The best roaster in your city is probably not the one ChatGPT names. It is more likely the one with the most words written about it, and those are different achievements that happen to wear the same clothes.
This is uncomfortable if you have spent a decade learning to roast. Quality was supposed to be the moat. In a search engine it partly was, because links followed reputation and reputation followed quality, slowly and imperfectly. In an answer engine the relationship breaks. A model recommending a roaster is not tasting anything. It is reading, and it can only read what somebody wrote down.
So the question stops being whether your coffee is good and becomes whether your coffee is legible.
Your cupping score is invisible to a language model
Specialty coffee runs on a scoring system. An SCA cupping protocol puts a number on a lot, eighty and above earns the specialty designation, and roasters who buy well can point to scores in the high eighties with justified pride.
None of that reaches a language model unless it appears in text it has read. Your Q grader’s notebook is not a public document. The score you paid for on a lot is usually mentioned once, in an email to a wholesale account, and never again.
Meanwhile the roaster three suburbs over, whose coffee you consider mediocre, has been written about in a city food guide, a regional business weekly, two subscription review sites and a Reddit thread with four hundred comments. When a model assembles an answer about roasters in your metro, it is assembling from those documents. Your scores are not in the corpus. Their name is.
The fix is not to abandon quality. It is to stop treating quality as self-evident and start treating it as a claim that needs publishing. Put the score on the product page. Name the grader. State the protocol. A sentence saying a lot cupped at 87.5 under SCA protocol is a retrievable fact. A tasting note saying “bright and complex” is not, because every roaster on earth has written it.
Name the origin, the process, and the varietal
Here is the framework worth stealing. Call it the Origin Triple: country and farm or washing station, processing method, and varietal. Three facts, stated as text, on every coffee you sell.
Most roasters publish one of the three. Nearly all say Ethiopia or Colombia. Fewer name the washing station or the producer. Fewer still specify washed, natural, honey or anaerobic. Almost none name the varietal, even though Gesha, Bourbon, SL28, Caturra and Pacamara are exactly the terms a knowledgeable buyer searches on and exactly the terms a model uses to distinguish one bag of Ethiopian from the four hundred others.
The Origin Triple matters because it is the only part of your catalogue that is genuinely unique and genuinely checkable. Roast level is a preference. Tasting notes are prose. But a washed Gesha from a named producer in Gedeb is a specific object in the world, and a model that has read about that producer elsewhere can connect your page to that knowledge.
Add a fourth field while you are in there, even though it breaks the name: harvest year. Coffee is agricultural and seasonal, and a page that says the 2026 harvest tells a retrieval system your information is current. A page with no date reads as possibly stale, and stale sources lose to fresh ones when a model picks between two otherwise equal candidates.
There is a practical objection to all of this, and it is worth answering because it is the real reason most roasters do not do it. Coffees rotate. A single origin might be on the menu for eleven weeks and then gone, and building a full page for something that short-lived feels like waste.
It is not waste, for two reasons. The first is that the page keeps working after the coffee sells out, because the terms on it are the terms buyers search: the producer, the varietal, the region, the process. A visitor who arrives looking for a washed Gesha and finds a sold-out one plus three current lots is a visitor you would not otherwise have had. The second is that a roaster with four years of archived origin pages has built something a competitor cannot assemble quickly: a substantial body of text demonstrating, in specifics, what kind of coffee this company buys. That body of text is what a retrieval system reads when it decides whether you are a serious specialty roaster or a company with a logo.
Keep the old pages. Mark them as past offerings with the harvest year attached. Do not delete them and do not redirect them to a category page, which is the most common and most costly housekeeping mistake roasters make with their catalogues.
Build pages that answer buying questions
Roaster websites are built as shops. The buyer arrives already knowing they want coffee and needs help choosing between twelve bags.
The person asking an answer engine is earlier and vaguer. They are asking which roaster ships freshest to their state, whether anyone local does a decaf worth drinking, who does a subscription that lets you skip weeks, what to buy for an espresso machine at home if they like chocolate rather than fruit, or which roaster will sell them five kilos a week for a café without a contract.
Every one of those is a page you do not have. Each one takes an hour to write and answers a question your shop pages ignore, because shop pages assume a decision already made.
Write the wholesale page with actual terms on it. Minimum order, lead time, whether you offer equipment support, whether you do bespoke blends and at what volume, what a café can expect to pay per kilo at three tiers. Cafés researching suppliers are a high value audience asking a model a procurement question, and the roasters who publish terms get shortlisted by systems that cannot phone anyone for a quote.
Write the shipping page with real numbers. Roast day. Cutoff. Which carrier. How many days to which regions. Freshness is the category’s central anxiety and most sites gesture at it instead of answering it.
Write the equipment guidance. Grind settings for common home grinders, dose recommendations, what a given coffee wants on espresso versus filter. This is the content roasters give away for free in conversation every day and never publish, and it is precisely the material that gets a page cited as a source rather than skimmed as a shop.
Get into the sources AI already reads
You can do everything above and still lose, because the heaviest weight in an AI answer sits on sources you do not control.
For coffee, that set is fairly legible. City and regional food media. Established coffee publications and competition coverage. Enthusiast forums and large subreddits where recommendation threads run for years. Subscription and review services that maintain roaster directories. Local business press covering food manufacturing. Award lists from recognised competitions.
Being present and accurately described across that set is the work. It is slower than a website change and it is the thing that actually moves whether a model names you.
Competitions are the underrated lever here, because they generate third-party text automatically. Placing in a regional or national roasting competition produces coverage, lists, and mentions that persist for years, all of which carry your name into the corpus attached to a credential. The entry fee buys you a shot at text you cannot otherwise purchase.
Local press is the other one, and roasters underuse it badly. A roastery is a manufacturing business with a story a business desk can run: green coffee prices, tariffs, what happened to margins, a sourcing trip, a producer relationship spanning years, the economics of a café supply contract. Business editors want operators who talk in numbers. Most roasters only ever pitch the opening of a new café, which is the least interesting story they have.
Enthusiast communities are the third pillar and the one that behaves least like the others. Large coffee forums and subreddits contain years of recommendation threads, and those threads are heavily represented in the text that models learned from. You cannot market into these spaces, and attempting to will get you removed and remembered badly. What you can do is be the kind of roaster people bring up unprompted, which in practice means shipping fast, being honest about roast dates, handling a bad bag without argument, and publishing enough detail that a knowledgeable drinker has something to point at when recommending you.
The uncomfortable truth underneath all three pillars is that this work has a lag measured in quarters. A press mention published in March is not fully reflected in what an assistant says about you in April. Sources get crawled, indexed and incorporated on schedules nobody outside those companies can see. Roasters who expect a website change to move an AI answer within a week conclude the whole exercise is nonsense and stop. The ones who treat it as a two-year accumulation of third-party text are the ones who end up being named.
Measure it by asking
There is no rank tracker for this. There is only the question.
Once a month, take the ten questions a real buyer would ask, phrased the way a real buyer phrases them, and put them to the answer engines your customers use. Who roasts the best coffee in your city. Which roaster should I use for a home espresso setup. Who supplies wholesale beans to cafés in your region. Best subscription for someone who likes natural process Ethiopians.
Record what comes back. Which roasters were named, in what order, described how, and whether anything said about you was wrong. Wrong facts are more urgent than absence, because they propagate and because they are easier to correct once you know they exist.
Over a few months that log tells you something no analytics dashboard will: whether the work is landing, and which of the five signals moved it. Most roasters have never run the test even once, which means they are guessing about the channel that increasingly decides who gets discovered.
One warning about the test, because roasters who run it once tend to draw the wrong conclusion. Answers vary between sessions, between users and between regions, so a single query is a sample rather than a verdict. Run each question three times, note the pattern rather than the instance, and pay more attention to whether your competitors appear consistently than to whether you appeared once.
So before you change anything on your site, go and ask. What does the machine currently say about your coffee, and would you sign your name to it?