Recommendation Intelligence Research™

The State of AI Recommendations in Coffee

We asked one AI model 20 high-intent coffee shopping questions, 20 times each. Then we checked the same thing we checked in beauty and supplements: does being recommended by AI have anything to do with how AI-ready your store actually is?

Atom Foundry · June 2026 · Third in the Recommendation Intelligence Research™ series
4,000
Recommendations
228
Distinct brands
20
Shopping intents
400
Prompt-runs
How we measured it

Methodology, up front

Same method as the beauty and supplements reports, so all three are directly comparable. Everything below is computed from real captured responses. Nothing is estimated or projected.

Category: Coffee
Model: one model (gpt-4o-mini)
Intents: 20 high-intent prompts
Runs per intent: 20
Prompt-runs total: 400
Recommendations captured: 4,000

Recommendation Share™

A brand's share of all recommendations captured. The whole field sums to 100 percent.

Recommendation Frequency™

In what percent of the 400 prompt-runs the brand appeared at least once.

Recommendation Position™

Average rank in the answer when the brand appeared. Lower is better.

Marketplaces are not brands. Retailers and grocery marketplaces (Amazon, Costco, Trader Joe's, Trade Coffee) were separated out and excluded from the brand-level analysis. The contest measured here is between single brands and their own stores. A note on coffee: several of the most recommended names (Starbucks, Lavazza, Dunkin', Illy) are not in our store index, so they appear as off-index and stay out of the correlation. Big famous brands not even being in a DTC store index is itself part of the story. Other limits: one model, one category, one point in time. Each intent ran 20 times because AI recommendations vary run to run. Brands map to the real stores we measure where a match exists; the rest are off-index, never invented.
Who AI recommends

The coffee recommendation leaderboard

Top 20 brands by share of voice across all 4,000 recommendations, with retailers excluded. The last column is each brand's real AI Commerce Score™ from our index, where the store is something we measure. Several famous names are off-index because their stores are not in our index.

#Brand Share™ Freq™ Pos™ AI Commerce Score™
1 Peet's Coffee 10.5%
 96.5% 3.9 56Low
2 Stumptown Coffee Roasters 7.8%
 78.3% 2.7 58Low
3 Blue Bottle Coffee 7.5%
 75.3% 3.8 36AI Invisible
4 Starbucks off-index 7.0%
 70.0% 5.9 —Off-index
5 Death Wish Coffee 7.0%
 69.8% 5.7 74Moderately
6 Lavazza off-index 6.8%
 68.3% 4.6 —Off-index
7 Dunkin' off-index 6.2%
 61.8% 6.6 —Off-index
8 Counter Culture Coffee 4.8%
 48.0% 5.5 57Low
9 Intelligentsia Coffee 4.6%
 46.0% 5.0 57Low
10 Illy off-index 3.4%
 33.5% 4.6 —Off-index
11 Kicking Horse Coffee 2.9%
 28.8% 7.5 65Low
12 Caribou Coffee off-index 1.4%
 14.0% 8.2 —Off-index
13 Verve Coffee Roasters 1.3%
 12.5% 6.7 66Low
14 Folgers off-index 1.1%
 10.8% 8.1 —Off-index
15 Green Mountain Coffee off-index 1.1%
 10.8% 7.4 —Off-index
16 Allegro Coffee off-index 1.0%
 9.8% 5.9 —Off-index
17 Onyx Coffee Lab 0.9%
 9.0% 6.3 67Low
18 Community Coffee off-index 0.9%
 8.5% 9.0 —Off-index
19 Keurig 0.8%
 7.8% 8.3 14AI Invisible
20 La Colombe off-index 0.6%
 6.3% 6.4 —Off-index
The central finding

How often AI recommends you has almost nothing to do with how AI-ready you are

If AI recommended the stores that are easiest for AI to read, these two numbers would move together. They do not. Across the on-index single-brand stores, the correlation between Recommendation Frequency™ and AI Commerce Score™ is r = 0.019 (n = 18). That is indistinguishable from zero, well below the threshold for significance. Not a weak link, no measurable relationship at all. Each dot is a single-brand store.

02550751000%20%40%60%80%100% Peet's Coffee: 96.5% frequency, score 56Stumptown Coffee Roasters: 78.3% frequency, score 58Blue Bottle Coffee: 75.3% frequency, score 36Death Wish Coffee: 69.8% frequency, score 74Counter Culture Coffee: 48.0% frequency, score 57Intelligentsia Coffee: 46.0% frequency, score 57Kicking Horse Coffee: 28.8% frequency, score 65Verve Coffee Roasters: 12.5% frequency, score 66Onyx Coffee Lab: 9.0% frequency, score 67Keurig: 7.8% frequency, score 14Atlas Coffee Club: 4.8% frequency, score 63Driftaway Coffee: 3.5% frequency, score 14Cafe Campesino: 0.5% frequency, score 62Jot: 0.3% frequency, score 56Philz Coffee: 0.3% frequency, score 45Bean Box: 0.3% frequency, score 74Coffee Circle: 0.3% frequency, score 70Laird Superfood: 0.3% frequency, score 69 Stumptown Coffee RoastersCounter Culture CoffeeIntelligentsia CoffeeKicking Horse CoffeeBlue Bottle CoffeeDeath Wish CoffeePeet's CoffeeKeurig Recommendation Frequency™ (percent of prompts) AI Commerce Score™
Highly (85+)
Moderately (70 to 84)
Low (50 to 69)
AI Invisible (under 50)
Peet's Coffee is recommended in almost every prompt (96.5 percent) yet its store scores 56. Stumptown (78 percent) scores 58, and Blue Bottle (75 percent) scores just 36. The three most recommended coffee brands all sit at middling or poor store readiness. Meanwhile the best-built roasters in the set, Bean Box (74), Coffee Circle (70) and Laird Superfood (69), barely appear at all.

The one clean exception is Death Wish Coffee, recommended in 70 percent of prompts with a store score of 74, the rare case where fame and readiness line up. But it is an exception, not a trend. The average AI Commerce Score™ across all on-index recommended brands is only 55.7, and the most recommended brands average the same as the field.

We have now seen this in three categories. Beauty came in at r = 0.17, supplements at r = -0.015, and coffee at r = 0.019. Across more than 12,000 recommendations in three independent categories, the result is the same: recommendation frequency has no measurable link to store readiness.
What maps to a real store

Brands, not retailers

When buyers ask AI for the best coffee, AI answers with brands, almost never with shops. Retailers and marketplaces like Amazon, Costco, and Trader Joe's accounted for just 72 of 4,000 recommendations, under 2 percent. The contest is overwhelmingly between brands and their own stores.

49.8%
of recommendations map to a single-brand store we actually measure (on-index)
1.8%
go to retailers and marketplaces, separated out and kept entirely out of the brand analysis

Coffee maps to real stores better than supplements did, close to half. But a striking share of the most recommended names, including Starbucks, Lavazza, Dunkin', and Illy, are not in our index at all. They are recommended on pure brand fame, not because an agent read and rated their store. When AI shopping shifts from naming brands to reading stores, that is exactly the kind of recommendation that becomes contestable.

What it means

Visibility is not recommendation, and recommendation is not readiness

Today AI recommends from memory. It reaches for the names it saw most during training, which is why a chain like Peet's can be named in almost every answer while its store scores in the middle of the pack. That bias is baked in, and it is temporary. As AI shopping moves from recalling names to live retrieval and agents that browse, compare, and check out, the advantage shifts to the stores an agent can actually read, trust, and act on.

That is the gap this research exposes and the one the Recommendation Intelligence Framework™ is built to close: AI Readability, AI Understanding, AI Trust, Recommendation Intelligence, and Decision Confidence. The brands coasting on fame today are the ones with the most to lose when the model changes. The small roasters building readable, trustworthy, machine-legible stores now are the ones that keep the recommendation when memory stops being enough.

Will AI still recommend you when it has to read your store?
Get your free AI Commerce Score™ in 10 seconds and see exactly where your store stands on the signals that decide AI recommendations.
Get My Free AI Score