This is the extension of the Webinar I did with Milestone on How AI Chooses Your Brand” and a deeper dive into fan-out, grounding, and the eligibility gates they create using the same prompt from the webinar, start to finish.
A family is planning a spring break trip. They open an AI assistant and type: “best all-inclusive resorts in Cancun for families.”
Four seconds later, they have an answer. Three resorts, each with a reason attached — one for its kids’ club, one for its swim-up suites, one for its beachfront location. No blue links. No scrolling. Just a decision, already partly made for them.
If your resort isn’t one of the three, the honest question isn’t “why not us?” It’s “what actually happened in those four seconds?” Most marketers have never seen the machinery, so they reach for the closest mental model they have — rankings — and assume there’s a list somewhere with their name too far down it. There isn’t. What happened is a short, mechanical pipeline, and every stage in it is something you can actually work with, once you know it’s there.
This is that pipeline, walked through once, start to finish.
The shape of it
Six things happen between the prompt and the answer:
- The prompt — a person asks a real question, with real intent buried inside it.
- The ingestion check — before anything else, the system asks whether it can even reach the content it wants to reconsider.
- Fan-out — the system rewrites and defines the query before it ever expands it, then splits it into the narrower questions a full answer would need to cover.
- Grounding — the system goes back out and checks its facts against the world.
- Eligibility Gates — everything that fan-outs and grounding surfaces gets turned into a scorecard, and every candidate is measured against it.
- The answer — whatever survives gets written into a sentence.
None of this happens because a person configured it that way for your industry. It’s the general shape of how a modern AI system answers a question it isn’t certain about. Hospitality queries just happen to be a particularly clean example, because “best,” “family-friendly,” and “all-inclusive” are exactly the kind of soft, comparative language that forces the whole pipeline to run in full.
Stage one: the prompt
“Best all-inclusive resorts in Cancun for families” looks like one request. It’s actually three claims stacked on top of each other. Best is a superlative which demands a winner. All-inclusive is a category; it excludes anything that doesn’t qualify. For families is a modifier, and it silently attaches a whole set of unstated requirements (kids’ club, pool depth, room configuration) that the person never had to type out because they assumed the system would know what “for families” implies.
That gap is created between what was typed, what was meant, and how the LLM interprets it. Understand why the next five stages exist.
Stage two: the ingestion check
Before the system does any reasoning about which resort is best, it has to answer a much more boring question first: can it actually get back to the content it wants to reconsider, right now, at the moment of the prompt?
This matters because a live AI answer isn’t only drawing on what a model memorized during training. For a query this time-sensitive — rates change, resorts renovate, availability shifts — the system typically reaches back out to the live web through retrieval-augmented generation, or RAG. That live reach-back is gated by three checks, and all three run before the system has spent a single token reasoning about your resort’s merits:
- Crawl capacity — every site has a standing budget for how much of it gets crawled, and that budget is shared across every crawler hitting it, not allocated per-bot. A slow, unstable, hard-to-parse site quietly eats into its own budget until less of it gets seen at all.
- Crawl demand — capacity has to actually get spent on you, and demand is shaped by how much of your site looks duplicated, how often it’s updated, and how relevant it looks next to everyone else’s.
- Information gain — even with capacity and demand both in your favor, a page can still be skipped if it doesn’t say anything the system doesn’t already know from elsewhere. Passing every other check and still saying nothing new is, on its own, enough to be quietly set aside.
Think of this as the RAG system doing a background check before it lets a source back in the room. It’s not evaluating whether your resort is good yet — it’s evaluating whether your resort is even reachable and worth re-checking right now. Everything from here on only applies to whoever clears this first.
Stage three: fan-out
With a set of reachable sources established, the system turns to the actual question — and the first thing it does with “best all-inclusive resorts in Cancun for families” is refuse to take it at face value.
Before any sub-questions get generated, the query itself usually gets rewritten. Not corrected for spelling — reinterpreted for inference, because the sentence as typed is doing far less work than it looks like it’s doing. Three separate problems are hiding inside those nine words, and each one has to be resolved before a scoring model can even be built.
“All-inclusive” has to be defined. The phrase means something different depending on who’s using it — sometimes it’s drinks and buffet dining only, sometimes it’s full à la carte dining and excursions, sometimes tips and airport transfers are folded in and sometimes they’re a separate line item. The system can’t score any resort against “all-inclusive” until it has settled on a working definition of what that label actually promises. That definition doesn’t come from the prompt — the prompt just used the word and assumed everyone agreed on what it meant.
“For families” has to become a persona. Nobody typed “with a kids’ club that accepts four-year-olds” or “with a pool shallow enough for a toddler” — but that’s what “for families” is quietly asking for. The system has to unpack the modifier into the actual, concrete needs it implies: age-appropriate programming, pool depth and safety, room configurations that fit more than two people, proximity of activities to where parents can still see them. None of that was in the query. All of it has to be inferred before “for families” means anything scoreable.
“Best” has to be resolved, not assumed. This is the one that does the most damage if it’s skipped. “Best” doesn’t point to one clear standard — it silently branches into at least two different philosophies of value, and the system has no way of knowing which one this particular person meant. Is “best” a five-star property with a premium price tag and top-tier amenities? Or is it a well-run, reliably clean, decently-appointed resort that gives a family real convenience and good value without the luxury markup? Research on how AI systems handle superlative queries draws exactly this distinction: an objective superlative has one explicit, measurable standard behind it — “highest-rated” points straight at a rating field. An implicit superlative, which is what “best” almost always is, carries multiple unstated criteria that the system has to infer on its own, usually by weighing whatever context clues the rest of the query offers. Here, “for families” is doing double duty — it’s not just adding kid-specific requirements, it’s also quietly nudging the interpretation of “best” away from adults-only luxury and toward convenience, reliability, and value for a group.
Only once those three things are settled — a working definition of the category, an inferred persona, and a resolved reading of what “best” means for this particular asker — does fan-out do the part that’s easier to picture: spinning the now-clarified question out into the narrower ones that will actually get checked.
- Which resorts have a strong kids’ club and activities program, matched to the inferred age range?
- What’s actually included at each all-inclusive tier, measured against the definition the system just settled on?
- What’s the beachfront and pool access like — direct beach, or a shuttle away?
- What’s the realistic price per night for a family of four, and does it read as good value or as a luxury premium?
- How far is each option from the airport, and what does the transfer actually involve?
This is the part most people picture when they hear “fan-out,” and it’s real — but it’s the back half of the process, not the whole thing. The front half, where the query gets rewritten, ambiguous terms get defined, and the superlative gets resolved, is what actually produces the scoring model everything downstream will be measured against. Skip that part of the story and fan-out looks like simple keyword expansion. Include it, and fan-out looks like what it actually is: the system deciding, on its own, what a good answer would even need to look like.
Stage four: grounding
Fan-out produced a working definition, an inferred persona, a resolved read on “best,” and five questions built on top of all three. Grounding is what goes and finds real answers to them.
Where fan-out is wide and exploratory, grounding is narrow and exacting. For each sub-question, the system isn’t looking for the ten best articles on the topic — it’s hunting for the single most current, most citable data point that lets it commit to a specific claim in writing. What’s the actual nightly rate right now, not “starting from” copy in a hero banner? Does the kids’ club have a published age range, or is “family-friendly” just marketing language with nothing under it? Is the resort still operating under this name, or did it rebrand last year?
This is also where the system is actively trying to avoid making things up. Grounding queries exist specifically to pull in verified, structured, often institutional information — official rate sheets, booking-engine availability, tourism board data — precisely because loose, unstructured prose is easy to hallucinate around and hard to verify against. A resort that only exists in flowery blog copy gives grounding nothing to hold onto. A resort with a rate published in structured, dated, easily-parsed form gives it something to cite.
Here’s the part worth sitting with: grounding sources often never appear as a visible citation at all. A page can shape the system’s confidence in a claim — “yes, this resort’s kids’ club really does start at age four” — without ever being named in the final answer. The visibility layer and the trust layer have quietly split into two different things, and most brands are still only measuring the first one.
Stage five: fan-out and grounding create eligibility gates
Fan-out decided which questions mattered — and, underneath that, what the query’s own ambiguous terms actually meant. Grounding decided which facts were true. Fold both together and what’s left is a scorecard: a working definition of “all-inclusive,” an inferred family persona, a resolved read on whether “best” means luxury or value, and a set of verified facts checked against all three. This is the part worth being precise about, because it’s also the part that gets flattened into a soundbite most often.
There is no fixed list of ten universal gates that every prompt runs through. The scorecard’s size, anywhere from two gates to N, is a direct response to the complexity of the prompt and how much qualification, validation, fan-out, and grounding actually had to happen to answer it. A narrow, factual query might resolve cleanly against two gates. A loaded, comparative one like “best all-inclusive resorts in Cancun for families” — with a category to define, a persona to infer, and a superlative to resolve — needs considerably more scaffolding before it can be scored at all.
What follows, then, isn’t a universal checklist — it’s a representative example of the types of gates this particular prompt’s complexity produced, applying the eligibility-gates framework originally laid out by SEO strategist Bill Hunt. Most of these ten will show up in some form for most comparative queries. Some won’t apply at all to a simpler prompt. And a sufficiently different query, one involving, say, regulatory compliance or medical claims, will surface gates that don’t appear here yet, because this particular prompt never needed them:
- Context — does the content match all three things fan-out just resolved: this Cancún as the destination, this type of all-inclusive, and this reading of “family” — not a generic beach page, an adults-only property, or a different inclusive tier entirely?
- Measurement — this one struggles against a qualitative prompt like “best,” because there’s no single number to check it against. It ends up covering a range — a real per-night rate is measurable, but so is a review score or a repeat-guest rate. Either way, a claim needs a figure behind it, not just “great value.”
- Comparability — largely settled during fan-out itself, where properties already got sorted into comparable buckets — beachfront versus lagoon-side, family section versus adults-only. This gate just confirms a candidate is being judged against the right bucket, not a different one.
- Freshness — the prompt never specified a date range, which puts real weight here: is the pricing current-season, or could rates vary significantly between when this content was captured and when a family actually travels? Without an explicit window, the system has to account for that variance on its own.
- Qualification — broader than it sounds: minimum-stay requirements, per-day drink limits, age cutoffs for kids’ club or teen programming, guest-count restrictions per room. Anything that qualifies or disqualifies a family before the headline rate even applies.
- Availability — the same missing-date-range problem as freshness, from a different angle: with no travel window given, are there blackout dates that would quietly rule the resort out for the dates a family actually wants?
- Trust & quality — easy to assume for an established name like Four Seasons; harder, and more necessary, for a resort without that built-in benefit of the doubt. This gate has to check whether the claims actually hold up, and whether there’s visible criticism of what’s really included once guests arrive.
- Clarity — closely tied to qualification: all those minimums, limits, and cutoffs only count if they’re stated plainly enough to extract. Buried in a PDF or vague marketing copy, they don’t function as an answer to anything.
- Penalty & restriction — overlaps with availability and qualification from a different angle again: blackout dates, prepayment requirements, per-night minimums, and cancellation windows are all restrictions that need to be disclosed, not discovered after booking.
- Reputation — the aggregate version of trust and quality: taken as a whole, is this brand reputable enough, and is the content about it verified enough, to actually put in front of someone making a real decision?
A few of these gates deliberately overlap — qualification, availability, and penalty & restriction all touch date-based limits from different angles, and reputation is really trust and quality rolled up to the brand level. That’s not redundancy for its own sake. It means the same underlying fact — a blackout date, an unverified claim — gets checked more than once, from more than one direction, before a candidate is trusted enough to cite.
A resort clears this scorecard by passing every gate that applies to it. Miss even one — say, the price is real but buried three clicks deep and never resolves to a clean number — and the resort doesn’t get a partial score. It gets set aside, however strong everything else about it was. That’s the mechanism a person is actually watching when they ask an AI assistant directly, “why didn’t you mention my resort?” and get a specific, citable reason back. It isn’t the model being arbitrary. It’s the model reading off the scorecard fan-out and grounding just built.
Stage six: the answer
Whatever survives every gate that applied gets assembled into the sentence the family actually reads. Not the resort with the best marketing copy — the resort that could be found, was worth re-checking, answered enough of the fanned-out questions, had its facts independently verified, and cleared every gate this specific prompt required.
That’s a fundamentally different game than the one most hospitality marketing teams have been trained to play. Traditional SEO could be won, in part, by appearing authoritative — enough backlinks, enough keyword density, enough confident copy. This pipeline doesn’t leave room for the appearance of quality. It’s checking whether the thing is true, whether it’s current, whether someone else independently agrees, and whether it can be lifted out of your page cleanly enough to use. There’s no equivalent shortcut for faking your way through six sequential, mechanical checks.
What this means before the next prompt gets typed
Here’s the literal version, because the soft version keeps getting missed: your content has to contain, in a form the system can pull out cleanly, the actual answer to every gate that applies to your category. Not a version of the answer. Not an invitation to go find the answer somewhere else. The answer, stated, on the page.
“Call for rates” fails the measurement gate before anything else about the property gets considered. It doesn’t matter how strong the resort is on every other dimension — if there’s no number on the page, there’s nothing to ground and nothing to score. A booking flow that takes five clicks and a phone call to surface a price a family could have compared in the first sub-query does the same damage. A source that sends the system elsewhere for the number it needs stops functioning as a source at all.
Generic amenities lists fail for a quieter reason. Half the resorts in a market run the same bulleted list — pool, spa, kids’ club, beach access, restaurants — because it was written to sound complete, not to answer anything specific. It doesn’t say what age range the kids’ club serves. It doesn’t say what’s inside the all-inclusive tier and what costs extra. It doesn’t distinguish this property from the ten others running the identical list. A gate can’t check a fact that was never actually stated, and “kids’ club” without an age range isn’t a fact — it’s a category.
The same failure shows up dressed as good copywriting. “Your kids will have the time of their life” reads well and answers nothing. It doesn’t become “supervised activities for ages 4–12, 9 a.m. to 9 p.m., water sports and arts and crafts included” on its own — and until it does, there’s no fact behind it for grounding to verify or a gate to check. The same is true of “all-inclusive” used as a label instead of a definition, especially when a dozen actual exclusions are sitting in eight-point font at the bottom of the page. That’s not a content gap, it’s a trust gap. A resort that defines its own terms clearly is behaving differently than one that lets the fine print do the defining later, and that difference gets read as a signal in the trust & quality and clarity gates both.
And this scales. A narrow, factual prompt needs a narrow, precise answer. A loaded one — “best,” “for families,” any prompt carrying a superlative and an inferred persona — needs content wide enough to answer every gate that complexity produces. The more attributes a prompt is quietly asking about, the more of those attributes your content has to state outright, because nothing implied gets credit and nothing assumed gets scored.
None of this is one team’s job to fix, either. Ingestion is partly an IT conversation — nobody clears the crawl-capacity gate if security policy blocks AI bots by default. Grounding is partly a paid-media and revenue-management conversation — your Schema.org pricing needs to match what’s actually in the booking engine, not what a campaign page said last quarter. Trust and reputation are partly a PR and reviews conversation — an independent source corroborating your claim is worth more than restating the claim yourself, no matter how well it’s written.
But the content itself is the part no other team can do for you. The days of getting a mention in someone else’s “10 Best Family Resorts in Cancún” roundup and calling that a visibility strategy are over. That kind of inclusion was built for a system that ranked pages, not one that reads a page looking for one specific fact to verify. If the fact isn’t there, stated plainly, in your own words, on your own property, no amount of appearing in someone else’s list gets you cited.
