How to use AI to plan a trip (and what to double-check)

To use AI to plan a trip, give a chat AI your dates, who is going, your pace and a budget in CA$, ask for a day-by-day draft with times, then check every place, hour, travel time and price before you build a day around it. AI is a good first draft and a bad source of facts.

We draft with general chat models ourselves, and the draft is the easy part. The pages that rank for this question are tool landing pages or roundups of tools, and none of them says what to do once the draft exists. That second half is this post.

Can AI actually plan a trip?

The whole method, in the order this post follows:

1. Brief it.
2. Draft it, one day at a time.
3. Make it confess.
4. Move it onto a map.
5. Watch the price separately.

A chat AI is genuinely good at three things. It will shortlist destinations against your constraints. It will give a day a sensible morning, afternoon and evening shape. And it will rewrite that day in seconds when one fact changes. Two prompts show all three:

I am in Calgary and want somewhere warm in the first week of March, direct flights under 6 hours if possible, two adults, no interest in nightlife. Give me five destinations with one honest downside each.

Same trip, but we now have a rental car and one of us cannot walk more than 2 km a day. Rewrite the week and tell me what you changed.

It is bad at four things this post keeps coming back to: opening hours, closures, travel times, and prices. It writes what is plausible from text it has read; it has no clock, no map and no booking site.

How do I write a prompt that gets a usable itinerary?

Every blank you leave, the model fills with an average traveller's answer. A good brief has eight parts:

• Dates or a month. Exact if you have them; a month and a length if you do not.
• Departure city. Flight time shapes the shortlist.
• Party and ages. Two adults and a family with an eight-year-old are different trips.
• Car or no car. The single biggest assumption a model makes silently.
• Pace. Outings per day, and the earliest start you will tolerate.
• Budget in CA$ for extras, and what it covers.
• Must-dos and no-gos. Mobility, food, bus tours, early starts.
• Output format. This one is the whole trick.

The format clause, spelled out:

Give me day 1 to day 7. For each day: a one-line theme, then morning, afternoon and evening stops, each with a start time, a duration and the real place name if there is one.

Then three rails, appended every time:

Name real places only when you are confident they still exist, and flag the ones you are not sure about. Do not tell me anything is open daily. Do not include prices.

A model told to flag uncertainty is far more likely to; chat models tend to write "open daily" when they do not know the hours; an invented price leaks straight into a budget. One more line earns its place: "Ask me three questions before you draft anything." It surfaces the assumptions the model would otherwise make silently.

"Best things to do in Cancun" returns the same ten tourist stops every time. Constraints are what make the draft yours. Build one for your trip and paste it into whichever AI you use, including Ask Mara:

How do I use ChatGPT to plan a trip, step by step?

Written for ChatGPT because that is what most people open first; the steps work unchanged in Gemini and Claude.

1. Settle when and where before what. Two different questions. When to go is seasons, which a chat AI knows in the "typically" register. When to book is price timing, which a chat AI has no observed history for.

Rank the Riviera Maya, Punta Cana and Los Cabos for early March on weather, sargassum risk and flight time from Calgary. Say "typically" when you are generalizing, and give no prices.

2. Skeleton first. Seven finished days in one go is the fastest way to a plausible week that falls apart on contact.

Give me one line per day for the 7 days: the theme, and whether it is a resort day or an outing day. No venues yet.

3. Expand one day at a time.

Expand day 3 only: morning, afternoon, evening. One named place per slot, the neighbourhood it is in, an honest guess at travel time from the resort, and a flag on anything you are not sure still exists.

4. Iterate with constraints, not compliments. General models often rewrite the whole week when you ask for one swap, and the rewrite is where earlier fixes quietly vanish. Say what to leave alone, and park cuts instead of deleting them; that parked list is what Mara calls Ideas.

Cut day 5 to two outings and put whatever you removed in a spare-ideas list. Add 30 minutes of buffer between everything on day 3. Leave every other day exactly as it is.

5. Make it confess. The most useful prompt in the post, and no tool roundup gives it to you.

List every named place in this itinerary and tell me how confident you are that it is open in February 2027. Mark anything you cannot verify.

6. Export. A flat list pastes cleanly into any planner, including Ask Mara with a free account.

Give me the final version as a plain numbered list, one stop per line, with the start time first.

The free tiers all do this fine. In a plain chat turn, none of them sees a live package quote, checks availability or measures a routed travel time; browsing modes can look things up, but you are still reading a summary you have to check. Stop iterating when every day has a morning, an afternoon, an evening and slack between them. A plan with no slack fails at the first closed door.

How do I check an AI itinerary before I trust it?

An AI itinerary is a list of claims in the shape of a plan. The shape is usually fine. Check the claims, in the order of what it costs to be wrong: timed entries and deposits first, then whatever anchors a day, then the nice-to-haves.

• Does it still exist? The listing's most recent reviews and photos; nothing in eighteen months is a warning. The model's knowledge stops at a training cutoff it may not volunteer, and it will not flag which places have changed since.
• Is it open that day? The place's own website, checked for the weekday you will be there: Monday museum closures, siesta hours, seasonal schedules. Models tend to write "open daily" when they do not know.
• Is the travel time real? A routed map, never the model's number. Models do not compute routes; they guess from text, and the guess runs short. Two stops across town at 9:00 and 9:30 is a tell.
• Did it invent a price or a rule? A price from an AI is a memory, not a quote. Entry rules and advisories come from dated government sources; for Canadians, travel.gc.ca.

We and others who have tested chat AIs on city itineraries have found closed restaurants recommended with full confidence. If your model can browse, make it show its sources and read those; browsing lowers the risk, it does not remove it.

An AI itinerary is a list of claims in the shape of a plan. Check the claims. Keep the shape.

Inside a Trip in Mara the same checks are close at hand. Open a place-backed Stop and the Details from Google card sits at the top of its editor, open on its own or one tap away: name, rating, hours and photos, as Google's snapshot, up to a month old, not a promise about your date. The Stop's pin status says how sure the match is: confirmed, unconfirmed and waiting for your "is this it?" tap, or a generic Stop because that beats a wrong pin. Mara's suggestions come from general knowledge too; the difference is that the map and the details card make her claims checkable in place. The test for any line in any draft is the same: would I put a non-refundable deposit on this sentence?

How do I get the draft out of the chat and onto a map?

Day 3 has three places that read fine in prose and sit 40 km apart on a map, and verifying them means re-typing every one into a map anyway. So the draft should live where the checks happen. Scope first: this needs a free account (the signed-out demo has no maps and cannot write an Itinerary), a destination in Mara's catalog of all-inclusive regions (some marked coming soon), and one active Trip at a time on a free account. A chat AI will plan Tokyo. Mara will not, yet.

Plan a trip to the Riviera Maya, 12 to 19 February, two adults, budget CA$4,500.

She shows an editable draft form in the chat; submitting it saves the Trip on its own Trip page, with the Itinerary beside the map. Then the Itinerary, from scratch or by pasting the step 6 export after "Save this as my Itinerary:" (keep prices out; a line with a dollar amount is refused):

Build me a 7-day itinerary for this trip: two resort days, one cenote day, one Tulum day.

Days of Stops land in the Itinerary. A real named place is pinned when the match is confident; an unsure match is marked unconfirmed rather than faked, and she says so; anything she cannot place stays a generic Stop. On the Stops Mara writes, the label is your words or her title, never a search result's name. Small edits happen in the Mara panel on the Trip page: "Add a dinner spot to day 3." "Save these as ideas." "Swap day 2 and day 3." A rework keeps Ideas, keeps the Stay, and carries your costs, reservations and hand-picked pins onto matching Stops. Every write Mara makes leaves a one-level Undo.

Travel time between Stops starts as a straight-line estimate marked ~. Show route asks Google for the Day, and a Day whose longest estimated leg runs past 30 minutes is routed for you once the map is open; Google's driving or walking time and distance then replace the estimate. Say "Optimize day 4" and the flexible Stops reorder for the shortest estimated travel while every Stop with a start time keeps its time; it never asks Google and it is one Undo away. The feature depth is on an itinerary planner with a map, including Google routes, Optimize and the read-only share link, which never carries your notes, costs, confirmation numbers, budget or Trackers.

What Mara never does is write a cost, a reservation or a price. When you book the cenote tour, you set the Stop to Reserved with the confirmation number; the product's own line under that field is "Only you mark a Stop reserved. Mara never does." Her other limits fit in one breath: she books nothing, and Book hands you to the provider's own site; she cannot know a place closed last month, only what Google currently lists. Reading that is still your check. The Plan pillar exists to make that check cheap, not to skip it.

Can AI tell me what the trip will cost?

No. A chat AI has no live feed of package prices. Today's quote for a February all-inclusive is not in its training data, so it produces a plausible number and presents it with the same confidence as everything else. The prompt not to trust:

What does a 7-night all-inclusive in the Riviera Maya cost in February for two adults?

Whatever comes back is a memory of numbers it has read, in whatever currency and year, not a quote. Mara never invents a price. The number comes from a Tracker: one watched trip configuration at a major travel provider, checked daily on a free account, every quote recorded. That is the idea behind the vacation package price tracker.

Set up a tracker for this Trip.

She shows a prefilled Tracker draft form and creates nothing until you confirm. From then on the Trip page's Overview carries the linked Tracker's observed price beside the Itinerary, and when there is not enough history to say anything useful she says so. The principle is one line: use AI for the shape, use a record for the money.

Plan your next trip with AI this weekend

Four moves, all of which fit in an evening:

• Brief it. Use the builder above, with real dates and a real party. A draft for a trip you will never take is trivia.
• Draft it one day at a time. Skeleton, expand, constrain.
• Make it confess. The step 5 prompt, before you believe a single venue.
• Move it onto a map, then watch the price separately. A Trip in Mara with a free account, and a Tracker on the package price.

Undecided on where? Ask Mara: "Somewhere hot in March, under 6 hours from Calgary, two adults, about CA$2,500 a person." Explore ranks her catalog on fit, with observed prices alongside when she has them, and when she has none she says so rather than estimating. Decided? "Plan a trip to Cancun, first week of March, two adults." A free account takes a couple of minutes to create. The AI will still be wrong about something. The method is what makes that cheap, and more is still coming.

AI trip planning FAQs, answered honestly

The questions people type before they trust a draft:

Two prompts worth keeping in your clipboard: "Flag anything you are not sure still exists." and "Do not include prices." AI will happily plan a trip for you. The only question is whether you check its work before or after you land.