No PriorsNo Priors Ep. 49 | With Shopify VP of Core Product Glen Coates
CHAPTERS
- 0:00 – 2:21
Glen Coates’ founder journey and scope at Shopify (Handshake → Core Product)
Sarah introduces Glen Coates and his remit at Shopify, including the core developer platform and AI products. Glen recounts his path from computer science and video games to founding Handshake, its evolution into a B2B Shopify-like platform, and his move into broader core product leadership at Shopify after the acquisition.
- •Career arc: games → eco-friendly bag company → insight for Handshake
- •Handshake’s evolution into wholesale/B2B commerce platform
- •2019 acquisition by Shopify and transition into Shopify leadership
- •Current scope: online store, checkout, admin/back office, developer platform, app store, AI products
- 2:21 – 3:24
What a Shopify “Code Red” means—and why checkout hit emergency mode in 2020
Glen explains Shopify’s “code red” mechanism: a top-down company-wide priority reset driven by Toby (Tobi) Lütke. He describes the 2020 checkout code red during the pandemic, the scale of the response, and how the incident triggered deeper organizational reflection.
- •Definition: CEO-declared #1 priority; anyone asked must drop work to help
- •Checkout failures in multiple modes made it mission-critical
- •Large cross-company mobilization (hundreds of people) over ~a year
- •Postmortem thinking: code reds often signal systemic issues
- 3:24 – 4:04
From firefighting to structure: reorganizing to reduce fragmentation
After the checkout stabilization, Shopify examined how internal structure contributed to the problems. Glen describes the shift from many small, fractured business units to a few larger ones, and how leadership must adapt when teams and ownership areas become broader.
- •Root-cause focus: how problems arose, not just fixing symptoms
- •Reorg: ~12–15 small units consolidated into ~3–4 larger groups
- •Alignment with how the product actually works end-to-end
- •Leading larger orgs requires different coordination and decision-making
- 4:04 – 7:12
Why Shopify retains acquired founders: decisiveness, POV, and avoiding committee design
Elad asks why Shopify is a strong home for founders post-acquisition. Glen attributes much of this to Tobi’s explicit point of view and a cultural shift toward decisive direction-setting—traits that resonate with founders used to accountability and execution pressure.
- •High concentration of founders across leadership layers
- •Tobi’s strong opinions set a clear directional “vector” for teams
- •Founders bring POV + persistence to “manifest” products in the world
- •Goal: accelerate decisions and avoid design-by-committee stagnation
- 7:12 – 10:30
Integrating acquisitions and systems: preventing duplicate layers in the stack
Glen explains integration as a “stack-layer” problem: duplicates at the same layer create compounding costs. He uses commerce examples (checkout vs invoicing vs order editing) to show how small features can quietly evolve into parallel engines that slow shipping and confuse customers.
- •Integration strategy depends on where duplication exists in the stack
- •Example: order editing/invoicing can become a shadow ‘checkout/negotiation engine’
- •Duplication forces teams to implement features multiple times, ballooning timelines
- •Customers also pay the cost via inconsistent product behavior
- 10:30 – 12:12
Consolidation discipline: building intuition early and ‘eating the vegetables’ later
Collapsing duplicated systems is compared to rebuilding a house’s foundation while living in it—painful but often necessary. Glen highlights two skills: spotting duplication early and being willing to do the hard consolidation work before costs become permanent.
- •Early detection: develop intuition for accidental platform duplication
- •Hard choice: consolidate even when it’s disruptive
- •Avoid perpetual tax on every higher-layer team and workflow
- •Consistency matters: incomplete parity leaves long-tail users with broken experiences
- 12:12 – 15:50
Shopify’s AI adoption lens: simplify entrepreneurship beyond ‘switches’
The conversation turns to AI: how Shopify experiments while also needing centralized infrastructure. Glen frames AI as moving Shopify from “imperative mode” (users operating many switches) to a co-pilot model that helps merchants achieve outcomes without mastering every control.
- •Mission focus: reduce friction to start/run a business
- •Classic Shopify model: simple UI that reveals complexity as needed
- •AI opportunity: a ‘driver/co-pilot’ that knows the switches on the user’s behalf
- •Concrete wedge: helping merchants overcome hurdles like product copywriting
- 15:50 – 17:32
When to ship AI: human-in-the-loop guardrails and feedback-driven iteration
Sarah probes how Shopify ships non-deterministic AI features quickly despite evaluation challenges. Glen explains a key principle: AI can propose but not commit, keeping merchants in control while generating rich feedback signals for continuous improvement.
- •Principle: propose changes, don’t commit without explicit user action
- •Examples: generated product text; suggested customer replies requiring “send/enter”
- •Risk mitigation plus faster learning cycles
- •Feedback signals: accept, accept-with-edits, reject → improves models and UX
- 17:32 – 18:49
Risk orientation at Shopify: take risks for Shopify, not for merchants
Glen describes a ‘risk-hungry’ founder culture balanced by a strong duty not to harm merchant businesses. The team embraces approaches that put risk on Shopify’s side while limiting blast radius for merchants, with a longer-term path toward more automation as confidence rises.
- •Founder mindset: discomfort with excessive safety/slow pace
- •Ethic: avoid risks that could break merchants’ businesses
- •Guardrails as a way to innovate aggressively with limited user harm
- •Future possibility: higher auto-mode as acceptance rates approach near-perfect
- 18:49 – 21:40
Foundation work: modernizing Shopify’s product data model to unlock AI and discovery
Glen spotlights ‘boring’ but foundational work: evolving Shopify’s core product model to handle complex catalogs (many options/variants). He explains why this is hard—ecosystem-wide breaking changes like pagination—and how taxonomy + attributes enable better AI-driven merchandising and distribution.
- •Core bottleneck: legacy product model struggles with complex/large variant sets
- •Technical challenge: unpaginated → paginated APIs implies ecosystem breaking change
- •Standard taxonomy: categories + attributes become structured product metadata
- •AI enriches data: auto-detect category and attribute values from text/images
- •Downstream impact: better storefront search + improved ranking on Google/Meta/Amazon
- 21:40 – 22:31
AI image editing for merchant-grade product photography
Glen describes Shopify’s AI image capabilities aimed at making small merchants look as polished as major brands. By enabling tasks like background replacement and image enhancement, Shopify reduces the need for expensive photo shoots and raises the baseline quality of storefront assets.
- •AI image editing: background replacement and professionalizing product imagery
- •Cost/effort reduction vs traditional photo shoots
- •Goal: level the playing field for ‘my mom starting a business’ merchants
- •Quality improvements translate to better conversion and brand trust
- 22:31 – 24:05
Semantic storefront search: from literal keywords to intent, synonyms, and multimodal relevance
Shopify search historically behaved like literal keyword matching, leading to zero-result failures (e.g., “sweater” vs “sweatshirt,” “LBD”). Glen explains the shift toward semantic understanding, including intent queries (“something to wear to a wedding”) and compositional concepts (“Christmas-themed shoes”).
- •Problem: literal search fails on synonyms/abbreviations and intent queries
- •Examples: LBD, sweater vs sweatshirt, ‘wedding’/‘beach’ outfits
- •Goal: make small-store search feel closer to Google-grade intelligence
- •Semantic understanding unlocks better buyer experience and product discovery
- 24:05 – 26:04
How it’s built: embeddings, fine-tuning, and balancing text, images, and attributes
Elad asks about technical approaches; Glen notes the role of embeddings, fine-tuned models, and category-specific datasets. He highlights the emerging importance of multimodal models and the ongoing work to weight signals from descriptions, images, and structured taxonomy for best retrieval quality.
- •Use embeddings + fine-tuning for commerce categories (apparel, gifts, homewares, etc.)
- •Dataset assembly and evaluation as core differentiators
- •Multimodal experimentation: text vs image vs taxonomy attribute weighting
- •Merchant-side value: automatic backfilling of attributes for large catalogs
- 26:04 – 28:42
What’s still missing: the brutal climb from 75% to 95% reliability (Sidekick and beyond)
Glen describes the current frontier as reliability and consistency, not just demos. He notes that LLMs reach a compelling prototype quickly, but making an agent trustworthy at scale requires sustained iteration, better data, stronger evals, and benefiting from rapid external model progress.
- •Agents get ‘pretty good’ fast; last-mile quality is slow and hard
- •Hallucination framing: models always generate; usefulness depends on correctness control
- •Improving via user feedback conversations, training sets, and evaluations
- •Step-function gains also come from new frontier models arriving frequently
- 28:42 – 38:48
Industry watch: Rabbit’s GUI-based agents and the battle for LLM-native search
Glen points to Rabbit as an intriguing bet: training agents to operate GUIs directly rather than relying on APIs—potentially yielding ‘the web’ as an app store if it works. He also discusses the emerging triangle of Google, ChatGPT, and Perplexity in redefining what users want from search: answers, synthesis, and the tradeoff between compression and accuracy.
- •Rabbit strategy: treat GUI/pixels as the interface instead of APIs
- •Upside: instant access to the ‘web’ as an app ecosystem; downside: very hard to execute
- •Search landscape: Google (IR), ChatGPT (LLM UX), Perplexity (hybrid)
- •User needs: facts vs opinions; tolerance for lossiness vs desire for direct answers
