Chart a Word, Then Read the Page

When Google's Ngram Viewer appeared in 2010, it gave everyone a new kind of historical instrument: type a word, and watch it rise and fall across two centuries of print. It was mesmerizing, and it changed how a generation of researchers asked questions. But anyone who used it seriously ran into the same two walls. You cannot lay two languages on the same chart, so a concept that lived in Latin before it lived in English simply falls off the left edge of history. And the curve is a dead end — you can see that a word surged in the 1650s, but you cannot click the surge and read what people were actually writing.

Source Library now has its own ngram viewer, at sourcelibrary.org/ngrams. It is much smaller than Google's — 17,819 books and about 3.4 billion words, against Google's half a trillion — and it is built over a deliberately curated corpus: alchemy, Hermetica, Kabbalah, natural magic, and early modern science, with the surrounding literature those traditions argued with. Within that world, it does the two things the giant one can't.

Start with the classic experiment

Paracelsus taught that all matter is composed of three principles — mercury, sulphur, and salt, the tria prima. If that doctrine really conquered European natural philosophy in the sixteenth and seventeenth centuries, the words themselves should show it. They do:

That chart is drawn from the English corpus — which needs a word of explanation, because it is the quiet trick behind everything else here. Source Library translates its holdings into English, whatever language they were written in. So the English corpus is not "books published in English": it is Latin treatises, German tracts, French dialogues, and Greek fragments, all counted through one language. A single curve for mercury tracks the concept across every tradition we hold at once. Google's ngrams cannot do this, because nobody translated Google's corpus into one language first.

Follow a concept across languages

Sometimes you want the opposite: not one merged curve, but the original languages side by side. The viewer takes a term:language syntax — mercurius:la charts the Latin word in the Latin corpus, quecksilber:de the German word in the German corpus — and overlays them on whatever else you're charting:

Now you can watch a concept migrate between languages: mercurius carrying the load through the Latin sixteenth century, the vernaculars taking over as alchemy moved out of the universities and into print for apothecaries and mine-owners. For common pairs the viewer suggests these equivalents itself — chart mercury and it offers the Latin, German, French, and Italian forms as one-click chips, from a small hand-curated lexicon of terms whose equivalence in early modern usage is actually unambiguous.

Click the curve, read the page

The second wall is the one that matters most. A frequency chart is a claim about thousands of pages you haven't read; if you can't reach those pages, the chart is an argument from authority. Here, every point on every curve is a link. Click the 1650s on this chart:

— and you land in a search scoped to that term and that window of years, listing the actual passages, in books you can open to the actual page, facsimile beside translation. Two clicks from a trend line to a paragraph like this one, from the English printing of Sendivogius that helped carry Paracelsian doctrine into the language:

That round trip — curve, to search, to page, to quotation with a stable citation link — is the point of the whole tool. The chart is not the evidence. The chart is a map of where the evidence is.

How to read these curves honestly

Every frequency chart is an argument, and this one has assumptions you should know before you trust a curve with a conclusion.

• The corpus is curated, not representative. These frequencies describe the Source Library collection — a library deliberately centered on alchemy, Hermetica, and early science — not print culture at large. A term's rise can mean the idea spread, or that we acquired more books that use it. Under every chart a coverage panel compares our per-year holdings against the Universal Short Title Catalogue's record of European print, so you can see what share of the printed record the corpus actually holds in the years that matter to your question.
• Thin years are thin evidence. The gray backdrop behind the curves shows how much text each year contributes; years with almost no text are dropped from the series entirely rather than plotted as false zeros or wild spikes. A dramatic peak standing on a few thousand words is an anecdote, not a trend.
• Years are edition years, not composition years. A 1650 reprint of a fifteenth-century text counts as 1650. This is the honest choice for a library of physical editions — it charts when words were circulating in print — but it means the curves measure publishing, not first thoughts.
• The English corpus counts a translator's choices. Our translations are AI-generated (reviewed and corrected over time, but machine-made at scale), so the English curve for a term partly reflects how the translator renders concepts, not only how authors used them. That is exactly why the original-language overlays exist: when a conclusion matters, check it against mercurius and Quecksilber, where the words are the authors' own. OCR of early print has errors too, and an unrecognized word is an uncounted word.
• Spelling is normalized before counting. Long ſ, ligatures, and — in Latin — the u/v and i/j pairs are folded together, so vniuersum and universum chart as one term. Genuinely distinct early modern spellings are not merged; if a word you expect seems undercounted, try its variants.

The counting itself involves no AI at all — it is plain tallying over the corpus, which means a curve you see today is exactly reproducible tomorrow. Charts live in the URL, so any comparison you build can be shared as a link; a toggle switches from raw frequency to the share of books that mention a term at all, which is often the sturdier question; and the editorial apparatus our AI adds around page text (summaries, keywords, marginal descriptions) is stripped before anything is counted, for the same reason we never let it be quoted as source text: counting a summary's words would fabricate frequency data the way quoting one fabricates a citation.

Go count something

The viewer is live at sourcelibrary.org/ngrams, with example charts one click away. Chart the words your own question turns on — and when a curve surprises you, don't stop at the curve. Click it, and read what they wrote.