Persian poetry · a rule you can run · 60,059 ghazals

The Word That Comes Back

Most Persian ghazals end every couplet on the same word or words, after a rhyme. That repeated ending is the radif, and it is a rule simple enough to run. Here it runs on every ghazal Ganjoor holds, 60,059 of them by 146 poets, and on any ghazal you choose. Seven in ten carry one. The typical poet's share rises from about three in five before 1500 to more than four in five for poets who died between 1500 and 1800, the age of the Indian style, then falls back after 1800, once Iranian poets had set out to write like Saadi and Hafez again.

A ghazal is a chain of couplets, most often between five and a dozen. Its first couplet, the matla, rhymes in both halves. Every later couplet rhymes only in its second half, on the same rhyme. And very often, after the rhyme, the same word or words come back exactly, every time, to close the line. The rhyme is the qafia; the word that comes back is the radif. Here is Hafez's third ghazal in Ganjoor's numbering, the one about the Turk of Shiraz. Press line up the endings and the radif stacks into a column.

In the Turk of Shiraz the rhyme is ـا, a long ā (ما، بخارا، مصلا، یغما…), and the radif is را, rā, the little word that usually marks a definite object (in two of these lines it means "to" or "for"). Ten rhyme lines, ten times rā. It is not a refrain in the English ballad sense of a line that returns; it is the last word of the line, fixed in advance, that every couplet has to find a new way to arrive at.

The rule, and the one hard part

Written as a procedure: take the rhyme lines (both halves of the matla, then the second half of every other couplet), and find the longest run of whole words that all of them end with, provided at least one rhyming letter is still shared in front of it. That is the radif. If nothing qualifies, the ghazal has none.

The hard part is whole words. Persian editions space the same words in three ways: with a space, with an invisible zero-width non-joiner, or not at all. In Hafez 86 above, the verb گرفت (gereft, "took") ends every line, but the edition writes it as برگرفت, درگرفت and سر گرفت, so a rule that trusts the spaces sees three different words and no radif. So the rule here compares letters, not spaces, and then asks where the words start. The refrain may start at a written boundary, or after a verbal prefix or head that editions write joined (بر، در، سر، فرو، فرا، وا، باز، ور), or after three or more letters that occur at least 50 times as a word on their own in the corpus. And it must be written as a separate word in at least one of its lines.

Then there are the endings. Plural ‑hā, the personal endings (‑am, "I am"; ‑ī, "you are"), the pronoun endings and the clitic ‑st (a contracted ast, "is") are glued to the word in front of them. Shams-e Qays's thirteenth-century handbook defines the radif as a word independent of and separate from the rhyme, and counts appended letters like these with the rhyme. Other accounts, both older and newer, count a repeated attached pronoun or ending as a radif. This page counts only free words, so Hafez's first ghazal, whose lines end مشکل‌ها، دل‌ها، محمل‌ها (moshkel‑hā, del‑hā, mahmel‑hā) on the plural, has no radif here. Its very first line does not even use the Persian plural: ناولها is Arabic, "hand it over". The last section shows what the other convention would do to the numbers.

There are three versions of the rule on this page, and the tables name them. As written: whole words exactly as the edition spaces them. Letters: the rule above, the page's main measure. Tolerant: the letters rule, with one rhyme line allowed to differ, which forgives a variant reading or a loose rhyme.

Run it on any ghazal

Choose a poet and the page picks one of their ghazals at random, fetches its text live from Ganjoor, the open library of Persian poetry, and runs the rule on it in front of you. Or paste the address of any Ganjoor poem. Beside the live verdict is the one this study stored when it measured the whole corpus on 25 September 2026. If Ganjoor's editors have corrected the text since then, the two can differ, and you will see it.

Sixty thousand ghazals

Ganjoor publishes its whole library as one small database per poet for its desktop and phone apps. From the 281 files listed on 1 September 2026, this study took every section titled as ghazals. It left out unfinished ghazals, excerpted couplets, mixed sections, and the Turkish ghazals some poets also wrote. That leaves 60,059 poems of three or more couplets, by 146 poets, in 517,492 couplets.

The tolerant rule adds only about two points, so almost every radif in the corpus holds in every rhyme line of its poem. The letters rule and the as-written rule disagree more for some poets than others, because some editions space more loosely. Bidel's is the loosest: as written, 54% of his ghazals have a radif, and read by letters, 74%.

Eight and a half centuries, one poet at a time

Each dot is a poet with at least a hundred ghazals in the corpus, placed at the year Ganjoor gives for their death and at the share of their ghazals that carry a radif. Its size shows how many ghazals they left. The band is the median poet of each century, with a 95% bootstrap interval for that median.

Three spans, drawn from the chart: the century medians sit at 61 to 67% from the 1100s to the 1400s, at 82 to 84% from the 1500s to the 1700s, at 58.5% in the 1800s, and at 67% in the 1900s.

The rise shows first in poets who died early in the sixteenth century, among them Baba Faghani (died 1519, 85%), whom literary historians count as a forerunner of the Indian style (sabk‑e hendi). The style itself is usually dated from the later sixteenth century to the eighteenth, in Safavid Iran and Mughal India, and the high rate holds through its seventeenth-century height. Its poets, Saeb above all (Ganjoor's biography reports that he has been given the title شاه سبک هندی, "king of the Indian style"), were famous for elaborate conceits, and the radif, a fixed word the whole poem must keep reaching in new ways, is a natural engine for them. Saeb, with 6,994 ghazals, carries a radif in 86% of them.

The fall comes with what Persian literary history calls the bazgasht, the Return. In the middle of the eighteenth century poets in Isfahan turned away from the Indian style, back to the older masters. Ganjoor's biography of Moshtaq Esfahani, who died in 1169 or 1170 AH (about 1756; other sources give 1757 or 1758), calls him از بنیان‌گذاران سبک دوره بازگشت, one of the founders of the Return style. In the few poets who can be placed in it, the Return shows in the numbers only a generation later. Moshtaq and the poets Ganjoor names as his students still put a radif in 74 to 81% of their ghazals, close to the median of 83 for 1500 to 1799. Two poets of Fath-Ali Shah's court, Neshat and the king himself, drop to 57 and 59%, below the median poet before 1500; the king's panegyrist Sahab, Hatef's son, stays at 73%. Five or six poets cannot carry a movement, and the chart's lower spread after 1800 is the better evidence.

Two cautions. First, the three spans were chosen by looking at the century medians, so the permutation test in the check at the bottom describes how large the step is; it is not an independent test of a prediction. Second, one number per poet treats a poet with 100 ghazals and one with 7,000 the same, which is the point (it asks what the typical poet did), but it hides that the late 1600s are dominated by a few enormous divans. The step survives moving either boundary by 25 years, and it survives every one of the four rules. One point is suggestive and no more than that: Ghalib, who died in Delhi in 1869 and wrote Persian in the Indian manner long after Iran had turned back, has a radif in 84% of his ghazals, far above his Iranian contemporaries.

The refrains themselves

The 41,888 ghazals with a radif use 7,219 different ones. The commonest is است, ast, "is", which is what the Persian Wikipedia article on the radif says too. Most are one word. Long ones are rare, and the longest are a pleasure.

The single longest radif in the corpus, by words, is a count. In a ghazal in Ganjoor's selection from Amir Khosrow of Delhi, every rhyme line ends یک دو سه چار و پنج و شش: "one, two, three, four, and five, and six". The rhyme in front of it is ‑ān.

Find a radif

What counts as a word decides the number

A published count exists, and it is higher than this page's. Ahmad Mohseni's Radif and the Music of Poetry (Mashhad, 2003), as the Persian Wikipedia article on the radif reports it, puts both Saadi and Hafez above 80%, Jami near 66%, and Saeb above 94% of 7,015 ghazals. We have not seen the book, only that summary of it.

The gap is not an error on either side. It is the endings. Saadi's ghazal 46, the last example at the top of the page, has thirteen rhyme lines, and all of them end on ‑āst. Thirteen lines: برخاست، بیاراست، زیباست، تواناست، راست، خواست… In zībāst ("is beautiful") the ‑st is the verb "is", glued on. In barkhāst ("rose") and rāst ("right") it is part of the word. Letters cannot tell these apart, and neither can this page's rule, so it counts no radif.

The fourth column adds every ghazal whose rhyme lines all end in the same one of ‑ast/‑st, ‑am/‑m, ‑īm, ‑at/‑t, ‑ash/‑sh or ‑ī, whether that ending is really a clitic there or part of the word. It is a rough upper reach for those endings, not a strict ceiling: it leaves out plural ‑hā, ‑ān, the verb endings ‑and, ‑īd, ‑ad and the plural pronouns, so Hafez 1 is not in it. Mohseni's figures for Hafez, Saadi and Jami are consistent with a count lying between the two columns. Saeb's is a point above the fourth column, and within reach if plural and verb endings are counted too (our fact-checker's rerun on the committed sample took Saeb from 92% to 95% that way); his edition also has 21 more ghazals than this study measured. The chart's fourth button draws this column over time. Counting endings lifts every century and flattens the rise and fall, because what grew in these three centuries was the free-word radif in particular.

English has a ghazal too, and it mostly arrived without its radif. Agha Shahid Ali's anthology Ravishing DisUnities: Real Ghazals in English (2000) was built on the argument that the form should keep both the rhyme and the word that comes back.

-- verify.mjs: reruns the page engine on a fixed sample of 3,908 ghazals whose rhyme lines are committed,
   checks each against the stored measurement, recomputes every number on this page from the
   stored measurement, and runs planted-fault checks. mutate.mjs breaks the engine in eleven small ways
   and confirms the verifier notices each one.