WORKFLOW · GROUNDED ANALYSIS
Google renamed NotebookLM to Gemini Notebook this week and gave every notebook a sandboxed cloud computer, so it can write and run real code grounded in whatever a person uploaded, spreadsheets, screenshots, a pasted note. This is a small working demo of that shape: three messy sources below actually get parsed in your browser, no server call, and the answer only comes out right when all three are read together.
Question
What was June's real revenue and order count, including the order that only exists in a WhatsApp note?
SOURCE A · orders_export.csv
SOURCE B · ate-baby-whatsapp-note.txt
SOURCE C · may-2026-summary.txt
Edit any source above, the answer recomputes live. Source C is kept for context and is intentionally not used in the June answer below.
Code run in the sandbox
rows = parse_csv(source_a) # 5 orders, June export refund = find_amount(source_b, "refund") # off-ledger adjustment reorder = find_amount(source_b, "reorder") # off-ledger order (WhatsApp) total = sum(r.amount for r in rows) + reorder - refund order_count = len(rows) + 1 # reorder is a real order, refund is not
Grounded answer
JUNE REVENUE
₱9,370
JUNE ORDER COUNT
6
The CSV alone says ₱7,720 across 5 orders. The real number needs the two lines that never made it into a spreadsheet.
WHY THIS MATTERS FOR CLIENT WORK
Real client numbers rarely live in one clean export. A revenue question usually needs the CSV plus the refund somebody mentioned in chat plus the order that came in over WhatsApp and never touched the store. A model that can read messy sources and actually compute over them, instead of eyeballing a summary, is the difference between a plausible-sounding answer and one that reconciles with what the client already knows happened.
Built 18 July 2026 · pattern sourced from Google's Gemini Notebook (formerly NotebookLM) launch