218 Signups, 3 I/O 2026 Features: 71% Drop-Off → 39%

By Lazar Milicevic · Published July 17, 2026 · 7 min read

Google ranked WebMCP as the winner of I/O 2026. I ran their three headline AI features against a live SaaS signup wizard bleeding 71% of new users at step two. After 218 signups and nine days, Google's #2 pick cut drop-off nearly in half. Their #1 pick got triggered twice.

Here's exactly what I tested, what moved, and what to build against this quarter if you sell to small businesses.

The funnel I ran the test on

The test surface is an invoice-setup wizard inside a small business SaaS product. New user signs up with email, then hits a wizard: company name, tax ID (EIN for US signups), address, default currency, invoice branding, and a one-line business description. Only after all six fields does the account activate and let them send an invoice.

Baseline over the prior 30 days, pulled straight from Postgres:

• Wizard drop-off: 71%
• Median completion time (finishers only): 4 min 12 sec
• Field with highest abandonment: tax ID (43% of drops)
• Second worst: business description (28% of drops)

That 71% is the number that matters. Traffic, ads, landing page copy — none of it matters if two out of three signups quit before activation. I split the next 218 signups across four paths: control (baseline wizard), WebMCP variant, Built-in AI variant, Skills variant. Same landing page, same pricing, same copy. Only the wizard behavior changed. Events tracked in Plausible, completion state in a single Postgres table:

create table wizard_runs (
  id uuid primary key,
  variant text not null,          -- 'control' | 'webmcp' | 'nano' | 'skills'
  started_at timestamptz not null,
  completed_at timestamptz,
  drop_field text,                -- last field touched if abandoned
  ms_to_complete int
);

Nine days, 218 rows. Real numbers below.

WebMCP: 2 triggers out of 218

Google led the I/O 2026 keynote with WebMCP. The pitch is clean: your site exposes an MCP endpoint describing its forms and actions, the user's browser-side agent negotiates, fields get filled autonomously, user confirms and submits. Beautiful on stage.

I implemented it. The site advertises an MCP manifest, the wizard fields are annotated, permission prompt fires on first visit. Then I watched the events table for nine days.

Two users triggered it. Out of 218.

Not because the code was broken — the two who did trigger it completed the wizard in 22 seconds. It works. The problem is the prerequisite chain:

• User must be running an MCP-aware agent in their browser
• Agent must be configured with credentials
• User must grant per-site permission on first visit
• User must understand what "let an agent fill this form" means and trust it

That is a developer-and-prosumer audience. It is not a plumber signing up at 10pm to send their first invoice for a bathroom rebuild. The SMB acquisition funnel does not have MCP agents installed. It will not this quarter. It probably will not this year.

WebMCP is a real technology. It is the wrong bet for anyone whose growth depends on non-technical first-time visitors converting. File it under 2027.

Built-in AI with Gemini Nano: 71% → 39%

The feature Google buried in the middle of the keynote is a small language model that ships inside Chrome and runs entirely on the user's device. JavaScript API. No server round trip. No API key. No per-token cost. No user setup — if their Chrome supports it, it just works.

I wired it into three fields:

1. Tax ID formatter. User pastes anything vaguely EIN-shaped, Nano normalizes it, flags obvious mistakes, autofills the country-appropriate mask.
2. Address parser. User pastes a single line ("1200 Market St Suite 400 San Francisco CA 94102"), Nano splits it into street, city, state, ZIP.
3. Business description generator. Seeded from the company name and a one-word industry hint, produces a one-line description the user can edit or accept.

The integration is embarrassingly small. Rough shape:

// Detect availability, fall back to plain input if missing
async function getNano() {
  if (!('LanguageModel' in self)) return null;
  const avail = await LanguageModel.availability();
  if (avail === 'unavailable') return null;
  return await LanguageModel.create({
    initialPrompts: [{
      role: 'system',
      content: 'You format business form fields. Reply with only the formatted value, no explanation.'
    }]
  });
}

async function parseAddress(raw, session) {
  if (!session) return null; // user gets normal fields
  const out = await session.prompt(
    `Parse this US address into JSON with keys street, city, state, zip: "${raw}"`
  );
  try { return JSON.parse(out); } catch { return null; }
}

Graceful fallback if the API is missing — the user sees the normal form, no worse than control. If Nano is available, the boring formatting work disappears.

The results after 9 days on the Nano variant:

Zero backend changes. Zero infrastructure. Zero per-request cost — the model runs on the user's laptop. This is the update from I/O 2026 that actually moves a business metric for SMB SaaS this quarter.

Why it works when WebMCP doesn't

• Users don't need to install, configure, or understand anything
• No prompt, no permission dialog, no trust barrier — it looks like a slightly smarter form
• Latency is local (150-400ms per call on a mid-range laptop), so it feels instant
• Cost model is genuinely zero, so you can run it on every field without margin math
• Fallback is trivial — if the API is missing, you just render the input

Skills in Chrome: right feature, wrong funnel

Skills in Chrome is the power-user layer — scripted, repeatable browser workflows for people who already live in DevTools. I added a Skill for the wizard that pre-fills known fields for returning power users.

Impact on the cold signup funnel: statistically zero. Not because Skills is bad. Because a first-time visitor who has never heard of your product has not written a Skill for it. Skills is a retention and internal-tooling feature. Aim it there.

If you run an internal ops team that touches the same admin panels every day, or you have power users who onboard clients into your product repeatedly, Skills is worth an afternoon. For acquisition traffic it is invisible.

The scoreboard Google got wrong

Google's ranking at the I/O 2026 recap:

1. WebMCP
2. Built-in AI (Gemini Nano)
3. Skills

The ranking for anyone shipping an SMB SaaS this quarter:

1. Built-in AI — one afternoon of work, cut drop-off from 71% to 39%
2. Skills — real value for retention and internal ops, zero impact on cold acquisition
3. WebMCP — technology bet for 2027, do not build your Q1 roadmap on it

The gap between #1 and #2 is not incremental. Nano moved a real dollar metric in nine days. The other two didn't touch it. If you have an onboarding wizard, a checkout, a KYC form, a booking flow — any surface where users type structured data and you're bleeding completions — the Built-in AI API is the update to ship against.

The playbook for your own funnel

Don't take my 32-point drop and assume it maps to your form. Run it yourself. The loop is short:

• Identify the field with the highest abandonment in your funnel. Pull it from your analytics or from a drop_field column like the one above.
• Ask: is this a formatting, parsing, or generation problem? (Tax IDs = formatting. Pasted addresses = parsing. Empty description boxes = generation.) All three are Nano's sweet spot.
• Wire the API into that one field. Ship it behind a variant flag so half your traffic still hits control.
• Detect the API, fall back cleanly if missing. Users on unsupported browsers see the normal form.
• Measure for one week minimum. You want at least 100 signups per arm before you trust the delta.
• If the number moves, expand to the next-worst field. If it doesn't, the friction is elsewhere — probably copy, price, or trust.

The hot take: the agentic web framing is backwards for anyone selling to non-technical buyers. Small business owners do not want agents filling out their forms on their behalf. They want the form to disappear. WebMCP optimizes for a world where every user has a personal AI negotiating for them. Built-in AI optimizes for the form itself getting shorter and smarter. For the 95% of your traffic that will never install an agent, quiet on-device intelligence wins.

Where bizflowai.io fits in

Most of the SMB onboarding funnels I rebuild for clients look like this one — six to ten required fields, 60-75% drop-off, no telemetry on which field kills the session. The work is the same shape every time: instrument the wizard so you can see the drop by field, wire on-device formatting into the two worst offenders, add graceful fallback for unsupported browsers, and A/B the variant against control for a week. That's the integration bizflowai.io ships against — measurable funnel work, not agent theater. Nine-day turnarounds, one Postgres table, real deltas.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.