When a traveller asks an AI assistant where to stay in a city they have never visited, something answers on their behalf. That answer is not a search results page. It is a short, confident recommendation with no visible ranking, no disclosed criteria, and usually no way to tell what was left out.
I measure what those answers say about places in Asia Pacific. The measurement is real, the method is published, and the per-place results publish free to the place measured once they have passed native and expert language review.
Here is where I have to stop, and it matters more than anything I could add.
Two claims, and only one of them is mine
The first claim is that AI recommendations can be measured as skewing toward chain and platform-bookable inventory rather than locally owned businesses. That is a measurable property of a set of outputs. It is the kind of thing an instrument can settle.
The second claim is that this skew changes where visitors actually spend money, and therefore how much economic value stays in a place. That claim is much more useful. It is the one that would put a number on a slide and move a budget. It is also the one I cannot support, because supporting it needs traveller behaviour data that I do not hold and cannot get by asking an assistant questions.
The mechanism is plausible. Recommendation shapes consideration, consideration shapes booking, booking shapes revenue. Plausible is not demonstrated. Anyone who tells you the second claim is established is either working from data they have not shown you or has quietly promoted the first claim into the second.
Why say this out loud
Because the alternative is worse in a specific way. If I lead with the strong version, the first serious economist in the room takes it apart, and when it goes the credible part goes with it. Conceding the boundary in advance costs a slide and buys the rest of the argument.
There is also a governance point. AI-mediated discovery has no accountability mechanism attached to it. If the people arguing for one are careless about evidence, the people who prefer no mechanism at all will win on rigour, which is not a fight worth losing on those terms.
The test that would settle it
A destination-level study pairing three things: measured AI recommendation outputs over a period, a panel of arriving visitors reporting which sources shaped their decisions, and merchant-level settlement data separated into chain and independent inventory. Run it in two comparable cities. Look for whether shifts in what AI systems surface precede shifts in the settlement mix.
That study needs a partner who holds payments data and a destination willing to be measured over more than one period. I do not have either yet. If you do, that is a conversation worth having, and it does not involve buying anything from me.