Long-Tail Phrase Builder

Most advice about long-tail phrases begins with a number. A share of all searches, a rate of conversion, a promise that the specific phrase is the easy one. This essay begins somewhere less comfortable: the long tail is the name of a shape, and a shape promises nothing about any single point on it. That matters for anyone holding a tool that can write phrases by the thousand. Writing them is cheap. Knowing which of them a person has ever typed is a different kind of knowledge, and no pattern contains it.

Where the phrase comes from

In statistics, a long tail is the part of a distribution that holds many occurrences far from the head. Wikipedia’s account notes that the term is often used loosely, although precise definitions exist, and that distributions of this kind have been studied by statisticians since at least 1946. The business sense is much younger.

The phrase was made popular by Chris Anderson in a magazine article of October 2004, later expanded into a book whose subtitle gives the argument away: the future of business is selling less of more. The same account traces part of the idea to a February 2003 essay by Clay Shirky about weblogs, where a handful of sites drew most of the links and millions drew a few each.

Anderson leaned on research published in 2003 by Erik Brynjolfsson, Yu Jeffrey Hu and Michael Smith. Their abstract reports that one online bookseller offered more than 23 times as many titles as a typical large superstore had on its shelves, and it argues that this added variety was worth more to readers than lower prices were. The claim is about access to rare things, taken together.

A later paper by Brynjolfsson, Hu and Duncan Simester compared the internet channel of one retailer with its traditional channels and found the internet channel’s sales significantly less concentrated. The authors tie part of that to lower search costs, and they hedge: the trend holds if the relationships they found persist. Careful work on the long tail tends to sound like that.

A distribution, not a forecast

The mathematics behind the picture is the power law. A review by M. E. J. Newman defines it plainly: a quantity follows a power law when the probability of measuring a value varies inversely as a power of that value. The same idea carries the names Zipf’s law and the Pareto distribution, and Newman notes that its origin has been debated for more than a century.

Search behaves in a way that at least resembles this. A study published in 2001 analysed over one million queries sent to one search engine of the time. Its authors found that most people used few search terms, that a small number of terms were used with high frequency, and that a great many terms were unique. Few common things, very many rare ones.

Here is the honest limit. A statement about a distribution is a statement about the whole. It says rare queries are numerous in aggregate. It does not say that a particular rare phrase you have just written exists in anyone’s mind, let alone in a search box.

Even the shape is harder to pin down than its popularity suggests. Aaron Clauset, Cosma Shalizi and Newman showed that detecting a power law is complicated by large fluctuations in the tail, and that a common method, least-squares fitting, can give substantially inaccurate estimates while offering no sign of whether the data follow a power law at all. The tail is exactly where the evidence is thinnest.

So a page that writes phrases should not print volumes beside them, and this one does not. It has no search data. Any figure it showed would be invented, and an invented figure beside a plausible phrase is the most persuasive kind of mistake.

Why a specific phrase can still be worth writing

None of this makes specific phrases pointless. It changes the reason for writing them. A specific phrase is useful because it describes a specific need, and a need is something you can serve whether or not a tool has measured it.

Andrei Broder’s 2002 taxonomy is a good corrective to thinking of every search as a question. He observed that the need behind a web search is often not informational at all. It may be navigational, where the person wants a particular site, or transactional, where the person wants to do something such as shop or download a file.

One search engine’s own guidance for site owners, read on 12 October 2026, makes a related point about vocabulary. People who know a subject well may use different words from people who are new to it, and anticipating that difference while writing for your readers could help. The same page adds that you need not anticipate every variation, because its language matching can relate a page to queries that do not use the page’s exact terms.

Questions are the most practical form of specificity. English builds them from a small set of interrogative words, among them what, which, when, where, who, why and how, alongside the yes-no questions that open with a verb such as is, has or does. Each one frames a different need around the same subject: what a thing is, how it works, when to replace it, whether it is worth having.

Used this way, a list of questions is a list of things to write about. It is a planning aid for a person who knows the subject, and its value lies in the gaps it shows you in what you have already explained.

What a pattern can and cannot do

A pattern expander is combinatorial, and it should say so. Take six seeds and fourteen frames, where four frames name a set of four words, one names a set of three, and one needs a plural and names two sets. Five seeds have a plural; one is uncountable. The count is fixed before anything is written: 6 times 27, plus 5 times 12, which is 222 phrases.

Grammar is where plain joining fails. English chooses its indefinite article by sound, not by spelling: an before a vowel sound, a before a consonant sound. That is why the phrase is an hour but a user, an MP3 but a one-armed bandit. A government style manual gives the same rule for initials, which follow the name of their first letter, and for numerals, which follow how the number is said, as in an 11-year-old.

Spelling cannot settle every case. Speakers differ over herb, which takes an where the h is silent and a where it is sounded. Writing was slow to follow speech as well: an stayed usual in print before words such as unique and utility until the nineteenth century. A rule built on letters has to carry short lists of exceptions and has to admit doubt.

It also has to be measured. Scored on a list of 200 nouns, a third of them chosen to be hard, a first-letter rule gets 163 right. The rule used here, with its lists, gets 197 of 200. Against the 117,467 lower-case alphabetic entries of a free pronouncing dictionary, the first letter alone agrees with the first sound 99.53 percent of the time and the fuller rule 99.84 percent, by our own measurement on 12 October 2026. The three misses that remain are sets of initials typed in lower case. Ordinary words are easy. The hard ones are the ones people notice.

Plurals are left to the writer entirely. Guessing a plural produces confident nonsense, so a frame that needs one uses only the form you typed and skips any seed without it.

Then comes the limit no rule removes: grammatical does not mean asked. A frame filled with every seed will write correct sentences nobody would say. The same guidance that encourages writing for readers lists warning signs, among them content made primarily to attract visits from search engines and extensive automation used to produce content on many topics. Its spam policies name scaled content abuse: many pages generated mainly to manipulate rankings, with little value to readers, however they are made.

Reading the list before using it

A generated list needs an editor, and the tool’s job is to make editing quick. Exact duplicates are easy to define: two phrases that are equal once case is ignored. Near-duplicates need a stated measure. One simple choice compares the sets of words two phrases hold, in the spirit of the Jaccard index, which divides the size of the intersection of two sets by the size of their union.

Lengths deserve the same care. A character, as a reader counts it, is what the Unicode standard calls a grapheme cluster, a user-perceived character that may be several code points. Exported tables follow the published CSV format, and a cell that begins with an equals sign, a plus, a minus or an at sign is guarded so that a spreadsheet does not read a phrase as a formula.

What remains is judgement, which belongs to you. Keep the phrases that name a need you can actually meet. Strike the ones that are only arithmetic. The long tail is real as a shape, and the way to reach it is still to answer one real question well.