Input
Approved fictional facts: two-bedroom condo, covered parking, balcony.
Google I/O 2026 · an evidence-review lesson
How do I tell whether a new AI workflow actually works?
Compare the result with its source, record what you observed and separate an announced capability from something tested in your account. This May 2026 class is useful because the demos reveal missing inputs, invented details and scores that deserve scrutiny.
Ryan Wanner · on-demand class recording published May 2026
All classes and lessonsWatch an imperfect demo become a useful lesson
Historical recording: announcements, forecasts, fictional demo characters and guessed product costs are preserved as spoken. Check current documentation and your account. The original guide is source history; the revised workbook below is the current practice companion.
Every class on this page is free to read and watch. No account needed.
Current discovery guidance: ordinary SEO and useful accessible content still matter. Google’s FAQ rich results have retired; neither FAQ markup nor an AI-only text file guarantees discovery. Read Google’s AI features guidance and its documentation updates.
Three checks. One observable result.
Choose an output you can inspect. You can complete the property exercise with a textual storyboard; no new paid render is needed.
Result to inspect
Intentionally flawed storyboard: three bedrooms and a pool.
The check
Reject the extra bedroom and pool. Compare every other depicted feature with the source.
This is an inspection exercise. No page is crawled, no account is connected and no result is fabricated by this tool.
Read the working request
Help me inspect one AI workflow. First ask which result I can access: a property asset with source photos/facts, a public page with a factual business record, or a feature with current official documentation and actual account access. Return a table: observed evidence | AI-added claim or change | source/date | keep/change/remove | reason | next check. Never invent access, a crawl result, property features or a visibility score. For a product feature, distinguish announced, available in my account and tested. Start with one small example and stop for my review. Use a fictional storyboard if tools or images are unavailable. No external edits, publication or messages are authorized by this exercise.
Make the evidence decide
Choose one check
Inspect a property asset, audit a real page or evaluate a new feature. Choose a result you can compare with actual evidence.
Collect the source
Use permissioned images and a fact sheet, a real accessible page, or current official product documentation plus actual account access.
Separate observation from inference
Write what the source shows and what the AI added. Mark unverified statements explicitly; a polished output does not make them true.
Make a specific correction
Reject invented property features. Replace an unexplained visibility score with a page finding and evidence. Use a manual practice case when a tool is unavailable.
Repeat one bounded test
Use the same criteria and inspect whether the corrected output works. A product announcement is not proof of account availability or a completed workflow.
Record the adoption decision
Choose adopt, pilot, watch or reject with the evidence and a review date. Keep a draft separate from any approved publication or send.
Notice where the demo fails its own test
Uploaded images are the only source of truth.
as a whole, the video is not great.
a four is honestly a little bit concerning.
Keep these distinctions visible
Treating an announcement as available access
Check the actual account and one bounded test.
Accepting an unexplained score
Ask for observable findings, source evidence and unknowns instead.
Publishing an attractive fabrication
Compare property features and claims with the approved source before use.
Use the revised evidence workbook
Create a free student portal account to download this class's files.
Download the revised workflow evidence workbookRead all three work areas and checks
# Workflow evidence workbook Revised companion to the Google I/O recorded class · November 8, 2026. This is not the unchanged historical guide. ## A. Inspect a property asset Fictional approved facts: two-bedroom condo, covered parking, balcony. Deliberately flawed storyboard: three bedrooms and a pool. Reject the extra bedroom and pool. For your own asset, compare every feature and alteration with permissioned source photos and verified facts. Preserve layout and condition; disclosure requirements depend on intended use. ## B. Audit an actual page Exact accessible URL and checked date: Observed identity, location, services, evidence and next action: | Finding | Exact page evidence | Missing evidence | Proposed repair | Result after checking | |---|---|---|---|---| | | | | | | Do not substitute an unexplained AI visibility score for observed findings. ## C. Evaluate a feature Official announcement/date: Available in my actual account: Bounded test and input: Observed output: Accuracy/access limit: Adopt / pilot / watch / reject, with reason: Owner and next review date: ## Extend only after one useful test Content: draft from a verified brief and inspect claims. Staging: compare with the original. Listing copy: use property facts and stated needs. Buyer comparison: use buyer-supplied criteria and weights, expose unknowns and leave the decision with the buyer. Follow-up: retain actual contact preferences and keep drafts unsent until reviewed. Tool access and prices must be checked at use time.
The revised companion replaces unexplained visibility scores with findings, keeps buyer criteria explicit and separates a draft from an external action.
Questions before adopting a workflow
Editorial learner questions.
Are all the recorded features available to me now?
The video preserves a May 2026 discussion. Check current official documentation and your own account before choosing a practice path.
What if a generated video changes the property?
Reject or revise the misleading part. Visual polish does not establish factual accuracy.
Can AI give my site a meaningful visibility score?
A number without a defined method and observed evidence is not reliable measurement. Ask for specific findings and verify them.
Does FAQ schema guarantee AI discovery?
No. Questions and answers help readers, but Google's FAQ rich-result feature is retired and AI inclusion is not guaranteed.
Do I need to run every original workflow?
No. Choose one relevant job, supply its sources and verify the result.
Is this the original Google I/O guide?
The recording is historical. The workbook on this page is a revised current companion; the original guide is retained as source history.
Selected passages from the demonstrations
Selected caption excerpts from the historical recording. These are deliberately non-contiguous passages, not the full transcript. Current practice and corrections are written separately above; the original recording is unchanged.
17:56 · Check the source constraint
Uploaded images are the only source of truth.
18:48 · Inspect the actual result
There would be a couple usable things from this, but as a whole, the video is not great.
23:11 · Question the generated score
So you can see here there's no verified testimonials. There's no transaction history.
a four is honestly a little bit concerning.
24:20 · Work from source materials
Last little tool that I want to go over is something that's probably my favorite thing and that is notebook LM.
So like you can see here like there's the ability to load up your sources and then what you can do is you can actually turn things into infographics.
Strengthen the next test
Six Mental Models for Judging AI Work
Use source facts and explicit criteria to judge the result.
Build a Context Card That Stays Useful
Keep relevant facts current and resolve conflicting notes.
From Asking Questions to Delegating AI Work
Define a bounded job before asking AI to produce it.
Past the feature announcement
Build the workflow you can actually verify.
The evidence desk separates a feature announcement from an observed result. Review AI Acceleration: Build if you want help turning a specific business task into a working system.