I recently spoke to Ed Schrager, who has spent the last few years building Summari, a company that creates automatic summaries from content. In a fast talking 45 minutes with me firing questions at him, he kindly shared his top learnings from his years of looking into this. Here are some exceptionally practical helpful insights and best practices in creating summaries. 

A couple of things to set the scene ...

There are two main types of summaries: 

• Extraction, where a part of the text is taken and promoted at the top of the page (easier, but rarely gives an accurate picture of the whole).

• Abstraction, which conveys the meaning of the entire piece using different words and phrasing.  

The challenge of the latter, which is what we are focusing on today, is that you are dependent on large language models understanding the text and synthesizing an accurate summary of that text correctly and — as we have seen from generative AI — that may mean hallucinations. 

So lesson No. 1, which I repeat below, is that there should always be an editor. Ed points out to me that it’s likely a good strategy to check work before publishing, and this applies to AI as well. AI is a copilot. A fairy on your shoulder helping you. It shouldn’t be viewed as a replacement, more so as an enhancement to your workflow. It is not a full-fledged tool that can do everything without some oversight. At least not yet. 

The other thing to think about is that there is always going to be a slight tension with summaries because of the amount of information you want to convey vs. the desire to be as concise as possible. Bear this in mind as you embark on this journey to find the right tension between the two for your audience.

With these things is mind, here are some best practices: 

• Whatever you create must be accurate: You have to give people the most important information as concisely as possible. People can be lazy and just want the highlights.
• Entice people to go deeper: Think of a summary as a movie trailer. It can help readers decide if the full piece of content is worth their time. In many instances, time on page increases 20+% for people that read the summary vs. those that didn’t. 
• Always use bullets (like these!): Ed told me that if you put the same five points in a paragraph vs. bullets, it feels like reading a mini article and is overlooked and less impactful or useful. Bullets are easier to read for the human eye and therefore easier to digest. For standard news articles, use three to five bullets, approx 120 words, and think about the formatting so a human can comprehend what’s going on as fast as possible. Summary length should be between 8%-12% of the total length of the source document — 8% if it's a long piece, 12% if it's shorter.
• Make it clear that it is a summary: Ed praised CNBC’s use of summaries: clearly marked as ‘key points,’ below the headline, above the image. Other examples I have seen aren’t explicitly marked and in some cases look like a sub header rather than a summary.

• Start with context: It’s wise to give readers an introduction or context into what they’re reading otherwise they will feel lost. That should always be your first bullet unless the headline more than explains the article.
• Summarise everything unless it’s less than 250 words: There isn’t a downside to having the summary (apart from time spent producing it) as this seems to appeal to all age groups and demographics. 
• Consider using across mediums other than text: A podcast can be summarised with a few bullets of text. See note above referencing a movie trailer. An effective summary can encourage a reader to view a video and that can have an impact on pre-roll ad impressions.

If you’d like to subscribe to my bi-weekly newsletter, INMA members can do so here.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .