Text Summarisation

Text summarisation is the natural language processing task of compressing a long text into a short one while keeping its central idea. MLforSEO uses it as the classic example of text-to-text content transformation: turning page copy into meta descriptions, video transcripts into show notes, or long reports into briefs.

How it works

There are two main approaches, and the difference between them decides which one you should use.

• Extractive summarisation selects the sentences that best represent the text and returns them unchanged. Models such as BERT score each sentence, often by comparing sentence embeddings with the document as a whole, and keep the top N. Because it only reuses your own wording, it cannot invent facts.
• Abstractive summarisation writes the key points in new words. Sequence-to-sequence models such as T5 and decoder large language models do this. The copy reads better, but the model can add details that are not in the source.

Summaries are usually evaluated against human-written references with overlap metrics such as ROUGE, and in practice by a person reading a sample.

How it’s used in SEO and marketing

• Generating meta descriptions in bulk for pages that are missing them.
• Drafting title and H1 variants, or short abstracts for category and listing pages.
• Condensing video or podcast transcripts into descriptions, chapters or blog outlines.
• Summarising competitor content, reviews or audit outputs so a person can review them faster, as in MLforSEO’s YouTube audit workflow, where transcript summaries are fed to an LLM to compare outlier videos.
• Writing self-contained summary passages that are easier for AI search systems to extract and quote.

Example

MLforSEO’s meta description tutorial scrapes the text of pages with missing descriptions, then runs a BERT extractive summariser over each page in Python. The 2026 refresh adds embeddings, a small open-source LLM and LLM APIs, and shows a real pitfall: in the original code, max_length=150 did not limit output length (it filters which sentences are eligible), so some “descriptions” came back several sentences long. The fix is to control the number of sentences and check lengths before anything goes live.

Limitations and things to watch

• Abstractive methods can produce hallucinations. Google’s guidance on generative AI content says it is critical to fact-check and review AI-generated content, metadata included, before publishing.
• Extractive output can be clunky or pick the wrong sentence (a cookie notice or a generic intro).
• Google rewrites many snippets anyway, so a summary does not guarantee what appears in results. Its snippet documentation still encourages programmatic descriptions for large sites as long as they are human-readable and diverse.
• Templates that produce near-identical summaries across pages defeat the purpose. Keep a human in the loop for brand voice and accuracy.

Learn it on MLforSEO

• How to Automatically Generate Meta Descriptions with Python: From BERT to LLMs
• How to Find Machine-Learning-Enabled Automation in Your SEO Workflow uses meta descriptions and titles as worked examples.
• How to do a YouTube SEO Audit uses transcript summaries in an audit workflow.
• Introduction to Machine Learning for SEO course, content transformation module.