An AI Built a Full Conversion Funnel with 116 GA4 Events. It Got 8,367 Users and $0.
Xiaomi’s AI agent in our $100 AI Startup Race just filed a help request for its own Google Analytics data. What it shared is one of the most detailed conversion funnels I have ever seen from a solo product, let alone one built entirely by an AI coding agent running MiMo V2.5 Pro.
The numbers: 8,367 unique users, 16,970 pageviews, 85,762 tracked events across 116 custom event types, 5 simultaneous A/B tests, and exactly zero dollars in revenue.
This is the story of what happens when an AI optimizes the wrong thing perfectly.
What the agent built
APIpulse is an AI API pricing calculator. You pick models, estimate your usage, and it tells you what you will spend. The agent built it from scratch in 10 weeks with a $100 budget, deployed on Vercel, and powered every session through OpenRouter with MiMo V2.5 Flash (cheap) and Pro (premium).
By session 1,041, the agent had instrumented the entire product with custom GA4 events. Not just pageviews, but everything:
- Scroll depth tracking at multiple thresholds
- Every calculator interaction (model comparison, chatbot, RAG, embedding, finetuning, legal, marketing, translation, content generation, data extraction, support cost, SaaS cost, agent cost, budget planner)
- Pricing page views and Pro button clicks
- A/B test assignments for 5 simultaneous experiments (popup timing, gated recommendations, pricing variant, calculator engagement, design treatment)
- Exit popup shown, dismissed, and variant tracking
- Sticky CTA visibility and hover tracking
- Deal banner impressions
- Deprecation banners (for a Claude 4 shutdown hook the agent built as a traffic play)
- Flash sale and urgency popup tracking
- Trial starts and email subscribes
That is 116 distinct event types. For context, most SaaS startups with real engineering teams track 20 to 40 custom events. This AI built more instrumentation than most funded companies ship.
The funnel, honestly
Here is what those 116 events actually tell us when you read them as a conversion funnel:
| Stage | Event | Users | Conversion |
|---|---|---|---|
| Land on site | page_view | 8,362 | baseline |
| Scroll past fold | scroll_depth | 4,036 | 48% |
| See sticky CTA | sticky_cta_shown | 5,284 | 63% |
| View pricing page | pricing_view | 911 | 10.9% |
| Use a calculator | model_comparison_calculated | 359 | 4.3% |
| Click Pro button | pro_button_clicked | 8 | 0.1% |
| Start a trial | pro_trial_started | 5 | 0.06% |
| Pay | (none) | 0 | 0% |
| Subscribe to email | email_subscribe | 1 | 0.01% |
The drop-offs tell the story. The agent got traffic (8,367 users is respectable for a free tool), and people engaged (4,036 scrolled, 911 viewed pricing, 359 used a calculator). But only 8 people in the entire history of the product clicked the button to go Pro, and not a single one converted to a paying customer.
What the agent optimized vs what it should have optimized
Here is the core lesson. Starting around session 900, the agent entered a pure conversion optimization phase. It ran 5 A/B tests simultaneously. It added countdown timers to 533 pages. It built exit popups, urgency banners, flash sales, bonus packs, social proof blocks, and payback period calculators. Every session for the final 100+ sessions was about squeezing more clicks out of the funnel.
The problem: the funnel was not leaking at the CTA stage because of bad button color or missing urgency. It was leaking because nobody wants to pay for an AI API pricing calculator.
The tool is genuinely useful. 359 people used the comparison calculator. But “useful for free” and “worth paying for” are different things. An AI API cost estimate is something you check once, get your answer, and leave. There is no recurring value, no workflow integration, no pain that justifies a subscription. The product has a conversion problem that no amount of A/B testing can fix, because it is a product-market-fit problem.
The AI optimization paradox
This is something we have seen across multiple agents in the race but Xiaomi demonstrates it most clearly: AI agents are extremely good at optimizing metrics within a framework they understand, but they cannot step back and question whether the framework itself is wrong.
The agent’s logic made perfect sense at each step:
- Traffic exists (8,367 users). Good.
- People use the product (359 calculator uses). Good.
- Nobody pays (0 revenue). Problem.
- Solution: optimize the conversion funnel.
A human founder at step 3 might ask: “Is this something people would ever pay for?” The agent never asked that question. It jumped straight to funnel mechanics because that is a solvable engineering problem. Questioning the business model is a judgment problem that requires market intuition, and no amount of GA4 data gives you that.
The instrumentation itself is remarkable
Let me be clear: what the agent built is genuinely impressive from a technical perspective. 116 custom events, properly named, correctly scoped, with A/B test bucketing and variant tracking, all implemented by an autonomous agent with no human guidance on analytics strategy. If you asked a junior developer to instrument a product this thoroughly, it would take a week. The agent did it incrementally across hundreds of sessions.
The per-calculator tracking is especially thorough. It tracks not just “calculator used” but specific calculators: chatbot, RAG, embedding, translation, finetuning, legal, marketing, content generation, data extraction, support cost, SaaS cost, agent cost, MCP, coding assistant, migration, budget planner. That is 17 distinct calculator types, each with its own event. The agent was clearly building toward understanding which calculators drive the most engagement and optimizing from there.
The 5 simultaneous A/B tests (ab_popup_timing_assigned, ab_gated_recs_assigned, ab_pricing_variant_assigned, ab_test_assigned, dt_ab_assigned) show the agent understood experimental design. It did not just change things randomly, it set up proper test/control groups and tracked assignments separately from outcomes.
What this means for the race
With 8 days left, Xiaomi (APIpulse) has the most traffic and the most sophisticated product instrumentation in the race. It also has the clearest demonstration that traffic and conversion optimization cannot save a product nobody wants to pay for.
Compare with DeepSeek (Spyglass), which built 201 SEO pages and a content flywheel but also has $0 revenue. Or GLM (FounderMath), which got 12 real users and then stalled. The pattern is the same: AI agents can build products, drive traffic, and optimize funnels, but none of them have solved the “why would someone pay” question.
The race was always going to surface this. The question was never “can AI build a product?” It was “can AI build a business?” The answer, with 8 days left and $0 across all 7 agents, is looking like no, not without a human making the core product-market-fit judgment call.
Frequently asked questions
How many users did Xiaomi’s AI agent get? 8,367 unique users and 16,970 pageviews, tracked via GA4 with 116 custom events.
How many paying customers? Zero. Five people started a trial. None converted to paid.
What A/B tests did it run? Five simultaneous tests: popup timing, gated recommendations, pricing variant, calculator engagement, and a design treatment test.
What is the product? APIpulse, an AI API pricing calculator that compares costs across providers and models.
Why did conversion optimization fail? Because the core problem was product-market fit, not funnel friction. Nobody wants to pay a subscription for a pricing lookup tool they use once.
The bottom line
Xiaomi’s agent built one of the most instrumented products in the race, with world-class analytics that most real startups would envy. The tragedy is that all those events are measuring a funnel that ends in zero, not because the funnel is broken, but because the product at the top of it does not solve a problem people will pay to keep solving. For the full race, visit the live dashboard and daily digest.