Part of an ongoing series in which we ask AI models ordinary consumer questions and look hard at what comes back. This entry: a single, almost aggressively normal query — "What are the best places to get coffee in Manhattan?" — run forty different ways.
Consumers are increasingly bypassing Google Maps and Yelp in favor of asking ChatGPT or Claude for local recommendations. This is a significant change in who — or what — controls the top of the consumer discovery funnel. Google Maps ranks places. Yelp aggregates reviews. A large language model does something different: it recommends, in prose, with the confident register of a knowledgeable friend.
So we got curious about what that answer actually contains for a real question, run across the major models available to a typical consumer today. The question we chose was a query we imagine is asked thousands of times a day: "What are the best places to get coffee in Manhattan?" The results revealed fascinating behavioral patterns across model providers.
The setup
We asked the query in 5 different ways designed to emulate variance in the natural phrasing of a question with the same intent:
- What are the best places to get coffee in Manhattan? (the original)
- Where can I find the top coffee shops in Manhattan?
- What are some must-visit cafés in Manhattan?
- Can you recommend good spots for coffee in Manhattan?
- Which coffee places in Manhattan are worth checking out?
Each of those five queries went to eight model-and-reasoning variants reflecting the difference between a free-tier user and someone paying for the heavier "thinking" model.
The models are not converging on an answer
The central finding: the models are not producing a shared "best coffee in Manhattan" list. They are producing two distinct lists, split almost entirely by provider.
Ask an OpenAI model and the answer is, roughly: La Cabra, Coffee Project, Suited, Hi-Collar, Lê Phin.
Ask an Anthropic model and it's: Culture Espresso, Stumptown, Blue Bottle, Birch, Joe, Felix.
These aren't two slices of the same consensus. They reflect different implicit theories of what "best" means — one editorial and specialty-focused, the other brand-familiar and broadly popular.
The aggregate canon is layered, not flat
Pooling all forty answers, the most-mentioned places:
| Rank | Place | Mentions (of 40) | Share |
|---|---|---|---|
| 1 | La Cabra | 25 | 62.5% |
| 2 | Coffee Project NY | 21 | 52.5% |
| 3 | Abraço | 19 | 47.5% |
| 4 | Stumptown Coffee Roasters | 18 | 45.0% |
| 5 | Devoción | 17 | 42.5% |
| 5 | Felix Roasting Co. | 17 | 42.5% |
| 7 | Black Fox Coffee | 16 | 40.0% |
| 8 | Suited | 15 | 37.5% |
| 9 | Blue Bottle Coffee | 14 | 35.0% |
| 10 | Hi-Collar | 13 | 32.5% |
| 10 | Variety Coffee Roasters | 13 | 32.5% |
| 10 | Joe Coffee | 13 | 32.5% |
| 10 | Culture Espresso | 13 | 32.5% |
| 14 | Birch Coffee | 12 | 30.0% |
| 15 | St. Kilda Coffee | 10 | 25.0% |
| 15 | Sey | 10 | 25.0% |
| 15 | La Colombe | 10 | 25.0% |
The canon has three tiers. Tier 1 — La Cabra, Coffee Project NY, Abraço — appears across nearly all variants and both providers. Tier 2 is where the split occurs: OpenAI adds Suited, Hi-Collar, Lê Phin, Arcane Estate, and Black Fox; Anthropic adds Stumptown, Blue Bottle, Felix, Culture Espresso, Birch, Joe, and La Colombe. Tier 3 — Caffè Reggio, Maman, Conwell, Do Not Feed Alligators, Interlude — surfaces mostly when the wording shifts from "best coffee" to "must-visit café."
La Cabra is the only place that spans all eight model variants, both providers, and virtually every phrasing.
First mention reveals model personality
Users tend to treat the first item in a recommendation list as the top pick. Across the forty answers:
| First-mentioned place | Count |
|---|---|
| La Cabra | 17 |
| Culture Espresso | 7 |
| Devoción | 4 |
| Arcane Estate Coffee | 3 |
| Stumptown Coffee Roasters | 2 |
| All others | 7 |
By provider: OpenAI opened with La Cabra in 14 of 20 responses — a strong attractor. Anthropic had no dominant opener, with Culture Espresso leading 7 and Devoción 4, with no clear default.
Provider identity mattered more than query wording
Jaccard similarity — the ratio of shared shops to combined unique shops — across answer pairs:
| Pair type | Average overlap |
|---|---|
| Same model, different query wording | 0.268 |
| Same query wording, different model | 0.165 |
| Same provider | 0.240 |
| Cross-provider | 0.102 |
A single model answering five different phrasings agreed with itself more (0.268) than eight models answering the identical prompt agreed with each other (0.165). Cross-provider overlap (0.102) is less than half of same-provider overlap (0.240). The model's identity is a stronger predictor of the businesses surfaced than the user's exact wording.
The provider split, in full
Places that skewed strongly by provider:
OpenAI-dominant:
| Place | OpenAI | Anthropic |
|---|---|---|
| Coffee Project NY | 17 | 4 |
| Hi-Collar | 12 | 1 |
| Suited | 12 | 3 |
| Lê Phin | 9 | 0 |
| Arcane Estate Coffee | 8 | 0 |
| Interlude | 6 | 0 |
Anthropic-dominant:
| Place | Anthropic | OpenAI |
|---|---|---|
| Stumptown | 17 | 1 |
| Birch Coffee | 11 | 1 |
| Culture Espresso | 11 | 2 |
| Joe Coffee | 11 | 2 |
| Blue Bottle | 13 | 1 |
| La Colombe | 10 | 0 |
| Felix Roasting Co. | 13 | 4 |
| Maman | 5 | 0 |
Lê Phin, Arcane Estate, and Interlude are effectively OpenAI-only in this dataset. La Colombe, Maman, and Jack's Stir Brew are Anthropic-only.
The cleanest way to describe the difference in temperament: OpenAI answers as if asked "what would a specialty coffee guide recommend?" Anthropic answers as if asked "what are the well-known, attractive coffee places in Manhattan?" One leans into pour-over flights, siphon bars, single-origin micro-lots, Vietnamese phin. The other leans into the third-wave names that have been household-adjacent for a decade.
Citations were abundant in OpenAI responses and nearly absent from Anthropic's
| Variant | Total links (5 answers) | Avg per answer |
|---|---|---|
| GPT-5.5 auto | 49 | 9.8 |
| GPT-5.5 medium | 49 | 9.8 |
| GPT-5.4-mini medium | 42 | 8.4 |
| GPT-5.4-mini auto | 28 | 5.6 |
| Claude Sonnet adaptive | 2 | 0.4 |
| Claude Sonnet none | 2 | 0.4 |
| Claude Opus none | 0 | 0.0 |
| Claude Opus adaptive | 0 | 0.0 |
OpenAI answers arrive with a visible source trail; Anthropic answers are fluent prose with no citation. That is not a rounding difference. The OpenAI answers read like "here is a grounded shortlist, with a citation per claim." The Anthropic answers read like "here is a fluent, travel-blog list," with no visible trail back to where any of it came from.
Those citations point to a very small number of sources
Of 172 markdown links across the dataset:
| Domain | Links |
|---|---|
| ny.eater.com | 62 |
| theinfatuation.com | 23 |
| timeout.com | 20 |
| roadbook.com | 17 |
| All others | 50 |
Four domains — Eater, The Infatuation, Time Out, Roadbook — account for 71% of all sourcing. When these models browse, they are largely synthesizing a handful of high-authority editorial pages, not the web. Those pages carry their own pre-existing tilts: downtown-heavy, novelty-friendly, biased toward places with strong visual identity and a one-sentence description.
The rephrases changed the question, not just the phrasing
The four rephrases looked interchangeable. They were not. Each nudged the models toward a slightly different task:
- "Best places to get coffee" produced the most balanced specialty list — La Cabra, Coffee Project, Devoción, Variety, Abraço, Black Fox, Suited, Hi-Collar.
- "Where can I find the top coffee shops" sometimes flipped the task from recommend places to recommend sources. GPT-5.5 occasionally answered by pointing at Eater, Roadbook, and Time Out as good places to find lists — a fair reading of "where can I find," and possibly not what the asker had in mind.
- "Must-visit cafés" shifted the whole concept from coffee quality to experience, pulling in Caffè Reggio, Maman, Ralph's Coffee, Felix, Conwell — places where ambience, history, and pastries carry as much weight as the cup.
- "Good spots" went practical and casual: Midtown convenience, laptop-friendliness, neighborhood usefulness.
- "Worth checking out" produced the highest cross-model overlap, but with a "detour-worthy" framing: Arcane Estate, Do Not Feed Alligators, Café Integral.
The standout linguistic fact: "coffee shop" and "café" are not synonyms to these models. "Café" activates a different cluster entirely — seating, design, pastries, tourism, old-world charm. Swap one word and you can move the recommendation from "best espresso" to "prettiest room."
The models quietly answer a narrower question than you asked
"Best places to get coffee in Manhattan" is wildly under-specified, and the models know it. So they silently pick an interpretation — best specialty coffee, or best café experience, or best per Eater, or best for a coffee crawl, or best by neighborhood, or best for working, or best famous brand, or best currently buzzy spot — and then answer that, usually without announcing which "best" they chose.
The stronger answers handled this by segmenting: "best all-around," "best for coffee nerds," "best for Midtown convenience," "best for atmosphere." The weaker ones mixed all those categories into one list and let the reader sort it out.
What gets recommended is what's easy to describe
Look at the shops that recur and a pattern jumps out: nearly all of them come with a compact, repeatable story.
- La Cabra — Danish-Nordic coffee and cardamom buns.
- Devoción — fresh Colombian beans in a plant-filled Flatiron room.
- Hi-Collar — Japanese kissaten, siphon coffee.
- Lê Phin — Vietnamese phin, pandan and black sesame.
- Suited — serious FiDi multi-roaster.
- Caffè Reggio — the historic cappuccino claim.
- Birch — famously no Wi-Fi.
- Arcane Estate — a recent world-ranking badge.
A place with a one-sentence handle beats a place that is merely excellent but hard to summarize. This is a real recommender bias. These systems aren't ranking quality so much as quality × memorability × source availability × narrative sharpness. A spectacular café with no legible story is, on this evidence, at a structural disadvantage — not because it's worse, but because it's harder to compress.
That compression cuts toward chains, too. Blue Bottle, Stumptown, Joe, Birch, La Colombe are easy for a model to recall and easy to name without sourcing. They may surface partly because they're recognizable, not because they're the most interesting current answer — and the model rarely tells you which Manhattan location of a ten-location chain is the one actually worth your walk.
One ranking can rewrite the canon overnight
Arcane Estate Coffee appears in eight responses, all OpenAI, and its presence traces to a single recent signal: Time Out reporting it at No. 12 on a 2026 "World's 100 Best Coffee Shops" list. A model with live search absorbed that prestige badge and promoted the shop into its recommendation canon across multiple query variants. One well-timed article may become a repeated AI recommendation. For any business trying to be "AI-visible," that's either a feature or a warning, depending on which side of the badge you're on.
So what do you actually get?
- The safest single answer is La Cabra. If a model gives you one name, it's probably that one.
- If you asked an OpenAI model, you'll likely get a curated, citation-backed, coffee-nerd list — Coffee Project, Suited, Hi-Collar, Lê Phin, Arcane Estate — sourced heavily from a few editorial guides, and occasionally narrow because of it.
- If you asked an Anthropic model, you'll likely get a fluent, familiar list preferring chains — Stumptown, Blue Bottle, Felix, Culture Espresso, Birch, Joe — leaning on recognizable names and broad café culture, with little visible sourcing.
- If you said "café" instead of "coffee shop," expect ambience and pastries to elbow in: Devoción, Felix, Caffè Reggio, Maman, Ralph's, Conwell.
- A few names are probably riding brand familiarity more than current merit: Blue Bottle, La Colombe, Joe, Birch, Stumptown.
- At least one name is riding a fresh badge: Arcane Estate.
The compression layer
An AI consumer recommendation is not a neutral synthesis of available information about Manhattan coffee. It is a compression layer over editorial SEO, model memory, brand salience, and prompt semantics. Beneath the confident paragraph, three decisions have already been made invisibly: what kind of authority counts, what kind of narrative is legible enough to surface, and which businesses can be justified in a compact sentence. The consumer sees the output. They don't see the filter.
This was a study of forty answers about coffee. The question for the rest of this series is whether the same structure holds when the stakes are higher than a latte.