Why AI Recommends and Erases Your Property at the Same Time
A traveler opens ChatGPT and types a question. The best honeymoon resort in Oahu. The best luxury hotel in the valley. The best place for a milestone dinner. Before they hit enter, the answer does not exist. It is not sitting in a file waiting to be read. It is assembled at the moment the question is asked.
The model produces a list, names a handful of properties, and leaves out every other. For the hotels that get named, they are present in that answer. For the hotels that do not, they are absent in the one place that traveler asked for a decision.
Which hotel gets named and which gets left out is not decided only by which hotel is better. The public information available about each property can differ in accuracy, consistency, specificity, and authority. A property can be excellent, fully booked, credentialed, and beloved, and still not appear in a particular answer while another property does.
In 1935 Erwin Schrödinger described a cat sealed in a box, alive and dead at once until someone opens it. He did not believe a cat could be both. That was his point. He used an impossible image to show that some outcomes are not settled until they are observed.
A hotel in AI search behaves enough like that cat to make the analogy worth using. Until a traveler runs the query, whether the property appears in that answer is unsettled. The query is the box opening. What the box contains is not arbitrary, but an outside observer cannot see the complete internal process that produced it. What can be examined is the public information environment around the property and the answer the system actually returned.
This is an analogy, not a law of physics. AI outputs are not quantum states. But the cat is still useful for seeing the practical problem: the answer is generated when it is requested, and the result can vary across systems, queries, users, and time.
The same cat explains a separate problem. A buyer evaluating AGR opens the box with the wrong instrument and finds a marketing agency, when what is inside is an infrastructure intervention. That argument is made in full in AGR Is Not a Marketing Agency, And That Is the Point. This article applies the same thought experiment to a different box: the one a traveler opens when they ask AI for a hotel.
Six Engines, Six Answers, One Hotel
There is not one box. There is a wall of them.
ChatGPT is one system. Gemini is another. Perplexity, Grok, Copilot, Google’s AI answers, each can return a different result from the same question. Their underlying products, models, retrieval behavior, ranking systems, and available sources differ, but an outside observer cannot see the complete internal process behind any one answer.
We ran the audits and watched it happen. The same hotel leads on one system and is absent on another, for the same category, on the same day. A property is the number one answer on six platforms for one phrasing of a question and absent on all six for a slightly different phrasing. There is no single state called “my AI visibility.” There are as many states as there are systems, and they contradict each other.
It gets more complicated. Even one system can return different answers across repeated queries or different conditions. Depending on the product, factors such as location, account state, session context, product changes, and query wording can contribute to variation. The property does not have one fixed state per system. What can be observed is a distribution of answers across defined test conditions and over time.
Many systems, many query phrasings, and personalization on top mean there is no single answer to check and no fixed target to optimize. There is a moving distribution. You cannot manage a state that will not hold still.
Present Is Not the Same as Represented
The obvious fear is that AI leaves a hotel out. That is real, and it can be expensive, because absence is silent. The property can lose an opportunity without ever seeing the moment it was omitted.
The opposite failure is worse, and most measurement tools cannot see it. A hotel can be named in the answer and described wrong.
In one AI visibility audit, a property that holds a Forbes Five-Star rating for both its hotel and its restaurant was not just left out of a dining answer. Two separate systems stated, as fact, that a competitor held the only rating of that kind in the market. The audited property holds the identical rating. The model did not omit the hotel. It named a rival and handed that rival the property’s own credential.
A tool that counts citations scores that as a win. The property appeared. The instrument was built to detect presence, not to read what the answer said. It cannot tell the difference between being recommended and being misrepresented, because it was never looking at the content of the recommendation, only at whether the name showed up.
For a hotel, those two failures are not the same size. Absence costs a booking. Misrepresentation can hand a competitor your strongest credential in front of the exact traveler who was looking for it. One loses demand quietly. The other transfers it.
You Cannot Observe Your Way to AI Visibility
The common response to all of this is to measure it. Run the queries, see where you appear, count the mentions, track the citations, and adjust toward the next answer. That is the logic behind answer engine optimization and generative engine optimization.
Measuring is necessary. It is not sufficient on its own, because the thing being measured has no single state and does not hold still.
A mention count in one system says nothing by itself about the other five. A citation win records that you appeared, not whether the answer about you was true, which is the misrepresentation failure that does the most damage. And any change tuned to one observed answer is fitted to one observation that may move the next time the question is asked under different conditions. You can measure the output all day. Measurement tells you what happened. It does not, by itself, establish why it happened.
AEO and GEO are real disciplines, and the serious versions do more than count mentions. They can test across systems, cluster queries, improve content, and address technical retrieval conditions. That work has value. The distinction here is narrower: this article is concerned with the coherence of the public source environment around the property and whether observable AI answers reproduce it accurately.
The Collapse
For a hotel, the box opens again and again, across multiple systems, queries, travelers, and dates, and the answers keep moving. You cannot stand at each box and force the result. There are too many observations, they can disagree, and no outside operator controls the output.
The leverage a hotel can actually control is not the individual answer. It is the public information environment around the property: the property’s own pages, structured data, listings, editorial coverage, and other third-party references. When those sources are fragmented, contradictory, outdated, or generic, the property has a source-environment problem that can be identified and corrected. Whether those corrections change future AI answers, and by how much, has to be measured rather than assumed.
Tuning work to one answer addresses one observation. Correcting an inaccurate or contradictory public source record changes the information environment that a hotel can control. The effect on future AI answers is then an empirical question: measure whether the distribution moves toward more accurate, consistent reproduction across relevant queries and over time.
Americas Great Resorts calls that work Knowledge Formation Optimization. KFO structures, sequences, distributes, corroborates, and corrects intellectual frameworks and entity definitions across the public information environment and measures whether AI systems reproduce them accurately across relevant queries and over time. It does not promise a specific placement in a specific answer, and it does not claim control over hidden model state or proprietary selection logic.
A hotel can keep treating the symptom, one system and one answer at a time, while inaccurate or contradictory public information remains untouched. Or it can correct the source environment it can actually control and then measure whether the answers change. The first observes the symptom. The second changes a controllable condition and tests the result.
