Nineteen citations, six domains, eleven hotels

What we ran, and what it cost

Fan-out scan #10, template "Dublin hotel questions", run against our own house study for The Shelbourne - the same public example behind our AI Vision writeup. Two seed questions: "best hotel in Dublin" and "best luxury hotel in Dublin city centre". Two engines: Google and ChatGPT. Four runs in total, all four completed, twenty-seven seconds start to finish, $0.184 of provider spend.

Two things about the method before any of the numbers mean anything. The ChatGPT side is the OpenAI Responses API with the web search tool forced on, running gpt-4.1-mini - not a person typing into the consumer app, which has its own retrieval stack and its own personalisation. The Google side is a live organic SERP fetch that harvests the People Also Ask box and the related searches at the foot of the page. It captures questions. It does not capture AI Overview citations, which is why every one of the nineteen citations below came from the ChatGPT side and Google contributed none. That split is our instrumentation, not a finding about Google.

The Shelbourne is not a customer. Nobody briefed this. It is a house study we keep so we have something public to point at.

Fourteen of nineteen citations are Google Maps

Nineteen citation rows, eighteen distinct URLs, six distinct domains. Here is the whole distribution:

Every google.com URL has the same shape: google.com/maps/search/<hotel name>,+Dublin,+Ireland. They are map lookups, not web results. The other five are the hotels' own sites, or in The Shelbourne's case its parent brand's site on marriott.com. Third-party publishers scored zero. No TripAdvisor, no Booking.com, no Reddit thread, no "15 best hotels in Dublin" from a newspaper travel desk.

If you sell hotel rooms in Dublin, the practical reading is short. Two assets carried this answer: your Google Business Profile and your own homepage. Everything a marketing team normally worries about - getting into the roundups, working the review platforms - contributed nothing to what the engine linked.

Two questions in, twenty-six out

The twenty-six follow-ups are all distinct. Twenty-four came from Google: eight People Also Ask entries and sixteen related searches. Two came from ChatGPT, and both are the seed question typed back word for word. Across three samples per question, the model never rewrote the query it searched for. Google surfaced twenty-four variations on these two questions; ChatGPT ran them exactly as asked. Clustered by theme:

Three of the twenty-six name a platform outright: "Best hotel in dublin reddit", "Best luxury hotel in dublin city centre reddit", "Best luxury hotel in dublin city centre tripadvisor". Google is telling you, in its own related searches, that a chunk of this audience wants to hear from Reddit and TripAdvisor. And the answer layer cited neither. The question layer and the answer layer are pointing in different directions, and a content plan built off keyword tools alone only ever sees the first one.

Two more clusters, and nobody optimises for either. "Where do celebs stay in Dublin?" and "Where do celebrities stay in Dublin?" both surfaced, from different seeds. So did two questions about which area to stay in - first-timers, and families. Those are choices a guest makes before they have any hotel in mind, and they are answered by whoever wrote about the neighbourhood, not by whoever owns the hotel. This is the same gap we went through in optimising for sub-questions: the money question has a dozen questions standing in front of it.

Same question, three runs, two different citation styles

Citations from all three samples are unioned; the answer text we keep is whichever came back first. That is why a seed's citation count can run ahead of the hotels its stored answer names.

For "best hotel in Dublin", the answer we kept names five hotels and links each one to its own website: The Merrion, The Shelbourne (via marriott.com), The Fitzwilliam, The Wilder Townhouse and Anantara The Marker. The full citation set for that same seed holds thirteen rows. The other eight are Google Maps links, and they cover eight hotels - the same five plus The Westbury, The Dean and Number 31.

So the same model, given the same four words, minutes apart, answered once by linking to hotel websites and at least once by linking to map entries, and the shortlist grew by three properties along the way. The second seed behaved differently again: "best luxury hotel in Dublin city centre" produced six citations, every one of them Google Maps, no hotel website at all.

Eleven distinct hotels appear across the two questions. Three appear in both: The Merrion, The Shelbourne and The Westbury. Those three are the only ones with a claim to being a stable answer here. The other eight showed up once, in one phrasing of one question, on one afternoon. Anyone reporting "we appear in ChatGPT for Dublin hotels" off a single screenshot is reporting a coin flip.

All nineteen citations

The whole set, grouped by the question that produced it. Titles are as the engine returned them, which is why the same hotel appears under slightly different names.

In priority order

On this evidence, for a Dublin hotel:

• Treat the Google Business Profile as a primary asset, not admin. Fourteen of nineteen citations were map entries. The name, category, address and photos on that profile are what the engine handed the reader.
• Make the page that says "this is the hotel" the strongest one you have. Five citations went to hotel-owned domains: three to a bare homepage, two to the property's page on a group site. No booking engine, no offers page, no room type. The engine linked whichever URL most obviously is the hotel.
• Keep the Maps entry and the site saying the same thing. When an engine pulls a name from one and a description from the other, any disagreement between them becomes the reader's problem. This is the citation supply chain in miniature.
• Do not read one answer as a ranking. Eight of the eleven hotels appeared once. Run the question repeatedly and watch which names hold.
• Write the neighbourhood pages. Four of the twenty-six questions are about areas and celebrity stays, not hotels. Nobody in the citation list is answering them.

The obvious follow-ups