Henry Williams

Product leader.Writes code.

At Reuters I lead engagement, mobile apps, partnerships and AI in the consumer portfolio. I prototype the idea before I pitch it.

Explore my work

Ask me a question

Powered by AI with my full career history. Try a suggestion or type your own.

Move the
number.

Interviews. Experiments. Prototypes.
Evidence first, then the decision.

173K
Paying subscribers in 18 months, from zero
+20%
Recirculation from Read Next recommendations
10%
CTR on AI push alerts, against a 1–2% benchmark

Group Product Manager
Mobile App & AI

173Kpaying subscribers in 18 months
$10M+ARR consumer portfolio

Right now, at Reuters.

Group Product Manager for Mobile App & AI at Reuters, leading engagement, mobile apps, partnerships and AI within a consumer portfolio. Launched subscriptions from zero to in 18 months. I figured out that Reuters is the #1 most-cited news source on ChatGPT (60% citation share) and turned that into a product strategy. Built AI notification systems hitting , which is 5-10x the industry standard.

Nights and weekends, I run Cleartext Labs, where I build and operate data products end to end. This summer that meant a subway delay model that beats the MTA's own alerts by a , retrained every night by an under evaluation gates, and a property-intelligence product for NYC brokers built entirely from open data.

Selected work

Four calls, and what happened next.

Each one is the problem, what the evidence said, the decision and the result. The full write-ups cover the options I rejected, who did what, and what I'd do differently.

01 · Recommendations · Reuters

A recommendation engine I prototyped alone, proved on live pages, then shipped with the team as Reuters.com's Read Next platform.

+20%
recirculation
+22%
subscription conversion, single-article visitors
Read the case study

Problem

Most readers arrived from search or social, read one article and left. The end of every article was a dead end.

Insight

Tag-based related links matched topics, not intent. What a reader did in this session said more than what the article was tagged.

Decision

Prototype before pitching: weighted embeddings, session signals and recency, proved on live pages through a Chrome extension.

Result

Shipped as Read Next. Recirculation up 20%, and subscription conversion among single-article visitors up 22%.

02 · AI strategy · Reuters

From search to conversation

I found Reuters was the most-cited news source on ChatGPT, and turned that into a distribution strategy and a paywall that works wherever readers are.

60%
citation share on ChatGPT
3×
conversion vs. Google search referrals
Read the case study

Problem

Search referrals were collapsing across the industry. The question was where readers were going instead.

Insight

A scraping and analysis framework showed Reuters was the #1 cited news organisation on ChatGPT, and that ChatGPT-referred readers converted at 3× Google search.

Decision

Treat AI assistants as a distribution channel, not a threat. Build Subscriptions on Rails: a paywall that travels with the content.

Result

The finding reset Reuters' AI distribution strategy. Subscriptions on Rails converts readers across partner publishers, the Reuters ChatGPT app and Reuters MCP servers.

03 · Experimentation · Reuters

The feature I stopped

AI article summaries shipped and engagement stayed flat. I stopped them and moved the investment to recaps for returning readers.

+15%
session frequency from returning-user recaps
20+
A/B tests run with kill criteria set upfront
Read the case study

Problem

AI summaries shipped behind an experiment with an explicit stopping rule. Engagement stayed flat.

Insight

Summaries answered 'what is this article' for readers who were already on it. The unmet need was 'what did I miss' for readers coming back.

Decision

Honour the kill criterion. Stop summaries and redirect the same team and models to recaps for returning users.

Result

Recaps lifted session frequency 15%. The kill-criteria habit became standard across 20+ experiments.

04 · Monetisation · Reuters

Subscriptions from zero

Launched Reuters mobile and web subscriptions and grew them to 173K paying subscribers in 18 months, with 94% sustained monthly retention.

173K
paying subscribers in 18 months
7.5×
iOS paywall conversion, 0.2% → 1.5%
Read the case study

Problem

Reuters had a huge free audience and no consumer subscription. The paywall had to earn money without driving readers away.

Insight

On iOS, the drop-off wasn't the price. It was the form: account creation and card entry before a reader could pay.

Decision

Take friction out rather than add incentives: Apple Pay and one-tap sign-up, with every change behind an A/B test.

Result

173K paying subscribers in 18 months, 94% sustained monthly retention, $1M ARR in week four, and iOS paywall conversion from 0.2% to 1.5%.

What shipped.

Four and a half years of shipped work, 16,000 changes in all, distilled to the initiatives that mattered. Each one went to production with the team.

34 shipped

2026

  • Free trial offer in the appA trial on the choose-a-plan screen, with the paywall held back for newly registered readers.Subscriptions
  • Reuters in ChatGPTReuters news tools and visual widgets for ChatGPT, built on the Apps SDK and MCP.Partnerships
  • AI code review gateAn AI risk verdict on every pull request decides how many human approvals it needs, and whose.AI
  • Continuous playA Listen Next queue that keeps audio going across stories, with AI-narrated articles badged as such.Engagement
  • Recommendations tuned to the readerA recommendation engine prototype that weights role, topics, regions and formats from onboarding.Engagement
  • One-tap Google sign-in on the paywallNative Google sign-in so a reader can get from paywall to account without a form.Subscriptions
  • Reuters News MCP serverSemantic search, entity extraction and relationship discovery over Reuters stories, for AI agents.AI
  • Apple Pay at checkoutApple Pay as a payment method on web subscriptions, rolled out behind a flag by country and device.Subscriptions
  • User intelligence for the product team41,000 app-store reviews across eight news publishers plus survey data, queryable by PMs through Claude.AI

2025

  • Dynamic paywallPaywall decisions driven by visitor type and location instead of a fixed meter.Subscriptions
  • Key points in articlesAI key points in the article body, with a feedback link and interaction tracking.AI
  • Vertical videoFull-screen swipeable short video in the app.Mobile app
  • AI-assisted development for the teamClaude agents, planning templates and MCP set-up shared across the engineering repos.AI
  • Subscriber-only newslettersPremium newsletters for subscribers, added automatically on sign-up and promoted to everyone else.Subscriptions
  • News AssistantAn AI assistant in the app, launched behind feature flags with in-app feedback and A/B-tested labels.AI
  • Offline readingSubscribers can read saved pages and their Today feed without a connection.Mobile app
  • Google and Apple sign-inSocial sign-on across web and app, with a path to set a password later.Subscriptions

2024

  • Magic-link sign-inPassword-free sign-in by email link, tracked step by step.Subscriptions
  • Follow topics and authors: My NewsReaders follow topics and authors and get a personal feed of saved and followed stories.Engagement
  • The new Reuters appiOS and Android rebuilt on one React Native codebase: Today page, topics, search, sign-up and accessibility throughout.Mobile app
  • Podcasts and video in the appPodcast directory, player and transcripts; a video tab with a mini-player.Engagement
  • Markets in the appStocks, currencies and commodities, company pages and a watchlist that prompts sign-up.Engagement
  • Notification preferencesReaders choose which alerts and newsletters they get, from preference cards in the feed.Engagement

2023

  • Dark mode and text densityReading settings in the new app: dark mode and a denser layout option.Mobile app
  • Onboarding in the new appSign-up and choose-your-topics screens, on the same account system as the web.Mobile app
  • Richer Apple News storiesTables, standalone slideshows and custom fonts, so Reuters stories read as well on Apple News as on Reuters.com.Partnerships
  • Custom partner feedsConfigurable RSS feeds with custom footers for syndication partners.Partnerships
  • Paywall on Reuters.comThe subscription paywall integrated into the web reading experience.Subscriptions
  • Apple News in more languagesLanguage attributes, translation and localised dates, starting with French.Partnerships

2022

  • Reuters on Apple NewsNative Apple News formatting and feeds, including live coverage, linking readers back to Reuters.com.Partnerships
  • Subscription management in My AccountSubscribers update billing, change or reset passwords and reactivate suspended subscriptions themselves.Subscriptions
  • Subscriptions in the Reuters appsThe web registration and subscribe flow brought into the iOS and Android apps, with deep-link support on iOS.Mobile app
  • Free registration and the registration wallA free sign-up tier, with a registration barrier and a sign-up path kept separate from subscribing.Subscriptions
  • New sign-in, registration and paymentAccount, sign-in and checkout rebuilt as one single-page app.Subscriptions

How I work.

  1. 01

    Set the kill criterion before launch.

    Every experiment gets a hypothesis, a primary metric and a stopping rule before it ships. It turns stopping a feature into a formality instead of a fight.

    See it in The feature I stopped
  2. 02

    Prototype before you pitch.

    A working prototype on live pages beats a deck. I build the first version myself so the roadmap argument starts from evidence.

    See it in Read Next
  3. 03

    Measure the channel nobody is measuring.

    The best opportunities hide in data nobody has collected yet. Build the instrument, then decide.

    See it in From search to conversation
  4. 04

    Remove friction before adding incentives.

    Discounts buy sign-ups that churn. Taking a step out of checkout buys subscribers who stay.

    See it in Subscriptions from zero

From the people I work with

“Henry is the rare PM who shows up with a working prototype and the data to back it. Roadmap debates got shorter.”
Engineering lead, Reuters
“He wrote the stopping rule before we launched, then held to it when the room wanted to keep going. That built a lot of trust.”
Data science partner, Reuters
“Editorial had a real veto and still shipped faster. Henry designed the process so both could be true.”
Editorial standards editor, Reuters

Nights and weekends

And I build.

Under Cleartext Labs I take products from raw public data to paying users on my own: the pipelines, the models, the interface and the on-call. It keeps my product judgement honest about what things cost to build.

Cleartext Labs

askprospect.com

Property intelligence for NYC brokers

Every parcel in the city from open data: 1.17M parcels, 2.07M sales, 16M+ deed and mortgage records. A likely-to-sell ranking across 1.24M residential units, graded nightly against recorded sales, plus an expected sale price for 1.48M parcels.

Open askprospect.com

Cleartext Labs

subway.fyi

NYC subway analytics & delay prediction

Polls all 9 MTA feeds every 30 seconds into TimescaleDB. A delay model that flags disruptions a median 43 minutes before official alerts, retrained nightly by an autonomous agent under replay and holdout gates. Over 1M visits in two weeks.

Open subway.fyi

More from Cleartext Labs.

nycwaittimes.com

Airport wait times with production MLOps

Live security wait times for 20 US airports, 12M+ observations. Nightly forecast retraining behind five promotion gates with automatic rollback. The archive is published as an open dataset on Hugging Face.

millionairesrownyc.com

1.2M-parcel public-records explorer

NYC's pied-à-terre surcharge roll, searchable and mapped, built in three days. A 140-field fixed-width government file reverse-engineered and validated against the official roll for 100.000% of 923,033 shared parcels.

CEDAR

Co-parenting email documentation

Syncs with Gmail, runs NLP over tens of thousands of emails, maps communications against 60+ custody provisions, generates court-ready PDF reports with SHA-256 verification.

Transcord

Call recording & transcription

AI-powered dual-channel capture with real-time transcription and searchable archives. Built for journalists and lawyers. Bootstrapped and running for 6+ years.

gist.

Natural-language Gmail search

Cmd+K overlay that finds emails by what they mean using local LLM indexing and hybrid vector/full-text retrieval. Chrome extension with MCP server for Claude.

userken.

Synthetic focus groups from app reviews

App store reviews turned into a vector-searchable knowledge base with synthetic personas, theme breakdowns, and a 2D semantic map. MCP server for Claude integration.

Wall Street Journal

Articles

Companies Turn to Internal Crowdsourcing to Pick Best New Ideas

By using a system modeled after websites such as Kickstarter, companies allow employees to have a say in which projects deserve to be funded.

NASA Administrator Jim Bridenstine on Life off Earth

The former Oklahoma congressman wants to make space accessible to everyone, and (possibly) discover life on other planets.

Forget Passwords. It's Time for Passphrases.

Two researchers say they have come up with a system that makes passphrases more secure and practical.

More articles 6

Artificial Intelligence May Make Traffic Congestion a Thing of the Past

Startups and car companies see AI as the answer to a lot of traffic troubles.

What's an Online Ad Worth? Blockchain Might Help With That

Three startups propose blockchain-based solutions to rampant fraud in digital advertising.

Blockchain May Offer a Resume You Can Trust

Colleges and tech companies are using the digital ledger to develop easily verifiable diplomas and employment records.

Inside Verizon Wireless's Hurricane Playbook

To maintain phone service, the chief network officer deploys drones, boats and a lot of fuel.

'Catch Me if You Can' Scam Artist Has a Warning for Today's Consumers

Frank Abagnale Jr., whose exploits inspired the Leonardo DiCaprio movie, says digital technology and social media only make things easier for swindlers now.

Living Wills: How Each Bank Performed

A look at how each bank was judged by regulators.

On camera

Technology, people, and the stories behind them. From my time at the Wall Street Journal.

More videos 6

GitHub projects

41 open-source repositories across MCP servers, web apps, desktop tools, browser extensions, and more.

Showing 6 of 41 repositories

The short version

Finance. Journalism.
Product. Code.

I started in finance, on the trading floors at Lehman Brothers (yes, that one), Barclays, and Morgan Stanley, building real-time systems where milliseconds actually mattered.

Then I switched to journalism. Got a Google Scholarship for a Master's at the University of Hong Kong, worked at the Wall Street Journal and Dow Jones, and spent years figuring out how to make newsrooms work like product organizations.

Now at Reuters, I lead consumer product for mobile and AI: the apps, the web, the things millions of people actually touch. I got there by shipping fast and measuring what worked: 173K paying subscribers in 18 months with 94% retention, a podcast on track for $9M ARR, AI features that outperformed everything before them.

I also build my own things. I founded Cleartext Labs because the data people need is usually already public, just not in a form anyone can search, rank, or predict from. Subway feeds, property records, airport queues, custody email. I write the pipelines, train the models, and run the infrastructure myself, and every product is reachable by AI agents through MCP servers I wrote.

Education

BSc Computer Science, Exeter

MJ Journalism, HKU

Citizenship

US & UK dual

Stack

Python, scikit-learn, SQL

Postgres, TimescaleDB, DuckDB

TypeScript, Next.js, MapLibre

Claude Code, MCP servers

Interests

AI product strategy

Production ML on public data

Real-time systems