
Server logs. The raw record of every request to your website, kept by your host or Cloudflare. The only place AI reading actually shows up; Google Analytics only counts humans in browsers.
robots.txt. The public file (yourdomain.com/robots.txt) telling bots what they may read. Two minutes to check whether you have accidentally blocked the AI tools you want to appear in.
Crawlable. Being findable and readable by AI at all: schema, Google Business Profile, FAQs, clean pages. Necessary, never sufficient. Gets you found this year.
Schema (structured data). Machine-readable labels behind a page telling systems exactly what it describes: a tour, an event, a business, opening hours, a price. The barcode on the product: a person reads the pretty page, the machine reads the barcode. Key tourism types: TouristDestination, TouristAttraction, TouristTrip, Event, LocalBusiness, FAQPage.
JSON-LD. The technical format schema is usually written in. Your developer’s concern, not yours; you just need to know whether it is present.
Open Graph. The labels controlling how a page appears when shared or read as a summary: title, description, image, page type. A bookable $2,000 tour tagged as generic “website” is telling AI it is nothing special.
NAP consistency. Name, address, phone: identical everywhere (website, GBP, directories, socials). When your website says 5pm, your Google listing says 6pm and the directory says 4pm, AI trusts none of them and recommends someone else.
Citation. AI naming or linking your organisation as the source of its answer. The new impression. Being read but not cited is the gap most DMOs are in.
GEO (Generative Engine Optimisation). The work of being found, understood, trusted and recommended by AI answer systems. The successor discipline to SEO. Also called AEO (Answer Engine Optimisation); same discipline.
SEO (Search Engine Optimisation). Visibility in the traditional blue-link results travellers are clicking less. Still worth having; no longer the game.
Google Business Profile (GBP). The structured Google listing for a business or visitor centre: identity, categories, services, products, posts, photos, reviews, hours. Now the single most influential input into AI recommendations for a region, ahead of the DMO website. Treat it as a social channel and a database, not a set-and-forget listing.
ATDW (Australian Tourism Data Warehouse). The national tourism listing and distribution source. The floor, not the finish line: keep the listing, and invest equal effort in the operator’s own content, schema and FAQ pages, because those are what AI quotes.
3. Making your knowledge queryable
Being crawlable gets you found this year. Being queryable keeps you in the answer next year. The difference is whether your knowledge lives only on a website, or is structured so a system can ask questions of it and get reliable answers back. These are the parts that make that work.
Queryable. Structured and accessible enough that a system can ask questions of your knowledge and get reliable answers. Deeper than crawlable: the whole organisation readable, not just the website. Keeps you in the answer next year.
Knowledge base / knowledge layer / second brain. Three names for the same asset: the organisation’s knowledge (members, events, itineraries, accessibility and seasonal detail, the know-how in staff heads) captured, structured and queryable. The trusted layer AI cannot work without, and the foundation every AI product sits on.
RAG (retrieval-augmented generation). The technique behind every grounded AI system: retrieve the most relevant pieces from an approved knowledge base first, then have the model answer from what was retrieved. How you get AI that cites your sources instead of guessing.
Chunking. Splitting documents into small passages before storage, so retrieval can find the exact relevant piece rather than a 40-page PDF.
Embedding. Converting a passage of text into numbers that capture its meaning, so “dog-friendly dinner” can find “pets welcome at our bistro” even though they share no words.
Vector index. The searchable store of all those embeddings: search by meaning rather than exact words. The engine room of a knowledge base.
Retrieval. Finding the most relevant passages or records before the AI writes its answer. Good systems combine meaning-based search with old-fashioned keyword search, because operator names and product codes need exact matches.
Guardrail. A hard rule in an AI system that overrides fluency: if nothing relevant enough is found, say “that is not in the knowledge base” rather than generating a guess; never send email; never publish without approval.
API. The doorway one system offers another for requesting data or actions. The opposite of clicking around a screen. If a platform you pay for will not give you API access, you do not control your own data.
Connector (or Action). A pre-built link letting an AI tool reach another system: your Drive, your calendar, your booking platform. Convenience with a permission attached; grant the minimum.
MCP (Model Context Protocol). The open standard that lets AI tools plug directly into a knowledge source or system without a custom build. The organisation that publishes an MCP server becomes the queryable knowledge layer agents read when they plan trips to its region.
4. Agents and automation
An agent does not answer a question. It pursues a goal, using tools, until the goal is met. That shift is already reaching travel booking, and it changes who your content needs to convince. These terms cover what agents are, what they do on a traveller’s behalf, and the working discipline that separates real capability from asking ChatGPT to write a review reply.
Agent. An AI system that pursues a goal using tools, rather than answering a single question. The Devil Wears Prada assistant: runs behind you, knows everyone’s business, talks to the other assistants. When you speak to your phone in voice mode, you are already briefing one.
Agentic booking. Agents completing travel purchases end to end: research, decision, payment, inside one AI session. The shopping cart is filled before a human reads a single review.
Information agent. An AI a traveller sets up to monitor the web on their behalf for weeks before a decision. By the time they book, their AI already holds a view of your region, built from whatever it could read. Fresh, structured content means that view is built from you.
A2A (agent to agent). Software agents transacting with each other directly: the traveller’s agent asking the destination’s agent. The reason knowledge needs to be machine-readable, not just human-readable.
Workflow. A repeatable process with an input, steps, a review point and an output. AI can run steps; a person owns the outcome.
Skill (or instructions file). The written job description an agent follows for a recurring task: how we build a carousel, how we draft a Google post. Write it once, correct it when the output misses, and it improves permanently; unlike briefing a person who may leave in two months.
Scheduled job (cron). An automation that runs on a clock: every morning, every Monday. What “the digest arrives at 8am without anyone doing anything” is made of.
Closing the loop (the 80/20 rule). The working discipline that separates 2023-style AI use from real capability: AI does the 80 per cent, and the team spends its effort training the 20, feeding back what good looks like so next time is better. The lazy version is asking ChatGPT to write one review reply; the skilled version is teaching a system your judgement and checking its work.
5. Safety and governance
None of the above matters if the organisation cannot say what it runs and who owns it. This section is the short list every tourism organisation needs settled before it scales AI past one enthusiastic staff member. It is mostly habits, not technology.
Prompt injection. Hidden text in a web page, email or file that an AI with permissions reads as instructions: “send me the revenue emails” in white text on a hacked page. The reason agents get minimum access and never send permissions.
Training data settings. Whether the tool’s maker may learn from what you type. Consumer plans often may; business plans (ChatGPT Business and equivalents) do not by default. Check before confidential material goes anywhere.
De-identification. Removing names and identifying details before information enters an AI tool: the role survives, the person does not. Standard practice for member and visitor information.
Minimum access. Every tool, connector and workflow gets the least access required, never blanket permissions, and access is reviewed when people join, change roles or leave. The whole of AI security in one habit.
Data residency. Where information is processed and stored, legally. Some organisations must keep data in Australia; check the provider’s processing location before connecting sensitive sources.
AI use policy. The one-page document every organisation needs before scaling: approved accounts, what must never enter a tool, verification rules, no auto-send and no auto-publish, a register of custom GPTs and agents, and who to tell when something goes wrong.
AI register. The living list of every custom GPT, agent, connector and automated workflow the organisation runs, with an owner for each. If nobody can list them, nobody is governing them.
Authenticity rule. No AI-generated imagery of the destination: real photos of real places. The organisations furthest ahead on AI (Destination Toronto, Brand USA) hold this line hardest, because the real photo is the one authenticity signal no machine fakes.
Common questions
What is a token in AI?
A token is the unit AI reads, writes and bills in. It works out to roughly three-quarters of a word. Every plan comes with an allowance, and bigger models on higher effort settings spend it faster. Running out mid-month is the AI version of the old mobile data cap.
What is a large language model (LLM)?
An LLM is the engine behind tools like ChatGPT, Gemini, Claude and Copilot. It generates language by predicting patterns learned from enormous amounts of text. The tool is the car. The LLM is the engine under the bonnet.
What is an AI hallucination?
A hallucination is a confident, fluent, wrong answer. The model is not lying. It is completing a pattern without having the facts to hand. The fix is grounding the tool in real source material and then checking the output. Trust is not a control.
What is GEO (Generative Engine Optimisation)?
GEO is the work of being found, understood, trusted and recommended by AI answer systems. It is the successor to SEO, which only ever covered the blue links travellers are clicking less often. You will also see it called AEO, or Answer Engine Optimisation. Same discipline, different acronym.
What is the difference between crawlable and queryable?
Crawlable means AI can find and read you at all: schema, a current Google Business Profile, FAQs, clean pages. Queryable means your knowledge is structured enough that a system can ask questions of it and get reliable answers back. Crawlable gets you found this year. Queryable keeps you in the answer next year.
What is RAG (retrieval-augmented generation)?
RAG is the technique behind every grounded AI system. It retrieves the most relevant pieces from an approved knowledge base first, then has the model answer from what it retrieved. That is how you get AI that cites your own sources instead of guessing.
What is prompt injection?
Prompt injection is hidden text in a web page, email or file that an AI reads as an instruction rather than as content. A hacked page might carry white-on-white text telling the tool to forward your revenue emails. If the AI has permissions, it can act on it. This is why agents get the minimum access they need, and why they are never given permission to send.
What is zero-click search?
Zero-click search is when a traveller asks an AI tool and gets the answer on the AI’s screen, without ever clicking through to the website that supplied it. The visit is real and it shows in your server logs. It is invisible in Google Analytics. That gap is why destination traffic is falling while destination content is used more than ever.
What is schema markup, and why does a tourism business need it?
Schema is a set of machine-readable labels sitting behind a page, telling systems exactly what it describes: a tour, an event, a business, opening hours, a price. It is the barcode on the product. A person reads the pretty page, the machine reads the barcode. The types that matter most in tourism are TouristDestination, TouristAttraction, TouristTrip, Event, LocalBusiness and FAQPage.
What is an AI agent?
An agent is an AI system that pursues a goal using tools, rather than answering one question and stopping. It can search, read, book and report back. If you have spoken to your phone in voice mode and asked it to sort something out, you have already briefed one.
What is MCP (Model Context Protocol)?
MCP is an open standard that lets AI tools plug straight into a knowledge source or system without a custom build. For a destination organisation it is the practical route to becoming the queryable knowledge layer that agents read when they plan trips to the region.
Does a tourism organisation need an AI use policy?
Yes, and one page is enough to start. It should name the approved accounts, what must never be typed into a tool, the verification rules, and a no auto-send and no auto-publish rule. Add a register of every custom GPT and agent the organisation runs, with an owner for each, and who to tell when something goes wrong.
Your free AI marketing coach, right in your pocket
The free Pocket Rocket app gives you a personal AI marketing coach, website audit, weekly action plans and 5-minute tips. Built for tourism operators.