Text summarization techniques: manual and AI methods
Short answer Summarize by hand first: read once for the point, mark the few lines that carry it, cut the rest, then rewrite in your own words and check it against the source. Use AI when the text is long, routine or arrives in bulk, but tell it to stay in the source and check every fact it gives you.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callHow to summarize text: the four-step method
Every good summary, by hand or by machine, does the same four things:
- Find the point. Read the whole text once without marking anything. Then say, out loud or on paper, what it is about in one sentence.
- Mark the carriers. Go back and mark only the lines that hold up that point: the decision, the number, the deadline, the reason. Topic sentences (usually the first line of a paragraph) and anything repeated are good bets.
- Cut the rest. Examples, stories, background and restatements go, unless the reader needs one to understand the point.
- Rewrite and check. Put the marked ideas in your own words, merging two or three into one sentence where you can. Then read your summary against the source and fix anything you bent.
A useful target: one sentence for an email or review, three to five for a report or contract, roughly a tenth of the length for anything longer.
The rewrite step is what separates a summary from a pile of highlights. If you can't say it in your own words, you haven't understood it yet.
Which summary pattern fits which text?
Different texts hide their point in different places. Pick the pattern that matches the shape of what you're reading.
| Text | Pattern | What you fill in |
|---|---|---|
| News, incident report, customer complaint | 5 Ws and H | Who, what, when, where, why, how |
| A customer story or case study | Somebody, wanted, but, so, then | The person, their goal, the problem, what they did, how it ended |
| A process or set of instructions | First, then, finally | The start, the key steps, the result |
| An opinion piece, proposal or pitch | Who says what, and why | The author, their claim, their best reason |
| An email or meeting you need to act on | The gist plus the ask | One sentence on the point, one on what you must do and by when |
Example of the story pattern, from a five-paragraph review: "A homeowner wanted her furnace fixed before a cold snap, but the first tech had no part, so the shop drove one over that evening, and she has used them every winter since." That one sentence is the review a prospect needs to hear.
Extractive vs abstractive: the two ways software summarizes
Software summarizes in one of two ways, and knowing which you're using tells you what can go wrong.
| Extractive | Abstractive | |
|---|---|---|
| What it does | Picks the most important sentences and copies them out | Writes new sentences that restate the meaning |
| Like a person who | Highlights a printout | Explains it to you over coffee |
| Strength | Every sentence is really in the source | Reads smoothly; can merge scattered points |
| Risk | Choppy; can miss a point spread across paragraphs | Can state things the source never said |
| Examples | TextRank, LexRank, older "summarizer" websites | ChatGPT, Claude, Gemini, Copilot |
The classic extractive methods treat a document as a network of sentences and rank each one by how strongly it connects to the rest. TextRank, published by Mihalcea and Tarau in 2004, borrowed this idea from how Google's PageRank ranks web pages (ACL Anthology). LexRank, by Erkan and Radev the same year, scores a sentence by how similar it is to the others, so the most "central" sentences win (Journal of Artificial Intelligence Research).
Today's chat assistants are abstractive. That is why their summaries read well, and why they need checking.
When does AI summarization help, and when should you do it yourself?
Use AI when the cost of a small miss is low and the volume is high. Do it yourself, or at least read the source, when one wrong word costs money.
| Hand it to AI | Do it yourself (or read the source too) |
|---|---|
| A 40-email thread you were copied on | A contract, lease or vendor agreement you're about to sign |
| 200 customer reviews, to find what people repeat | A legal notice, insurance policy or tax letter |
| Notes or a transcript from a long call | Anything with prices, dates or obligations you'll act on |
| An industry report, to decide whether to read it | A summary you'll publish under your name |
| Long documents you'll skim anyway | Anything where the exception is the point |
The reason is the risk OpenAI names in its own help center:
"This phenomenon is often referred to as a hallucination: when the model produces responses that are not factually accurate." (OpenAI Help Center: Does ChatGPT tell the truth?)
The same page advises: "Use ChatGPT as a first draft, not a final source." Anthropic's guidance for Claude says the same of its safeguards: they reduce made-up content but "don't eliminate them entirely. Always validate critical information, especially for high-stakes decisions." (Claude Platform Docs: Reduce hallucinations)
How to get an accurate summary from ChatGPT, Claude or Gemini
Run the manual method through the AI. Most errors come from asking for "a summary" with no reader, length or limit.
- Say who it's for and how long. "Summarize this for me, the owner, in five bullets. I need to decide whether to renew."
- Keep it in the source. "Use only this document. If something isn't in it, say so." Anthropic lists both moves, letting the model admit it doesn't know and limiting it to the documents you gave it, as ways to cut made-up content.
- Ask for quotes first on long files. Anthropic's advice for long documents is to "ask Claude to extract word-for-word quotes first before performing its task" (Claude Platform Docs: Reduce hallucinations). Then ask for the summary built only from those quotes, so every claim points to a line you can find.
- Ask for the numbers separately. "List every date, dollar amount and deadline, with the sentence it came from." Check each against the source. This is where errors hurt most.
- Split very long files. If a file is longer than the tool handles well, summarize each section, then summarize the summaries. Then spot-check the parts that matter.
For reusable prompts and a review routine, see AI marketing prompts.
Tools you probably already have
You likely don't need a separate summarizer. The apps you use daily now include one:
- Gmail: Gemini can summarize an email thread for you (Gmail Help: Collaborate with Gemini in Gmail).
- Word: Copilot can summarize a document on request, or show a summary at the top when you open it, depending on your license and settings (Microsoft Support: Create a summary of your document with Copilot in Word).
- ChatGPT, Claude or Gemini: paste or upload the text and use the steps above.
Before you paste client or customer data, check the privacy setting. OpenAI's data controls page says that on personal ChatGPT Free, Plus or Pro accounts, "data sharing is enabled by default," and you can turn off "Improve the model for everyone." It also states: "By default, OpenAI does not use content from ChatGPT Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces to train its models." (OpenAI Help Center: Data controls FAQ)
For which AI tool fits which job, see AI tools for small business.
The one-sentence summary you write most: your page description
The meta description under your page title in Google is a one-sentence summary. Google describes it as "a short, relevant summary of what a particular page is about" and warns that "Identical or similar descriptions on every page of a site aren't helpful" (Google Search Central: Control your snippets in search results).
Use the gist pattern: what the page offers, for whom, and why to click. Write a different one for each page. See how to write meta descriptions.
Example
A plumber summarizing reviews
A plumber has 150 Google reviews and wants to know what customers repeat. She pastes them into ChatGPT and asks for the five themes most often mentioned, with a count and two quotes each. It reports "on time" in 60 reviews and "price explained up front" in 35. She opens five of the quoted reviews to confirm they say that, then puts both lines on her homepage.
A dental office and a new lease
An office manager gets a 30-page lease renewal. She asks Claude for every date, dollar amount and obligation, each with the sentence it came from. It lists 14 items. She checks all 14 against the lease and finds one rent figure from the old term. She reads the renewal and escalation clauses herself before the meeting with the landlord.
A contractor's weekly email pile
A general contractor gets 300 emails a week. He uses Gemini's thread summary in Gmail for supplier updates and newsletters, and reads every message from clients and inspectors in full.
FAQ
What are the main text summarization techniques?
By hand: find the point, mark key lines, cut the rest, rewrite. Software either copies key sentences (extractive) or writes new ones (abstractive, like ChatGPT).
Is AI summarization accurate?
Often, not always. OpenAI and Anthropic both say their models can state false things, so check every number, date and quote against the source.
How long should a summary be?
Long enough to act on: one sentence for an email, three to five for a report, about a tenth of a long document.
Can I summarize a contract with AI?
As a first pass, yes. Don't sign on the summary alone; read the clauses on money, dates and obligations yourself.
Is it safe to paste customer data into ChatGPT?
Check settings first. Personal Free, Plus and Pro accounts share data for training by default (you can opt out); Business and Enterprise don't by default.
Different for you?
- Writing one-sentence summaries for your pages: see how to write meta descriptions.
- Using AI to draft content, not just summarize it: see AI content creation.
- Turning summaries of customer feedback into better pages: see the Conversion agent.
- Summaries for search and AI answers: see the SEO agent.