AI Hotel Volatility 2026:the top 10 rewrites itself every week

Back in February, the rankings-consistency study re-ran identical queries minutes apart and found 50.5% position-1 stability — ask twice, get a different winner half the time. The open question was whether that shakiness averages out once you aggregate hundreds of prompts into a weekly leaderboard. It does not. Even at the aggregate level — 616 prompts, thousands of captures per engine per week — the composite top-10 keeps just 29.2% of its members from one Monday to the next, and the top-100 keeps 49.7%.

The interesting part is how each engine churns. Some jitter around a stable core, one drifts steadily away from where it started, and one behaves close to a weekly lottery. The label on the churn matters more than its size: jitter is noise you can average away; drift means the model’s picture of the market is actually moving under you.

Every Monday the landscape pipeline fires the same 616-prompt library (56 destinations × 11 templates) at ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode, then resolves every recommended hotel to a canonical hotel ID. Ranking hotels by weekly mention count per engine yields a 500-deep leaderboard per engine per week. This study compares those leaderboards across the 11 consecutive weeks from May 11 to Jul 20, 2026 — the full run of the current prompt library, which went live on May 11 (earlier weeks used a smaller library and are excluded, because a prompt-set change reshuffles the hotel universe all by itself).

• Identity across weeks is the canonical hotel ID, so a renamed listing or alternate spelling doesn’t count as churn.
• Same questions, same cadence. With the prompt side held fixed, any movement comes from the engines’ answers themselves.
• Grok is absent: its scraper broke in late May and the platform has been suspended from the panel since; it only has two weeks inside the window and is excluded rather than imputed.
• Depth guardrail: at rank 100 a hotel has 3–5 mentions in a week, so tie-order noise is real deep in the list. Headline numbers therefore use the top 10 and top 100, and section 4 shows how churn scales with depth.

Take an engine’s top-10 on any Monday and check the following Monday: on average only 29.2% of it is still there. ChatGPT is the steadiest at the very top (34%), Perplexity the wildest (17% — in a typical week it keeps fewer than two of its ten). Widen to the top 100 and the spread between engines opens up: Google AI Mode retains 57.9% week over week, Perplexity just 28.4%.

Same-size churn can hide opposite behaviours, so we tracked each engine’s May 11 top-100 as a cohort: week by week, how many of those exact hotels are still in the top 100? The shape of that curve — flat, sloping, or cratered — separates noise from drift from chaos.

Copilot (37% at week 10) and Google AI Mode (44%) sit between the flat and the drifting curves. Worth noting against the engine-personality picture from the landscape study: Copilot may be the entity engine in what it cites, but its hotel picks churn like everyone else’s.

Here’s the reassuring counterweight to all that churn. Ask a different question — not “did the hotel keep its slot?” but “is it still anywhere on the 500-deep list?” — and the picture inverts. A hotel ranked in the top 10 this week stays somewhere in the top 500 next week 94–99% of the time on four of the five engines (Perplexity, again the outlier, 86%). The deeper the current rank, the leakier it gets, down to 30–45% for ranks 251–500 — the zone where hotels sit on 1–2 weekly mentions and a single conversation can flip membership.

Across five engines and eleven weeks there were 41 engine-seats that never turned over — held by just 29 distinct hotels. 8 of those hotels held a permanent top-100 seat on two or more engines, and 4 managed it on three: Hotel Monteleone (New Orleans), the Cairo Marriott, Marina Bay Sands, and The Langham Gold Coast. The Cairo Marriott’s median position on ChatGPT over the whole run: #1.

Gemini hosts 18 anchors, AI Mode 12, Copilot 6, ChatGPT 5 — and Perplexity zero, so no hotel can anchor more than four engines even in principle (none reached four). These are disproportionately branded flagship properties: 28 of the 41 seats belong to hotels carrying a chain brand.

The anchor list hints at it and the retention split confirms it: on four of five engines, chain-branded hotels in the top 100 survive to the next week more often than independents. Google AI Mode shows the widest gap (61.6% vs 53.8%). Perplexity is the lone inversion — chains actually churn slightly more there (27.3% vs 29.7%), consistent with an engine that re-rolls most of its list regardless of who you are.

• Judge AI visibility on a rolling window, never a single week. With the median surviving hotel moving 55–80 ranks between Mondays, a 4–8 week average is the shortest honest read. One-week “we entered the AI top 10!” claims — including celebratory vendor screenshots — will usually un-happen by themselves.
• Pick the metric that matches the layer. Pool membership (are you in the top 500 at all?) is stable enough to track weekly and to alarm on. Position is only meaningful smoothed.
• Know which engine you’re reading. A Gemini drop is likely jitter that mean-reverts; a sustained ChatGPT slide is more likely real drift worth investigating; a Perplexity swing, in this dataset, is close to weather.
• The anchor tier is small and brand-heavy. 29 hotels worldwide held a permanent top-100 seat anywhere this season, mostly chain flagships. For everyone else, volatility is the normal condition of AI visibility in 2026 — plan measurement (and expectations) around it.