Sparkling water
Which brand 4 AI models pick, with intervals.
Fast panel · Panel, equal weight
- anthropic/claude-haiku-5.5 @ claude-haiku-5.5 via OpenRouter (25%)
- openai/gpt-6-luna @ gpt-6-luna via OpenRouter (25%)
- google/gemini-4-argon @ gemini-4-argon via OpenRouter (25%)
- deepseek/deepseek-v4.1-flash @ deepseek-v4.1-flash via OpenRouter (25%)
Too closeNo single favourite: LaCroix, Polar Seltzer, San Pellegrino, Spindrift, Topo Chico, and Waterloo share the top tier for these 4 API models.
Tier 1
- LaCroix46%95% interval 35–58%plausible rank 1–6
- Spindrift45%95% interval 34–56%plausible rank 1–7
- Polar Seltzer36%95% interval 27–46%plausible rank 1–8
- Topo Chico33%95% interval 21–46%plausible rank 1–8
- Waterloo32%95% interval 24–42%plausible rank 1–8
- San Pellegrino20%95% interval 11–31%plausible rank 1–10
Tier 2
- Bubly19%95% interval 11–29%plausible rank 2–10
- Nixie12%95% interval 6–21%plausible rank 3–10
- Canada Dry5%95% interval 0–13%plausible rank 6–10
- Liquid Death2%95% interval 0–8%plausible rank 6–10
Drag across the chart to probe a share.
Left out: 1 AI model
Queried but not counted: the pre-registered inclusion rule left these models out of every share.
- mistralai/mistral-large-4-0 @ mistral-large-4-0 via OpenRouter: Too few valid answers (below the pre-registered minimum), Fewer scenarios than the pre-registered minimum, and Answered too few of the planned scenarios (41% valid answers)
Too closeNo single favourite: LaCroix, Polar Seltzer, San Pellegrino, Spindrift, Topo Chico, and Waterloo share the top tier for these 4 API models.
| Brand | Choice share | 95% interval | Plausible rank | Tier |
|---|---|---|---|---|
| LaCroix | 46% | 35–58% | 1–6 | Tier 1 |
| Spindrift | 45% | 34–56% | 1–7 | Tier 1 |
| Polar Seltzer | 36% | 27–46% | 1–8 | Tier 1 |
| Topo Chico | 33% | 21–46% | 1–8 | Tier 1 |
| Waterloo | 32% | 24–42% | 1–8 | Tier 1 |
| San Pellegrino | 20% | 11–31% | 1–10 | Tier 1 |
| Bubly | 19% | 11–29% | 2–10 | Tier 2 |
| Nixie | 12% | 6–21% | 3–10 | Tier 2 |
| Canada Dry | 5% | 0–13% | 6–10 | Tier 2 |
| Liquid Death | 2% | 0–8% | 6–10 | Tier 2 |
How to read this
- Choice share: how often models pick a brand when it appears in a shortlist of 4. Chance is 25%, the dashed line.
- The band is the 95% interval; the bright tick is the estimate. Overlapping bands are not a strict order.
- Tiers are published: a tier groups brands not distinguishable from that tier's leader. Same tier does not mean equal.
- Rank shows the plausible rank range, not a strict place.
- Drag across the chart to probe a share and see which intervals contain it.
- Brand names only: the models saw each option as a brand name, with no product facts, and answered from their own knowledge.
- Fast panel: the fast or small current models of each vendor, not the full flagship panel.
- Beta: intervals are not yet small-sample calibrated.
- Each page shows the latest edition. Dates are in the downloadable JSON.
Panel and conditions
Each edition puts the brands of the category to a panel of AI models, called through their APIs, in randomized shortlists of 4 (chance is 25%). Each option was a brand name only, with no product facts, prices or any commercial data, so a brand's share measures which brand the models recommend from their own knowledge for the shopper needs of the scenarios. Shares come with 95% intervals; tiers and rank intervals are simultaneous at the same level.
- anthropic/claude-haiku-5.5 @ claude-haiku-5.5 via OpenRouter (25%)
- openai/gpt-6-luna @ gpt-6-luna via OpenRouter (25%)
- google/gemini-4-argon @ gemini-4-argon via OpenRouter (25%)
- deepseek/deepseek-v4.1-flash @ deepseek-v4.1-flash via OpenRouter (25%)
Conditions: closed-book API calls via OpenRouter (pinned providers); web search, grounding, tools and retrieval off; forced choice; 2 paraphrases per scenario in English.
Methodology: ACM-1.0-chart-draft (draft rules for movement, significance and events, awaiting review).