DART: The four metrics that actually measure AI agents

Ship faster, measure better: experimentation in the age of AI

Summary How do you know if the thing you just shipped actually worked? On this episode of The Experimentation Edge, host Ashley Stirrup, CMO of GrowthBook, sits down with Kevin Yang, Executive Director and Head of Experimentation at JPMorgan Chase, who has spent six years building experimentation across Chase's digital platforms. Kevin shares how his team turned experimentation into more than a billion dollars of estimated value, why the losing experiments matter more than the winners, and the simple chart exercise he uses to prove that a million-dollar change is invisible without a control group. He and Ashley also dig into measuring engagement without chasing vanity metrics, planning for failure to defeat confirmation bias, and why AI is pushing experimentation into a golden era. It's a practical look for product managers, data scientists, and engineers at how a bank operating at massive scale makes better decisions. Chapters 00:00 Welcome to the experimentation edge 01:45 Kevin's role leading experimentation at chase 04:15 Why chase invested in experimentation 06:45 A billion dollars and the value of losers 12:45 Plan for failure to beat confirmation bias 14:30 The million dollar change you can't see 18:45 Sharing learnings and experimentation wrapped 20:45 Engagement without vanity metrics 22:00 Experimentation's golden era with AI 23:30 Why AI needs more experimentation, not less Takeaways * Chase estimates over a billion dollars of value from experimentation, and most of the lasting learning comes from the losing tests, not the winners. * A control group is non-negotiable: at scale, a change worth millions is invisible under noise and seasonality, and no one can spot it by eye. * Treat engagement carefully. For a bank, more time in the app isn't a win; trust, fast task completion, and healthy repeat engagement are. * Plan for failure before you run a test. A pre-built playbook for a loss prevents confirmation bias and keeps teams from gaming the metrics. * AI is ushering in a golden era for experimentation, because shipping faster only compounds mistakes unless you measure what you ship. Connect with the Guest LinkedIn: https://www.linkedin.com/in/kevintyang [https://www.linkedin.com/in/kevintyang] Website: https://www.jpmorganchase.com [https://www.jpmorganchase.com] Sponsor Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts. Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse. With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction. See a demo at https://www.growthbook.io/ [https://www.growthbook.io/]

False negatives are killing your best product ideas

Summary  How do you make a high-stakes product decision when the safe choice is to never test it at all? In this episode of The Experimentation Edge, host Ashley Stirrup talks with Arun Bodapati, director of data science at Twitch, about the discipline behind trustworthy experimentation. Drawing on his experience at Schwab, Uber, and Twitch, Arun explains why false negatives are the most dangerous result a team can produce, what hygiene to nail before you push play, and how Twitch used geo-fenced experiments and causal inference to finally settle a pricing question it had avoided for years. It's a practical conversation for product managers, engineers, data scientists, and growth leaders who want experiments that hold up  and earn executive trust.   Chapters 00:00 Welcome and introduction 01:15 Arun's background and marketing experimentation at Schwab 04:15 Uber's mature, experiment-driven culture 06:30 Coming to Twitch: from Python notebooks to a shared standard 08:30 The pricing problem Twitch had long avoided 10:30 Geo-fenced experiments, matched markets, and elasticity 13:15 The gifted-subs surprise and testing promotions 16:15 The discipline that matters before you push play 18:15 Why false negatives are worse than false positives 20:05 Enrollment triggers and broad explore experiments 22:45 AI, the Kiro tool, and what's next for experimentation Takeaways  * False negatives are more dangerous than false positives — they get institutionalized as "we tried that, it didn't work" and quietly kill good ideas for years. * The most valuable experiment work happens before you push play: clear enrollment logic, a plain-English hypothesis, and no optimizing ahead of the test. * If an intervention sounds weak when you write it out in plain English, don't run the experiment — you're just wasting time. * Run a broad explore experiment first; small, over-narrowed populations lack power and raise the odds of a false negative. Find the responsive segment with heterogeneous treatment effects afterward. * Twitch used geo-fenced experiments with matched markets and causal inference to measure true price elasticity, turning a feared pricing decision into a measured, accretive one. Connect with the Guest  LinkedIn: https://www.linkedin.com/in/abodapati/ [https://www.linkedin.com/in/abodapati/] Website: https://www.twitch.tv [https://www.twitch.tv] Sponsor Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts. Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse. With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction. See a demo at https://www.growthbook.io/ [https://www.growthbook.io/]

Squarespace killed its blank template and built something better

Summary What do you do when your big launch increases engagement and tanks conversion? On this episode of The Experimentation Edge, host Ashley Stirrup talks with Lina Blackman, Director of Product Analytics at Squarespace, about the blank template launch that flopped — and how its learnings became Blueprint, Squarespace's AI-guided website builder. Lina explains how her embedded analyst team runs 150–200 experiments a year for 3 million customers, the two questions she asks every time a test loses, why teams only need one or two big wins a quarter, how Squarespace calibrates statistical certainty to business stakes, and where AI belongs (and doesn't) in the A/B testing workflow. For product managers, data scientists, and experimentation leaders who want to extract more learning from every test. Chapters 00:00 Introduction: Lina Blackman, Director of Product Analytics at Squarespace 01:45 Squarespace's business and 3 million website customers 02:30 Decentralized analysts, centralized experimentation program 04:15 150–200 experiments a year: onboarding, mobile, checkout, pricing 04:55 The blank template disaster that became Blueprint AI 07:45 Two questions for every losing test 09:30 Moving ship-first teams up the experimentation maturity curve 12:30 A/B test logs and insights rituals 13:30 North Star metrics and the KPI tree 16:35 AI in the A/B testing workflow — and what stays manual. Takeaways * Stated preference lies: users asked for a blank canvas, but behavior demanded guided design — and only the experiment could referee. * Close every losing test with two questions: did it work for a granular segment, and is the idea worth further investment? * One or two big wins a quarter is a healthy hit rate when you run 150–200 experiments a year. * Calibrate certainty to stakes — tight bounds on revenue and pricing tests, wider bounds on engagement tests so teams don't spin on noise. * Hand AI the mundane parts of the workflow (tracking, assignment setup), but if AI runs the brief and the analysis, ask why you're running the test at all. Connect with the Guest LinkedIn: https://www.linkedin.com/in/linanguyen [https://www.linkedin.com/in/linanguyen] Website: https://www.squarespace.com [https://www.squarespace.com] Sponsor Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts. Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse. With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction. See a demo at https://www.growthbook.io/ [https://www.growthbook.io/]

The "View All" page that made more money by showing less

Summary Craig Kistler, VP of Experience Design, Personalization, and Experimentation at Signet Jewelers (the parent company of Kay, Jared, Zales, Peoples, and Banter), joins host Ashley Stirrup on The Experimentation Edge to unpack how a hybrid online-and-in-store jewelry retailer runs experimentation at scale. Craig shares the counterintuitive "view all" experiment where his team blocked the product grid, added friction on purpose, and grew revenue; why he optimizes for revenue per visitor instead of conversion rate; and how Signet deliberately traded a high-volume testing program for fewer, higher-value experiments. A practical listen for product managers, designers, and experimentation leaders building programs that compound. Chapters 00:45 What Signet Jewelers actually is (Kay, Jared, Zales, and more) 01:45 From art school to UX to experimentation: Craig's background 04:45 How experimentation is organized: a centralized model across brands 06:45 From 40–50 tests a quarter to 15–25 value-driven experiments 08:45 The "view all" experiment: adding friction to grow revenue 12:45 One product page, many stakeholders: financing, warranties, chat 15:45 Extracting learnings when an experiment loses 18:45 Why revenue per visitor beats conversion as the north star 20:15 Intent-based personalization and "engagement season is every day" 24:45 Bringing the whole org along by tying insights to dollars. Takeaways * Friction can increase revenue. Blocking the "view all" grid and forcing a style choice sent shoppers deeper and lifted conversion and revenue, because the extra click added value. * The three-click rule is conditional. Clicks only hurt when they're empty; a click that narrows thousands of options to dozens is a feature, not a cost. * Revenue per visitor is the honest north star. Conversion rate can be gamed to 100% by making everything free or cutting bounce-heavy traffic; revenue per visitor can't. * Fewer, bigger experiments beat high volume. Signet went from 40–50 tests a quarter to 15–25 because complex, value-driven tests produce reusable insights that small tweaks don't. * Tie every result to dollars. Translating experiment outcomes into revenue is how Craig keeps financing, warranty, and chat stakeholders aligned and gets executives to act. Connect with the Guest LinkedIn: https://www.linkedin.com/in/craigkistler/ [https://www.linkedin.com/in/craigkistler/] Website: https://www.signetjewelers.com [https://www.signetjewelers.com] Sponsor Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts. Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse. With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction. See a demo at https://www.growthbook.io/ [https://www.growthbook.io/]

DART: The four metrics that actually measure AI agents

Summary RingCentral's Director of Product Management for AI Products, Mayank Agarwal, joins host Ashley Stirrup to dismantle the metrics most teams use to judge AI agents. Drawing on his background founding an AI-first quantitative trading firm and scaling Groupon's bookable marketplace, Mayank explains why accuracy and thumbs-up/down feedback both mislead, and introduces DART — a four-metric behavioral framework (decay, acceptance, relevance, task completion) ported from how he measured trading strategies. He also breaks down a Groupon flash-discount experiment that backfired and the scarcity pivot that fixed it. Essential listening for product managers, engineers, and data scientists building or measuring AI features. Chapters 00:00 Welcome and Mayank's path from quant trading to RingCentral AI 02:45 Why experimentation has to be owned cross-functionally 04:55 Small experiments that compounded to a 12% lift at Groupon 06:45 Why accuracy and thumbs-up/down fail for AI agents 08:15 The DART framework, metric by metric 12:45 Applying DART to AI-generated smart notes 14:55 The Groupon flash-sale that dropped conversion 16:45 Swapping price urgency for scarcity and social proof 19:45 North Star metrics, guardrails, and Goodhart's law 26:45 The future: experimenting on — and for — AI agents Takeaways * Accuracy is a comfortable lie. It grades a narrow test set and can stay high while the agent fails real users. * Thumbs-up/down feedback is sparse and skewed. Unhappy users rarely rate — they just quietly stop using the product. * DART measures behavior, not opinions. Four signals read off logs and transcripts: decay, acceptance, relevance, and task completion. * Acceptance rate is the trust metric. The share of output users keep without editing is the strongest available proxy for trust. * A losing experiment is paid-for information. Groupon's flash-sale flop revealed the lever was wrong, not the goal — scarcity beat price-based urgency. Connect with the Guest LinkedIn: https://www.linkedin.com/in/mayank-agarwal-6223b04a/ [https://www.linkedin.com/in/mayank-agarwal-6223b04a/] Website: https://www.ringcentral.com [https://www.ringcentral.com] Sponsor Growthbook helps you ship features with confidence by bringing experimentation and feature flagging into one open-source platform. No more guessing whether that new checkout flow actually moved the needle, waiting weeks for data team bandwidth, or flying blind on rollouts. Growthbook gives you a single place to run A/B tests, manage feature flags, and analyze results against your existing data warehouse. With powerful stats built in, it takes the complexity out of experimentation, helps you catch regressions before they hit every user, and makes it easy to test ideas that keep your product improving and your metrics moving in the right direction. See a demo at https://www.growthbook.io/ [https://www.growthbook.io/]