Thirty minutes to know who was turning on Noogle

At 1:00 PM GMT on 22 October 2019, an open letter in France over the EU Copyright Directive started to move against Noogle. In 48 hours the system logged 278 posts in Europe with 13.6 million potential impressions, and nine in ten of them came from the press. The alert went further than volume. It named the key opinion formers in the conversation, the politicians and journalists among them, and showed how each one felt about Noogle.

That list was the point. Noogle's team could see who was driving the story and who was least favourable towards the company, and avoid exposure it did not need with them. Inside the system the report was filed as "EU vs. Noogle 9.7 billion fine". The fight it tracked was the EU Copyright Directive in France.

Reputation is an asset you notice only when it drops. The system handled billions of posts a year, so the question was never whether people were talking. The question was which voices could move policy, the press or customers, and how much time the team had before a post spread. The further a story travels, the less room there is to answer it.

The ideaWatch people, not keywords

The brief asked for a global monitoring system for key influencers on Twitter. It had to catch sudden spikes in a post's virality, a topic's momentum or its press coverage, and keep a sentiment score for every person of interest towards Noogle.

That turned the product around. Counting mentions tells you a topic is loud. We kept a list of hundreds of thousands of key opinion formers instead, and watched what they posted. A politician with a low score towards Noogle posting about a fine is a different event from the same words posted by an account nobody follows.

A mention count tells you a topic is loud. A person of interest tells you who to call.

The alarmAn alert people would trust at 3 AM

An alarm that fires too often gets ignored, and one that fires late is useless. I researched the human factors of alarm design, and three decisions came out of it. Every alert carries one of three levels, set by potential impressions. Every alert says in a line why it fired. And every alert sets itself against past crises, so an executive can see at a glance whether this is a the leaked memo or a Cambridge Analytica.

With the yardstick, executives did not need to learn what 13.6 million potential impressions meant. They could see it sat far below Cambridge Analytica's 300 million in its first day, and far above the 4.9 million of the Noogle Photos "gorillas" story.

The newsroomA human on every alert

Machine learning sorted each alert by how exposed Noogle could be, measured against hundreds of past scandals, from Cambridge Analytica to the Noogle Photos tagging failure. Then a person checked it. The team that did this was called the Newsroom: data scientists, IT engineers, sociologists and journalists, working with Noogle's Social Media Lab. I wrote the editorial policy they followed for every alert:

1. Is it new? Check it against the database of alerts already sent.
2. Is it in the right language? Only the languages each stakeholder chose.
3. Is it ours? Discard anything posted by the brand's own accounts.
4. Is it about the topic? Keyword rules remove most noise, but not all of it.
5. Is it true? Fact-check the post before anyone acts on it.
6. Is it real? Drop ads, clickbait, link-only posts and broken links.

Only then did the alert go out. The whole run, from the post to the inbox, fit inside 30 minutes.

The recordFour alerts that paid off

The buildConcept to MVP in three months

We shipped in four versions. The first won the client's buy-in. The second was the first automated prototype. The third gave the alerts their information architecture and moved them to Material Design. The fourth made them readable on a phone.

The system's layers, from sources to reports.

The same system later watched for Wayvo, VidTube and Allied Leaf Tobacco. My research on human factors in alarm design was fundamental to the product's success and to the contract's renewal.