What a fake deodorant brand tells craft distillers about where discovery is heading. Welcome to the AI drinks shelf.
Earlier this year the marketer and writer Deana Burke, who publishes the technology and business newsletter Boys Club, built a natural deodorant brand called Morrowen.
It did not exist. No factory, no inventory, no product. There was a cheap domain, a simple website, copy written partly by AI, and one self-described low-effort Substack article.
In this case, she seeded that copy with a few specific ingredient terms (for example, colloidal oatmeal, magnesium hydroxide) to see whether she could catch long-tail queries. She deliberately didn’t spam Reddit, write fake reviews, testimonials, or buy coverage.
Three weeks later, ChatGPT with browsing switched on began recommending Morrowen to people asking what to buy. Sometimes ahead of brands with clinical trials behind them. Inc. covered the experiment in August. Worth mentioning for fairness here, Claude and Gemini never recommended it.
Burke calls what she found the AI shelf.
Picture a real shelf in a real shop. You ask the assistant what to buy and they reach for a shortlist. That shortlist is the shelf. She had already mapped it once, running close to 9,000 purchase queries across ChatGPT, Claude and Gemini to see which products kept reappearing. Morrowen was the follow-up question to that research: can you get onto the shelf from nothing?
The answer is a qualified yes.
Native still dominates the category overall, named first in 99 of 180 queries in her wider mapping of the space. Morrowen never touched that. On the seventeen queries she tracked specifically for the fake brand, the split was stark: zero appearances across the thirteen asking for the “best,” “cheapest” or “aluminum-free” option, and 4 out of 4 first-place mentions on the need-state questions about baking soda irritation and sensitive skin. Burke calls what she landed on “a dusty spot at the back of the store no one really goes.” But critically, she was in the store.
That is the finding that caught my attention and where I see huge ramifications for the drinks industry. Not that AI is broken or that it can be fooled.
That the ‘AI shelf’ is real, that it is reachable, that it is going to be the single biggest curator of what gets recommended to the public at large.
That, and the fact almost nobody in drinks is looking at it.
The web the models are actually reading
Burke’s experiment sits inside a wider shift worth understanding before I share explicit craft distilling insights by viewing the topic from a drinks industry lens. The evidence behind it is firmer than most people assume.
Automated traffic overtook human traffic on the web this year. Cloudflare Radar recorded bots at 57.5% of HTML traffic against 42.5% human in June 2026.
I can see this on Everglow Spirits. There isn’t a 1000+ person distilling community in the middle of China’s countryside looking for cooperages hundreds of times a day, nor UK-specific compliance advice. There are data centres though.
If you work in editorial, bot traffic and access is nothing new. Publishers have responded. Around four in five top news sites now block at least one AI training crawler, and Reuters and Time moved to blocking by default in May.
So when someone asks an AI what to drink, the model is not reading the whole internet. It reads the part that still lets crawlers in.
What remains there is instructive, and two separate studies tell the story. An academic audit, Synthetic Sources?, scraped around 26,000 sources cited by generative search engines and found roughly one in six had itself been written by AI.
Separately, Ahrefs analysed 26,283 URLs that ChatGPT cited across 750 search terms and found 43.8% were “best of” listicles.
Stack those and the compounding problem becomes clear. The single most-cited format is the cheapest one to manufacture. A meaningful slice of what the models read is machine-written.
And, as Burke points out, some of those lists are published by the brand that ranks itself first.
That is the part that should bother us most. Not the fake brands. The real ones marking their own homework, and a system that reads it as an acceptable verdict.
The current lag is the opportunity
There is a second thing worth saying plainly. The noise around agentic AI has run well ahead of the reality. If you’ve tried it, you know what I mean.
Discovery has clearly moved. Transaction, however, largely has not.
OpenAI paused Instant Checkout in ChatGPT, refocusing on product discovery and comparison. Google’s Universal Commerce Protocol and OpenAI’s Agentic Commerce Protocol are being built now, not finished. Conversion through agentic channels still trails established affiliate routes by a wide margin, because merchant infrastructure was never designed for machines.
For a craft producer, or those working to help distillers make more D2C sales, that gap is a gift. It means that we have a window to work with.
The recommendation layer is live and already shaping intent. The purchase layer is still being wired. Anyone who does the groundwork in the next twelve months arrives before the category does.
A basic problem that is specific to craft distillers
Every distillery site carries an age gate. Or at least it should. It is a legal necessity and it is not going away.
It is also, in many builds, a wall crawlers cannot climb.
Google has said for years that its own bot does not click buttons or complete date-of-birth fields, and that badly configured age interstitials can stop content being crawled at all. The same mechanics apply to GPTBot, ClaudeBot and the rest.
So a distillery can hold the richest content in its category, be it their cask programme, their botanical reasoning, or their eight years of graft outlined in monthly blogs, and be functionally invisible to the systems now making recommendations.
Almost no other retail category carries this handicap universally. Ours does. Server-side age verification or content that renders before the interstitial keeps you compliant and visible at once.
Check this first.
There is no point spending all the time working on how to be recommended if the entire thing is invisible anyway. It is the highest-value hour you will spend on the subject.
Trusted voices, and a tension that’s only getting more polarised
There is an apparent contradiction at play. Institutional journalism is walling itself off, expertise is perceived to be less valuable than it used to be and the gatekeepers hold less power than ever. Yet the advice for how to get recommended is still to get other people talking about you.
Both are true, and it’s as much to do with discovery as it has to do with validation.
Traditional media is closing. That includes yours, if it sits behind a paywall or subscriber only. For example, a Substack post reads to a crawler exactly like a paywalled newspaper article: teaser text, and nothing more.
For a writer, that can feel like a genuine advantage. The people who click through are real readers, not bots. It builds a direct line to a community you can actually see. That you can be a part of. There no vanity-only numbers of clicks and uniques and you can really understand the true reaction that something gets.
I understand that pull and have long been drawn to it as a writer.
But for a brand deciding where to put a sponsorship budget, the calculation looks different. A newsletter capped at a few hundred or a few thousand subscribers delivers exactly that many readers, once, and nothing beyond it.
An open platform, a public post, a video, a forum thread, reaches the stated audience first, but then gets picked up by the crawlers and goes on to a second life inside AI recommendations. That second life is hard to buy anywhere else right now, and my guess is that it will be why some marketers will argue that the same spend will be worth more on an open platform than behind a gate.
Community content is wide open.
Peec AI’s analysis of 30 million sources ranked Reddit the single most-cited domain across major engines, ahead of YouTube and LinkedIn.
On Perplexity it can account for close to half of all citations. Notably, ChatGPT and Claude lean on it far less — so the picture varies sharply by platform, and a strategy built on one channel is fragile.
What this means in real terms is that a dozen genuine, detailed Reddit threads may now carry more machine weight than a single national newspaper feature.
For someone who has worked in PR and Marketing as well as journalism, this feels wild. Think about it for a second here and about the journalists who are genuinely well versed in their subject areas. They are specialists. They do their homework. Yet, their opinion (that influences so many who read the actual words they put out and click through to those publications which also have a genuine authority of their own), is worth less and less to an algorithm that’s becoming increasingly important as it’s simply invisible.
Meanwhile Dave, Peter and Karen’s 11pm Thursday night chat about how smooth the 12yr Speyside is carries genuine algorithmic weight.
What can you do about it?
Ideally, your brand would get featured universally.
But if you have been treating enthusiast forums, YouTube reviews and long-form community discussion as the soft end of your PR effort, that hierarchy has not just softened since the rise of the influencer, it’s increasingly accelerating towards total inversion.
The move therefore is to diversify the voices you are enlisting to help with discovery, all the while making sure that you still build the right kind of relationships with experts who can validate quality.
Where i’d start is different however. There is a further move available to drinks producers, and it suits this industry unusually well.
Reciprocal category content with adjacent brands seems to have more weight than ever. A gin working with a vermouth producer on owning the negroni. A rum maker and a ginger beer brand on the occasion rather than the bottle.
Each publishes, each references the other, and both appear in third-party content rather than self-published lists only. It is honest, it is genuinely useful to the reader, and it produces exactly the co-citation pattern these systems reward.
Clarity over occasion beats process specification
One detail from Burke’s experiment deserves more attention when viewed with a drinks lens than it has had from the AI-focussed coverage so far.
The query that surfaced Morrowen was not a straight ‘best’ or ‘cheapest’ product queries. Four out of four need-state queries named it, and named it first. Thirteen “best” and “cheapest” queries did not surface it at all. Think search queries that are along the lines of ‘an irritation from baking soda, sensitive skin, what should I buy’.
That’s a need state, written in natural language, at length. AI Mode searches now run roughly three times longer than traditional searches.
This is the part that should reassure anyone who has grown tired of fact lead content marketing. Where ABV, cask finish or botanical list is used up front and as primary USP.
The winning ground is the occasion, the moment, the problem and the language people actually use to describe it. Who is this for. When is it drunk. What does it solve or what does it replace. What does someone type when they do not yet know your category exists.
That is still specificity. Models reward detail over vagueness, and a detailed unofficial source will beat a vague official one every time. But the specificity is centred on the why, not the how or what.
Own the occasion in language, and you own the query (and currently – the AI recommendation that users get).
So what does the AI shelf look like in practice?
I ran a quick test, run cold in a single Claude session with browsing off, so treat it as illustrative rather than a study. For the non-tech-nerds, what it means is that the answers come from training data rather than a live search which makes a material difference from Burke’s test.
I asked for the best gin for a Negroni, then asked the same question as a need state: a gin for someone who finds most gin too piney, served in a Negroni.
The “best” answer led with Tanqueray and Beefeater, then Sipsmith, Ford’s and Bombay Sapphire. Diageo, Pernod Ricard, Suntory, Brown-Forman, Bacardi. Five names, five multinational owners. I’ve run it again since and it seems like quite a locked in selection.
The need-state answer moved differently: Brockmans, Ableforth’s Bathtub Gin, Four Pillars, Malfy, Gin Mare. The ownership picture barely shifts. This is still mostly multinational money with legacy content links and big partnerships or affiliations, but the brands themselves shift a long way, from five household staples to a set most drinkers would call more distinctive.
Compared to the established shelf on best or cheapest style search – a sharper, more specific question opened the door a little wider for (comparatively) less established gins. With browsing switched on, feeding on live reviews and forum threads rather than training data alone, I would expect that needle to move further still.
It doesn’t change what good brand strategy looks like for drinks producers
None of this is new thinking. It is the same discipline that has always sat behind good brand strategy, which is why I built chapter three of the Creative Strategy and Brand Planning workbook around it. The messaging hierarchy, the elevator line, the three story strands. All of them start from the same question: who is this for, and what moment does it belong to?
The workbook asks you to write the story before you write the spec sheet, and to test every claim against “so what?”
What has changed is the reward. That work used to pay off in a pitch meeting or on a shelf talker. Now it also determines whether a machine can find a reason to name you.
The producers who have already done the occasion work are the ones best placed for this, and they did not do it for AI. They did it because it was the right way to build a brand.
How can you prepare for the AI Drinks Shelf
Audit crawler access and your age gate build.
I’m deeply uncomfortable that so much of my original content gets summarised on the top of Google and that I never see a click while they benefit from my expertise, but resist the instinct to block.
That larger businesses are walling themselves off is understandable in publishing. But producers are not an editorial site. Openness on your own brand’s site is currently a producer advantage.
Publish the occasion, not just the process.
Earn placement in other people’s content rather than writing your own rankings. At the end of the day, it’s the right thing to do, and while it may work for AI, I can imagine there are others who would be instantly turned off by reading a listicle naming your brand over others as the ‘best’ for a Martini or Old Fashioned.
Build reciprocal category work with adjacent producers.
Lastly, run a monthly shelf audit. Ask ChatGPT, Claude, Gemini and Perplexity the questions your customers actually ask, and note who gets named, in what order.
Increasing literacy will change what you need to do over time
It’s important to add context around AI literacy when discussing buying patterns and discovery.
When Google hands you ten links, you apply judgement about who to trust. We’ve learned how to work our way through search, what’s paid, what’s optimised and what’s authoritative. But it’s taken years.
When an AI hands you a recommendation, it arrives in one confident voice that sounds vetted. We’re still figuring out what to make of it, how to asses that, and unfortunately, some are not even questioning it.
Soon enough however, agents may well close the loop entirely. You’ll ask a question about the best gin to buy for your dad’s 70th with an instruction to get what it thinks as it knows you. Not only will it search and give you an answer, it will also buy it for you and have it delivered. That’s the end-to-end agent loop that is being built.
Human decision, filter, or oversight not needed.
And while it sounds alarmist and further away than tech bros seem to suggest, you can easily see it happen for low stakes cabinet staples.
Vodka for a party. Beers for the park. Sure, bigger ticket items may still involve a person getting the thrill of the hunt and wanting the joy of shopping, but there’s a whole load of items and situations that sit beneath that.
And even then – once the decision has been made, handing over the task of securing an item at best price, delivered to door is a tedium an agent will surely take over once they are competent enough. Which will open layers of optimisation for your own D2C store and an article I’ll no doubt need to type in the upcoming months.
Which brings me back to a point Burke makes in her reels. Several trillion dollars of commerce is drifting toward a shelf that is currently very hard to inspect or monitor, with very little transparency around it.
My bet is that the producers who understand it early, and who feed it real substance rather than manufactured signals, will be standing somewhere very good when the rest of the category catches up.
The shelf is being stocked now. It is worth knowing whether you are on it.
