What a summary is supposed to do
A summary is not a shorter version of a text. It is an answer to a question the reader has not asked out loud: do I need to read this, and if I do not, what did it say? Those are two different jobs, and most bad summaries fail because they attempt neither — they simply take the first few sentences and stop.
A good one keeps the claim, the evidence that supports it, and any figure a decision would turn on. It drops the throat-clearing, the examples that repeat a point already made, and the transitions that only exist to join paragraphs together.
Extraction and abstraction fail in opposite directions
There are two ways to shorten a document, and knowing which one you are getting tells you what to distrust.
Extraction selects sentences from the original and returns them verbatim. Nothing can be misquoted, because nothing is rewritten. What it loses is flow: the selected sentences were written to sit next to different neighbours, so pronouns lose their referents and a sentence that began with however now contradicts nothing.
Abstraction writes new sentences describing what the text said. It reads far better and carries the risk that matters: a model can phrase something the source did not quite claim, merge two hedged statements into one confident one, or attach a figure to the wrong subject.
The [msm_tool slug="text-summarizer"] does extraction in your browser and abstraction when you add an AI key, and it labels which one produced the result. That label is the useful part — it tells you whether to check the summary against the source or merely against your own reading of it.
How sentence ranking actually works
Browser-side extraction scores every sentence and keeps the strongest. The scoring is simpler than it sounds:
- Term frequency. Words that recur across the whole document are what the document is about. Sentences carrying more of them score higher. Stopwords are ignored, or every sentence would score for the.
- Length normalisation. Scores are divided by the square root of sentence length, so a long sentence cannot win by containing more words, and a three-word fragment cannot win by being dense.
- Position. Sentences near the opening and the close get a modest lift, because that is where written prose puts its thesis and its conclusion.
Selected sentences are re-emitted in their original order rather than by score, which keeps the argument running in the direction the author wrote it.
Text that summarises badly
Some documents defeat both methods, and it is worth recognising them before you trust the output.
- Dialogue and interviews. Meaning lives in the exchange. Any single line pulled out of a conversation loses who was pushing back on whom.
- Narrative. A story's point is usually its sequence. Ranking sentences by keyword density returns the most eventful moments in an order that no longer tells the story.
- Documents that are already dense. An abstract, a changelog or a specification has no padding to remove. Shortening it removes content.
- Anything with a table. Structured data does not survive being flattened into prose, and the numbers usually are the point.
- Legal and medical text. The hedges and qualifiers are load-bearing. A summary that drops may, generally or except where has changed the meaning, not condensed it.
Getting a better result
Paste clean text. Navigation menus, cookie banners and footer links copied along with an article all carry words that the ranking will treat as content, and a summary that mentions the newsletter signup is the result.
Give it enough to work with. Anything under a few hundred words has too few sentences for ranking to mean much — you will get back most of what you pasted, which is not a failure of the tool so much as a question that did not need asking.
With a key, be explicit about the shape you want. A line such as in three bullets, for a colleague who has not read the original at the end of your text changes the output far more than any setting will.
Check the summary against the source
This is the step people skip, and it is the one that matters. Read the summary next to the original and ask three questions: does every claim in the summary appear in the text, does the summary hedge where the source hedged, and has any number changed its units or its subject?
Abstractive summaries fail those checks quietly. The sentence reads well, the tone is right, and one figure has quietly become an annual total when the source said quarterly. Nothing in the output looks wrong, because nothing about generated prose ever looks wrong.
Everything happens in your browser or, with a key, between your browser and the provider you chose. The text does not pass through this site, which is what makes it reasonable to paste a document you would not upload somewhere.