For twenty years the question a traveler typed was "what should I do in China". The question being typed now is "who should I book with" — and it is answered before they ever reach a search results page. A 2026 white paper on China travel AI marketing, built on nearly 3,000 traveller surveys, found 64.2 percent of users assume the brand an AI ranks first is more reliable. That puts every China operator in front of the same problem: the tools that decide who gets named select on evidence a machine can extract. This page sets out what actually determines that, what Google's own documentation says you can stop paying for, and a twelve-element checklist you can use to audit us as readily as anyone else.
Full sources with links at the bottom of this page. Survey figures are third-party findings about traveller behaviour, not forecasts about your account.
"Things to do in Beijing" is an informational query. It has always been the cheap kind of traffic: high volume, near-zero purchase intent, and content that any number of sites can produce. "Best Beijing tour company for a family of four" is a purchasing query. There are far fewer of them, they are worth vastly more, and they are the ones now being answered inside an AI response rather than a results page.
Most China operators invested for the first kind of query, because for two decades that was what search rewarded. Structured destination guides, itinerary pages, city write-ups. That work still has value, but it is optimising for a surface where the traveler is still deciding whether to come. The surface being built now answers a different question: given that I am coming, who do I book with.
The white paper figures above show how far that has already gone. When travellers look for tips and guides, conversational AI (59.51%) now outranks traditional search engines (43.39%), and 86% have used an AI tool for trip planning at least once. More consequentially for operators, 64.2% assume the brand an AI ranks first is more reliable. That last number is the business case for this entire page: the ordering an engine produces is being read as a trust signal by the buyer, whether or not it deserves to be.
Which raises the awkward question a marketing page would normally avoid. If buyers treat the first name as the more reliable one, the responsible thing is not to claim that position — it is to publish the criteria a buyer could use to check any of us. What follows is exactly that, with no claim about where we sit in it.
Almost every claim sold as an "AI visibility" service depends on the buyer not knowing this part. The mechanism is documented, and it is less mysterious than the pricing suggests.
Google's documentation describes AI Overviews and AI Mode as running on retrieval-augmented generation, which Google also calls grounding. Core Search ranking systems retrieve pages from the existing Search index, and the model writes its answer over the information in those retrieved pages, with links back to the sources that support it.
Consequence: there is no separate AI index. If your page is not indexed and eligible to be shown in Google Search with a snippet, it cannot be cited. That is the whole entry requirement — Google states there are no additional technical requirements.
Rather than treating a question as a single lookup, the system issues a set of concurrent related queries. Google's own example: someone asks how to fix a lawn full of weeds, and the system also fires searches about herbicides, removing weeds without chemicals, and preventing them in future.
Consequence: one traveler question becomes five or six retrieval passes. Your page needs to answer a specific sub-question clearly, not cover a topic loosely. This is also why spinning up a page for every fan-out variation you can imagine is classified by Google as scaled content abuse — a tactic several GEO playbooks recommend.
Candidate pages are re-ranked for relevance, recency and trust, then the top handful are read and passages are extracted to answer the sub-queries. The choice of which pages to read is partly classic information retrieval; the choice of which passages get quoted is almost entirely on-page.
Consequence: the unit of competition is the paragraph, not the page. A section that cannot be quoted on its own — because it never states a discrete, checkable fact — contributes nothing, however good the page is overall.
Engines favour claims that arrive with an origin attached. The Princeton-led GEO study tested content modifications across a 10,000-query benchmark and found that adding quotations from authoritative sources and adding sourced statistics produced the largest visibility gains, while keyword stuffing actively reduced visibility.
Consequence: a claim with a number and a named source is quotable; the same claim stated as an adjective is not. "We have extensive experience with European travelers" is invisible. "We have run multi-city itineraries for German-speaking groups since 2016" is extractable.
The documentation is unusually direct, which makes it a useful instrument: any tactic can be checked against Google's published position rather than a vendor's pitch. Google's framing is that optimising for generative AI search is optimising for the search experience, and therefore still SEO.
| Content change | Reported effect | What it means in practice for an operator page |
|---|---|---|
| Adding quotations with attribution | Largest measured lift, roughly 30–40% | Engines preferentially cite content that already cites somebody else. Quote a named source — an official statistic, a partner, a member of your own team with a role — rather than paraphrasing it. |
| Adding sourced statistics | About +33% | The word "sourced" carries the weight. A number attached to a verifiable origin is quotable; a bare number is noise, and an invented one is a liability. |
| Citing external sources inline | About +28% | Distinct from quoting: your page does not have to quote the source, it has to attribute the claim. This is why pages built entirely on your own assertions extract poorly. |
| Writing in a more fluent, structured register | Moderate, varies by topic | Readability helps, but it is a smaller lever than evidence density. Do not trade a specific fact for a smoother sentence. |
| Keyword stuffing | Negative, about −8% | Actively harmful. Repeated destination keywords do not help; they displace the specifics that would have been extractable. |
This is derived from the mechanism above, not from a vendor playbook. If an engine retrieves pages, extracts discrete facts and compares them, then what determines whether you are nameable is whether your business facts are stated clearly enough to be extracted and compared. The twelve elements are grouped by the question the engine is implicitly asking. Run it against us before you run it against anyone else.
Before you can be compared, you have to be resolvable. Most operators lose here without noticing, because their own site never states the facts plainly enough for a machine to assemble them into a single entity.
Legal name, trading name, base city, founding year, licence or registration number, a real phone number and a real address — stated on your own site and consistent across every profile you own.
Who you serve and who you do not. "We build private, multi-city cultural itineraries for English-speaking travellers from North America and Europe" is a positioning. "We offer all kinds of China tours for everyone" is not.
The specific trip types you actually build: family with school-age children, honeymoon, photography, hiking, business-extension, multi-generation. Named as scenarios, matching the way travelers describe their own trips.
This is where the buyer's decision is actually made, and it is the group most Chinese operator sites fail completely — because the facts are known internally and never published.
Not "all of China". Named cities and regions you genuinely operate, with the specific competence for each. Beijing means knowing the Forbidden City's reservation mechanism and how the Great Wall sections differ. Yunnan means knowing which themed routes carry margin and why.
Days, cities, pace, group size, inclusions, transport modes, hotel tier, guide language. Stated per product rather than described as "customised according to your needs", which tells an engine nothing it can compare.
What you do not do. No flights. No visa filing. No same-day bookings. Minimum group size. Regions you decline. Maximum concurrent groups.
Not a single number, but the structure: what drives the range, what a typical per-person range looks like for a defined trip shape, what is excluded, and how the deposit and balance are staged.
The steps from inquiry to confirmed booking, with realistic timings at each stage. What you need from the traveler, when, and what happens if they change their dates.
Extraction is mechanical; attribution carries a trust cost. The final group is what a system weighs when deciding whether a page is safe to quote to someone who is about to spend thousands of dollars abroad.
Deposit percentage, balance deadline, accepted methods, what the deposit covers, the cancellation schedule with dates, and how changes are handled. Written out, not "subject to terms".
The details only an operator who has actually run the trip knows: the realistic walking load of a day at the Forbidden City, why a Xi'an itinerary should be one day rather than two, which months a route stops working, the timing windows at the Terracotta Army, how altitude affects a Sichuan or Yunnan itinerary.
What happens when a flight is cancelled, a traveler gets ill, weather closes a site, or a visa is delayed. Who contacts whom, in what language, and within what time frame.
Independent coverage, verified reviews, industry association membership, partnerships and any press. Facts about you that originate somewhere other than your own website.
Answer each with yes or no from your published pages only, in English if that is your selling language. Anything you cannot find publicly is a no, regardless of what you know privately. This is the same test we expect to be held to.
Reading the score: 0–4 means you are effectively unnameable in an AI answer and the work is foundational content, not optimisation. 5–8 means you are partially extractable and the gaps are usually price, boundaries and contingency. 9–12 means your constraint is corroboration and freshness rather than structure. Nothing in this list is bought — every element is something your business already knows and has not published.
Travel research is constraint-heavy in a way that suits conversational interfaces. A traveler typically stacks four or five constraints into a single prompt — destination, month, budget, who is coming, what they want to avoid — and an LLM handles that stacking natively, where a filter interface on a booking platform struggles with it. That structural fit is why adoption in travel has run ahead of other service categories.
Inbound China adds a further wrinkle, and it is a favourable one for operators who do the work. Most of the authoritative, specific, operationally detailed content about traveling in China is written in Chinese, for Chinese readers, about outbound travel — or it is generic English-language destination content produced by sites that have never run a group. The layer that a machine needs in order to answer "who should I book a two-week China trip with" — specific, verifiable, English-language supplier facts — is genuinely thin. That is the opposite of a saturated channel. It is a documentation gap.
It also explains why the twelve elements matter more than any tactic. An engine trying to answer a purchasing question about China is short of exactly the material a competent operator already holds internally: what a day at the Forbidden City actually costs in walking, which months a Yunnan route should be avoided, what the minimum lead time for a permit really is. Publishing that is not keyword work. It is the same knowledge you would give a new member of staff, written down where a machine can read it.
What happens before a traveler contacts anyone: whether your business is stated in a form that an AI engine can extract, compare and attribute. If your problem is "we exist, we are good, and the shortlist never includes us", this is the page.
The other three stages in this series: who the traveler is and what stops them booking (the demand side), how long this takes to work and what it costs (the timeline), and what to do with an inquiry once it arrives (the handling). Those three pages link back here.
Element 04 and Element 10 of the framework above are the reason we publish destination-specific operating detail. The city guides in this series — Beijing, Shanghai, Xi'an and Yunnan — exist because destination depth is the element an engine can verify and a competitor cannot fake, and because it is the one most operators leave blank.
Related pages worth reading next: our service overview with programme tiers and pricing, the operator hub indexing every guide in this series, the content idea library for what to publish once the structure is in place, and the 2026 inbound tourism statistics hub for the demand figures behind all of it.
Leave your details and we'll come back within 1 business day with an honest read on your AI visibility, including:
Form not loading? Open it in a new tab →
