A hotel question is the expensive kind for an AI to answer

A single web search costs an AI assistant a cent. Comparing hotels takes ten or more, and somebody has already decided how many your guest's answer is allowed.

by Markus Busch

For two years the industry has argued about how to get an AI to notice a hotel. Review volume, a file on the website, a monthly visibility score. All of it assumes the machine is looking. The published prices say something plainer: every time an assistant searches the web, somebody is billed, and the people building assistants have a dial to turn that number down.

A search costs a cent, and the machine is told how many it gets

Anthropic publishes its price for letting Claude search the web: $10 per 1,000 searches, plus a further charge for reading what comes back. OpenAI lists $10 per 1,000 for the same thing, again with the reading billed on top. Two companies, same number, arrived at separately. A cent to look once, then you pay again for what you found.

The second detail matters more than the price. The same page describes a setting that "limits the number of searches performed." Go past it and the search returns an error. Whoever builds the assistant sets that number before a traveler types anything.

So there is a number, set by somebody at every company building a travel assistant, deciding how many times the machine may go and look. Nobody has published theirs.

Why a hotel question is the expensive kind

Anthropic's own guidance tells you where a hotel sits on that dial: "Simple factual queries typically use 1–3 searches; comparative or multientity research can use 10 or more."

Read what a guest actually asks. A quiet boutique hotel in Bordeaux for the last week of October, walkable to the old town, under €400, with a pool. That is not one fact to find. It is a comparison across several properties, each needing its own look — the availability, the rate on those dates, whether the pool is open in October, what the last hundred guests said. By the vendor's own categories, your hotel sits in the ten-or-more bracket. It is the costly kind of question.

Now set the dial to five and ask it again. Something has to come out of the answer, and it will not be the property the model can name without checking.

The free door is closing at the same time

None of this would matter if there were a cheap way around it. There is less of one every month.

Google confirmed on August 26 that the links on its results page now point back through Google first, so a machine reading that page cannot see where a link leads without following it one at a time. Last year it changed Search so a simple automated fetch no longer works. In December it sued SerpApi, alleging the company got around the protections on its results. The paid search feed it used to sell to developers is closed to new customers and ends on January 1, 2027.

Microsoft got there first from the other direction. It shut down the Bing search feeds developers used to buy and now sells that index only inside its own AI products. You no longer buy a page of results. You buy each look, one at a time, inside somebody else's system.

Both have reached the same position from opposite ends. Looking things up on the open web used to be the free part of building software. It is now a line on an invoice.

Whose web your guest's agent is reading

Because looking is a purchase, it is also a choice. An assistant can buy its looks from Microsoft, from Brave, which built its own index rather than borrowing one, or from the newer search products made for machines rather than for people typing into a box. Each returns a different web. Each has seen your hotel to a different degree, or not at all.

That decision is made by a software company the hotel has never dealt with, in a contract the hotel never sees. It changes when the contract changes, and no page exists where a general manager can find out whose web the assistant recommending hotels in his city reads.

What it means for your hotel

Take the two together — a metered search, and a cap on how many — and the shape of the problem changes.

The old worry was ranking: sitting further down a list than the hotel across the square. The new one is narrower. It is whether your property is reachable inside the first few searches an assistant may run, because after that the budget is spent and the answer is written from what the model already carried.

What the model already carries is the loud part of the web: the chains, the OTAs, the properties named ten thousand times in the text it was trained on. That is not a conspiracy about placement. It is arithmetic about cost. The cheapest possible answer to a question about hotels in Bordeaux is the one that requires no searching at all, and an assistant that is charged by the look has a reason to prefer it.

This publication has argued for a year that an AI recommends the hotel it knows best rather than the best hotel. The pricing pages explain why: looking for a better one costs a cent, then costs again for the reading, and somebody has already decided how many cents this answer is worth.

What we don't know

No company building a travel assistant has published its search limit, so the step from price to behavior is ours, not anyone's reporting. It rests on solid ground — the setting exists, the price is public, and the vendors' own guidance puts comparison in the expensive bracket — but nobody has shown a product turning the dial down and a hotel dropping out of an answer.

Google has published no comparable price, only a contact form. Cheaper search products exist and would change the arithmetic for anyone who buys them. And prices fall.

None of that touches the structure. Whatever the number is, there is a number, and it is not yours.

There is one move the arithmetic rewards, and it isn't a visibility package. If an assistant gets three looks at a question about your city, the hotel that answers the whole question in one of them — the dates, the rate, whether the pool is open in October, in one place a machine can read without going elsewhere for the rest — costs less to recommend than the hotel that takes three. Being easy to check used to be a courtesy to your guest. It now decides whether you are in the answer.

Read also: AI recommends the hotel it knows best, not the best one · Your hotel is paying for an AI file that nothing reads

Enjoying this analysis?* Hospitality.today delivers daily insights on hotel distribution, AI trends, and travel commerce — straight to your inbox. Subscribe for free at Hospitality.today →