Editorial standards

TrackCaffeine is published by Vast Flow, LLP, and edited by one person, named in the byline on every page. It is a reference about a stimulant most people take daily, which puts it close enough to health for the wording to matter, and far enough from medicine that pretending otherwise would be its own kind of dishonesty. This page states what we publish, how it is checked, how AI is used, what advertising does and does not buy, and what happens when a reader finds us wrong.

What this site is for, and what it refuses to be

The scope is deliberately narrow: how much caffeine is in a drink, and what happens to that amount over the hours afterwards. Everything else is out of scope, and being out of scope is not an oversight to be filled in later.

We do not give medical advice, and no page here should be read as any. We do not diagnose, we do not treat, and we do not tell anyone whether they personally should drink coffee. We do not publish a figure for a child, because a body-weight calculation aimed at children is a medical decision dressed as arithmetic. We do not model interactions with medication. Where a topic genuinely requires a clinician — pregnancy, a heart condition, an anxiety disorder, a sleep disorder under treatment — the page says so in the place a reader is most likely to stop reading, which is early, not in a footer.

The same restraint runs through the machine-readable side of the site. The interface that answers questions from software declines medical questions outright rather than answering them with a confident milligram figure, and it declines before it looks anything up — because an answer arriving at all implies the question was one we should have taken.

The rules a sentence has to pass

Four rules govern every claim on this site, and they are the ones a draft is checked against before it is published.

• A number and its source belong in the same sentence. "The FDA cites 400 mg per day for healthy adults" is publishable; "experts recommend 400 mg" is not, because there is nothing in it to check.
• Dates are absolute and units are explicit. "In 2026", never "recently"; "mg", never "points"; "240 ml (8 fl oz)", never "a cup".
• A figure about a product names the product and the serving. A caffeine number without a serving size is not information.
• An estimate is called an estimate. Where the underlying source gives a range, or where a chain drink varies between shops, the page says so rather than printing a midpoint as though it were measured.

We also do not write in the register that makes a health page feel authoritative without being it. No exclamation marks, no superlatives we cannot substantiate, no "studies show" without a study, and no claim that caffeine causes or prevents a disease — which is a literature this site is not competent to summarise and does not try to.

How AI is used here, and how it is disclosed

This is the question most likely to be asked about a reference of this size maintained by one person, so the answer is on this page rather than in a footnote.

The guides are drafted by a large language model. That is disclosed, it is not hidden behind a fictional staff of writers, and every page carries the name of the person who is answerable for it rather than an invented byline. What the model is allowed to do is narrow. It writes prose against a brief. It does not choose the caffeine figures: every milligram value in a guide is rendered from the dataset, and every chart is drawn by the same engine the calculators run, so the prose and the picture on a page cannot disagree with each other. It does not choose which pages exist, and it does not choose what to link to — the internal links in a brief are given to it, and a draft whose links do not resolve against the pages that actually exist is rejected.

Before a draft is stored it passes a deterministic check: prose length, a banned-phrase list that catches the filler this class of model produces, the presence of a closing summary, resolvable internal links, and outbound links restricted to an allow-list of primary sources. A brief that cannot clear those checks in three attempts is recorded as a failure rather than published short. The twelve language editions are machine-translated from the English and checked for missing strings and for text that came back in the wrong language.

The limits of that arrangement are real and we would rather state them than be caught by them. A model-drafted page is fluent before it is right, and fluency is exactly what makes an error hard to see. The checks above catch a thin page, a broken link, a missing summary and a wrong number; they do not catch a paragraph that is well written and subtly mistaken about a mechanism. That is what the correction path below is for, and it is why a reader who finds one is doing the most useful thing anyone can do on this site.

Advertising, and the wall between it and the data

This site carries display advertising, served by Google, and the app it recommends is made by the same company that publishes the site. Both of those are commercial relationships and both are disclosed here rather than implied.

What advertising does not buy: no advertiser has ever seen a page before it was published, no advertiser has ever requested a change to one, and nothing about ad placement feeds into the dataset, into the ranking of drinks by caffeine content, or into the wording of a guide. The drinks are ranked by milligrams because that is the only ranking this site has. A brand's figure in the dataset is the brand's own published figure whether or not that brand advertises anywhere.

Ad slots are placed automatically and never inside a chart. That is not a courtesy, it is the site's own rule: colour here classifies caffeine against the sleep threshold and appears nowhere else, so an advertisement sitting where a figure's legend should be would read as part of the measurement.

The app is recommended plainly, as our own product, on pages where logging doses is the thing a reader is trying to do. Every calculation the app performs is also available free in the browser here, with no account and no install, and the published dataset is free to reuse under a CC BY 4.0 licence. We would rather be useful to somebody who never installs anything than convert them by making the website worse.

Corrections

We correct in public. A factual error — a wrong milligram value, a wrong serving size, a mis-stated guideline, a broken formula — is fixed on the page and recorded in the updates log with the date, what was wrong, how long it was wrong for, and what changed so that it does not recur. Corrections are not folded silently into a general "site improvements" entry; the log is only worth keeping if the embarrassing lines are in it, and there are several.

Two things are not corrections and are handled differently. A dataset revision — a manufacturer reformulating, a source publishing a new value — is a new dataset version, stamped with a date, and the footer of every page says which version you are reading. A change of model parameter would be a change to every time-based figure on the site simultaneously, and it would be announced as such, not slipped in.

To report something, write to the address on the contact page. A report that names the page and the figure gets fixed fastest. We do not require you to be right, and we would rather check ten reports and find one real error than miss it.

Who decides what gets published

One person, named on the about page, decides what this site publishes and is answerable for every page on it. There is no editorial board, and inventing one would be the single easiest way to make this site look more authoritative than it is. The author is not a clinician, not a nutritionist and not a pharmacologist, and the site is built so that this does not matter as much as it sounds: the figures are compiled from named public sources rather than asserted, the model is published in full so it can be checked by anyone who does have that training, and the scope stops at the point where expertise we do not have would be required.

What we do claim competence in is measurement discipline — keeping one copy of a number, stamping it with a date, citing where it came from, testing that the site still agrees with itself, and writing down what went wrong. The methodology page is that claim in detail, and the updates log is the evidence for it.

Bottom line

TrackCaffeine publishes caffeine figures compiled from named public sources and time estimates computed from one published formula, edited by one named person who is not a clinician and does not pretend to be. Guides are drafted with AI against briefs, with every number supplied from the dataset rather than by the model, and that is disclosed rather than hidden. Advertising funds the site and has no input into the data, the rankings or the wording. Errors are corrected on the page and logged with a date, and reporting one is welcome.