Five categories, doing five different jobs
Lumping these words together as "filler" hides the fact that they fail in different ways, so the tool separates them and you should treat the counts separately too.
Intensifiers — very, really, extremely, incredibly — modify an adjective upward and almost never succeed. Very cold is weaker than freezing, because the reader has to combine two vague words instead of receiving one precise one. The usual fix is not deletion but replacement of the adjective.
Hedges — I think, perhaps, somewhat, tends to — reduce your commitment to a claim. These are the most contested category, because honest uncertainty is a virtue and stripping every hedge produces writing that overstates what it knows. The question to ask of each one is whether you are uncertain, or whether you are certain and being polite about it.
Fillers — just, actually, basically, obviously — are discourse markers imported from speech. They regulate conversation, where they do real work, and in writing they mostly sit at the front of a clause taking up room. Obviously and clearly carry an extra hazard: they tell a reader who did not find the point obvious that they should have.
Padding phrases are the only category with a mechanical fix, and the page gives you the replacement. In order to is to. Due to the fact that is because. At this point in time is now. These are close to free wins.
Vague attribution — studies show, experts say, some argue — attaches a claim to nobody in particular. If the study is worth invoking, name it; if you cannot name it, the sentence is asserting something you have not verified.
The words that are load-bearing about half the time
Just deserves special attention because it is the single most common target of automatic deletion and it has at least four distinct meanings. It means only, as in just two of them survived — delete that and the sentence is wrong. It means a moment ago: I just called. It means exactly: just enough. And it is a discourse softener: I just wanted to check — that fourth one is the deletable one, and it is not the majority. This is precisely why the tool shows you the line rather than offering to strip them.
Actually works the same way. It genuinely means contrary to what you would expect, and used that way it is doing work. Used as a verbal shrug at the start of a sentence, it is not. Literally is worth checking every time, because when it is being used as an intensifier for something not literal, the sentence usually reads better without it and occasionally reads absurdly with it.
What the per-thousand figure is for
Raw counts are not comparable between a 400-word email and a 6,000-word report, so the summary normalises to matches per thousand words. That makes it usable in exactly one way: comparing a draft against your own earlier draft, or against a piece of your own writing you were happy with. It is not usable as a benchmark against other writers, because these word lists are a choice rather than a standard, and a different tool with a different list would produce a different rate on identical text.
The lists here run to roughly a hundred words and phrases and are deliberately conservative. Longer lists exist and they pick up words like that, then, some and make, which appear so often in ordinary correct prose that the output stops being readable. The trade is real: a shorter list misses some padding, and a longer list buries the padding it finds.
Matching, line numbers and what counts as a hit
Each entry is matched case-insensitively on whole words, with word boundaries at each end, so just does not match inside adjust and a bit does not match inside arbitrary. Multi-word phrases allow any run of whitespace between their words, so a phrase broken across a line break is still found. Overlaps between categories are possible in principle — a phrase can contain a word from another list — and each entry is counted independently, so the category totals can exceed the number of distinct spots in the text. Line numbers count physical lines in what you pasted, not sentences and not paragraphs.
Up to six example lines are kept per word, and the overall number of example lines printed is capped by the field above so a long document does not produce a page you cannot scroll. Narrowing to a single category is usually a better way to work through a long draft than raising the cap.
How to actually use the output
Work down the list in frequency order and stop early. The word appearing forty times is where the pattern is, and fixing it changes the texture of the whole piece; the word appearing twice is not worth the attention. For each frequent word, read three or four of its lines and decide whether they share a shape — hedges clustering in the introduction, intensifiers clustering in the conclusion, just clustering in the sentences where you were least confident. That pattern is more useful than the count, and it is the reason the lines are printed instead of only the totals.
When you have finished, the passive voice finder covers the verb side of the same problem, and word frequency will show repetitions these fixed lists cannot anticipate. If the draft is heading for a length limit, the text statistics tool gives you the word count in the same pass.
Questions people ask
Should I delete every match?
No, and a tool that offered to do it automatically would be dangerous. Padding phrases such as "in order to" and "due to the fact that" are near-automatic cuts and the page prints the replacement for each. Everything else is a judgement call made one line at a time, which is why the lines are shown. "Just" alone has four meanings and only one of them is deletable; "perhaps" is a hedge when you are stalling and an accurate statement when you genuinely do not know.
What is a normal rate per thousand words?
There is no published standard, and any figure quoted as one would just be that tool's word list restated. The rate is worth comparing against your own earlier drafts or against a piece of your own writing you liked, because both use the same list. Comparing it against a number from a different checker is meaningless, since the lists differ by dozens of entries and a longer list mechanically produces a higher rate on identical text.
Why is "clearly" on the filler list when the thing really is clear?
Because of who it is addressed to. If the point is clear, the reader already knows and the word is redundant. If it is not clear to them, the word tells them they have failed to follow something you consider obvious, which is a small insult delivered by accident. The same applies to "obviously" and "of course". They are not errors, and in a conversational register they can soften a transition. They are on the list so you see the count and can decide whether the tone is what you intended.
Can I use it on speech or interview transcripts?
You can, and the counts will be very high, because most of these words are legitimate features of speech rather than defects. Spoken language uses discourse markers and hedges to manage turn-taking and to signal attitude, and a transcript stripped of them reads as robotic. If you are editing a transcript into prose the counts are genuinely useful; if you are assessing how somebody speaks, they are not measuring anything meaningful.
Does my text leave the browser?
No. The word lists ship inside the page and every match is computed locally in JavaScript on your machine. There is no upload, no logging of the input and no network request in the code path, so a confidential draft stays on your computer.