Methodology

Source data

All popularity figures start with the US Social Security Administration’s public baby-name files. The national files cover births from 1880 through 2024; territory files are included as a second source. Each row contains a name, a recorded sex, a year and a count. The SSA suppresses names recorded fewer than five times for one sex in one year, so the source is strong for common names and deliberately incomplete for very rare ones.

The underlying rules are documented by the SSA on its data limitations page. This site is independent and is not affiliated with or endorsed by the SSA.

What happens before analysis

Two narrow cleanup rules run before names are grouped or ranked:

• Administrative placeholders. Entries such as “Unknown”, “Infant”, “Notnamed” and similar filing artifacts are removed.
• Likely Jr suffix artifacts. A value ending in “jr” is removed only when stripping those letters leaves a name that also exists in the data. The condition is intentionally conservative so ordinary names that happen to end in those letters are not discarded.

Nothing is removed because it looks unusual, unfamiliar or non-English. The source contains names from many languages and communities, and unusual spelling is not evidence of an error.

Grouping spellings by pronunciation

Official SSA counts treat every spelling separately. That can make one familiar sound appear artificially uncommon when births are split across several spellings. BabyNames therefore builds pronunciation groups and uses the most common spelling in each group as the page label. The other accepted spellings remain visible as variants rather than disappearing.

The grouping step is deliberately more conservative than the display pronunciation step:

1. A curated pronunciation override is used first when a known dictionary pronunciation needs correction.
2. Otherwise the name is looked up in the CMU Pronouncing Dictionary and represented in ARPABET.
3. If there is no direct CMU entry, the grouping code tries conservative subword combinations from the same dictionary.
4. The neural g2p_en model may supply a pronunciation for display, syllable and stress analysis, but its prediction is not used to decide spelling groups. That separation prevents English-centric guesses for unfamiliar names from merging names that are not actually variants of one another.
5. A small forced-merge table handles known pairs that the automatic grouping misses. These exceptions are explicit and reproducible rather than hidden in application code.

When a name has several grouping-safe pronunciations, the pipeline chooses the pronunciation shared by the largest cluster. A name with no safe grouping pronunciation remains by itself; it is never merged merely because another unresolved name also lacks a pronunciation.

Counts, ranks and peak years

• Total births is the sum of the recorded counts for every spelling in the pronunciation group, across the available years and included territory data.
• Rank is based on total recorded births within a gender after spelling grouping. It is therefore intentionally different from a single-year SSA ranking.
• Peak year follows the group’s primary spelling: the spelling with the largest all-time total. If two years tie for that spelling’s highest count, the later year is used.
• Unisex share is calculated from the recorded male/female counts. The source exposes only those two sex markers, so the statistic describes the dataset rather than gender identity.

The year-by-year curve

The chart is not a decorative trend line. For every canonical name, the build reconstructs a yearly series from the same raw files and sums all spellings that belong to that name. Missing years inside the range are stored as zero. The generated series is then checked against the aggregate total so a page cannot quietly show one number beside a chart that represents a different one.

This distinction matters when comparing a name that had one intense decade with a name that accumulated a similar lifetime total over a century. A single rank hides that shape; the series exposes it.

How trends and related names are derived

Trend labels use the yearly series rather than search volume or editorial guesses. Rising names compare recent use with the preceding period; revival lists require an old peak plus recent recorded use. Related-name suggestions are built from observable features such as dominant recorded sex, first letter, syllable count and peak era. These are heuristics, not claims that two names share an origin, culture or meaning.

The same principle applies to sibling suggestions. They are data-based similarity candidates, not a statement that one pairing is objectively better than another.

Which names get a page

The raw source contains tens of thousands of names once historical spellings are considered. Publishing a page for every sparse record would create pages with almost nothing to analyze. The production site therefore requires at least 5,000 recorded births for a name to enter the searchable page set. That leaves roughly 3,600 names with a long enough record to support a curve, comparisons and derived observations.

The threshold is a publishing decision, not a claim that rarer names are less valid or less interesting. They are simply where this particular dataset is least informative because the SSA’s own privacy threshold removes many of their small annual counts.

What the site deliberately does not infer

• No invented etymology. Popularity data cannot establish a name’s meaning, language of origin or cultural significance, so those claims are not generated from it.
• No prediction presented as fact. Recent movement can be measured; future popularity cannot be known from the historical record.
• No individual-level conclusions. The source is aggregate. It says nothing about a particular person who has a name.
• No global ranking. The site describes US records only. Popularity in other countries can be completely different.

Known limitations

• The five-child threshold. Rare names are undercounted. Repeated annual counts below five do not become visible simply because many years have passed.
• Pronunciation varies. Regional, linguistic and family pronunciation can make two spellings equivalent for one person and distinct for another. Automated grouping can therefore make both false merges and false separations.
• Recorded sex is binary in the source. The SSA files use M and F. That is a limitation of the dataset, not a statement about people.
• Recorded name is not necessarily used name. The files describe names in Social Security records, not what every person is called day to day.
• Historical coverage is not culturally neutral. Access to Social Security records, migration and reporting practices changed over time. Early-year counts should not be read as a complete census of every community in the United States.

Corrections and reproducibility

The data pipeline, exclusion rules, pronunciation overrides, forced merges and tests live in the same project as the site. Regenerating the data applies the same rules to every record. That does not make the method infallible, but it makes mistakes inspectable and correctable instead of arbitrary.

If you find a grouping error or a page whose chart contradicts its summary, use the contact page. Corrections are applied to the source rules and regenerated rather than patched into one page by hand.