Journal · White paper
What it takes to sell beef on the internet
A six-generation cattle farm in the Southern Appalachians sells to restaurants, universities, food hubs and a nationwide shipping trade. This is how the system underneath it works, and how it gets found.
In 1872, Nathan and Sarah Shipley bought land in the Southern Appalachians and began raising cattle on it. Six generations later their descendants are still doing it, on the same ground, as a North Carolina Century Farm. That phrase is a formal designation and it is rarer than it sounds. Most farms do not survive one generational handover. Shipley Farms has survived five.
What the farm does today would be a demanding operation for a company ten times its size. It sells wholesale to restaurants, universities and food hubs, each on its own terms and its own price list. It ships frozen beef to customers across the country. It runs a farm store with a till. It hosts tours and lodging. Every one of those is a different kind of transaction with a different set of rules, and until recently they were spread across disconnected tools and re-keyed by hand between them.
They now run on one system. We built it and we still run it.
This piece is in two parts, because the farm had two distinct problems and they are almost always confused with each other. The first is being found by the person who wants to buy. The second is being able to take the order once they arrive. A business can solve either one and still fail, and most of the advice available addresses only the first.
You can be the best-ranked beef in the country and still lose the sale, if the system behind the button cannot price the box.
Part one: being found
The three customers you lose before they see you
A storefront loses buyers in three places, and only one of them is visible from the inside.
The first is the shopper who waits. Someone on a phone with an ordinary connection taps a link and the page shows them nothing. Google’s own research put it starkly a decade ago: 53% of mobile visits are abandoned when a page takes longer than three seconds.1 The effect is not a cliff at three seconds either, it is a gradient. Google and Deloitte found that a one-tenth-of-a-second improvement in mobile speed raised retail conversion by 8.4% and average order value by 9.2%.2 Akamai measured the same relationship from the other direction: a hundred milliseconds of extra delay cut conversion by 7%.3
Those customers do not appear in your sales figures as lost. They do not appear at all. It is the most expensive category of loss in online retail and it is invisible from inside the business, because the only record of it is an absence.
The second is the shopper who cannot use the page. Low-contrast text, controls too small to hit accurately, images with no description, a checkout that cannot be completed with a keyboard. About one in four U.S. adults reports a disability.4 And this is not a rare failure: WebAIM’s annual survey of a million home pages found detectable accessibility failures on 95.9% of them, with low-contrast text alone on 83.9%.5
Read those two together and the conclusion is uncomfortable. A large minority of the market meets a barrier on almost every site it visits. For a business competing on anything other than price, that is a differentiator sitting unclaimed, and it is cheaper to fix than almost anything else on this list.
The third is new, and most businesses have not yet priced it. A growing share of shoppers now begin with an AI assistant rather than a search box. In the 2025 holiday season, traffic to U.S. retail sites referred by generative-AI sources rose 693% year over year.6 Those visitors behaved better than average once they arrived, being meaningfully less likely to leave immediately and converting at higher rates on the heaviest trading days of the year.
Then the ground shifted again. In March 2026 the largest assistant stopped completing purchases inside the chat and began sending shoppers to the merchant’s own site to finish.7 The assistant now does the finding, the comparing and increasingly the recommending, and then hands the customer over. The merchant’s page has to do the rest, and it can only do that if an assistant can read its stock, its prices and its availability in a structured form rather than inferring them from the layout.
Put plainly: there is now a customer who never sees your home page, never runs your JavaScript, and decides whether to recommend you based on whether your catalog is legible to a machine. Most storefronts are invisible to that customer, and they will not find out from their analytics, because a visit that never happened leaves no trace.
What was done, and what was deliberately not done
The storefront runs on a platform the farm did not build and could not rebuild. That is the normal situation for a small operating business, and it is usually treated as a dead end.
We did not rebuild it either. The work was done in place: the page made faster to first paint, the contrast and the hit targets corrected, the images described, the catalog made legible to a machine. This matters more than it sounds, because the usual advice to a business in that position is a replatform, which is expensive, risky, and interrupts trading in exactly the season a food business cannot afford to be interrupted.
Measured against a field of its named competitors, including national meal-kit brands with engineering teams, the storefront came out at the top of the field on phones and near the top on desktop. On accessibility, on technical best practice and on search fundamentals it reached the maximum score in both views.
The result we care about most is the one nobody was asking for a year ago. Of every site in that benchmark, the farm’s was the only one an AI assistant could act on rather than merely read. Everyone else’s catalog was a picture of a shop. Theirs is a shop.
What we will not claim
Two honest limits, stated here rather than buried.
Lab scores are noisy. Identical page bytes, measured minutes apart on the same device, produced a spread wide enough to change a ranking. That is a property of the measurement, not of the page. We therefore report medians of repeated runs rather than a best result, we treat the desktop position as settled and the phone position as indicative, and we assume any single run, ours or a competitor’s, is one draw from a distribution.
We are not claiming a revenue outcome here. Faster, more accessible and machine-readable pages are necessary conditions for being found and bought from. They are not a promise of traffic, and anyone who tells you they can forecast your search traffic from a page audit is selling you a number they do not have. What we can say is that the conditions are now met, and that they were not before.
Part two: what the system actually does
Being found is the easier half. The harder half is what happens when the customer says yes.
The problem almost no storefront can solve
Beef is sold by weight, and you do not know the weight until it is packed.
Sit with that for a moment, because it breaks the assumption underneath nearly every e-commerce system in existence. Ordinary online retail assumes a known price at the moment of checkout. You click, you are charged, the box ships. A cut of beef does not work that way. The customer orders a ribeye; the ribeye that actually goes in the box weighs what it weighs. The price is not knowable when the card is presented.
The conventional answers are all bad. You can sell fixed-weight portions and throw away the difference, which destroys margin on a product where margin is thin. You can charge an estimate and refund or rebill afterwards, which means a second transaction, a confused customer and a support burden on every order. You can take orders by phone, which is where most farms end up.
The system does it properly. The card is authorised at checkout for the expected amount, the order is picked and packed, the actual weight is captured, and the card is settled at the real figure. One transaction, no refund cycle, no phone call. The customer buys a cut of beef on a website the way they would buy a book, and the arithmetic underneath is doing something no book retailer has ever had to do.
That single capability is the difference between a farm that can sell nationally and a farm that takes orders by telephone.
One cost, many prices
A restaurant does not pay what a university pays, and neither pays what a shopper at the farm store pays. Multiple customer tiers run off a single baseline cost, so that when the underlying cost of a product moves, every tier moves with it correctly and at once.
The alternative, which is what the farm had before and what most operations still have, is a set of price lists maintained by hand in separate places. They drift. Somebody quotes last season’s number to a wholesale account. Nobody can say what the margin on a given tier actually is, because establishing it means reconciling several documents that were last agreed at different times.
Three businesses, one ledger
The online store, the register at the farm store and the wholesale accounts are the same system. A sale through any of them moves the same stock and lands in the same books.
This is the part that sounds least impressive and matters most. When those three are separate, every question that spans them becomes a reconciliation exercise: what did we actually sell this month, what is really in the freezer, is this account current. Each of those questions is answerable, at the cost of somebody’s morning, which means in practice they stop being asked. When the three run on one ledger the questions are just queries, and a business that can ask cheap questions makes different decisions from one that cannot.
Shipping is part of the same path. Carrier rates are fetched live and labels are produced from the order, so the rate the customer is quoted is the rate the carrier will charge, and nobody is copying an address into a separate shipping tool.
What is not finished
Receiving inventory is still an open problem. Billing still lives in a separate accounting package. The order-to-cash path, from a stranger finding the site to money settled at the correct weight, works end to end and runs daily. The rest of the operation is not yet on the same footing.
We state this for the same reason we report medians instead of best runs. An implementation story with no unfinished edges is a story that has been tidied, and the tidying is usually where the interesting information was.
What generalises, and what does not
The specifics here are agricultural. The shape is not.
Most owner-operated businesses that have outgrown spreadsheets have some version of the weight problem: a rule their trade takes for granted that no off-the-shelf system expresses. A configurator with dependent options. A price that depends on a grade only known after inspection. A unit of sale that is not the unit of stock. The usual outcome is that the rule moves into a spreadsheet and a person’s head, and from there it quietly sets a ceiling on how much of the business can be automated, how fast it can grow and how easily it can be handed to the next generation.
What does not generalise is the result. This is one engagement at one farm. It does not predict what your system will do, and the only honest way to find out what yours would do is to measure yours. We would rather describe one job in detail than show a wall of logos, and the detail is the part worth arguing with.
If any of this sounds like your operation, the useful first step is not a proposal. It is to write down the one question about your own business you currently cannot answer without asking a person, and bring that.
What we do, and how an engagement runs.
Sources
- DoubleClick by Google, “The Need for Mobile Speed” (September 2016): 53% of mobile visits are abandoned if a page takes longer than three seconds to load.
- Google and Deloitte Digital, “Milliseconds Make Millions” (2020): a 0.1 s improvement in mobile site speed raised retail conversion rates by 8.4% and average order value by 9.2%.
- Akamai, “The State of Online Retail Performance” (2017): a 100-millisecond delay in load time cut conversion rates by 7%.
- U.S. Centers for Disease Control and Prevention, Disability and Health data: about one in four U.S. adults reports a disability.
- WebAIM, “The WebAIM Million” (2026): 95.9% of one million home pages had detectable WCAG 2 failures; low-contrast text appeared on 83.9% of them.
- Adobe Analytics (Adobe Digital Insights), 2025 holiday season, reported 13 January 2026: traffic referrals to U.S. retail sites from generative-AI sources up 693% year over year, with AI-referred visitors less likely to leave immediately and converting at higher rates on Thanksgiving and Black Friday than other traffic.
- Reported March 2026: the largest assistant ceased in-chat checkout and began directing shoppers to the merchant’s own site to complete the purchase.
Sources are third-party research, cited as published. Claims about Shipley Farms in this piece are descriptive rather than numeric, by the firm’s own standing instruction, and the measured benchmark figures behind part one are held on file.