Getting Recommended by ChatGPT, Gemini, and Perplexity: Generative Engine Optimization for Hotels

When ChatGPT names a hotel, about nine times in ten its link goes to the hotel's own website. Getting named is the hard part. The models weigh rating, review volume and price, and they read a short list of sources most hotels have never checked. This piece covers what the research says about how AI assistants pick hotels, which sources each one trusts, and the system that fixes it: a weekly panel of real guest prompts run against each assistant, logged, checked against your facts, and turned into a fix list.

Ninety-one percent. That's the share of hotel links GPT-5.2 pointed at the hotel's own website in Nicolas Sitter's January index, which ran 2,500 guest-style prompts across 25 cities for 19,579 AI answers. Gemini sent 89.4 percent of its hotel links direct. Perplexity, 74.7.

For anyone who has spent a career fighting for the direct booking, that's an odd number to sit with.

I ran sales strategy and operations at Starwood across more than 800 hotels, and later owned ten hotels of my own. The argument in both jobs was the same one: be the name the traveler sees first, then make sure the click lands on your site and not someone else's. As an owner it showed up once a month as the OTA row on the channel report. That row was the one I read first, because when it crept up, margin went the other way.

AI assistants have quietly solved the second half of that problem for us. The link goes direct. The first half is still wide open, because the link only goes direct if the assistant names your hotel. And the assistant decides who to name from signals and sources most hotel teams have never looked at.

The Problem: An Answer Has Three to Five Slots

A search results page gave you ten blue links, a map pack, metasearch ads and room to buy your way onto it. An AI answer names a handful of hotels, in order, with a sentence on why. There's no page two.

Position inside that short list matters too. A June 2026 algorithm audit of twelve language models by Baig, Gillani and Ali found the first slot in a list was worth about $12 a night to the model's choice. That's a simulation with synthetic hotel cards, so read it as direction, not a price. The direction is clear enough.

The traffic is small and growing fast. Lighthouse data published in July put AI referrals at just under 1 percent of hotel website visits, and measured against organic search, AI's share went from about 2 percent to 3.1 or 3.2 percent in a matter of weeks. Adobe saw AI-driven visits to U.S. travel sites grow 539 percent year over year over the 2025 holidays, and those visitors stayed 65 percent longer on site than visitors from other sources. Phocuswright found 56 percent of U.S. travelers used AI on at least one trip in the past year.

One percent of visits doesn't sound like a crisis. The visit count also understates it. A traveler who asks ChatGPT for three boutique hotels in Santa Fe and gets three names has made most of the decision before any website logs a session. If you weren't one of the three, no report you own will tell you.

Then there's who gets named. 5W Public Relations ran ChatGPT and Perplexity on October 29 across six U.S. cities and published the split: the six biggest chains won 77.5 percent of budget picks and 39.2 percent of "best hotel" picks. Independents do well when the traveler asks for the best. Chains win when the traveler asks for cheap. A Cloudbeds study of 145 properties in six destinations, summarized by Stiplo, found 72.4 percent of AI-recommended hotels were branded or large-group properties, and nearly half the brands were misclassified by at least one AI platform.

Misclassified. Wrong star level, wrong neighborhood, a closed restaurant still listed. An assistant that describes your hotel wrongly is worse than one that skips it.

If you weren't one of the three hotels the assistant named, no report you own will tell you.

The Data: What the Models Weigh and Whom They Read

Two questions decide whether a hotel gets named. What does the model value when it compares hotels? And where does it read about them? The research now answers both, with caveats.

Start with what the model values. The June audit used a conjoint design, the same method market researchers use to price features. It showed models sets of five hotel cards with randomized ratings, prices, review counts and labels, 3,024 choice sets per model across twelve models, and measured which features moved the pick.

That last row will annoy a lot of reputation managers. Replying to reviews still matters to the next human who reads them. It didn't move the models. Rating and volume did.

The chain row is worth a second look too. In the simulation, the brand label itself gave no edge. In the live tests, chains win budget questions. Put those together and the chain advantage looks like coverage, not preference: the chains have more listings, more reviews and more pages for the model to find. Coverage is something an independent can build.

Now, where the models read. Sitter's index logged the domains each assistant cited and how often each one showed up across runs.

Every assistant has its own diet. Perplexity and Grok read Tripadvisor on nearly every hotel question. ChatGPT's main source moved from Wikipedia to Booking.com between two model versions a few weeks apart. Anyone selling you a single "rank on AI" fix hasn't looked at this table.

The type of question changes the sources as well. 5W sorted every source Perplexity linked in its October test.

Ask for the best and Perplexity reads editors. Ask for cheap and it reads booking platforms and chain sites. ChatGPT and Perplexity agreed on 63.3 percent of the "best hotel" picks and only 11.7 percent of the budget picks. Editorial coverage is how an independent gets into the prestige answers, and it's the one source on this list a hotel can't fake.

The last piece of the picture is the gap between what an assistant reads and what it shows. Sitter ran 30 hotel price prompts eight times each on June 25 and logged every page ChatGPT fetched against every page it cited.

Sitter's line on it is the best summary I've read: ChatGPT "treats online travel agencies as price oracles to read from," without showing them as sources. OTA pages were 46 percent of everything fetched. The model reads your Booking.com listing to learn your price and amenities, then cites a Reddit thread to explain why you're worth it.

So the OTA listing still matters, just in a different way. It's where the assistant checks your facts.

The Mechanism: An AI Visibility Monitor and a Fix Loop

You can't fix what you haven't measured, and nobody can measure AI visibility by typing a question into ChatGPT on a Tuesday. Answers vary run to run, by model version and by how the question is phrased. What works is a small piece of software that asks the same questions on a schedule, logs every answer and turns the log into a fix list. Six parts.

1. A prompt panel written in your guests' words. Thirty to fifty questions, built from how your actual guests search: "best boutique hotel in Santa Fe for a couple's weekend," "dog-friendly hotel near the plaza under $250," "quiet hotel with parking near the convention center." Include both the prestige version and the budget version of each question, because the 5W data shows they draw on different sources and reach different winners. Add a few branded questions too ("is the [hotel] worth it," "does the [hotel] have parking") to check what the assistant says when it already knows your name.

2. Scheduled runs through each assistant's API. A weekly job sends every prompt to ChatGPT with web search on, to Perplexity's Sonar API and to Gemini with Google Search grounding, three to five times each. Sitter ran each of his prompts eight times for a reason. One answer is an anecdote. Forty answers a week per assistant is a trend line. At current API and search-tool pricing, a panel this size costs tens of dollars a week to run, less than a single night in most of the rooms it is protecting.

3. A parser that logs what matters. For every answer: was the hotel named, in what position, which competitors appeared, which URLs were cited, and what facts the answer stated about the hotel. Those go into one table, by prompt, assistant and date. A model can do the parsing. The output should be plain rows a revenue manager can sort, not another dashboard.

4. A fact check against one master fact sheet. Every claim the assistants make about your property (star level, neighborhood, pet fee, parking cost, restaurant hours, distance to the thing people come for) gets compared against a single fact sheet the hotel owns. Given the Cloudbeds finding that nearly half of brands were misclassified somewhere, expect mismatches. Each one gets traced back to the cited source: an old OTA description, a stale Google Business Profile, a 2022 magazine piece about a restaurant that closed.

5. A source map that sorts every cited domain by who can change it. Some sources you control outright: your website and its schema.org markup, your Google Business Profile, the content on your OTA extranets. Some you influence: Tripadvisor and Google ratings through review volume, editorial coverage through PR, Wikipedia only if the property is notable enough to qualify on its own terms. Some you can only join honestly, like Reddit, where a general manager answering a question under their own name helps and a fake guest account will get caught and hurt you. The map tells you, per assistant, which of those to work on first.

6. A review engine wired to checkout. Rating and volume were the two biggest levers in the conjoint. So the most valuable automation in the whole loop is the dullest: a review request triggered by the checkout event in the PMS, timed for the next morning, with low scores routed to the duty manager for service recovery before anything else happens. Don't screen out unhappy guests before asking. The FTC's 2024 rule on fake reviews covers review suppression too, and the models read the whole distribution anyway.

The model reads your Booking.com listing to learn your price, then cites a Reddit thread to explain why you're worth it.

One more piece sits on the website itself, and it's where the original research on this topic is useful. The Princeton team that coined the term generative engine optimization tested nine ways of rewriting a source page and measured how much of the generated answer it won. Published at KDD 2024, their headline was a visibility gain of up to 40 percent. Adding quotations lifted their position-adjusted visibility score from 19.3 to 27.2. Adding statistics took it to 25.2, and citing sources to 24.6. Keyword stuffing dropped it to 17.7, below doing nothing.

That was a simulated engine built on GPT-3.5, so I wouldn't treat the percentages as gospel. The pattern holds up in every live study since, though. Models quote pages that state things plainly, with numbers and sources. For a hotel that means writing "twelve minutes' walk to the plaza, pets $40 a night, two dogs maximum" in place of "steps from everything, pet-friendly." It means the press quote on the about page with the outlet's name attached. It means schema.org Hotel and HotelRoom markup with the same facts, and rates an agent can see without running a script. Our earlier piece on agentic booking and the direct channel covers the rate and markup side in detail.

Last, the measuring stick. Set up a GA4 custom channel group for AI referrers (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com) and watch for the utm_source=chatgpt.com tag ChatGPT adds to many of its outbound links. The panel tells you if the assistants name you. GA4 tells you if being named turns into bookings.

The Framework: A GEO Audit Scorecard

This is the scorecard I'd put in front of an owner after the first four weeks of panel data. It fits on one page and it gets re-scored monthly.

The order of the rows is the order of the work. Fix wrong facts first, because a wrong answer does more damage than no answer. Then reviews, because they're the biggest lever. Then coverage and editorial, which take longest.

Implementation: Six Weeks From Baseline to First Fixes

None of this needs a new platform. It needs one owner on the team (usually whoever runs e-commerce or marketing), a few hours of developer time for the panel and parser, and the discipline to read the log every week.

Weeks 1 and 2: the baseline. Write the prompt panel with the front desk and reservations team in the room, since they hear how guests actually ask. Write the master fact sheet. Build the weekly job and run it twice before changing anything, so you know your starting mention rate, position and fact accuracy per assistant.

Week 3: wrong facts. Trace every mismatch to its source and fix the ones you control the same week: website copy, schema markup, Google Business Profile, every OTA extranet. Facts on your own pages can show up fast, because the assistants now search live for almost every hotel question. In Sitter's June test, 98.8 percent of price prompts triggered a web search.

Week 4: the review engine. Connect the checkout event to the review request, set the routing for low scores, and agree who answers them. This is the step that moves the biggest lever, and it's the one most often skipped because it isn't new.

Weeks 5 and 6: coverage and the site. Complete every listing on the top five cited domains for your market. Rewrite the room, location and policy pages in plain, numbered facts. Put the press quotes on the site with names attached. Start the editorial list: the three or four travel writers who cover your destination, and one honest reason for each to come and stay.

Then keep the job running. The scorecard after month two tells you whether the fixes worked. The scorecard after month six tells you whether editorial and reviews are compounding.

Two cautions.

Be wary of anyone who promises a guaranteed ranking inside ChatGPT. The answers change by model version, by run and by phrasing, and the most-cited domain for one assistant changed completely between two versions released weeks apart. A monitor shows you the trend. Nobody controls the output.

And don't treat this as a marketing side project. An independent's visibility has always been built one source at a time. Collective Retreats, which I founded and ran across ten locations until it was acquired in 2025, never had a global brand name to lean on, and the AI answers make that same math more visible: the models reward the properties with the most consistent facts, the best rating and the most places saying good things about them. That's an operations standard as much as a marketing one.

If you want an outside read on how the assistants see your hotel today and what that's worth in room revenue, our AI revenue optimization and forecasting work starts with exactly this prompt panel and scorecard. The reputation management piece covers the review engine in more depth.

The links already go direct. The work now is getting named.

Frequently Asked Questions