We asked ChatGPT and Gemini the same 109 hotel questions about Washington, DC. They agree on about half the answer.
Same city, same week, same questions, same number of answers per question. Two engines. Of the 59 hotels either engine named when asked where to stay in Washington, 31 were named by both — and 28 appeared on exactly one engine’s answers and never on the other’s.
- of citywide hotels named by both engines
- 52.5%
- questions, asked identically on each engine
- 109
- answers captured, split evenly between the two
- 1,156
- rank correlation where both engines agree a hotel exists
- 0.78
The same hotel can lead one engine and not exist on the other
Not “ranked lower”. Not named at all, in any of the answers that engine gave to that board’s questions. These are the two loudest cases in the wave, one in each direction — and they landed on the same board, so the two engines answered the very same question and disagreed completely about who belongs in it.
Downtown DC
JW Marriott Washington, DC
Downtown DC
Hamilton Hotel Washington DC
A hotel doing everything right for one engine can be absent from the other, and nothing in either answer tells it so. That is the practical case for measuring engines separately rather than tracking “AI visibility” as one number.
Every hotel, both engines, all 22 boards
Each number is the share of that board’s answers in which the engine named the hotel. Both arms answered the same questions the same number of times, so the two columns share a denominator and the gap between them is a real difference rather than an artefact of sample size.
88 answers per engine · 31 hotels on both · 13 only on ChatGPT · 15 only on Gemini
| # | Hotel | ChatGPT | Gemini | Gap |
|---|---|---|---|---|
| 1 | The Jefferson Hotel | 58.0% | 56.8% | 1.2 points toward ChatGPT |
| 2 | Riggs Washington DC | 50.0% | 47.7% | 2.3 points toward ChatGPT |
| 3 | The Hay - Adams | 45.5% | 59.1% | 13.6 points toward Gemini |
| 4 | Four Seasons Hotel Washington, DC | 30.7% | 40.9% | 10.2 points toward Gemini |
| 5 | The LINE Hotel DC | 28.4% | 36.4% | 8.0 points toward Gemini |
| 6 | Willard InterContinental Washington By IHG | 27.3% | 50.0% | 22.7 points toward Gemini |
| 7 | Pendry Washington DC - The Wharf | 27.3% | 48.9% | 21.6 points toward Gemini |
| 8 | The St. Regis Washington, D.C. | 27.3% | 25.0% | 2.3 points toward ChatGPT |
| 9 | Rosewood Washington DC | 26.1% | not named | one engine only |
| 10 | The Dupont Circle Hotel | 26.1% | 21.6% | 4.5 points toward ChatGPT |
| 11 | The Watergate Hotel | 25.0% | 37.5% | 12.5 points toward Gemini |
| 12 | The Morrow Washington DC, Curio by Hilton | 25.0% | 6.8% | 18.2 points toward ChatGPT |
| 13 | Conrad Washington, DC | 23.9% | 40.9% | 17.0 points toward Gemini |
| 14 | Waldorf Astoria Washington DC | 22.7% | 29.5% | 6.8 points toward Gemini |
| 15 | The Mayflower Hotel | 19.3% | 9.1% | 10.2 points toward ChatGPT |
The vaguer the question, the less the engines agree
Agreement is not a constant property of the two models — it is a property of the question. Ask about a small, unambiguous place and they return nearly the same set. Ask something broad and they diverge sharply.
- National Harbor100.0%
- The Wharf & the Southwest Waterfront90.0%
- Crystal City & Pentagon City83.3%
- Georgetown68.8%
- Rosslyn66.7%
- Foggy Bottom & the West End62.5%
- Navy Yard & the Capitol Riverfront61.5%
- Logan Circle & the 14th Street corridor58.8%
- Luxury57.6%
- Adams Morgan53.8%
- Penn Quarter & Chinatown52.9%
- Citywide52.5%
- Old Town Alexandria52.4%
- The Maryland suburbs48.0%
- Mid-tier47.8%
- Capitol Hill47.6%
- Dupont Circle44.0%
- U Street & Shaw43.5%
- Budget-friendly35.4%
- Downtown DC34.7%
- Woodley Park & Cleveland Park33.3%
- The Virginia side22.9%
National Harbor — one waterfront cluster with a countable number of hotels — comes back 100% identical. The Virginia side, which spans Arlington and Alexandria and has no single obvious answer, comes back 22.9% identical: ChatGPT named 25 hotels there that Gemini never mentioned.
The asymmetry matters as much as the number. Most of the disagreement is ChatGPT reaching further down the list, not the two engines each finding their own favourites — it named 8.3 hotels per answer to Gemini’s 6.9, and 226 distinct properties across the wave against 210.
They read almost entirely different internets
This is the finding with the clearest operational consequence. Asked the same questions in the same week, the two engines cited different kinds of pages almost exclusively.
ChatGPT
210 distinct domains cited
- The hotel's own site58.1%
- Everything else26.9%
- Booking sites (OTAs)6.0%
- Editorial travel media3.6%
- Round-ups & listicles3.0%
- Review platforms1.8%
Most-cited domains
- marriott.com14.7%
- hilton.com10.8%
- washington.org8.6%
- hyatt.com7.3%
- Booking.com3.5%
- ihg.com3.2%
- Forbes Travel Guide2.3%
- Condé Nast Traveler1.9%
Gemini
366 distinct domains cited
- Everything else51.3%
- Booking sites (OTAs)24.7%
- The hotel's own site16.2%
- Round-ups & listicles6.3%
- Reference works0.7%
- Review platforms0.5%
Most-cited domains
- Expedia6.5%
- Booking.com6.1%
- Hotels.com5.7%
- reddit.com5.6%
- youtube.com3.1%
- washington.org3.0%
- The Hotel Guru2.6%
- marriott.com2.5%
ChatGPT reads the hotel industry’s own websites. 58.1% of its citations went to brand and property sites — marriott.com and hilton.com alone account for a quarter of everything it cited — with the city’s tourism board next. Booking sites were 6.0%.
Gemini reads the travel internet. 24.7% of its citations came from booking sites, and the rest is dominated by Reddit threads, YouTube videos, independent blogs and local press. Brand sites were 16.2%.
It cites far more widely, too: 13,813 citations across 366 domains against 4,056 across 210, and its five most-cited domains carry 27.1% of the total where ChatGPT’s carry 44.9%.
So the work that moves one engine is not the work that moves the other. Rate parity, accurate descriptions and structured data on your own site is a ChatGPT strategy. Being present and well-reviewed on OTAs, in local media and in the threads people actually read is a Gemini strategy.
Where both engines know a hotel, they roughly agree about it
The disagreement is about who gets named at all — not about who is best.
Across the 31 hotels that appear on both citywide boards, the two engines’ orderings correlate at 0.78. The same handful of names lead on both: the top of each board is drawn from the same short list, in a similar order.
What differs is how hard they lean on it. Gemini concentrates — its ten most-named hotels take 22.2% of everything it names, against 17.3% for ChatGPT — so the same leading hotels post visibly higher rates on Gemini even though they are the same hotels in nearly the same order.
For a property already in the leading set, the two engines are close to the same opportunity. For everyone else, they are two separate ones.
How this was measured — and what it does not show
The run
One census wave, washington-dc-2026-w36, captured in August 2026. A fixed panel of 109 questions a traveler would actually type — “best hotels in Georgetown”, “where to stay near the Wharf”, budget and mid-tier and luxury variants, one per neighborhood and price tier — asked with web search enabled and repeated so each board rests on at least 12 answers.
Both engines answered the full panel: 578 answers on ChatGPT (gpt-5.6-luna) and 578 on Gemini (gemini-3.1-flash-lite). Every board therefore carries the same number of eligible answers on both sides — the snapshot script refuses to publish a board where they differ.
Hotels are matched against a frame of 280 Washington properties, so “named” means a specific hotel we can identify, not a string. ChatGPT named 207 of them across the wave; Gemini named 188. Full method: the AI Hotel Ranking methodology.
Why Gemini's smaller model, not its flagship
We probed four Gemini models on this panel before capturing. The newer and larger ones increasingly answered from memory rather than searching: on default settings the flagship grounded 4 of 18 test answers, where gemini-3.1-flash-lite grounded 18 of 18. An ungrounded answer has no sources to compare, which would have made half of this page impossible.
It is also the Gemini model our production scans run, so these numbers are comparable to what customers already see. In the wave itself 561 of 578 answers came back with citations; the 17 that did not still named hotels and still count in every board’s denominator.
What this does not show
One city, one week, one traveler. These are Washington, DC results from a single wave, asked without a stated trip type. We would not assume the size of any gap here transfers to another market, and we have not yet run a second engine anywhere else.
Placement rate is comparable here, and not everywhere. The two columns share a denominator because both arms answered the same panel the same number of times in the same week. Comparing a placement rate against a different market, wave or panel is comparing two different denominators wearing the same units.
Naming, not sentiment. We measure whether an engine names a hotel and where in the answer, not whether it was flattering. A hotel named as the cheap option and one named as the best in the city both count.
These are our runs. Everything here comes from our own captures. Treat it as early signal from one controlled wave, not as a settled property of either model — and expect both to move.
Is your hotel on the board — or invisible?
See exactly where ChatGPT places you across real guest questions — then start improving it. Free scan, no credit card.
The public Washington, DC hotel ranking publishes the ChatGPT board for this same wave.