What we cite, and for which claim

Everything on this site is either a figure taken from somewhere else or a figure computed from one. This page names the somewhere else. Each entry below lists the publisher, the year, a link, a digital object identifier where the work has one, the specific claims on this site it stands behind, and the date somebody last opened it and confirmed it still says what we say it says.

Why every citation carries a date it was last followed

A citation nobody follows is decoration. We have been caught by this in a neighbouring project, where two guideline documents had been cited by their publisher's web page for months: one answered nothing at all to an automated request and the other answered with a refusal, and neither had ever actually been fetched after the day it was added. The claims turned out to be correct, which is exactly the problem — nobody knew that, and finding out took ninety seconds.

So each entry here carries a checked date, and where the work is a journal article or an agency opinion with a digital object identifier, the identifier is what we cite rather than the publisher's web page. A publisher's page can be moved, paywalled, or configured to refuse anything that is not a browser; an identifier resolves to citation data that anybody can retrieve in one command, which means this list can be audited by a reader rather than taken on trust.

The kinds of authority a figure can rest on

Not every caffeine figure comes from the same kind of place, and which kind it came from changes how much weight it should carry. These are the four kinds, in order of preference; the next section shows how the 87 rows of the dataset divide between them, and every drink page prints its own row's class under the milligram reading.

A manufacturer's own published figure for its own product. For a branded, sealed product — an energy drink, a bottled coffee, a pre-workout powder — the company that formulates it knows what is in it, and its own statement outranks any third-party estimate. This is the strongest source we have and it covers most of the branded rows in the dataset. Its weakness is that it changes without announcement: a reformulation moves a number and nothing notifies anyone, which is why the dataset is versioned and dated rather than described as current.

A food composition database. For the generic entries — brewed coffee, a single espresso, black tea, green tea, cola — no manufacturer exists to ask, because nobody makes "brewed coffee". These come from the USDA's FoodData Central, which publishes measured values with the sample and the method attached. Their weakness is that they are averages over a genuinely wide range: the same 240 ml cup of drip coffee spans roughly 70 to 140 mg depending on the grind, the dose and the contact time, and the dataset stores one number where the truth is a distribution.

An agency document. For figures that are about intake rather than about a product — how much per day, how much in a single dose, how much during pregnancy — we cite the agencies that actually publish those figures: the U.S. Food and Drug Administration, the European Food Safety Authority, and the American College of Obstetricians and Gynecologists. These are the strongest sources on this page and the ones we quote most carefully, because an agency figure has a precise meaning that is easy to flatten. The FDA's 400 mg is a level not generally associated with negative effects in healthy adults; it is not a recommendation, not a target, and not a statement about you.

A peer-reviewed paper. We cite these sparingly and only for a claim the site's whole structure rests on, rather than to decorate a page with references. The clearest example is the sleep-timing work below: the reason this site is a reference about when rather than a table of milligrams is a controlled study finding that a dose six hours before bed still measurably disturbs sleep. That finding is load-bearing here, so it is cited.

What is deliberately not on this list

Three kinds of source are missing on purpose.

There is no citation for the 5.7-hour half-life, because it is not a cited constant — it is a modelling choice. It sits in the middle of a range that the pharmacology literature reports broadly, and an individual value can fall anywhere between about three and about nine hours. Attaching a single reference to it would imply a precision the choice does not have. The methodology page states the parameter, the equation it sits in, and what the model deliberately leaves out.

There are no citations to other caffeine websites, content aggregators or blog posts. If a figure can only be traced to a page that itself cites nothing, it does not go in the dataset. This costs coverage — there are drinks with no reliable public figure, and they are absent rather than estimated.

There are no citations supporting health claims about caffeine, because this site does not make any. We publish what is in a drink and what happens to it over the next twelve hours. Whether caffeine helps or harms a given person is a clinical question, the literature on it is large and contested, and summarising it is outside what this site is competent to do. The editorial standards set that boundary out in full.

What happens when two sources disagree

They do, regularly, and the rule is to name the disagreement rather than to quietly pick a winner. Where a manufacturer's figure and a database figure differ for the same product, the manufacturer's figure goes in the dataset and the page says who published it. Where two agencies differ, both figures are stated with the body that publishes each — which is why the pregnancy figure is given as 200 mg per day with two independent bodies named rather than as a bare number that appears to come from nowhere.

Where a published standard has changed, the page names the change and the old value, because a reader arriving with the old number in their head has to be told rather than silently contradicted. And where a figure cannot be traced to a source at all, it does not go in. There are drinks with no reliable public caffeine figure, and they are missing from the dataset rather than estimated into it.

Citing us in turn

The dataset is published under a Creative Commons Attribution 4.0 licence and is available as a download and as an open API that needs no key. Reuse it, including commercially, with a link back. The figures this site computes — residual caffeine, cutoff times, multiples of a standard cup — are reproducible from the published dataset and the published formula, so a reader who disagrees with a number here can rederive it rather than argue about it.