A long tail keyword is a query with low search volume and a very specific intent. Word count is a frequent side effect, not the definition: "led wall price" is three words and behaves like long tail, "how to make pancakes" is four and behaves like a head term.
Most guides on the subject open with the same figure: 70% of Google searches are supposedly long tail. None of them attributes it to a study you can actually read. Here the maths is redone on sources you can open, plus the numbers from a project we have been running since 2020.
What long tail keywords are
The term comes from Chris Anderson, who in 2004 described in Wired how, in digital markets, the sum of niche products can be worth as much as the bestsellers. SEO borrowed the curve: on the left a few queries with huge volumes, on the right a very long tail of rare queries.
An example from a sector we work with often. "Safety boots" is a head query, with 18,100 searches a month in the UK. "S3 waterproof safety boots" is long tail, with 20 (DataForSEO, Google UK, September 2026): few people search for it, but the ones who do have already decided what to buy. The first brings traffic, the second brings quote requests.
The distinction changes the kind of page worth writing. On a head query you compete with sites that have ten years of authority behind them; on a long tail query you compete with whoever has not written that page yet. That is why properly done keyword research does not stop at the list of highest volumes.
The difference from short tail is not the number of words
Counting words is convenient and misleading. "Led wall price" has three words and 30 searches a month in the UK: high commercial intent, a clear need, and it behaves like long tail in every respect. "Pumpkin soup recipe" also has three words, averages 22,200 searches a month and reached 135,000 in October 2025 (DataForSEO, Google UK, September 2026). In autumn it is a head query by any measure.
The criterion that holds up on its own is volume, read together with search intent. The others are correlations.
| Criterion | Short tail (head) | Long tail |
|---|---|---|
| Monthly volume in the UK | a few thousand and up | from 0 to a few hundred |
| Number of words | variable | variable |
| Intent | ambiguous, often multiple | explicit |
| SERP competition | domains with strong authority | specific pages, often few of them |
| Distance from purchase | high | low |
| How many exist | a few thousand per sector | practically infinite |
A practical consequence: if you are judging a keyword only by how long it is, you are using the wrong proxy. Open the SERP and look at who ranks. If page one shows five category pages from large marketplaces, that query is a head term even if it has seven words.
How much the long tail really weighs
The most solid publicly available data comes from the keyword database of Ahrefs, which reports over 110 billion keywords discovered and 28.7 billion kept after its popularity filter. On the US database alone, the distribution looks like this.
| Monthly volume band | Number of keywords |
|---|---|
| Over 100,000 searches a month | just under 18,000 |
| Fewer than 10 searches a month | 2.3 billion |
| Share of the database under 10 searches a month | about 93% |
Eighteen thousand keywords above one hundred thousand monthly searches, against 2.3 billion below ten. The ratio is one to one hundred and twenty-eight thousand. That is the real weight of the long tail, and it has nothing to do with the 70% that has been doing the rounds for years without a source.
The second data point comes from Google. In 2017 the company publicly confirmed that 15% of each day's searches had never been seen before. John Mueller repeated the same percentage at Search Central Live in New York in March 2025, noting that the figure has stayed stable even after the arrival of generative systems. A search engine that, after twenty-five years, still sees one search in seven for the first time describes a market where the tail never runs out.
Third data point, useful because it goes the opposite way from what you would expect. Constructor analysed over 70 million internal search requests on e-commerce sites and found that between 75% and 90% are shorter than four words. This is not Google search, it is search inside a site, and for that very reason it neatly takes apart the equation between long tail and long phrase: people type little even when they are looking for something very specific.
Three families of long tail, and only one deserves its own page
Treating them all the same way is the mistake that produces archives of thin articles. Ahrefs proposes a three-way split that works in practice.
| Family | What it is | What to do with it |
|---|---|---|
| Supporting | A less common variant of a more popular query ("led wall cost" compared with "led wall price") | No dedicated page. It ranks by itself when the main page ranks |
| Topical | The most common way of searching for that specific thing, with low volume because the thing is niche | Dedicated page. This is where the value is |
| Conversational | The natural-language question asked to ChatGPT, Gemini or AI Mode | No measurable volume. You cover it by covering the topic well, not the single phrase |
Telling the first from the second is the single check that saves the most time in a content plan. If two queries return the same first page of results, they are the same query and one page covers both, which is how a keyword cluster is built. We described the method for checking it in the guide on how to create SEO content.
LLM prompts are a different kind of long tail
This is where the classic model breaks, and it is the part most guides leave out entirely.
Semrush analysed almost 69 million desktop search sessions in the United States between 1 May and 5 July 2025 and measured average query length: 7.22 words in AI Mode against 4.0 in traditional search. Almost double.
The gap gets much wider on ChatGPT. Again from Semrush, on 80 million rows of clickstream data collected between July and December 2024, it measured an average of 23 words per prompt when the user does not turn on web search, and 4.2 words when they do. In the first case we are no longer talking about keywords: we are talking about whole sentences, often with context, constraints and a budget in them.
No keyword research tool will ever show you those 23 words, because they are never repeated identically. The practical consequence is that covering a topic matters more than matching a string, and that the work shifts towards how the page is read and cited by the models. On this we have written the guide to generative engine optimisation and the one on how to rank in answer engines; for the specific behaviour of the two most used systems there are the pieces on SEO for ChatGPT and on Google AI Mode.
The long tail converts, but one conversion at a time
The most honest data on the subject is eight years old and still hardly anyone cites it. In January 2018 Wil Reynolds of Seer Interactive analysed 1,698,576 unique search terms taken from one client's paid search data over a year. Of these, about 107,000 had generated direct or assisted conversions.
Within those 107,000: 81% had produced a single conversion in the whole year, and 95.3% had produced between one and four. The long tail is not a seam of keywords that each convert a lot. It is a seam of keywords that convert very little one by one and a great deal added together.
This changes how you measure. Asking a single long tail page to justify itself leads to shutting it down after three months. The return is read on the cluster, not on the row, and you need a horizon of at least six months before the total becomes readable in ranking terms.
How to find long tail keywords for the UK market
The UK market has a question of scale that needs saying before the methods. According to Ofcom, Google Search reaches 82% of UK adults (Online Nation 2025). That is high reach, but on a population about a fifth the size of the United States. Put simply: volume thresholds need rescaling. A keyword with 200 monthly searches in the UK is a meaningful figure, not something to discard. Most published keyword benchmarks are built on US data, and read literally they make almost every UK B2B query look too small to bother with.
The five sources that, in our experience, pay off most, in order of return:
- Search Console, filtered for high impressions and low clicks. These are queries the site already appears for without having been designed to. Zero cost, real data, no estimates.
- Emails to sales and customer service. Written questions are already phrased in natural language, which is exactly the shape of conversational queries.
- Google autocomplete, "People also ask" and related searches. Free, and they reflect how people in your market actually phrase things, including British spellings and terms that US-based tools often underweight.
- A tool with a UK database. Ahrefs, Semrush and similar tools give different estimates for the same terms: use just one for consistency and do not chase the decimal point.
- Forums, groups and subreddits in your sector. The place where you read what people call things before a tool records any volume.
What is worth avoiding is mass-generating variants with AI without filtering them: it produces lists of supporting keywords dressed up as opportunities, and the result looks a lot like keyword stuffing spread across several URLs rather than inside one page.
When a long tail keyword deserves a dedicated page
This is the grid we use before putting a line in the content plan. If a keyword does not pass at least three checks out of four, it becomes a paragraph inside an existing page.
| Check | How to verify it | Passes if |
|---|---|---|
| Different SERP | Compare the top 10 results with those of the parent keyword | At least 6 URLs out of 10 are different |
| Independent intent | Look at the dominant format on page one | It asks for a format the parent page does not have |
| Enough substance | Try to list the subheadings without looking anything up | At least four come out with real content |
| Commercial value | Ask sales whether that question comes up in negotiations | The answer is yes, even once a month |
The fourth check is the one most often skipped and the one that weighs most. A keyword with 90 monthly searches that sales hears in every negotiation is worth more than one with 900 that nobody has ever mentioned on the phone.
Macropix: the long tail in an Italian B2B market
Macropix is a Milan-based manufacturer of custom LED screens, with over a thousand installations worldwide. When we started, in 2020, the site did not appear for any of the sector's category keywords. The cluster we worked on is almost entirely long tail by Italian volumes.
| Keyword (Italian) | Monthly volume in Italy | Position 2020 | Position 2025 |
|---|---|---|---|
| monitor pubblicitario (advertising monitor) | 390 | 88 | 2 |
| ledwall outdoor | 260 | not ranking | 1 |
| totem led (LED totem) | 260 | 31 | 3 |
| ledwall indoor | 210 | not ranking | 1 |
The row we talk about most is the first: "monitor pubblicitario", from position 88 to position 2. Position 88 is the ninth page of Google, which means zero traffic.
Added together, these four queries of under 400 monthly searches each helped take macropix.it to a 25.55% share of voice on the ledwall cluster in July 2026, ahead of amazon.it at 18.08%. A niche Italian manufacturer ahead of Amazon on its own category keywords is the most direct proof of the principle: on the long tail, specialisation beats size. The full method is in the guide to SEO for manufacturing and in the broader reasoning on B2B SEO strategy.
Four mistakes we see again and again
- One page for every variant. It causes cannibalisation between URLs that answer the same question. Before opening a page, compare the two SERPs.
- Choosing by number of words. It leads you to discard short, specific, uncontested queries and to chase long phrases nobody types.
- Measuring at three months. With single conversions spread across hundreds of queries, three months is not enough to tell noise from signal.
- Optimising the page for the exact string. With generative systems rephrasing questions, literal matching counts for less than topic coverage. The updated rules are in the guide to on-page SEO.
If you want to know which long tail queries your site can realistically win in the next six months, it is the kind of analysis we open every SEO consultancy project with.
Frequently asked questions
There is no threshold. The criterion is low search volume combined with a specific intent, not length. "Led wall price" has three words and is long tail; "how to make pancakes" has four and is a head query.
The figure has been circulating for years without a study you can check behind it. The closest verifiable data comes from Ahrefs: in its US database about 93% of keywords have fewer than 10 searches a month, against just under 18,000 keywords above 100,000 monthly searches.
They convert more precisely but very few times each. In Seer Interactive's analysis of 1,698,576 search terms, 81% of the keywords that had brought a conversion had brought just one in the whole year. The return is read on the cluster total, not on the single query.
Google Search Console filtered for queries with many impressions and few clicks is the best source and costs nothing. Add Google autocomplete, the "People also ask" box, the related searches at the bottom of the page and the questions customer service receives by email.
They are a separate family. Semrush measured an average of 23 words per prompt when the user does not turn on web search, and 7.22 words per query in Google AI Mode against 4.0 in traditional search. They have no measurable volume and you do not capture them with exact matching, but by covering the topic thoroughly.
Something the industry rarely says: the economic value of the long tail lay in the clicks, and long tail clicks are the first thing generative systems take, because they are exactly the questions a model answers well without sending the user anywhere. Working on the long tail today no longer means buying cheap visits; it means buying the probability of being the cited source when that question is asked. They are two different goals and they are measured differently, and anyone still judging them by the old yardstick will shut down projects that are working.
