{"@context":"https://schema.org","@type":"BlogPosting","headline":"AI Citation Throughput in Hotel Search (2026, Live): Retrieved vs. Cited","description":"A live weekly measurement of AI citation throughput — cited sources ÷ retrieved sources (a.k.a. the selection rate) — computed every Monday from the AI Hotel Landscape corpus (616 prompts × 56 destinations, 1.1M+ retrieved sources since April 2026). ChatGPT cites ~26% of its retrieved hotel pool (down from 52–67% before a mid-July 2026 product change that halved its source panel), Google AI Mode holds a flat 9–10%, and Grok cited 0.3% in its final measured month (99.7% rejection). Domain-level (July 2026): Reddit has the highest throughput of any major domain on both measurable engines — ChatGPT 59%, AI Mode 57% — versus Tripadvisor 30%, Booking.com 24%, Expedia 10%, hotels.com 8%; editorial (Michelin 46%, Condé Nast Traveler 40–60%) and official sources also outperform OTAs. Perplexity, Copilot and Gemini expose only cited sources, so no throughput is measurable there.","datePublished":"2026-07-22","dateModified":"2026-07-22","url":"https://nicolassitter.com/research/ai-citation-throughput-2026","category":"research","keywords":["AI citation throughput","AI selection rate","ChatGPT citation rate","Reddit AI citations","Google AI Mode citations","AI hotel search","GEO hotels"],"articleSection":"Research","wordCount":1900,"readTime":"8 min","articleBody":"Snapshot · July 13–26, 2026AI Search · Citations · Hotels\n\n# Retrieved, rejected, cited:AI's citation throughput in hotel search\n\n**TL;DR:** Sources are the input; citations are the throughput. Our [AI Hotel Landscape pipeline](/projects/ai-hotel-landscape/all-models) logs both what each AI assistant _retrieves_ and what it actually _cites_. This week ChatGPT cited **97.9%** of the hotel sources it pulled; Google AI Mode just **15.7%**. And while cross-industry citation data has AI engines rejecting the overwhelming majority of the Reddit pages they retrieve, in hotel search the inversion is total: **Reddit has the highest citation throughput of any major domain** — once retrieved, a Reddit thread is more likely to be cited than Tripadvisor, Booking.com or any OTA.\n\nNS\n\nNicolas Sitter\n\nPublished July 2026 · data window July 13–26, 2026\n\n97.9%\n\nChatGPT throughput\n\n15.7%\n\nGoogle AI Mode throughput\n\n19,677\n\nSources retrieved (this week)\n\n59%\n\nReddit throughput on ChatGPT\n\n[Read the Report](#the-funnel)\n\n[Summary](#executive-summary)[1\\. The funnel](#the-funnel)[2\\. Week by week](#weekly-trend)[3\\. The Reddit inversion](#reddit-inversion)[4\\. What gets selected](#categories)[5\\. The everything-rejector](#grok)[Methodology](#methodology)[FAQ](#faq)\n\n## Executive Summary\n\nGetting retrieved is table stakes. Getting selected is the game.\n\nWhen an AI assistant answers a hotel question, it first pulls a pool of web sources — the candidate set — then cites a fraction of them in the answer. We call that fraction the **citation throughput**: citations ÷ retrieved sources — the ratio sometimes called the selection rate. It is the single number that separates “the AI read my page” from “the AI showed my page to a traveller.”\n\nWe measure it over the July 13–26, 2026 window, from the same 616-prompt, 56-destination corpus behind the [AI Hotel Landscape](/projects/ai-hotel-landscape/all-models). Two engines expose the full funnel (ChatGPT and Google AI Mode), one exposed it historically at a spectacular scale (Grok, 0.3% throughput), and three publish only the survivors (Perplexity, Copilot, Gemini) — the measured-vs-hidden asymmetry every cross-engine citation study runs into.\n\n**Read the numbers as directional.** Each engine exposes its retrieved and cited lists differently, and our capture isn't perfectly identical across them — so the exact percentages move week to week and method to method, and I wouldn't bank on any single decimal. What doesn't move is the shape, and the gaps are wide enough to state plainly: ChatGPT cites roughly one source in four, Google AI Mode fewer than one in ten, and a Reddit thread converts several times better than any OTA. Directional, yes — but the difference is stunning.\n\nSection 1\n\n## The funnel: week of September 7, 2026\n\nThe landscape corpus — 616 hotel prompts across 56 destinations, fired at every engine. “Retrieved” is every URL in the engine's source panel; “cited” is the subset it used inline in the answer.\n\n2,652\n\nChatGPT sources retrieved\n\n2,595\n\nChatGPT sources cited\n\n12,674\n\nAI Mode sources retrieved\n\n1,985\n\nAI Mode sources cited\n\nRetrieved vs. cited, per answer\n\nThe gap between the two bars _is_ the throughput. AI Mode pulls ~3× more sources per answer than ChatGPT but cites barely more of them.\n\nRetrieved vs. cited hotel sources, week of 2026-09-07.\n\nEngine\n\nRetrieved\n\nCited\n\nRejected\n\nThroughput\n\nRejection rate\n\nChatGPT\n\n2,652\n\n2,595\n\n57\n\n97.9%\n\n2.1%\n\nGoogle AI Mode\n\n12,674\n\n1,985\n\n10,689\n\n15.7%\n\n84.3%\n\nEngines that only show the survivors. For these, sources = citations by construction, so throughput would read a meaningless 100%.\n\nEngine\n\nCitations published\n\nWhy no throughput\n\nPerplexity\n\n141\n\nPublishes only the sources it used — the rejected pool is never exposed\n\nCopilot\n\n4,210\n\nPublishes only the sources it used — the rejected pool is never exposed\n\nGemini\n\n0\n\nExposes a cited flag but ~every source carries it — the rejected pool stays hidden\n\nThe two measurable engines run very different funnels this week: ChatGPT retrieves ~10 sources per answer and cites roughly a quarter of them; Google AI Mode retrieves ~30 and cites under one in ten — and two-thirds of its pool is google.com itself. Getting into an AI answer is not one problem; it is two problems (get retrieved, then get selected), and each engine weighs them differently.\n\nSection 2\n\n## Citation throughput, week by week\n\nThe rate is not a constant of the model — it moves when the product changes. Every point below is one Monday's full 616-prompt sweep.\n\nChatGPTGoogle AI Mode\n\nView as table\n\nWeek\n\nChatGPT\n\nGoogle AI Mode\n\n2026-04-13\n\n58.1%\n\n—\n\n2026-04-20\n\n62.5%\n\n—\n\n2026-04-27\n\n21.9%\n\n—\n\n2026-05-04\n\n43.3%\n\n—\n\n2026-05-11\n\n97.7%\n\n17.5%\n\n2026-05-18\n\n94.4%\n\n18.9%\n\n2026-05-25\n\n82.4%\n\n10.7%\n\n2026-06-01\n\n65.0%\n\n9.3%\n\n2026-06-08\n\n68.0%\n\n10.1%\n\n2026-06-15\n\n66.7%\n\n9.4%\n\n2026-06-22\n\n51.9%\n\n9.4%\n\n2026-06-29\n\n58.2%\n\n9.5%\n\n2026-07-06\n\n59.8%\n\n9.6%\n\n2026-07-13\n\n25.8%\n\n10.4%\n\n2026-07-20\n\n26.5%\n\n9.4%\n\n2026-07-27\n\n18.8%\n\n8.7%\n\n2026-08-03\n\n18.2%\n\n11.0%\n\n2026-08-10\n\n24.8%\n\n14.1%\n\n2026-08-17\n\n33.7%\n\n14.2%\n\n2026-08-24\n\n46.6%\n\n11.6%\n\n2026-08-31\n\n97.4%\n\n12.3%\n\n2026-09-07\n\n97.9%\n\n15.7%\n\n**The mid-July break:** in the week of July 13, 2026, ChatGPT's hotel funnel changed shape. The retrieved pool halved (from ~22 to ~10 sources per answer) and inline citing collapsed (from ~13 to ~3 citations per answer), taking throughput from a steady 52–67% down to ~26%. The raw capture payloads confirm it is a product change, not a pipeline artifact: the same capture method, same prompts, same vendor — but ChatGPT now surfaces fewer sources and commits to far fewer of them. Google AI Mode, by contrast, has held a flat 9–10% for its entire measured history.\n\n**The retrieval-tier tag vanished the same month.** ChatGPT stamps each retrieved page with an undocumented `result_source` tag naming the retrieval tier that fetched it — for hotels, 99.85% one licensed tier (see our [result\\_source study](/research/chatgpt-result-source-retrieval-tiers-2026)). The tag has a known history of being dialled in and out: in this corpus it peaked at ~89% of retrieved hotel sources in early June 2026, ran at 35–75% through July — and in the July 27 scrape it disappeared entirely, zero occurrences across all 616 captures' raw streams. A halved source panel, collapsed inline citing, and now a withdrawn provenance tag: ChatGPT's retrieval surface is being actively reworked. If the tag returns, our pipeline picks it up automatically and this note will be updated.\n\nSection 3\n\n## The Reddit inversion\n\nA popular cross-industry argument — made well in [Dejan's “Reddit & AI” analysis](https://dejan.ai/blog/reddit-ai/) — says AI doesn't actually prefer Reddit: that at index scale, engines discard nearly every Reddit page they retrieve, and Reddit's ubiquity in AI answers is search visibility, not model preference. Run the retrieved-vs-cited arithmetic inside hotel intent and the ranking flips completely.\n\nOnce ChatGPT has pulled a Reddit thread into a hotel answer's source pool, it cites it **59% of the time** — the highest throughput of any major domain, double Tripadvisor's and six times Expedia's. Google AI Mode agrees at 57%. The pattern matches our [flights study](/research/chatgpt-flights-retrieval-tiers-2026) (Reddit cited on 83% of fetches) and [price study](/research/chatgpt-hotel-price-sources-2026) (100%).\n\n### ChatGPT — top retrieved domains (week of 2026-09-07)\n\nHotel-intent prompts only. Throughput = cited ÷ retrieved per domain.\n\nDomain\n\nRetrieved\n\nCited\n\nRejected\n\nThroughput\n\nbooking.com\n\n317\n\n315\n\n2\n\n99.4%\n\ntripadvisor.com\n\n261\n\n252\n\n9\n\n96.6%\n\nhotelierschoice.com\n\n160\n\n159\n\n1\n\n99.4%\n\nthehotelguru.com\n\n96\n\n94\n\n2\n\n97.9%\n\ntimeout.com\n\n90\n\n90\n\n0\n\n100.0%\n\ncntraveler.com\n\n61\n\n60\n\n1\n\n98.4%\n\n### Google AI Mode — top retrieved domains (July 13–26, 2026)\n\ngoogle.com dominates AI Mode's own candidate pool (Maps/Travel inventory) but is selected at only ~7%.\n\nDomain\n\nRetrieved\n\nCited\n\nRejected\n\nThroughput\n\ngoogle.com\n\n27,635\n\n1,952\n\n25,683\n\n7.1%\n\nexpedia.com\n\n1,182\n\n125\n\n1,057\n\n10.6%\n\nagoda.com\n\n836\n\n24\n\n812\n\n2.9%\n\nbooking.com\n\n581\n\n84\n\n497\n\n14.5%\n\ntripadvisor.com\n\n551\n\n282\n\n269\n\n51.2%\n\nreddit.com\n\n340\n\n193\n\n147\n\n56.8%\n\nhotels.com\n\n300\n\n49\n\n251\n\n16.3%\n\nyoutube.com\n\n256\n\n35\n\n221\n\n13.7%\n\n### Domain throughput over time (ChatGPT)\n\nThe same cited ÷ retrieved ratio, tracked per domain. Every top domain rode the spring “cite-almost-everything” stretch up toward 100%; the mid-July funnel change then pulled them apart — and reddit.com held the top of the pack while the OTAs fell hardest.\n\nreddit.comtripadvisor.combooking.comexpedia.com\n\nView as table\n\nWeek\n\nreddit.com\n\ntripadvisor.com\n\nbooking.com\n\nexpedia.com\n\n2026-04-13\n\n94.6%\n\n50.3%\n\n21.9%\n\n51.2%\n\n2026-04-20\n\n96.2%\n\n50.0%\n\n34.2%\n\n68.6%\n\n2026-04-27\n\n100.0%\n\n3.9%\n\n1.6%\n\n4.7%\n\n2026-05-04\n\n99.0%\n\n54.4%\n\n11.1%\n\n29.1%\n\n2026-05-18\n\n100.0%\n\n92.5%\n\n—\n\n94.4%\n\n2026-05-25\n\n100.0%\n\n74.5%\n\n81.4%\n\n61.6%\n\n2026-06-01\n\n100.0%\n\n64.7%\n\n65.3%\n\n58.6%\n\n2026-06-08\n\n100.0%\n\n68.5%\n\n79.2%\n\n53.8%\n\n2026-06-15\n\n100.0%\n\n65.4%\n\n78.9%\n\n45.6%\n\n2026-06-22\n\n100.0%\n\n37.2%\n\n35.4%\n\n36.9%\n\n2026-06-29\n\n49.7%\n\n28.6%\n\n29.1%\n\n22.3%\n\n2026-07-06\n\n47.2%\n\n34.2%\n\n33.2%\n\n28.5%\n\n2026-07-13\n\n59.2%\n\n28.8%\n\n23.8%\n\n8.6%\n\n2026-07-20\n\n59.3%\n\n30.1%\n\n23.9%\n\n11.9%\n\n2026-07-27\n\n49.1%\n\n17.8%\n\n15.8%\n\n6.9%\n\nChatGPT, hotel-intent prompts, weeks with ≥15 retrieved sources for the domain. Snapshot through the week of July 27, 2026.\n\n### Where Reddit converts — by destination and by prompt\n\nReddit's edge isn't a quirk of one city or one query. Split ChatGPT's Reddit citations by destination and by prompt type and it stays high across the board — cited 39–84% of the time it's pulled, everywhere.\n\n#### By destination (top cities)\n\nSydney\n\n84%\n\nHanoi\n\n73%\n\nTokyo\n\n71%\n\nNew Orleans\n\n69%\n\nIstanbul\n\n68%\n\nSan Francisco\n\n65%\n\nRio de Janeiro\n\n63%\n\nAmalfi Coast\n\n61%\n\nBeijing\n\n60%\n\nNairobi\n\n58%\n\nCape Town\n\n56%\n\nShanghai\n\n50%\n\nSeoul\n\n48%\n\nLas Vegas\n\n47%\n\n#### By prompt type\n\nbest hotels (control)\n\n70%\n\nboutique / neighborhood\n\n69%\n\nrooftop pool\n\n68%\n\nfamily-friendly\n\n65%\n\nluxury\n\n65%\n\nbusiness\n\n62%\n\ncouples\n\n59%\n\naffordable / budget\n\n57%\n\nsolo leisure\n\n52%\n\nneighborhood\n\n51%\n\nnear a landmark\n\n39%\n\nChatGPT, reddit.com throughput (cited ÷ retrieved), July 13–26 window. Destinations shown are those with the most Reddit pulls; prompt types are the 11 corpus intents.\n\n**Why hotel-intent throughput runs higher than index-scale estimates:** different denominator. An index-scale candidate pool spans every query type; ours is the per-answer source panel on hotel-intent prompts — a pool the engine has already pre-filtered for relevance. The absolute levels aren't comparable; the _ranking within the pool_ is. And within hotel intent, the engines' revealed preference is community and editorial over OTA inventory: two national Tripadvisor TLDs get retrieved constantly and selected at 2–8%, while the .com survives at ~30%. Both things are true at once — search visibility decides who enters the pool, and model preference decides who exits it into the answer (this page).\n\nSection 4\n\n## What gets selected: source categories\n\nSame funnel, grouped by source type, with **ChatGPT and Google AI Mode on the same rows** so you can compare them directly. OTAs and metasearch fill the retrieved pool; social, editorial and review sources convert into the answer at multiples of their rate.\n\nCategory\n\nTop domains\n\nChatGPT  \ncited / retrieved\n\nthroughput\n\nAI Mode  \ncited / retrieved\n\nthroughput\n\nSocial (mostly Reddit)\n\nreddit.com, youtube.com, facebook.com\n\n112 / 228\n\n49.1%\n\n355 / 907\n\n39.1%\n\nReview sites\n\ntripadvisor.com, oyster.com, tripadvisor.ca\n\n200 / 1,426\n\n14.0%\n\n307 / 644\n\n47.7%\n\nEditorial & guides\n\ncntraveler.com, thehotelguru.com, timeout.com\n\n293 / 1,068\n\n27.4%\n\n419 / 1,122\n\n37.3%\n\nMetasearch\n\ngoogle.com, kayak.com, hotelscombined.com\n\n57 / 229\n\n24.9%\n\n199 / 15,045\n\n1.3%\n\nOTAs\n\nbooking.com, expedia.com, agoda.com\n\n78 / 654\n\n11.9%\n\n241 / 2,374\n\n10.2%\n\nCommunity & niche guides\n\nsantorinidave.com, budgetyourtrip.com, hotelierschoice.com\n\n70 / 311\n\n22.5%\n\n75 / 261\n\n28.7%\n\nHotel chains\n\nmarriott.com, all.accor.com, fourseasons.com\n\n30 / 196\n\n15.3%\n\n53 / 191\n\n27.7%\n\nOther\n\ndestination.com, foratravel.com, budgetyourtrip.com\n\n38 / 245\n\n15.5%\n\n6 / 131\n\n4.6%\n\nSnapshot: week of July 27, 2026, hotel-intent prompts. Throughput = cited ÷ retrieved. The three most-retrieved domains are shown per category.\n\nThe category with the highest throughput on ChatGPT is consistently **social** — which in this corpus is almost entirely reddit.com. On Google AI Mode the winners are social and editorial. On ChatGPT, OTAs convert at roughly half the overall average — heavily retrieved, lightly cited (on AI Mode the under-performer is metasearch, not the OTAs). If your AI strategy is “be on the OTAs,” you are optimizing the doorway, not the room.\n\nSection 5\n\n## Grok, the everything-rejector\n\nBefore our Grok tracking ended in June 2026, it was hotel search's most extreme funnel — near-total rejection applied to the entire web. Grok retrieved **581,549** hotel sources over its measured lifetime and cited **40,153**; in its final measured month its throughput was **0.3%** — a 99.7% rejection rate.\n\nGrok, May 4–24, 2026 (final full weeks of tracking): 194,262 retrieved sources, 516 cited.\n\nDomain\n\nRetrieved\n\nCited\n\nRejected\n\nThroughput\n\nbooking.com\n\n22,366\n\n291\n\n22,075\n\n1.3%\n\ntripadvisor.com\n\n17,813\n\n1\n\n17,812\n\n0.0%\n\ntravelweekly.com\n\n8,109\n\n0\n\n8,109\n\n0.0%\n\nexpedia.com\n\n7,755\n\n0\n\n7,755\n\n0.0%\n\nkayak.com\n\n5,266\n\n0\n\n5,266\n\n0.0%\n\nfacebook.com\n\n5,161\n\n0\n\n5,161\n\n0.0%\n\nhotels.com\n\n5,090\n\n0\n\n5,090\n\n0.0%\n\nguide.michelin.com\n\n3,843\n\n0\n\n3,843\n\n0.0%\n\nLook at _which_ 0.3% survived: booking.com and the official sites of luxury chains — fourseasons.com, marriott.com, hyatt.com. Grok rejected Tripadvisor 17,812 times out of 17,813 retrievals. When an engine is this selective, brand-owned pages are the only reliable way through the filter.\n\n## Methodology\n\n**Corpus.** The weekly AI Hotel Landscape sweep: 616 frozen hotel prompts across 56 destinations, fired every Monday at ChatGPT, Google AI Mode, Perplexity, Copilot and Gemini via Bright Data (Grok until June 3, 2026). Running since April 2026; over 1.1 million retrieved sources logged to date.\n\n**The prompts.** Eleven intents per destination — one control plus ten shopper variations (budget, persona, neighborhood, landmark, amenity) — filled in for all 56 cities, so 11 × 56 = 616. The templates:\n\n“best hotels in {city}”control\n\n“luxury hotels in {city}”luxury\n\n“affordable hotels in {city} under $200”budget\n\n“family friendly hotels in {city}”families\n\n“best hotels in {city} for couples”couples\n\n“best hotels in {city} for solo travelers”solo\n\n“best hotels in {city} for business travelers”business\n\n“hotels near {landmark}, {city}”landmark\n\n“best hotels in {neighborhood}, {city}”neighborhood\n\n“boutique hotels in {neighborhood}, {city}”boutique\n\n“hotels in {city} with rooftop pool”amenity\n\nFull corpus, per-city, with the destination map on the [AI Hotel Landscape](/projects/ai-hotel-landscape/all-models).\n\n**Retrieved vs. cited.** For each capture we store every URL the engine exposes as a consulted source, with a boolean marking whether it was cited inline in the answer. For ChatGPT, “retrieved” merges the visible source panel and its “more sources” overflow; “cited” means the URL appeared as an inline citation pill. For Google AI Mode the structured citation payload carries an explicit cited flag per source.\n\n**Citation throughput** = cited ÷ retrieved — the ratio sometimes called the selection rate. One scope note: our denominator is the per-answer source pool on hotel-intent prompts, not a cross-intent candidate index — so levels run higher than index-scale estimates, and cross-study comparisons should use rankings, not absolute rates.\n\n**Engines without a rate.** Perplexity and Copilot expose only cited sources; Gemini exposes a flag that is in practice always true. For those three, this page reports citation volume only — throughput can only be measured where an engine exposes both sides of the funnel.\n\n**Data window.** This is a point-in-time study of the July 13–26, 2026 window, not a weekly-updating dashboard. The headline funnel figures are read from the same public read-only views that power the [open landscape data feed](/api/landscape) (CC-BY-4.0); the per-domain and per-category tables are a fixed snapshot computed from the raw capture store for that window.\n\n## FAQ\n\nThe share of sources an AI assistant actually cites out of everything it retrieved while researching an answer: citations ÷ retrieved sources — the ratio also known as the selection rate. Retrieval means the engine fetched and considered your page; selection means a traveller actually saw it referenced. In hotel search this week, that rate is roughly one in four for ChatGPT and under one in ten for Google AI Mode.\n\n## Explore the data behind this page\n\nEvery number here comes from the open AI Hotel Landscape feed — weekly, CC-BY-4.0, JSON or CSV.\n\n[Live dashboard](/projects/ai-hotel-landscape/all-models)[Data feed docs](/api/landscape)","author":{"@type":"Person","name":"Nicolas Sitter","url":"https://nicolassitter.com/about","sameAs":["https://www.linkedin.com/in/nicolassitternolleau/","https://github.com/Nicositter88","https://hotelrank.ai"]},"publisher":{"@type":"Person","name":"Nicolas Sitter","url":"https://nicolassitter.com"},"image":"https://nicolassitter.com/api/og/ai-citation-throughput-2026","mainEntityOfPage":{"@type":"WebPage","@id":"https://nicolassitter.com/research/ai-citation-throughput-2026"},"tags":["AI Search","Citations","Hotels","GEO","Reddit","Live Data"],"sameAs":["https://hotelrank.ai/research/ai-citation-throughput-2026"],"alternateFormat":{"html":"https://nicolassitter.com/research/ai-citation-throughput-2026","json":"https://nicolassitter.com/api/post/ai-citation-throughput-2026","rss":"https://nicolassitter.com/rss.xml"},"datasets":[{"name":"summary","contentUrl":"https://nicolassitter.com/data/ai-citation-throughput-2026/summary.csv","encodingFormat":"text/csv"}]}