Shutting Down Wordsworth AI

What we built, why it worked, and why we shut it down anyway.

1. What Wordsworth AI Was

Wordsworth AI was a closed-loop landing page generation and optimization platform. We built the self-optimizing loop: every section on a page is measured, the underperformers are identified, changes are made, and the cycle repeats.

We were AI-powered and served enterprise brands ($50M–$500M+) across e-commerce, financial services, and edtech. We worked with 10 paying clients with active pilots across categories. We reached six-figure revenue.

We wound down the e-commerce business in December 2025, pivoted to an automated website generation product for small businesses, spent a few months testing through scalable channels - and decided to sunset the company at the end of March 2026.

2. Our Core Thesis

Two ideas drove everything we built:

The section is the atomic unit

A landing page is composed of sections - hero, testimonials, product grid, FAQ, social proof. Across 200+ brands, sections overlap heavily in structure. What changes is the content. So we built every system around this unit: build sections with AI and a human-in-the-loop to ensure production readiness, deploy them, measure engagement at the section level, feed data back, and improve. Same unit throughout.

The closed loop

The industry is fragmented: a page builder to create, a heatmap tool to measure, a CRO platform to test, and Slack to coordinate the handoffs between them. 5–9 tools, none of which talk to each other. Result: 11–16 weeks from hypothesis to deployed change. Only 1 in 7 A/B tests produces a winner.

We built a system where the unit you build in is the same unit you measure and optimize. That’s the closed loop - and it’s why the results compound.

3. What We Achieved

Revenue impact across clients

We worked with mid-market and enterprise clients across consumer brands, insurance, and education.

• 100+ websites shipped, 500+ iterations

• 1.2M+ live sessions captured

• > $10M in ad spend optimized

• > $1.5M in estimated annual revenue impact

• +16% to +190% CVR improvements

• 30 - 75% reductions in customer acquisition costs for some clients

The pattern repeated across every engagement: section-level data surfaced the insight; the insight pointed to the change; the change moved conversion.

The canonical example

A health brand’s FAQ section was getting disproportionately high engagement - users were scrolling past product features to reach it. Page-level analytics would have shown “people scroll to the bottom.” Section-level data showed which section pulled them there and why. Fix: move top FAQs into the hero. Conversion improved - and the fix took hours, not the usual 11–16 weeks.

How it compounds

The system studied competitor pages, analyzed customer reviews to understand what buyers actually care about, and cross-pollinated learnings across brands. A structural insight from a skincare brand - simplify the PDP into one viewport - proved equally effective for a supplements brand weeks later. Every new engagement started with accumulated intelligence from every previous experiment - not from zero. That’s why results compounded over time: each test made the next one smarter.

4. Design Principles

Determinism over generation. Rather than letting AI generate arbitrary code, constrain AI to filling known fields within a human-approved, typed schema. AI populates content within fixed boundaries - it doesn’t invent structure. Enterprise brands required pixel-perfect design fidelity, complex e-commerce functionality, and performance under load. That’s what made it production-ready and enterprise-grade.

Human-in-the-loop at structural boundaries. Agents handle high-volume constrained work. Humans review the structure - a short list of element types - not 4,000 lines of generated code. That’s where judgment matters most.

Measure at the unit you control. Section-level analytics - not page-level - enables the build-measure-improve loop. No existing tool (Hotjar, Clarity, GA, Contentsquare) captures at this level. The gap is structural.

5. Why We Shut Down

We optimized for production-quality output with human oversight. The market rewarded fast, self-serve generation. Tools like Lovable, Bolt, and v0 scaled by being good enough for most and accessible to everyone. We built for the 5% who needed enterprise-grade; the 95% went elsewhere.

The “services as software” trap. Everyone talks about AI-enabled services as the new model - deliver outcomes, not features. We lived it. Two things have to be true for it to work: your ACV has to be high enough to absorb the cost of delivery, and the outcome you’re selling has to be verifiable quickly. We had neither. Our contract sizes were $25K–$40K, and proving conversion lift required significant traffic over weeks - long feedback loops before the client could see the value. When you combine low ACV with slow verifiability, every engagement becomes a long, expensive proof-of-concept.

Value creation without value capture. The results speak for themselves - six and seven-figure annual revenue impact for clients. But when we tried value-oriented pricing early on, one client - more SMB than enterprise - simply copied our landing page instead of paying more. Clients didn’t churn because the product didn’t work. They just weren’t willing to pay proportional to the value they received.

Scope creep is structural when you sell outcomes. When you commit to outcomes, clients want you to do everything. One engagement started as landing pages and within weeks expanded to ad creative alignment, email campaigns, and retail marketing. You start focused on the core and end up spread across the entire surface area.

Scalable demand generation was hard. E-commerce is a crowded market with many incumbents across page building, analytics, and CRO. Breaking through required a neighborhood-level go-to-market - the kind where trust and word-of-mouth drive adoption. We learned the hard way why neighborhood matters so much. We tried LinkedIn outreach for months - zero conversions. Every paying client came through a warm intro.

Landing page optimization looks easy until you’ve tried it. Everyone assumes it’s straightforward - build a page, run a test, get a winner. In reality, you need significant traffic to prove value, which means pilot timelines are long. Getting pilots was hard because the perceived simplicity made it difficult to justify the investment upfront.

The pivot didn’t close either. We wound down the services business and pivoted to a self-serve automated website generation product for SMBs. Willingness to pay was $10–$20/month, but cost per acquisition was >$50. We spent a few months testing through scalable channels. Lead capture economics didn’t work, and we didn’t see a credible path to fixing them with the runway we had.

6. What We Left Behind

All code is open-sourced under the MIT license.

github.com/wordsworthai

We’re proud of our small but mighty team - they shipped under genuine uncertainty, and our early customers got real value out of what we built. The market shifted a lot since we started, and so did our understanding of the problem. We couldn’t scale the way we imagined, but we’re walking away with many learnings.

If you’re building in this space, working on Applied AI, or solving similar problems - I’d like to hear from you.