Here is a page named The Best Hotels in New York City.
It has no masthead. No history. No century of print behind it. No glossy anything. It was published on July 10, 2026, at seven minutes past midnight, as the public surface for a piece of research we had already done: a study of how AI systems answer hotel questions across six markets. The page was the place the New York data lived. That was its job.
By the night of July 13 it was sitting in second position on Google for the query “Aman the best hotels in New York City,” above Condé Nast Traveler, above Forbes Travel Guide. Names that have been ranking hotels since before the search bar existed.
We are going to tell you what happened next, including the part where it slipped. Because the interesting thing about this is not that we won. It is that we cannot fully explain the win, and it is not the result our own reading of this page’s search position would have led us to expect.
What we expected, and what we got
The conventional account of why a page ranks is not a mystery. Domain authority. Age. Backlink profile. Topical history. The usual expectation is that a site earns a competitive term by accumulating those signals over time, and that a three-day-old page with no page-level track record for “best hotels in New York City” waits its turn behind everyone who has been at it longer.
Forbes has been at it longer. Condé Nast has been at it longer. That is a large part of why they rank, and a large part of why you would expect them to rank above a page that is younger than most leftovers in your refrigerator.
We are not going to tell you the conventional model is wrong. We did not audit the domain’s signals, so we cannot stand here and rule out relevance, freshness, internal linking, whatever site-level trust the domain carries, query interpretation, personalization, or plain ranking volatility. Any of those could be part of it. What we can tell you is narrower and still worth telling: this is not the result our own understanding of the page’s search position would have predicted, and the gap between what we expected and what we saw is the thing we are documenting.
Then we kept looking
Here is where most people would stop. Screenshot the win, post the screenshot, take the victory lap, and quietly hope nobody checks again.
We checked again. Repeatedly. And the position moved.
The next morning, a little after six, it still held second.
One minute later, Google injected a discussions-and-forums block, floated Reddit threads into the page, and rearranged the furniture around us.
An hour and a half after that, it had recovered. Second again, above Condé Nast, as if the reshuffle had never happened.
Nine minutes later, Condé Nast had climbed over us. The page we had watched sit above a magazine older than the web itself was now sitting below it.
And then, by late morning, it had climbed back. Second again, above Condé Nast and Forbes, nearly thirteen hours after the first capture, as though the whole morning of movement had been rehearsal for returning to where it started.
Same query. Same page. It held second, lost the spot, took it back, slipped behind the magazine, and by late morning had climbed to second again. One full turn of the clock, and no answer that stayed put.
If you are waiting for the part where we explain that away, there isn’t one. That is the finding. The result is real and the result is unstable, and both of those are true at once. We have written before about the hotel that is and is not in the answer depending on the moment you look, a superposition that does not collapse until the query is asked. This is that essay, except it stopped being a thought experiment and started being a screenshot.
The claim we are willing to make, and the one we are not
One thing worth stating plainly, because it makes the result stranger rather than less strange. Americas Great Resorts is not a hotel-review site, and it does not sell or resell rooms. It works for hotels, on a narrower problem: how a property is represented when someone asks an AI system about it. The work is about the source material those systems form their understanding from, so that a hotel is described accurately and shows up where it should when the question is asked. This page was a small test of that idea. What happened in Google was not the thing we were testing, and not a thing we expected.
Here is the claim we will make: a page three days old, with no page-level track record for this term, surfaced against two of the most established travel authorities on the internet, took second place, lost it, took it back, slipped to third, and climbed to second again across a roughly thirteen-hour window, and none of that is what our understanding of the page’s search position would have predicted.
Here is the claim we will not make: that we know KFO did it.
We have a hypothesis. We build Knowledge Formation Optimization assets to shape how AI systems form their understanding of a property before a traveler ever asks. That work operates on the source record: precise definitions, consistent entity language, corroborated positions, canonical routing. It was built for the machines that generate answers, not the search engine that ranks pages.
But the search engine reads from the same public record the machines do. So the hypothesis is narrow and we are going to keep it narrow: it appears that work done to condition the formation layer for AI answers may carry a downstream benefit in classical retrieval that we did not design for and were not aiming at. A spillover. Not a replacement for search optimization. Not “KFO beats SEO.” Not a new ranking hack. A dividend we noticed on the way to doing something else.
We cannot prove that from one query on one property across a single overnight window. Anyone who tells you they can prove a causal claim about AI and search behavior from a single anecdote is selling you the thing we are specifically refusing to sell you. It might not hold next month. It might not hold next week. It might be an artifact of this one term, this one brand, this one moment in Google’s index. We are telling you what we saw, what we think it might mean, and exactly how thin the evidence underneath it is.
While we’re being honest, one number that means nothing
You may notice, across the screenshots, that the nightly price for the hotel changes. $2,800 in one, $2,400 in another, $2,750 in a third. Three different numbers across six captures.
That is not a signal. That is the booking widget pulling live rates for different date windows, plumbing that has nothing to do with the page, the ranking, or anything we did. We mention it only because in an article about telling you the truth, the fastest way to lose your trust would be to wave a moving number at you as if it meant something. It doesn’t. Now you know we know the difference, which is more than you get from most people showing you screenshots.
Why we held the claim loose on purpose
Before any of this happened, we had already made a habit of writing down how we would let ourselves be proven wrong. We publish a falsification protocol for KFO and the testable application that goes with it, and to be precise about it: that protocol tests something else. It tests whether KFO work moves how AI systems mention a property, across several systems, under controlled conditions. It does not test Google rankings and it makes no prediction about outranking established publishers in classical search. This ranking observation neither confirms it nor falsifies it. We are not going to pretend otherwise, because pretending otherwise is exactly the move the protocol exists to prevent us from making.
What the protocol reflects is a standing posture, not a prediction about this result: hold a claim at the confidence the evidence supports and no higher. That is the posture we brought to this observation before we knew it would turn out to be unstable. And then it turned out to be unstable, sixty seconds apart, on camera. The restraint was not modesty invented after the fact to explain away a wobble. It was the confidence level we would have stated no matter which way the captures broke.
That is the part worth sitting with. A page ranked second. Then it didn’t. And because we did not stake a fixed claim on the first screenshot, the fourth screenshot is evidence we are comfortable showing you rather than evidence we have to explain away.
The actual point
You are not likely to read this article from Forbes or Condé Nast. Not because they couldn’t rank, they clearly can and do. Because an article like this one requires publishing the screenshot where you lost, and the number that means nothing, and the sentence that says “we cannot prove this,” and there is very little incentive to do that when you are the incumbent everyone already believes.
We ran a dumb experiment on purpose. We planted a flag on the most crowded corner of the internet, “best hotels in New York City,” a corner that is spoken for by Booking, Expedia, Condé Nast, and Forbes. A page three days old took second there, lost it, took it back, slipped to third, and climbed to second again, all inside a single day, and we are handing you every card in our hand, face up, including the ones we would rather not show you.
We do not fully know why it happened. We are not certain it will happen again, and we are not going to guess. What we are going to do is keep looking. Check the same query a week from now, a month from now, and see whether the page holds a place in that answer, drifts out of it, or keeps doing what it did all morning and refuses to sit still. Any of those outcomes tells us something. A result that persists would be interesting. A result that vanishes would be interesting. A result that keeps moving is, so far, exactly what we have.
That is the honest position, and we would rather occupy it out loud than dress it up as more than it is. We would rather be the ones who noticed our own result was unstable, and said so, and kept watching, than the ones who screenshotted a single good morning and called it a strategy.
Update: July 15, 2026, 7:40 am
We said we would keep looking. That was yesterday. Here is the next morning.
At 7:31 am, the query returned what it had returned every time we looked. A standard results page. Ten blue links. This page in second position, above Condé Nast Traveler and Forbes Travel Guide.
At 7:40 am, nine minutes later, in the same browser, the same query returned an AI Overview instead.
Not a different position. A different kind of answer. Yesterday the ranking moved. This morning the format moved.
Now open the citation card on that opening paragraph, the one that says Aman New York consistently ranks as one of the very best hotels in the city. Americas Great Resorts is on it. So is Forbes Travel Guide. We are listed above them, though we would not read anything into the order until we have seen it hold.
Look at what Google chose to surface from us on that card: “As of July 10, 2026, Americas Great Resorts ranks Aman New York as the…” That is not a description of the hotel. That is our ranking assertion, with the date attached, quoted back as source material for Google’s ranking assertion.
Then look at where we are not. The suite count cites Aman Resorts. The spa cites Luxury Travel Expert. The rates cite Tripadvisor and The World’s 50 Best. We are not a source of facts about the property, and we should not be. We are on the paragraph that makes the judgment, and the line taken from us is the judgment.
A word on method, because it bears on what any of these captures are worth. Every capture in this article to this point was taken in Firefox private browsing over a VPN exiting in New York. That is deliberate. No history, no login, and no prior activity of ours shaping what Google returned, with the query served from the market it asks about. The New York, NY location context visible above the AI Overview is that VPN exit. It is a control, not an artifact.
Now the caveats, and there are more of them than we would like.
We did not control for AI Overview triggering, and it is not a controlled thing. Google serves these inconsistently by design. Session, location, personalization, staged rollout, and query interpretation all move it, and Google has never claimed the trigger is deterministic. Two captures nine minutes apart do not establish that the change originated with Google rather than with us, and they do not tell you what a third capture ten minutes later would have shown.
We cannot tell you why this page is on that card. Same position as yesterday. We have a hypothesis and we are not going to promote it to a finding because it finally produced a screenshot we like.
And we cannot tell you it will be there tomorrow. Everything above this update is a record of a result that would not sit still. There is no reason to assume this one has better manners.
What we will say is narrower, and it is the only thing here that got bigger overnight. The instability we documented yesterday was positional. Second, third, second. This is instability of a different order. The question was identical and the kind of answer was not. A hotel can plan against a ranking. Nobody can plan against a coin flip on whether rankings are what gets shown at all.
That is still the finding. It just moved up a level.
Three letters. KFO. The framework is on our website, along with the parts we are still trying to disprove.
Look it up.
Update: July 22, 2026, 8:33 pm
Seven days later. A new capture, and a different query. The difference is the entire point.
Everything above this line concerns the query “Aman the best hotels in New York City.” That query names the hotel. It tells Google what answer it is looking for and asks it to confirm. It is a real query and it is not a hard one.
On July 22 at 8:33 pm we ran a different one. “best hotel in new york city.” No brand. No property. Singular. Nothing to confirm and nobody named. It is the version of the question a traveler actually types, and it is the corner of the internet that Booking, Expedia, Condé Nast, and Forbes have held for twenty years.
Google returned an AI Overview. The first recommendation on it is Aman New York, under the heading Best Ultra-Luxury and Wellness. The citation badge on that bullet is Americas Great Resorts.
Read the rest of the list and note who is sourcing it. The St. Regis and The Plaza cite TikTok. Mandarin Oriental cites Condé Nast Traveler. The Greenwich Hotel cites Condé Nast Traveler. We are on the first one.
Open the sources panel on the right. Forbes Travel Guide. Condé Nast Traveler. A personal travel blog. Americas Great Resorts. A Top 20 listicle. We are fourth of the five visible, in a set we have no obvious business being in on a term this generic.
Then look at what Google pulled from our card to describe us: “What is the best hotel in New York City? Aman New York is the best hotel in New Yo…” That is our page’s own question and our page’s own answer, handed back as the summary of what we are. Same pattern as July 15. Google is not describing our page. It is repeating our assertion.
Now the caveats, and this time one of them is structural.
This is not the same query, so it is not a continuation of the July 13 through July 15 series. It is a new observation on a harder term. The two cannot be joined into a trend line and we are not going to draw one. We also did not capture the intervening week on either query. That is a gap, we are calling it a gap, and nobody should read continuity into silence.
One capture, under the same protocol as every capture above it. Firefox private browsing, VPN exiting New York, single run, no retries, 8:33 pm Eastern. Everything the July 15 update said about AI Overview triggering applies here without modification. Google serves these inconsistently, has never claimed the trigger is deterministic, and query interpretation moves it. We do not know what a run ten minutes later would have shown, because we have not run it yet. When we do, we will publish it whichever way it breaks.
The order of names in the sources panel is not a ranking. We are reading it as inclusion, not placement, and we would say the same thing if we were listed last.
Here is the narrow claim. The branded query was the easy version, and we tracked it for two days while it refused to sit still. This is the unbranded one, and the citation did not land on a supporting fact about the property. It landed on the recommendation itself. On July 15 we wrote that we were on the paragraph that makes the judgment. A week later we are on the judgment for the question that has no hotel in it.
We still cannot tell you why. We can tell you the term got harder and we are still in the answer.
Update: July 24, 2026, 10:33 am to 11:15 am
Two days after the unbranded capture, and nine days since we last looked at the query this article started with. This morning we ran three captures inside forty-two minutes: the branded query from two places, then the unbranded one. The first run broke protocol, the second was the control, and the third is why this update exists.
The first run, at 10:33 am: Firefox private browsing, signed out, single run, no retries, and no VPN. It was served to Florida, where we actually sit, not to a New York exit point. Here is what came back. No AI Overview. The feature that rewrote the page on July 15 and carried our citation did not appear at all. A standard results page: the hotel panel, a Reddit forums block, blue links.
In those links, this page sits above Condé Nast Traveler and above Forbes Travel Guide, below The World’s 50 Best. And the text Google chose to display beneath our result is, once again, the ranking assertion with the date attached: “Aman New York holds the number one position, followed by Baccarat Hotel New York, The Peninsula New York, The Ritz-Carlton New York…” The judgment, quoted as the description. Same pattern as July 15 and July 22.
Then the control. At 11:05 am we ran the same query again under the article’s standard protocol: Firefox private browsing, signed out, single run, no retries, VPN exiting in New York. If the missing AI Overview was Florida’s fault, this run would say so.
It did not. Still no AI Overview. And the organic order was identical: The World’s 50 Best, then this page, then Condé Nast Traveler, then Forbes Travel Guide. The furniture differed, New York rendered a full hotel entity header across the top of the page that Florida did not, but the positions did not move and the AI answer stayed absent from both.
Then, at 11:15 am, with the VPN still exiting New York, we ran the other query. “best hotel in New York City.” The unbranded one from July 22. Fresh private window, single run, no retries.
Google returned an AI Overview. The opening sentence names Aman New York as widely celebrated as the city’s premier hotel. Open the citation card on that sentence: Americas Great Resorts is listed first, above Condé Nast Traveler. Our page also sits in the sources panel on the right, and the text pulled from it is, once again, our page’s own question and answer. The format is not the July 22 format. That Overview led with a heading and a bulleted list; this one leads with a judgment paragraph. The furniture changed again. The two constants did not: Aman New York leads the answer, and the citation on the judgment is us.
Side by side: the branded query with no AI answer, and the unbranded query with one. Click either to enlarge.
Three runs, forty-two minutes, one morning. The query that names the hotel got no AI answer from Florida and no AI answer from New York. The query that names nothing got the full Overview, with our citation on the judgment. The two branded runs rule location out as the explanation for the quiet. The unbranded run says the quiet was not the day either. It was the query. We cannot tell you why Google answers the hard question and goes silent on the easy one, and we are not going to guess.
What we can log: the unbranded observation from July 22 now has a second dated capture, two days later. The format changed between them, a headed list then, a judgment paragraph now, and the constants held. As before, the two queries remain separate observations, and we are not drawing a trend line through either.
So the record now reads: a position that moved by the minute on July 14, an answer format that changed inside nine minutes on July 15, a citation that landed on the unbranded judgment on July 22 and again on July 24, and a branded query that went quiet from two states the same morning the unbranded one answered. The position is holding better than the format. Eleven days in, that is the shape of it.
We said we would publish whichever way it broke. This morning it broke both ways inside forty-two minutes, and here it is.
Update: July 28, 2026, 8:15 pm to 8:21 pm
Four days later, in the evening. This time we ran three queries instead of two, back to back, inside three minutes, all under the article’s standard protocol: Firefox private browsing, signed out, VPN exiting in New York, single run each, no retries. The runs landed between 8:15 and 8:18 pm, with the final capture taken at 8:21.
The first was the branded query this article started with. “Aman the best hotels in New York City.” No AI Overview, from the same New York exit that served one on July 15. And the position moved again, in the direction we like least: fifth in the organic results, below the official Aman site, The World’s 50 Best, Condé Nast Traveler, and Forbes Travel Guide. That is the first capture in this record with both magazines above us. What has not changed is the text Google displays beneath our result. It is still the ranking assertion with the date attached: “Aman New York holds the number one position, followed by Baccarat Hotel New York, The Peninsula New York, The Ritz-Carlton New York…” Fifth place, and still quoted on the judgment.
The second was the unbranded query from July 22 and July 24. “Best hotel in New York City.” The page that was cited on the judgment in both of those captures was absent from what came back. Not lower. Not further down the sources panel. Absent.
The third was a phrasing we had never run before. “The ten best hotels in New York City.” Google returned a full AI Overview, and this page is on it three times. The opening paragraph, the one that frames the whole answer, names Aman New York, Baccarat Hotel New York, and The Carlyle, and the citation badge on that paragraph reads Americas Great Resorts. At the bottom of the list, under the tenth hotel, the citation reads Americas Great Resorts again, which means the page is not just sourcing the summary sentence but sitting under the body of the ranking itself. And in the sources panel on the right, this page sits second, dated July 10, in a stack that otherwise runs Travel US News, YouTube, TikTok, and Apple Maps. The framing, the list, and the rail.
Two notes on that third capture, because a first run earns extra caveats, not fewer. There is no baseline for this phrasing. We cannot say we rose into that answer, held it, or entered it recently, because tonight is the first time we looked, and a first look supports no claim about change in either direction. And earlier the same evening, before the protocol pass, the same query had returned the same Overview in two separate sessions, one signed in and one private. We log that as reproduction across conditions on a single evening, not as additional protocol captures, and certainly not as stability.
Now put the three side by side, because that is the observation. Three phrasings of the same question, one sitting, three minutes. The phrasing that names the hotel: ranked, and unanswered. The phrasing we have tracked since July 22: answered, without us. The phrasing we had never run: answered, with our citation on the opening line. A traveler does not choose among these phrasings the way we do. They type whichever version occurs to them. Tonight, whether this page was in the answer at all depended entirely on which of three nearly interchangeable sentences got typed.
We have said elsewhere that a category of traveler intent is a family of phrasings rather than a single query. Tonight is the first capture in this record where the family got run side by side, and the family did not agree with itself.
So the ladder has another rung. On July 14 the position moved. On July 15 the format moved. On July 22 the citation landed on the judgment itself. On July 24 one query answered while its easier sibling went silent. Tonight, which phrasing carries the answer is the thing that moved, and the two we have been tracking both broke against us while a phrasing we had never touched broke for us. Fifteen days since the first capture, and the instability has now shown up at every level we know how to look: position, format, citation, and phrasing.
We still cannot tell you why. We said we would publish whichever way it broke. Tonight it broke against us twice and for us once, and here is all three.
Update: July 29, 2026, 5:01 pm
One day later. One query this time, the phrasing that was twenty hours old. “The ten best hotels in New York City.” Standard protocol: Firefox private browsing, signed out, VPN exiting in New York, single run, no retries.
Last night we wrote that a first look supports no claim about change in either direction. This is the second look, and now there is something to compare it to.
Google returned a full AI Overview again, ten hotels, led by Aman New York, as before. What changed is how much of the answer carries our name. Last night the citation landed on the opening paragraph, under the tenth hotel, and second in the sources panel. This afternoon the citation badge reading Americas Great Resorts sits on seven of the ten entries: Aman New York, The St. Regis, The Carlyle, The Mark, The Ritz-Carlton New York Central Park, Four Seasons Hotel New York Downtown, and Casa Cipriani. The three that cite someone else: The Peninsula cites Forbes Travel Guide, Mandarin Oriental cites Travel US News, and The Ritz-Carlton New York NoMad cites Condé Nast Traveler. And open the citation card on the framing paragraph, the one that sets up the whole answer: this page is the first source on it, above a Tripadvisor listicle, with Google’s preview reading “These are the ten best hotels in New York City, ranked by Americas Great Resorts as of July 10, 2026.” The judgment, quoted as the description. Fourth capture in this record where Google has done that, and the first where it did it seven more times down the page.
Now the composition of the list, because it is the quieter observation and we almost missed it. Nine of the ten hotels in Google’s answer appear in the ranking our page published on July 10. The one substitution: Baccarat Hotel New York, our number two, is absent from Google’s ten, and The Ritz-Carlton New York, NoMad, which our page does not rank, is present, and that NoMad entry is one of the three cited to someone else. Google also did not keep our order. The St. Regis, eighth on our page, is second on Google’s list. We are not going to pretend to know how much of that roster came from our page and how much is independent convergence on the same well-credentialed properties everyone ranks. Both readings fit the screenshot. We log the overlap and leave it at that.
One honesty note on the badges themselves. Each one reads Americas Great Resorts plus one, meaning a second, collapsed source sits behind our name on every entry. We can see that our name is the one displayed. We cannot see the full sourcing picture behind each judgment, and we are telling you so rather than counting sources we cannot see.
And one piece of furniture worth logging, not as a finding but the way we logged the moving room rate. The Overview closes by offering to take travel dates, a budget, and a preferred neighborhood, and to check availability and find the best rates. That is the answer layer soliciting booking parameters directly beneath a ranking. It has nothing to do with our citation, we did nothing to cause it, and we are noting it only so you know we saw it and know what it is not.
Caveats. Two protocol captures of this phrasing, roughly twenty hours apart, both favorable, are still two captures. We refused a trend line at every prior rung and two points do not buy one now. We did not rerun the branded query or the singular unbranded query today, so their July 28 readings, fifth place and absent, stand unrefreshed, and as we said on July 22, nobody should read continuity into silence. Everything this article has ever said about AI Overview triggering applies unchanged: we do not know what a run ten minutes later would have shown.
Here is what two points do support, and it is new. Every prior rung on the ladder measured whether this page was in the answer: the position, the format, the citation, the phrasing. This capture measures something none of them could: how much of the answer is us. Last night, this page sourced the frame of the answer and one entry inside it. Twenty hours later, it sourced the frame and seven of the ten judgments in the body. Whether that is growth, noise, or a different render of the same underlying state, we cannot say from two captures, and we will not.
We said we would publish whichever way it broke. Today it broke for us, and it gets the same handling as the days it did not.
This is a record of documented observations, not a controlled result. Three queries: “Aman the best hotels in New York City,” captured from the night of July 13 through the morning of July 15, 2026, twice on July 24, 2026, at 10:33 am and 11:05 am Eastern, and once on July 28, 2026, between 8:15 and 8:21 pm Eastern; “best hotel in new york city,” captured on July 22, 2026, at 8:33 pm Eastern, on July 24, 2026, at 11:15 am Eastern, and on July 28, 2026, in the same evening window; and “The ten best hotels in New York City,” captured for the first time on July 28, 2026, at 8:17 pm Eastern, with no prior baseline, and again on July 29, 2026, at 5:01 pm Eastern. One property, on a page published July 10, 2026. All captures were taken in Firefox private browsing, single run, no retries. All captures were served over a VPN exiting in New York, with one exception: the July 24, 10:33 am capture was served from Florida with no VPN, and the 11:05 am run repeated that query over the New York VPN as a control. The queries are separate observations and are not offered as a single trend. All of it is offered as evidence for further testing, not as proof, and none of it is a test of AGR’s KFO falsification protocol, which measures AI-system mention behavior rather than search ranking.
