ChatGPT 5.3 Halved Its Hotel SourcesWhat changed on March 5, 2026
TL;DR: Across 9,722 hotel responses from 4 country locales, the day ChatGPT's UI switched to GPT-5.3 is a cliff. URLs per answer fell 49% (24 → 12), unique domains per answer fell 46% (18 → 10), and the pool of domains per prompt fell 54% (132 → 61). Booking −82%, Expedia −76%, Reddit −93%, Wikipedia −93%. A Stockholm-operated network of ~17 Booking-affiliate listicle sites gained the freed-up citation slots.
Executive Summary
A model change, not a geography rollout.
Since late December 2025 we've run the same 140 world-wide hotel discovery prompts on ChatGPT's UI every day, from 4 country locales (US, GB, DE, ES). On 2026-03-05 OpenAI quietly switched the UI from GPT-5.2 to GPT-5.3 (GPT-5.4 in the API). The same prompts, unchanged, now produce a very different answer: half as many URLs, half as many domains, and a radically different source mix.
The drop is uniform across all 4 locales (−47% to −53%), so this is a model/pipeline change — not a rollout that hit one country first. The breadth of the source pool per prompt shrank from 132 distinct domains to 61. And the inline-citation rate — the share of answers where at least one URL is linked inside the response text — collapsed from 100% to 24%.
The halving
Overnight, ChatGPT cut its hotel sources roughly in half. Every headline metric moved together on March 5.
| Metric | Pre (GPT-5.2) | Post (GPT-5.3) | Δ |
|---|---|---|---|
| Captures | 6,414 | 3,308 | — |
| URLs per capture (mean) | 23.99 | 12.21 | −49.1% |
| Unique domains per capture (mean) | 18.43 | 10.03 | −45.6% |
| URLs per unique domain | 1.31 | 1.24 | −5% |
| Inline-cite rate (% captures with cited=true) | 100% | 24% | −76pp |
| Unique domains pooled per prompt | 132 | 61 | −54% |
Uniform across locales
If the drop were a geographic A/B rollout, we'd expect country deltas to differ. They don't.
| Country | Captures (pre / post) | URLs/answer | Domains/answer | Δ URLs |
|---|---|---|---|---|
| Germany | 1,249 / 739 | 24.67 → 12.12 | 18.75 → 9.82 | -50.9% |
| Spain | 1,232 / 944 | 26.24 → 12.44 | 19.25 → 10.03 | -52.6% |
| United Kingdom | 631 / 641 | 23.11 → 12.10 | 17.99 → 10.06 | -47.6% |
| United States | 3,302 / 984 | 23.07 → 12.13 | 18.09 → 10.17 | -47.4% |
The source mix tilted
Every category shrank in absolute terms, but some shrank far faster than others — and the tiny seo_directory bucket held flat, nearly tripling its share of URLs.
| Category | Pre | Post | Δ pp |
|---|---|---|---|
| OTA | 99.3% | 96.4% | -2.9pp |
| Meta | 100.0% | 98.6% | -1.4pp |
| Independent | 100.0% | 100.0% | 0.0pp |
| Editorial | 92.9% | 87.9% | -5.0pp |
| Chain | 90.7% | 63.6% | -27.1pp |
| UGC | 94.3% | 32.9% | -61.4pp |
| SEO Directory | 67.9% | 70.0% | 2.1pp |
| 43.6% | 1.4% | -42.2pp |
The inline-citation collapse
A separate — and arguably larger — behaviour change sits on top of the halving.
Every answer cited at least one URL inline — a linked phrase or number pointing into the sources panel.
Three in four answers surface a sources panel with no link to any of its entries inside the response text itself.
Winners & losers
Ranked by total URL delta across the full sample. OTAs and UGC lost hardest; a handful of listicle directories and independent editorial gained share.
| Brand | Category | Pre URLs | Post URLs | Δ |
|---|---|---|---|---|
| tripadvisor.com | Meta | 7,714 | 2,397 | -68.9% |
| expedia.com | OTA | 7,259 | 1,753 | -75.9% |
| booking.com | OTA | 5,983 | 1,076 | -82.0% |
| thehotelguru.com | Meta | 3,667 | 899 | -75.5% |
| hotels.com | OTA | 3,008 | 684 | -77.3% |
| wikipedia.org | Independent | 2,822 | 186 | -93.4% |
| Brand | Category | Pre URLs | Post URLs | Coverage pre→post |
|---|---|---|---|---|
| luxuryhotel.guide | SEO Directory | 24 | 88 | 7.9% → 17.9% |
| all-boutique-hotels.com | SEO Directory | 6 | 60 | 0.7% → 11.4% |
| boutiquehotel.guru | SEO Directory | 24 | 62 | 8.6% → 17.1% |
| couples-hotels.com | SEO Directory | 9 | 49 | 1.4% → 10.7% |
| business.reddit.com | UGC | 0 | 55 | 0.0% → 9.3% |
| hotelierschoice.com | SEO Directory | 0 | 28 | 0.0% → 5.7% |
The Stockholm listicle network
The biggest surprise is who took ChatGPT's freed-up citation slots. Not Michelin, not Condé Nast, not Marriott — a cluster of thin-content, single-operator Booking-affiliate directories.
Sample sites
All share a hero search backed by Booking.com's affiliate widget, a templated destination grid, a Scandinavian curator persona (Ted Valentin, Maja Holm, Elain Olsson, David Bachmann), and byte-identical footer boilerplate.
Infrastructure fingerprint
| Fingerprint | Value |
|---|---|
| Registrar | Gandi SAS |
| Cloudflare NS (primary cluster) | ANDY.NS.CLOUDFLARE.COM + RITA.NS.CLOUDFLARE.COM |
| Cloudflare NS (secondary cluster) | BRENDA.NS.CLOUDFLARE.COM + GRAHAM.NS.CLOUDFLARE.COM |
| Image host (shared across sites) | images.luxuryhotel.guru |
| Bulk .guide registrations | 5starhotels / poolhotels / beachhotels / luxuryhotel — all 2014-05-08 |
| About-page slug | /about-us/ identical across inspected members |
The operator
Ted Valentin, named as curator on all-boutique-hotels.com, is a Stockholm-based entrepreneur publicly known for directory sites (hitta.se among others). The other curator names — Elain Olsson, Maja Holm, David Bachmann — do not correspond to verifiable individuals and appear to be personas reused across the templated sites.
Luxury queries hit hardest
We classified each prompt into an intent tier by keyword. Luxury and high-amenity queries — which previously leaned on OTAs, chains and editorial — took the biggest hit.
| Tier | Prompts | Pre URLs | Post URLs | Δ |
|---|---|---|---|---|
| Rooftop / Pool | 5 | 26.32 | 11.30 | -57.1% |
| Luxury (5-star) | 30 | 27.74 | 12.64 | -54.5% |
| Beachfront | 7 | 27.98 | 12.91 | -53.9% |
| Family | 14 | 23.34 | 11.90 | -49.0% |
| Boutique / Design | 28 | 23.02 | 11.89 | -48.3% |
| Other | 30 | 23.50 | 12.42 | -47.2% |
| Affordable | 12 | 19.83 | 11.69 | -41.0% |
| Romantic | 14 | 20.34 | 12.15 | -40.3% |
City hotspots
St Barts and the Alpine ski markets were hit hardest — both previously pulled 30+ URLs per answer because the destinations are OTA-thin and required more sources to triangulate. Post-cutover every city lands in the same ~12-URL band.
| City | Pre URLs | Post URLs | Δ |
|---|---|---|---|
| St Barts | 33.59 | 12.28 | -63.5% |
| Courchevel | 26.67 | 12.16 | -54.4% |
| Istanbul | 25.67 | 12.02 | -53.2% |
| Barcelona | 25.27 | 12.02 | -52.4% |
| Amsterdam | 23.78 | 11.47 | -51.8% |
| Los Angeles | 23.68 | 11.42 | -51.8% |
| Rome | 23.46 | 11.67 | -50.2% |
| Saint-Tropez | 25.81 | 13.27 | -48.6% |
| Berlin | 23.65 | 12.28 | -48.1% |
| San Francisco | 22.61 | 11.84 | -47.7% |
| Megeve | 23.82 | 12.45 | -47.7% |
| Dubai | 23.30 | 12.45 | -46.5% |
| New York | 23.45 | 12.64 | -46.1% |
| Las Vegas | 23.24 | 13.05 | -43.8% |
Weekly view of the cliff
The drop isn't drift — it's a step change. Weekly means sit in a ~23-URL band for all of January and February, then collapse to ~12 in the week of March 9 and stay there.
Weekly means weighted by capture count across all 4 country locales. Cutover week is 2026-03-02. The tough week of 2026-03-23 (mean 8.95) is rollout noise — the following weeks return to a stable ~12-URL equilibrium.
How we ran the study
Data collection
- Same 140 world hotel discovery prompts, run daily from 4 country locales (US, GB, DE, ES)
- Capture the ChatGPT UI (not the API) — including the full sources panel per response
- Per URL: domain, position, cited=true/false (was it linked inside the answer text?)
- Coverage window 2025-12-25 → 2026-04-17 (87 distinct capture days, 9,722 responses)
- Residual gpt-5-2 rows after Mar 5 are dropped as rollout noise
Processing
- Brand rollup. Localised TLDs (tripadvisor.es, expedia.de, hoteles.com, fr.wikipedia.org, maps.google.com) are canonicalised to a single brand.
- Taxonomy. Every domain is bucketed into OTA, meta, chain, editorial, UGC, Google, independent, or the new
seo_directorybucket for programmatic-SEO hotel listicles. - Two views.
citation_count= every URL in the sources panel.cited=true= URLs also linked inline in the answer text. - Listicle forensics. WHOIS + Cloudflare nameserver lookups + image-host and template inspection, not just scrape signals.
Caveats
- Post-cutover sample is lighter (3,308 vs 6,414) because of the shorter post window. Country-level deltas are consistent, which mitigates this.
- The ChatGPT UI may have product-level changes stacked on the model change. We treat them as a single “March 5 intervention” because they shipped simultaneously.
- WebFetch 403'd on several listicle homepages; WHOIS and DNS evidence confirmed the network irrespective of that.
- 140 prompts is a finite set chosen for breadth of destinations and intent tiers, not statistical exhaustiveness.
Frequently asked questions
Citations per answer dropped 49%. Before Mar 5, 2026 each hotel answer consulted 24 URLs from 18 distinct domains on average. After Mar 5 the same prompts returned just 12 URLs from 10 domains. The pool of unique domains across all 140 world prompts also halved: 132 → 61. The drop is consistent across all 4 country locales we tested (US −47%, GB −48%, DE −51%, ES −53%).