Did GPT-5.6 really kill listicles?Not in hotel search.
TL;DR: A widely shared analysis (Tomek Rudzki / Peec AI, based on a 1M-prompt, 180M-source panel) found that after GPT-5.6, ChatGPT ran far fewer “best/top/vs” fan-out searches, adopted site: search at scale, and cut listicle citations by half. We re-ran the same comparison — fan-outs, citations, content type, before vs. after — on our own AI Hotel Landscape corpus (616 destination-hotel prompts, weekly, since December 2025). In this vertical, none of it replicated. Fan-outs per chat: flat. Citations per chat: up 1.9%, not 25%. site: searches: zero, before and after. Listicle citation share: up 7.6%, not down 50%.
Executive Summary
A real, widely cited effect — that doesn't show up in this vertical.
In late August 2026, an analysis based on Peec AI's data (shared publicly by Tomek Rudzki) circulated widely: after ChatGPT's GPT-5.6 rollout, fan-out searches per chat roughly doubled, site:-operator searches went from near-zero to 18.4% of chats, and listicle/comparison pages' share of citations fell by roughly a third to a half. The interpretation — ChatGPT retrieving more, more precisely, and going straight to source over GEO-optimized listicles — is a plausible and important story for anyone doing AEO/GEO work.
We have our own weekly, first-party ChatGPT hotel-search corpus running since December 2025 — a natural place to check whether the same effect shows up. It doesn't. Not in fan-out volume, not in site: adoption, not in citation content mix. If anything, several numbers move in the opposite direction. This isn't a rebuttal of the original finding — a 1M-prompt, 180M-source panel is a much bigger and more diverse sample than ours. It's evidence that whatever changed in GPT-5.6's retrieval behavior, it didn't change hotel destination-discovery search the same way.
One methodology note before the numbers: July 9 is GPT-5.6's developer/API release date, not when it became ChatGPT's consumer default — that happened August 6. Our own capture pipeline's observed transition lines up with the consumer date, not the developer one. Worth checking which date a claim is actually anchored to before you slice your own data on it. See Section 1.
When GPT-5.6 actually arrived — in our data
Bright Data's ChatGPT scrape reports a model field directly in its response envelope — not something we compute, a value it observed. We checked it week by week across our hotel corpus, expecting to see a switch around the widely cited July 9 date.
| Week | Model observed | n captures |
|---|---|---|
| 2026-07-13 | gpt-5-5 | 616 |
| 2026-07-21 | gpt-5-5 | 614 |
| 2026-07-27 | gpt-5-5 | 616 |
| 2026-08-03 | gpt-5-5 | 616 |
| 2026-08-10 | gpt-5-6 | 615 |
| 2026-08-17 | gpt-5-6 | 582 |
Fan-outs: essentially unchanged
Two weeks of GPT-5.5 (Jul 27, Aug 3) vs. two weeks of GPT-5.6 (Aug 10, Aug 17), same 616-prompt hotel library, same destinations.
| Metric | Pre (GPT-5.5) | Post (GPT-5.6) | Change | Peec AI claimed |
|---|---|---|---|---|
| Fan-outs per chat | 0.959 | 0.965 | +0.7% | +154% |
| "best" share of fan-outs | 46.2% | 79.7% | +72.3% | −56.4% |
| "top" share of fan-outs | 8.7% | 8.3% | −5.2% | −75% (share); −36.5% (per-chat) |
| "vs" / "comparison" fan-outs | 0% | 0% | — | not directly reported |
| Chats with a site: fan-out | 0% | 0% | — | 18.37% (+20,018% relative) |
The site: operator result is the starkest: zero occurrences in any of the four weeks we checked, before or after. For a destination-discovery prompt (“best hotels in Bangkok”), there's no obvious single domain to target with site: the way there might be for a branded or comparison query — so this may be a query-shape effect as much as a model-version effect.
Citations: more of them, and listicles didn't lose ground
| Metric | Pre (GPT-5.5) | Post (GPT-5.6) | Change | Peec AI claimed |
|---|---|---|---|---|
| Citations per chat | 9.85 | 10.03 | +1.9% | +25% |
| Listicle share of citations | 33.8% | 36.4% | +7.6% | −50.5% (15.77% → 7.80%) |
| Comparison share of citations | 0.31% | 1.03% | +237%* | −32.1% (9.08% → 6.17%) |
Real examples of what our classifier calls a listicle, still showing up in GPT-5.6 citations: Tripadvisor's “THE 10 BEST Business Hotels in Gold Coast 2026,” Trip.com's “10 Best Hotels with Rooftop Pools in Beijing,” Latitude's “The 10 best luxury hotels in Bangkok in 2026.” These are exactly the page type the original analysis described as losing ground — and in our corpus, this category's citation share is slightly larger after GPT-5.6, not smaller.
A hypothesis for the gap, not a conclusion
The most likely explanation isn't that the original finding is wrong — it's that hotel destination-discovery is a specific, narrow query shape that may not be where GPT-5.6's retrieval changes concentrate:
- Listicles are the destination-discovery answer format, not a GEO workaround, in travel. “Best hotels in Bangkok” genuinely is a listicle-shaped question. In software/B2B GEO — the world the original analysis's examples (“top 10 [brands]”, “[X] vs [Y]”) point at — listicles and comparison pages are more often a scaled tactic competing with a vendor's own site. ChatGPT reducing reliance on them there doesn't necessarily generalize to a category where the listicle is the primary real source type.
site:search needs an obvious target domain. “Site: marriott.com best hotels in Paris” isn't a natural sub-query for an open discovery prompt the way it would be for “[brand] pricing” or “[brand] reviews.”- Our prompt library doesn't include comparison-style prompts at all. “X vs Y” fan-outs were 0% in every week we checked, before and after — there may be nothing for GPT-5.6's retrieval change to act on if the underlying query never invites a comparison framing.
None of this is proven — it's the most plausible reading of why a widely reported, real effect on a 1M-prompt panel doesn't reproduce on a 616-prompt hotel-only one. The honest takeaway for hotels specifically: don't assume the “listicles are dying” narrative applies to your category without checking your own vertical's data.
Methodology
Corpus. The AI Hotel Landscape ChatGPT corpus: 616 destination-hotel prompts (56 destinations × ~11 templates), captured weekly via Bright Data since December 2025. Four weeks analyzed here: 2026-07-27 and 2026-08-03 (GPT-5.5, “pre”), 2026-08-10 and 2026-08-17 (GPT-5.6, “post”) — the two weeks immediately either side of our own observed model transition (see Section 1), giving the “week before/after and two weeks before/after” view.
Model field. fanout_captures.model, a top-level field in Bright Data's own response envelope (not derived by us). Checked week over week from mid-July through late August to locate the actual transition.
Fan-out terms. Each capture's fanout_queries array was scanned for whole-word matches of “best”, “top”, “vs”, “comparison”, and “review”. Share of fan-outs = matching fan-out queries ÷ total fan-out queries issued that week (same framing as the original analysis's “share of total” metric). site: adoption = share of captures where has_site_operators (an existing pipeline flag) is true.
Citation content type. Our own classifier, not Peec AI's: a citation is tagged listicle if its title or URL matches a “top/best + number” pattern, and comparison if it matches “vs”, “versus”, “comparison”, or “alternatives”. Spot-checked against real citation titles before trusting it (examples in Section 3) — precise enough to trust the direction of change, not necessarily calibrated to match any other team's classifier in absolute terms.
External data. The GPT-5.6 findings this article tests against are from a publicly shared analysis by Tomek Rudzki, based on Peec AI's data (a 1M-prompt panel, ~180M sources, tracked one week before and one week after the rollout the source dated to July 9, 2026). We did not have access to Peec AI's underlying data or classifier — the numbers attributed to them above are taken from their public post and are shown for contrast, not independently verified by us.
Access. Read-only throughout: existing stored fields (model, fanout_count, fanout_queries, citation_count, citations, has_site_operators), no new captures, no writes.
FAQ
Every number here is from our own AI hotel-search captures
AI Hotel Landscape corpus, weekly since December 2025 — CC-BY-4.0. Using this data? A citation or link back to nicolassitter.com is always appreciated.
Live dashboard