We asked a current AI assistant 12 of the questions travelers actually type before a Vietnam trip — what it costs per day, whether an American needs a visa, when to go to Hoi An, how much a bowl of pho runs — and then graded every factual claim against primary sources and this site's own sourced pages. 10 of the 12 answers contained at least one claim that was outdated, misleading, or wrong. The single most common failure was not a hallucinated fact but a stale one: every price, fee, and exchange rate we checked was quoted too low, because the model's sense of Vietnamese prices was frozen a year or two before a stretch of real inflation and a weakening dong.
That is the headline, and it is worth sitting with. The advice was rarely nonsense. It was usually the right shape — go in the dry season, eat where the locals queue, take the overnight train north — wrapped around numbers that had quietly gone out of date. Which is arguably the most dangerous kind of wrong, because it reads as authoritative and slots neatly into a spreadsheet.
One thing this test is not: a claim about any named product. The answers here were generated by a current large language model — the same class of assistant a traveler would open — and we grade them as representative AI output, not as ChatGPT, Gemini, or anyone else. We ran a model. We did not run a competitor and put words in its mouth. The full method is below, and the whole exercise is built to be rerun on any assistant you like.
What we tested, and how it scored
Twelve questions, each posed the way a traveler would type it, with no follow-up and no instruction to cite sources — the default single-turn experience. Each answer was graded on a four-point rubric: Correct (every material claim matches a current source), Outdated (was true once, no longer), Misleading (not flatly false, but likely to cause a bad decision), and Wrong (a material claim contradicts the current value).
| Question | Verdict |
|---|---|
| Daily mid-range budget | Misleading |
| Visa for a US citizen | Wrong |
| Cheapest month to fly | Misleading |
| Hanoi to Sapa, and the cost | Misleading |
| Ha Long Bay cruise cost | Outdated |
| ATM fees and cash limits | Wrong |
| Best time for Hoi An | Wrong |
| Cost of a bowl of pho | Outdated |
| Is the tap water safe | Correct |
| Dong-to-dollar exchange rate | Outdated |
| Is street food safe | Correct |
| How far ahead to book a cruise | Misleading |
Two answers correct, three outdated, four misleading, three wrong. The two the model nailed were both safety questions, and both were qualitative rather than numeric. Every question that turned on a current number scored below Correct. The full graded dataset has the exact wording, the correct value, and the source for each row.
The errors, by type
Stale prices and rates — the dominant failure. Four of the ten flawed answers were simply behind on money, and all four erred in the same direction: too cheap. Mid-range daily spend came back as $50-100 when the sourced 2026 figure is $80-150, and $100-200 in central Hanoi and District 1 Ho Chi Minh City, per our Vietnam travel cost index. An overnight Ha Long Bay cruise was quoted at $80-120 a person when the budget tier now starts around $105 and the mid-range volume tier runs $145-250 before add-ons, per our Ha Long cruise cost breakdown. A bowl of pho came back at 20,000-30,000 dong — a real price, in about 2022 — when the local bowl now runs 30,000-50,000, per the Pho Index and VietnamNet's reporting on the fading 30,000-dong bowl. And the exchange rate came back at 23,000-24,000 dong to the dollar, a 2022-2023 level, when the dong has since drifted to about 26,000, roughly 26,250 in mid-July 2026 per Trading Economics. None of these is a wild invention. Each is a snapshot that expired.
Visa oversimplification. Asked whether a US citizen needs a visa, the model said Americans get 15 days visa-free and can use a visa on arrival. This was the clearest single error in the run. In 2026 the US is not among the 24 nationalities granted 45-day visa-free entry, no 15-day US exemption is in force, and tourists apply for the online 90-day e-visa ($25 single-entry, $50 multi-entry) on the official portal rather than a visa on arrival, as our Vietnam visa and entry atlas documents from the underlying resolutions and decrees. Immigration rules move faster than training data, and it shows.
Over-optimistic and imprecise logistics. Two answers assumed the friction out of Vietnam. One sent travelers on an overnight train "straight to Sapa" — but no train reaches Sapa. The line ends at Lao Cai, followed by a roughly one-hour road transfer up the mountain, and most people now take a limousine van or sleeper bus door to door in 5-6 hours for $20-35, per our transport guide. The other said you can book a Ha Long cruise once you arrive, because "there's always availability" — true for a rushed budget boat in low season, false for reputable mid-range and luxury boats on peak dates, especially after the 2025-2026 safety re-inspection thinned the licensed fleet.
Seasonality misfires. Asked the best time for Hoi An, the model recommended October to December as "cooler and pleasant." That window is central Vietnam's wet season, and late October to mid-November is the single most-avoided stretch, with Cua Dai beach flooding and cruise cancellations, per our best-time-to-visit guide. The cheapest-month answer made a subtler version of the same mistake, folding the expensive June-August school-holiday peak into a blanket "low season is cheapest," when the sourced cheapest-month analysis puts the real troughs in September, October, and May.
Where it did not go wrong. Worth stating plainly, because a test that only reports failures is its own kind of dishonest: the model did not invent an operator name, did not manufacture a fake statistic, and did not over-hedge the safety questions into uselessness. The tap-water answer ("do not drink it, use bottled or filtered, city ice is fine") and the street-food answer ("eat where locals queue, cooked to order and hot") were both accurate and appropriately calibrated. The failures clustered in money and logistics, not in safety.
How to sanity-check AI travel advice
You do not need to distrust everything an assistant tells you. You need to know which claims to check, and the pattern from this test points cleanly at them.
- Re-verify every number. Prices, fees, exchange rates, and daily budgets are the first thing to go stale and, in this run, they were wrong every time. Cross-check against a dated source from the current year before you build a budget on them.
- Treat visa and entry rules as volatile. Anything about exemptions, e-visa costs, ports of entry, or overstay fines should be confirmed against an official government portal, not an assistant's memory. These changed materially in Vietnam within the past 18 months.
- Ask "how do I actually get there," not just "can I." Logistics answers tend to smooth over transfers and availability. If a route sounds frictionless, look for the connection the model skipped — the Lao Cai transfer, the sold-out cruise, the booking window.
- Distrust confidence on regional weather. Country-level seasonality heuristics break on Vietnam's three climate zones. Check the specific region and month, not "the best time to visit Vietnam."
- Give more weight to qualitative safety guidance. It held up here. But still confirm anything medical or legal against a primary authority such as the CDC's Vietnam page.
- Ask for the date. A quick "what year is this priced for" often surfaces the staleness on its own.
Methodology
What was run. In August 2026 we posed 12 realistic Vietnam trip-planning questions to a current large language model and recorded its answers verbatim. Each question was asked as a plain traveler would type it — single turn, no system prompt, no persona, no instruction to cite sources.
Who generated the answers. The answers came from a current large language model, the same class of general-purpose assistant a traveler would use. They are disclosed as generated AI output and are not attributed to any named commercial product. We consider it a matter of integrity not to invent what ChatGPT, Gemini, or any competitor "said" when we did not run it; putting fabricated wording in a named product's mouth would be exactly the kind of error this piece exists to catch.
How answers were graded. Every material factual claim in each answer was scored against the four-point rubric above, using primary sources — the official e-visa portal, Trading Economics for FX, the CDC for health — and this site's own sourced reference pages for prices and logistics. Each row in the dataset records the claim, the verdict, the correct value, and the source.
The exact prompts and the answer key are published in full in the open dataset, so the run is reproducible end to end. USD conversions use a rounded 26,000 dong to the dollar.
Limitations
- One model, one run, one moment. This is n = 1: a single model queried once in August 2026. A different assistant, a browsing-enabled model, a follow-up question, or the same prompt a month later would score differently. This is a framework, not a leaderboard.
- Our own pages are part of the answer key. Several correct values are drawn from Day Trips Vietnam guides. We disclose that rather than hide it, and every figure on those pages is itself sourced to named menus, official documents, or dated reporting — but a reader who distrusts the publisher should check the primary sources directly.
- Grading involves judgment. The line between Misleading and Wrong, or Outdated and Correct, is a call. We erred toward the gentler verdict when a claim was defensible, and the dataset shows our reasoning per row so you can disagree.
- Prices move in both directions of error. We caught the model quoting low today; a future run could catch a different model quoting high. The finding is about staleness, not about a permanent bias.
- The questions are curated. Twelve questions cannot represent every trip-planning query. We chose common, high-stakes ones spanning budget, visas, transport, seasonality, and safety, but a different battery would surface different strengths and gaps.
Rerun this on any assistant
This is designed to be copied. Take the 12 questions from the dataset, paste them into whatever assistant you use, and grade each answer's claims against the correct-value column — or against fresh sources of your own. If you are a journalist or researcher writing about AI travel accuracy, the framework, the rubric, and the answer key are open under Creative Commons, and we would rather you rerun and challenge this than cite it uncritically. Point-in-time snapshots age; the method does not.
How to cite this
Nguyen, J. (2026). How Accurate Is AI Vietnam Trip-Planning? A 2026 Test. Day Trips Vietnam. Retrieved from https://daytripsvietnam.com/guides/ai-vietnam-trip-planning-accuracy-2026/
For a specific finding, cite the row and its verdict — for example, "Day Trips Vietnam's 2026 test found a current AI assistant quoted a mid-range Vietnam daily budget at $50-100, against a sourced 2026 figure of $80-150."
Published under Creative Commons BY 4.0. The full 12-question grid, verdicts, and sources: ai-vietnam-trip-planning-accuracy-2026.json. Editorial enquiries: info@daytripsvietnam.com.
Related research and reference
- Vietnam trip-planning checklist 2026 — the human-checked version of what to nail down before you go
- Vietnam travel cost index 2026 — the sourced budgets the AI answer got low
- Vietnam visa and entry requirements 2026 — the rules the AI answer got wrong
- The Pho Index 2026 — the bowl the AI answer under-priced
- Cheapest month to fly to Vietnam — the seasonality the AI answer blurred