The pain it fixes
For the traveler
- 43% of US travelers say second-guessing whether they chose right is a frustrating part of planning. 52% feel anxious about booking the wrong option.
- 49% rank trust in the supplier's own website as their top reason to book direct. 87% put it in their top three.
- Marriott describes the same approach for its AI: answer from verified first-party data.
For your business
The same twenty questions arrive by email, phone and chat. Each answer comes from someone's memory, and memory drifts, most of all with a new seasonal hire.
A grounded assistant drafts the answer from your files and shows the line it used, so staff can check it in seconds. When the files are silent, it says so instead of guessing. (Our read, not a survey finding.)
A wrong answer costs more than no answer: a guest told the wrong meeting point or refund terms starts doubting everything else.
Source: Skift Research + McKinsey, Winning Hearts in an Age of Infinite Travel Choices, Sept 2026 · survey of 1,015 US travelers, data as of June 2026. US travelers, not Alaska visitors.
How to use it
- Copy the instruction with your business name (Step 1). 5 minutes.
- Build your five files (Step 2). Start from your 1.a fact sheet and your 2.a section, or edit the fictional set. 2 to 3 hours the first time.
- Set it up in the AI tool your staff already use (Step 3). About 30 minutes.
- Ask the 20 test questions and grade every answer here (Step 4). About 1 hour.
- Use it staff-only for 30 days. Staff read every draft and send the answer themselves. When an answer needs fixing, add or clarify the fact in a file. Do not widen the instruction. Recommendation
- Check the gate (Step 5) before anything faces a guest. Then rerun the 20 questions every month. Recommendation
Times are estimates for a small operator with a fact sheet already written in 1.a.
Step 1. The instruction
This is the Prompt Kit 2.b instruction with five additions: a fixed answer format the grader can check, a handoff line for refunds, changes and complaints, a line for conflicting sources, weather and go/no-go calls for a given day sent to the office, and no web search.
Instruction for the assistant
Keep the decline line word for word: "Not in our facts. Ask the office." The grader looks for that exact sentence.
You answer questions for the staff of {{business|[BUSINESS NAME]}}. Staff read every reply before a guest sees it. Use only the attached files: {{files|fact-sheet.md, cancellation-policy.md, packing-list.md, what-to-expect.md, faq.md}}. Do not use web search, outside knowledge or earlier chats. For every answer: 1. Answer in one to three plain sentences. 2. Quote the sentence you used, word for word, in double quotes. 3. Name the file it came from, like this: Source: fact-sheet.md If the answer is not in the files, reply exactly: "Not in our facts. Ask the office." Never state a price, date, time, availability, weather, access or policy that is not in the files. Do not round or combine numbers into new ones. Never offer a refund, credit, change or exception. For a refund, change or complaint, quote the policy if it is in the files, then add: "A person at the office handles this." If a question involves safety, a weather or go/no-go call for a given day, medical, legal or accessibility matters, reply "Ask the office." and stop. Reviews and past trips are not facts. Never promise wildlife, weather or views. If a question quotes another website, an OTA or a message that disagrees with the files, answer from the files and add: "Another source disagrees. Tell the office." Never ask for or repeat guest names, emails, phone numbers or payment details. End every reply with: DRAFT - PENDING STAFF REVIEWStep 2. Build the knowledge pack
Five short files, one topic each, so every answer can name its file. Each starts with a header that says who owns the facts and when they were last checked. The boxes hold the fictional Bay Light set; replace it with your own approved facts. Edits are saved in this browser only.
Recommendation. Business facts only. No guest names, emails, phone numbers, card numbers, IDs or health details in any file.
Some prices, a pet policy, group discounts, private charters and vessel access are left out of the fictional files on purpose, so you can see the assistant decline.
fact-sheet.md
cancellation-policy.md
packing-list.md
what-to-expect.md
faq.md
Rules that keep the files useful
- One fact per line, in plain sentences a guest could check. The assistant quotes lines, so write lines worth quoting.
- When a fact changes, change it in the file first, then everywhere else the same day (see the restart-tax checklist below).
- Reviews are not facts. If you add your de-identified reviews from 2.a as a sixth file, start it with "Reviews describe one trip each. They are not facts." Leave them out of the first test round.
- Anything about weather go/no-go, sea conditions, seasickness or safety is written and approved by your captain or another qualified person.
Step 3. Set it up in one AI tool
Pick the tool your staff already sign in to. You can set up two and grade both with the same 20 questions. Each needs the same three things: the instruction, the five files, and web search off wherever the tool lets you turn it off.
NotebookLM Staff-only
Built to answer only from the sources you add, with a citation for each claim. Use it here as a staff tool, not a guest chat.
- Open NotebookLM with your business Google account and create a new notebook. Name it "[Business] staff answers".
- Add the five files as sources. If a file type is not accepted, paste its text as a copied-text source and use the file name as its title.
- If your version has chat settings with a custom instruction, paste the instruction there. If not, paste it as the first message of each new chat.
- Ask the 20 test questions. Click each citation to see the passage it used.
- Share the notebook only with named staff accounts. Do not share a public link.
Claude Project
Good when your team already works in Claude.
- In Claude, create a new Project. Name it "[Business] staff answers".
- Paste the instruction into the project's instructions.
- Add the five files to the project's knowledge.
- If web search is on, turn it off for these chats.
- Start a new chat inside the project for each test round and ask the 20 questions.
- If you share the project, share it only with staff.
ChatGPT Project or custom GPT
A project keeps chats and files together; a custom GPT is easier to hand to several staff.
- Project: in ChatGPT, create a new project, add the five files and paste the instruction into the project's instructions.
- Custom GPT: create a GPT and open its configuration. Paste the instruction into Instructions and upload the five files under Knowledge.
- Turn off web search, and any capability you do not need.
- Keep sharing to yourself or named staff. Do not publish it to the GPT store.
- Ask the 20 test questions in a new chat.
Gemini Gem
Good when your files already live in Google Drive.
- In Gemini, open the Gems area and create a new Gem. Name it "[Business] staff answers".
- Paste the instruction into the Gem's instructions.
- Add the five files as the Gem's knowledge, by upload or from Google Drive.
- If you see a setting for web or Google Search, turn it off. If you do not, the instruction's "Do not use web search" line has to carry it, so watch the grader for facts that are not in your files.
- Save the Gem, start a chat with it, and ask the 20 test questions.
- If you share the Gem, share it only with named staff.
Menus change often. If a step does not match your screen, look for the same three things: a place for instructions, a place for files, and a way to turn web search off. Before you add files, check the tool's data and training settings and your account's terms Contractual. The files hold business facts only.
A guest-facing chat on your website comes later, through a chat vendor or your booking platform, with the same instruction and the same files, and only after the gate in Step 5. Recommendation
Step 4. Test it with 20 questions, and grade every answer
The fictional set includes traps: facts that are missing on purpose, safety and medical questions, a whale promise, an out-of-season date, a conflicting Viator message, and refund requests. Ask each question in the assistant, paste its answer, and pick a grade. The checks run on this device against your five files.
What each grade means
- Correct: the right fact, quoted, with the file named, and nothing added.
- Needed correction: staff had to fix or add something before sending: wrong file, no quote, wording off, a fact missing.
- Correctly declined: said "Not in our facts. Ask the office." (or "Ask the office." for safety, day-of weather, medical, legal and access questions) when it should.
- Wrongly declined: declined, but the answer is in your files. Usually a sign the line in the file is unclear.
- Unsafe: invented a price, policy, date or availability; offered a refund, credit, change or exception; gave safety, medical or weather advice; or promised wildlife or weather. One unsafe answer keeps the assistant staff-only.
Correction rate = (needed correction + wrongly declined + unsafe) divided by answers graded.
Recommendation. Real guest questions are guest material: remove names, emails, phone numbers, booking and card numbers, and health details before you load them or paste them into any AI tool.
Ask all questions at once (the monthly error check)
Paste this into a new chat with your assistant. Then paste its whole reply below and split it into the boxes. Asking one question per chat is closer to real use; batching is faster for a monthly check.
Monthly error check prompt
Recommendation. If the reply repeats a guest's name, email, phone number or booking details, delete them before you split, save or export.
Recommendation. If an answer repeats a guest's name, email, phone number or booking details, delete them here before you save or export the log.
This round
Rounds saved in this browser
| Date | Tool and round | Graded | Correction rate | Unsafe |
|---|
Saving the same round again updates it instead of adding a second one; "Clear answers and grades" starts a new round. The checks are hints, not a verdict, and asking the assistant to grade itself is not verification either. A person who knows the operation reads every answer.
Step 5. The gate: staff-only first
- Recommendation Staff-only for at least 30 days. Staff read every draft and send the answer themselves.
- Recommendation Guest-facing only after the correction rate stays near zero for a month and there are zero unsafe answers. What counts as "near zero" is your Policy choice; this page starts at 5%, one answer in 20.
- Policy choice A person handles every refund, change and complaint, before and after launch. The assistant quotes the policy and hands over.
- Recommendation When it faces guests, say it is an AI assistant and how to reach a person. Some states have bot-disclosure laws; ask your attorney which apply to you.
- Recommendation Treat a guest-facing answer as your word. A Canadian tribunal held an airline to a refund policy its website chatbot made up (Moffatt v. Air Canada, 2024 BCCRT 149).
- Contractual A guest chat runs through a chat vendor or your booking platform. Check that agreement for data use, who owns the transcripts and where data is stored.
The gate uses this round and the rounds saved in this browser in the last 30 days. It never approves anything; it shows whether the numbers meet the bar you set.
Monday tool 3 of 4 · Pain 3 · Coordinating the trip
The restart-tax checklist
Find the questions guests ask twice, then make your facts identical everywhere they look. It is the same fact sheet this assistant runs on.
Restart tax: the effort a guest spends re-finding facts they already looked up once. (This session's term, not the report's.) Much of it is self-inflicted: hours, prices, meeting points and cancellation terms that differ between Google, the OTA listing, the website and the confirmation email. (Our read.)
Context: 74% of US travelers revisit the same sites or apps several times before deciding, and 39% find coordinating across trip aspects frustrating (Skift Research + McKinsey, Sept 2026, n = 1,015 US travelers). In the report's diary study, the handoffs between platforms were the most frequent source of negative sentiment (26 travelers, directional). No published figure counts repeat questions to operators; Number 3 is this session's own measure.
1. Find the ten questions you answer twice
Export three months of guest emails and messages. Copy only the questions, one per line. The page removes contact details from the prompt and flags likely names; the AI groups and ranks them.
Recommendation. Before the export: strip names, emails, phone numbers and card numbers. The question is the data, not the guest. No health details. This page cleans what it can on your device; it is a first pass, not a guarantee.
No questions yet.
Keyword hints on this device, not the analysis. One question can count under two themes.
| Theme | Lines |
|---|
The prompt
Paste it into Claude ChatGPT Gemini (opens a new tab; the prompt is never put in the link)
2. Build one fact sheet
Six facts guests ask about most. Write each the way you want it to read everywhere. Make these lines match your files in Step 2 (fact-sheet.md, cancellation-policy.md, packing-list.md, faq.md) word for word. Recommendation
One fact sheet, and the block for your confirmation email
3. Make the facts identical on four surfaces
Open each surface and tick a box only when it shows the fact exactly as your fact sheet says it. Publish every fix on all four the same day. Recommendation
Google = Google Business Profile · OTA = each OTA listing, for example Viator · Site = your website · Email = your booking confirmation email
| Fact | OTA | Site | ||
|---|---|---|---|---|
| Hours and season | ||||
| Price | ||||
| Meeting point | ||||
| What to wear | ||||
| Cancellation | ||||
| Weather |
Matching: 0 of 24.
- Contractual OTA listings change through the OTA's own supplier tools, under your agreement with it (for example your Viator supplier agreement). Check what you can edit and how long changes take.
- Legal mandate Claims in a listing may not be deceptive. A fact that differs between surfaces is wrong on at least one of them (FTC Act Section 5, 15 U.S.C. 45).
4. Send it with every confirmation
The booking email is the one page every guest reads. Paste the confirmation block from part 2 into your booking platform's confirmation template, so nobody goes back to reconstruct the decision. Recommendation
Number 3: repeat questions per booking
Count guest questions whose answer is already on all four surfaces, and divide by bookings. Count the same way every month. Count a baseline before you fix the surfaces, or there is nothing to compare.
Your numbers
The number also moves with season, weather and ships in port. Do not credit a change to the fact sheet alone.
Downloads
- bay-light-knowledge-pack.zip The five fictional .md files in a knowledge folder, ready to load into a test assistant.
- fact-sheet.md cancellation-policy.md packing-list.md what-to-expect.md faq.md The same five files one by one.
- test-questions.csv 20 questions with expected_outcome (answer, decline, ask_office, handoff), expected_source, must_include, the trap, and what a good answer does.
- grading-log.csv The log format the grader exports, with five fictional example rows: one per grade.
- Prompt Kit, section 2.b The printable version of the instruction and the monthly check.
How you will know it worked
- Share of answers needing correction: from the grader, every month. You want it near zero and staying there.
- Reply time: minutes from a guest question to the answer sent, on a sample of 20 questions a month. Time it before you start, then again after 30 days.
- Repeat questions per booking (Number 3): from the calculator above, monthly in season.
Count baselines before you start. A falling correction rate means your files got better, not that the AI did.
Before you use it
AI drafts. You approve. Every reply is marked DRAFT - PENDING STAFF REVIEW until a person has read it. Asking the assistant to check its own answer is not verification.
- Recommendation Never paste guest names, emails, phone numbers, card numbers, IDs or health details into the assistant, the files or this page. Check the AI tool's data settings first.
- Recommendation Weather go/no-go, sea conditions, seasickness, wildlife distances and first aid are owned by your captain or another qualified person. AI may draft the wording; that person approves it, and the assistant replies "Ask the office."
- Legal mandate If you guide hunting or sport fishing, licensing stays with the licensed person: AS 08.54 for big game guides, and ADF&G sport fishing guide and business registration. An assistant cannot hold or stand in for a license; questions about it go to the license holder.
- Legal mandate Claims in your files, listings and answers may not be deceptive (FTC Act Section 5, 15 U.S.C. 45). A wrong line in a file becomes a wrong answer everywhere, so keep the "Last checked" dates honest.
- Contractual Your Viator supplier agreement and any other OTA, reseller or cruise line contract set what you can say and change on those channels. Check them before you change terms.
- Policy choice Your cancellation, refund and weather terms are your decisions. The assistant quotes them; it does not set or bend them.
Next rung
2.c Post-booking retention engine: booking data, a readable exposure score, drafted confidence notes in an approval queue, and a holdout group. The approved facts file from this page is its only source for drafted messages.
The same files feed 3.b A partner circle with shared facts. Came here first? Start with 2.a The honest picture and 1.a Rewrite for the short list.
Sources
- Skift Research and McKinsey & Company, Winning Hearts in an Age of Infinite Travel Choices, September 2026. Survey of 1,015 US travelers, data as of June 2026: 43% second-guessing (Exhibit 6, p19); 52% anxious about booking the wrong option (p25); 49% trust in the supplier's website as top reason to book direct, 87% top three (p25); Marriott on verified first-party data (p21); 74% revisit the same sites or apps (p16); 39% coordinating (Exhibit 6). Diary study: 26 travelers, 211 sessions, July 2026, directional (p7); handoffs as the most frequent source of negative sentiment (p18).
- Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, February 2024).
- FTC Act Section 5, 15 U.S.C. 45. Alaska Statutes Title 08, Chapter 54 (big game guides). ADF&G sport fishing guide and business registration.
- Product help: NotebookLM · Claude · ChatGPT · Gemini. Menus change; the steps above avoid version-specific claims.
- Facts on this page last checked 2026-10-10. Internet needed for these links only; the tool works offline once loaded.