2026-07-02 · Calc-Tools Editor

Keyword Density Analyzer: Key Metric for SEO Optimization

Keyword density is a fossil metric. It was born in an era when search engines ranked pages by counting words, and it stopped being a ranking factor long ago. Throwing it away entirely is still a mistake, though: it remains the cheapest diagnostic available for three engineering jobs — checking whether a draft has drifted off topic, keeping terminology consistent across technical documentation and translations, and producing the term-frequency baseline that topic extraction and TF-IDF are built on. This article covers what the number is still good for, and why it should never become a writing target again.

One boundary first. Our word counter now includes keyword density analysis and reports characters, words, and frequency shares in a single pass. The separate question of how many different "word counts" exist, and whether a platform limits characters or words, belongs to the complete word count guide.

1. The Era When Word Frequency Decided Rankings

Around the year 2000, relevance was computed almost entirely from an inverted index plus term counts. The more often a document contained a phrase, the more likely it was about that topic. At the time this was a reasonable approximation: there was not enough text available to learn synonyms, and no practical way to model context.

If frequency is the ranking signal, then raising frequency is the optimization. An entire toolkit grew up around counting: repeating a phrase in titles and body copy, stuffing keywords into image alt text, appending a keyword list at the bottom of a page, hiding keywords in text colored to match the background, and generating a separate doorway page for every phrase. Keyword density turned from a descriptive statistic into an operating target during this period. It was never anything more than a number that described text, and it was treated as a dial to be turned.

The "2% to 5% is ideal" advice that survives today came from observing pages that ranked well. It was empirical feedback, not a finding from information retrieval or linguistics. It had no theoretical basis, and it could not explain why the same number means something different in another language, another length, or another genre. Section 4 shows the arithmetic.

2. Why It Is No Longer a Ranking Factor

Three generational shifts removed its signal value permanently.

1. Semantic understanding replaced literal matching. Modern retrieval models map terms into vector space, where "affordable" and "cheap", or "laptop" and "notebook computer", already sit close together. The same idea expressed differently is still recognized. Literal frequency stopped being necessary information.
2. Entity recognition replaced string counting. Engines first resolve terms into entities — brands, products, places, standards, people — then judge which entities a document covers and whether the relationships among them are explained. The unit of measurement moved from "how many times did this word appear" to "was this explained properly".
3. Intent matching replaced keyword matching. The same query can carry informational, navigational, or transactional intent, and the result page has to supply the matching content format. That judgement happens on the query side and has nothing to do with how often a word appears in your document.

The conclusion: in modern ranking, term frequency is not a weak signal, it is noise. Deliberately inflating it buys no ranking and costs readability. The table below places the metric next to what replaced it.

Notice what the last three rows have in common: none of them outputs a percentage. What replaced density is not a better density, but a set of coverage questions that require human judgement.

3. Three Places Where It Is Still Useful

Now the part that matters. In the three situations below, term counting remains the least expensive approach available.

Use 1: Self-checking a draft for topic drift

The point is not to treat density as a pass mark, but as an existence check. If a two-thousand-word article never mentions its subject in the title, the opening paragraph, or a single subheading, then what the piece is about becomes fuzzy for readers and machines alike. That is not a low density problem; it is a genuinely missing topic signal.

Conversely, if the phrase appears in every paragraph while the article answers nothing specific, a tidy percentage changes nothing. So the right question is never "did we hit 2 percent". It is three questions: do the title and opening state the subject; does at least one subheading carry the subject or an equivalent phrase; does the body contain at least one concrete, checkable answer.

Use 2: Terminology consistency in documentation and translation

This is the most underrated use of a frequency tool. In technical documentation and localization projects, one concept written several ways is a persistent hazard: a single document mixing "callback" and "call-back", or a translation rendering "token" as both "令牌" and "凭证", leaves readers assuming these are different things.

The method is to run a frequency count on the source and the translation separately, then align both against a glossary. If a source term occurs 40 times and its approved translation occurs only 12, where did the other 28 go? Either they were rendered with a different term, or whole passages were rewritten or dropped. The frequency gap is a locator. It points at passages that need human review; it is not a score to be optimized.

Use 3: Text analysis, where frequency is the starting point

On the data side, term frequency is the first step of every lightweight text pipeline: filter stop words, sort by count, and you have candidate topic terms. The first version of a feature set for feedback clustering, ticket triage, or review summarization usually starts from exactly that list. TF-IDF does nothing more than multiply frequency by a rarity weight — it does not discard term frequency, it gives it a denominator.

Getting that pipeline right is worth far more than arguing about a density number. Cleaning invisible characters and stray whitespace is the prerequisite step, and the text cleaner handles it; when the matching rules get more specific, such as counting terms with a particular suffix or filtering sentence patterns, validate them first in the regex tester before running them at scale.

4. Worked Example 1: One Text, Density Readings 2.56x Apart

This is the most direct evidence that density is not a comparable quantity. Take a Chinese body of text: 1000 CJK characters, 1050 characters once punctuation is included, with the target phrase appearing 12 times. Change how the denominator is defined and the reading changes with it.

The first two rows follow directly from the premise above. The 620 and 410 in the last two rows depend on the specific segmenter and stop-word list, so they are marked illustrative. What is certain is the direction: switching to a segmented or stop-word-filtered denominator always shrinks the denominator and inflates the reading, and those two choices are the main reason Chinese tools disagree with each other.

The phrase is unchanged, no sentence was added, not a single character moved, and yet the reading climbs from 1.14% to 2.93% — the highest value is 2.56 times the lowest. More importantly, under the old guidance 1.14% signals "too low, weak topic signal" while 2.93% lands neatly inside the "2% to 5% ideal range". The same article both fails and passes, depending entirely on the denominator.

Chinese makes this worse. There are no spaces between words, so any "total word count" depends on a segmentation engine, and engines disagree about whether a compound phrase is one word or four. Counting by character is perfectly reproducible, but the denominator is visibly larger and the reading visibly smaller. Chinese and English density readings are therefore not comparable as numbers — English splits on spaces, so its denominator is unambiguous. Note: any claim that Chinese density "should stay under 3%" quietly assumes a denominator it never states.

5. Worked Example 2: High Frequency Does Not Mean Useful

The 58-character product description below is unremarkable. Counting adjacent two-character pairs — an unambiguous, fully reproducible method — the most frequent pair is "支持" (supports), while the actual subject of the passage, "离线" (offline), appears twice.

Text: 本工具支持离线处理，支持批量导入，支持导出报告。所有数据支持本地计算，不会上传服务器。离线处理意味着断网也能用，适合处理敏感数据。 (This tool supports offline processing, supports bulk import, supports report export. All data is processed locally and never uploaded. Offline processing means it works without a network, which suits sensitive data.)

Pick a primary term by density and you would choose "supports" at 6.90%, double the 3.45% of "offline". Yet "supports" can be swapped for "provides" or "offers" with no loss of meaning and contributes nothing to the topic, while deleting "offline" would leave the passage with no claim at all. Density measures how many times something appeared. Topical relevance measures whether removing it would damage the meaning. These two have never been equivalent.

As an aside, 6.90% already approaches the old warning line of "above 8% may be treated as stuffing". A generic verb being flagged as suspicious stuffing is a good indication that the threshold itself carries no information.

6. Worked Example 3: How Modern Metrics Correct It

TF-IDF, and the BM25 variant used in retrieval, do exactly one essential thing: multiply term frequency by a rarity weight so that terms found everywhere are automatically devalued. A tiny corpus demonstrates it. Suppose a site has only 3 documents and the one under analysis contains 200 words.

Both terms occur exactly the same number of times in this document, yet "offline" ends up weighted 1.69 times higher, entirely because of cross-document rarity. The smoothed IDF formula ln((N+1) / (df+1)) + 1 with N = 3 yields 1.00 when df is 3 and 1.69 when df is 1.

That correction explains why keyword stuffing cannot work. No amount of repetition inside your own page changes how rare the term is across the corpus. The only lever you can pull is the numerator, and the numerator is precisely the part designed to be corrected by the denominator.

7. What Stuffing Actually Costs

There was a clear turning point when keyword stuffing shifted from "technique" to "cheating": engines began detecting hidden text, keyword lists, doorway pages, and auto-generated low-quality pages at scale, and around 2011 a cluster of algorithm updates specifically targeted low-quality content and ranking manipulation. Pages of that kind were pushed down systematically afterwards. Deliberate stuffing went from diminishing returns to a certain risk.

But three costs arrive earlier than any ranking penalty, and none of them depends on an algorithm update.

• Readers leave first. The immediate result of stuffing is stiff prose with one phrase recurring inside a single paragraph. By the third repetition the reader knows exactly what you are doing.
• Rewriting costs are irreversible. Stuffed text has almost no reuse value; fixing it means rewriting whole passages, whereas naturally written paragraphs need a few word swaps.
• It hides the real content problem. Treating density as proof of optimization makes it easy to stop at a passing number instead of supplying the missing answer — which is usually the actual reason a page does not rank.

Modern detection does not stop at frequency either: duplicated passages, semantic jumps, and templated structure are evaluated together. In other words, even detecting stuffing no longer relies on the density reading.

8. Four Common Misconceptions

Misconception 1: density is a writing KPI. 2% is not a pass mark and 5% is not a ceiling. The range came from observing early ranking results; it has no theoretical basis and it drifts with the denominator (see section 4). Writing it into a style guide gives authors a target unrelated to the goal.

Misconception 2: low density means the topic is unclear. Not necessarily. An article about generating pages at scale can explain the subject thoroughly without ever using the phrase "programmatic SEO", and the topic stays perfectly clear. Judge clarity by title, opening, subheadings, and answer paragraphs, not by how many times a phrase appears.

Misconception 3: a synonym will fool the engine, or the exact phrase is mandatory. Both ideas are outdated. Semantic models already handle synonyms and broader or narrower terms, and the literal phrase was never a required incantation.

Misconception 4: high density only affects rankings. It affects people first. Ranking movement is a probability; losing readers is a certainty.

9. Frequently Asked Questions

What keyword density should I aim for?

There is no universal threshold. The 2% to 5% range came from observing rankings in the 2000s, has no theoretical basis, and drifts with the denominator: the same 1000-character text reads 1.14% against total characters and 2.93% against content words after stop-word removal. Only three checks are actionable: does the subject appear in the title and opening, do the subheadings carry it, and does the body contain a concrete, checkable answer.

Should I still optimize density deliberately?

No. Inflating frequency buys no ranking and costs readability, and stuffed passages are far more expensive to rewrite than naturally written ones. What actually works is covering the topic: address the subtopics, the relevant entities, and the questions readers genuinely ask, then let the phrase and its synonyms appear where they belong.

How can I tell whether an article has drifted off topic?

Check three things: whether the title and opening describe the subject in one sentence, whether the subheadings form a continuous line of reasoning, and whether each subheading is followed by a specific, checkable answer. A frequency table can only flag hard failures such as the subject never appearing at all; drift more often shows up as subheadings that do not connect.

Can Chinese and English density readings be compared directly?

No. English splits on spaces, so the denominator is unambiguous. Chinese has no inter-word spaces, so any total word count depends on segmentation, and engines disagree about how many words a compound phrase contains. Counting by character is reproducible but inflates the denominator and deflates the reading. Each set is fine for tracking its own change over time, not for comparison against the other.

What is the difference between density and TF-IDF?

One denominator. Density counts how often a target phrase appears on the page. TF-IDF multiplies that count by how rare the term is across the whole corpus. As a result two terms with identical counts score differently: the one found only in this document is weighted higher (1.69 times higher in the example above). This also means repeating a phrase on your own page cannot raise its cross-document rarity, which is why stuffing cannot work.

10. Six Practical Recommendations

• Establish the denominator first. Whenever you are handed a density reading, confirm whether the denominator is total characters, CJK characters, or segmented words. Readings with different denominators should never be compared.
• Trade repetition for placement. Put the subject phrase in the title, the opening, and at least one subheading. In the body, use pronouns and equivalent phrasing whenever that reads better instead of forcing the literal phrase back in.
• Manage consistency with a glossary. Technical and localization projects should keep a glossary and align the source and target frequency lists; every position where they diverge is a position that needs review.
• Check topic coverage across articles, not density within one. Articles on the same subject should be checked against each other for distinct subtopic coverage. Inflating one phrase in every article only sets them up to compete.
• Treat the frequency list as a diagnostic, not a scorecard. Filter stop words first, then read the distribution of content words. Use the text cleaner for the cleanup step, and validate pattern rules in the regex tester before applying them.
• When a reading looks wrong, suspect the denominator before the text. Density rarely doubles because the content changed; it doubles because the measurement did.

Put together, the workflow is short: finish the draft, run it through the word counter to inspect the frequency distribution, clean up noise with the text cleaner if needed, and when planning publishable length, measure it with the reading time tool and see the reading time guide for the conventions behind it. Density shows up only at the first step, where it answers exactly one question: has this drifted off topic?