Text Summarizer

What is a text summarizer?

A text summarizer takes a longer passage and produces a shorter one that carries the same information. There are two fundamentally different ways to do that, and this summary generator gives you both. An abstractive summary is written fresh by an AI model: it reads your text and produces new sentences in its own words, which lets it compress hard and read like prose, but the wording is the model’s, not yours. An extractive summary does something more literal: it picks whole sentences straight out of your original text, copied character for character in the order they appeared, so every word you read was already written by you — just fewer of them.

This tool defaults to the AI summary, because for most text that is the one that reads best. But it also ships two extractive modes, TextRank and word frequency, that run entirely on your device with no model call at all. Which one is right depends on what you are summarizing and what you cannot afford to have changed, which is what the comparison below is for. (Some people search for this as a “summerizer”, a misspelling that turns up more often than you would expect.)

It works natively in five languages: English, Spanish, Portuguese, French, and Arabic, each using its own stop-word list and Unicode-aware sentence splitting rather than an English pipeline with translated button labels bolted on top.

How to summarize text

1. Paste your text. Drop the article, report, or email into the text box. In the two extractive modes the summary recalculates as you type or paste, with nothing sent anywhere.
2. Choose a method and a length. Pick AI summary, TextRank or Word frequency, and drag the slider to choose how many sentences to keep (it starts around a quarter of your document, never fewer than three). In AI mode press “Summarize with AI” when you are ready, since each press is a real model call rather than something that can fire on every keystroke; the two extractive modes keep updating live.
3. Read, check, and copy. Review the summary alongside the statistics row below it, and the line underneath telling you which engine produced it. In the extractive modes, glance at the warnings panel for any sentence that lost its context, then copy the result with one click.

AI or extractive: the honest comparison

The AI summary reads better and can compress much harder, because it is free to merge three of your sentences into one clause. That is also its risk: because the wording is the model’s, it can drift from what you actually said, so anything you plan to rely on is worth checking against the original. An extractive summary can never misrepresent you this way, since every sentence is copied verbatim, but it reads like a list of highlights rather than connected prose, and pulling a sentence out of order can strand it without the context that made it make sense — which is exactly what the “lost their context” check is looking for. That check only runs in the extractive modes; an AI summary has rewritten the text, so there is nothing to detect.

The modes also differ in where your text goes, and that is a real difference in the privacy promise rather than a detail. AI mode sends your text to this site’s own server, which calls the model on your behalf, and is capped at 6,000 characters per request. TextRank and word frequency never leave your browser and have no length cap at all, because there is no server involved. In either mode, asking for a specific number of sentences is an instruction rather than a guarantee — a model may land a sentence over or under.

The AI mode also has a ceiling the extractive modes do not: a daily budget shared across every AI tool on this site, plus a per-IP limit. If the model is rate limited, out of budget, or simply fails, the page does not leave you empty-handed or tell you to come back tomorrow — it falls back to the TextRank summary, which was already computed on your machine, and says plainly that this is what you are reading. You get a summary either way, and you are always told which engine produced it.

How the two extractive methods work

TextRank treats your document as a graph: each sentence is a node, and the strength of the connection between two sentences is how much vocabulary they share. Running PageRank over that graph (Mihalcea and Tarau, 2004, with the standard 0.85 damping factor) scores every sentence with the same circular logic search engines use for web pages — a sentence is important if the sentences that resemble it are also important — and the highest-scoring sentences, up to however many the slider allows, become the summary. Because that computation reduces to solving a linear system, we checked it against an exact answer rather than trusting the iteration to converge on its own: a direct Gaussian solve of the same equations PageRank approximates agreed with this implementation to within 3.45×10⁻¹⁰ across 220 random graphs, and to 3.08×10⁻¹⁰ on graphs containing isolated sentences — the case where probability mass leaks away if the math is wrong.

Word frequency is the older approach, close to the method Hans Peter Luhn described in the 1950s: strip out stop words, count how often each remaining content word appears in the document, then score each sentence by the average frequency of its words. Averaged, not summed — summing would just reward whichever sentence happens to be the longest, regardless of what it says.

These are genuinely different algorithms, not two names for the same idea, and they produce genuinely different summaries: across article-length test documents, TextRank and word frequency picked different sentences 60% of the time — six summaries out of ten disagreed on at least one sentence. Neither is more “correct”. There is no single right extractive summary of a document, which is exactly why this tool lets you toggle between the two instead of choosing one for you.

What the extractive modes cannot do

The most important thing to understand about this tool is what it does not do: it does not write. It cannot rephrase a clumsy sentence, merge two related facts into one clause, or drop a redundant subordinate clause the way a person editing the same passage would. If a sentence in your original document runs sixty words, that same sixty-word sentence is what shows up in the summary — extraction can only choose sentences, it can’t improve them.

Pulling a sentence out of its surroundings creates a specific, predictable failure: anything in that sentence that points backward loses what it was pointing at. “This confirmed the survey” means nothing once the sentence describing what “this” refers to has been left out. That’s exactly what the warnings panel is built to catch, flagging any summary sentence that opens with a demonstrative like “This” or “That”, a third-person pronoun like “He” or “It”, or a discourse connective like “However”, “Therefore”, or “Instead”, whenever the sentence it originally followed wasn’t selected. At the default length, about one in five article-length documents (20%) produces at least one such warning — worth reading the flagged sentence back in its original context before you rely on it.

The ranking also has no idea what your text is about: it scores a sentence by how much vocabulary it shares with the rest of the document, so a sentence that reuses common words scores well whether or not it carries the actual point. The mirror image of that problem is worse — a single crucial point, stated once in unusual words, shares vocabulary with nothing around it and therefore ranks low, which means extraction is systematically bad at exactly the rare but important sentence. The length control is also quantised to whole sentences rather than to a percentage: asking for exactly two sentences returned anywhere from 41% to 76% of the original words in testing, depending only on how long those particular two sentences happened to be. It also can’t answer a question about your text, follow an instruction, or adopt a requested tone; it only ranks and selects sentences that already exist. Dialogue, bulleted lists, tables, and poetry all break the sentence-boundary assumptions the ranking depends on, and summarizing a document that’s already tightly written won’t save you much, because there’s no padding left to cut.

Why is it free?

The two extractive modes run entirely in your browser tab. Pasting your text into the box doesn’t upload it anywhere — the sentence graph, the PageRank iteration, the word-frequency counts, and the ranking all happen in JavaScript on your own machine, and closing the tab leaves nothing behind on any server. Those modes cost this site nothing to serve, however much text you put through them.

The AI summary is the one mode that costs money per use, which is why it draws on a daily budget shared across every AI tool here rather than an unlimited one — and why it is built to fall back to the free extractive summary instead of refusing you when that budget runs out. A tool that degrades is better than a tool that stops. Either way there is no account to create, no watermark stamped on your text, and no cap on how many documents you can shorten with the extractive modes.