Using Google Gemini
Use Gemini for a bounded task, understand grounding and connected tools, and distinguish a generated answer from verified evidence.
Before you start: Generative AI. Reading time is an estimate; allow extra time for practice.
What you’ll learn
- Use Gemini Apps to analyse a small supplied dataset with a clear question.
- Distinguish missing values, zero values and unsupported explanations.
- Check cited sources and uploaded material against the original evidence.
- Choose appropriate privacy and connected-app boundaries for a task.
On this page
Begin with a question the data can answer
A learning centre wants a simple attendance summary for four practice sessions. Tuesday has 12 recorded attendees, Wednesday 18, Thursday has no recorded value, and Friday has 0. You want to know the total recorded attendance and the mean across sessions with known counts. You also want a readable note explaining what is missing. This task exposes an important distinction: no recorded attendance value is different from a recorded zero.
Gemini Apps provide Google's conversational assistant experience, including ways to work with supported files and other inputs. Gemini also names models available through developer products. This guide uses the user-facing app, not the Gemini API. Copying a prompt into the app does not configure a model endpoint or integrate an assistant into your own website.
Start with this small fictional dataset before uploading a real workbook. It gives you an answer you can calculate independently. You are testing whether the assistant preserves the meaning of the data, not whether it can produce a polished chart. Features and limits can vary by account and interface; the text-only exercise remains usable without a connected app.
Choose what to share before you attach it
A real attendance workbook might contain names, email addresses, ages, support needs or notes about individual learners. None is needed to compute an aggregate count. Create a small extract containing session labels and authorised counts, remove irrelevant sheets and comments, and check whether the resulting group sizes could still identify someone in your context. Use the fictional data for initial practice.
Google's Gemini Apps Privacy Hub explains the data involved in chats, uploads and connected services. Review the settings applicable to your account, including activity controls, before using real organisational data. A work or school account may be governed by administrator settings and different terms. Turning off an activity feature should not be interpreted as a general guarantee of no processing or retention.
A Google account alone does not mean Gemini can read every Drive file or email. Connected-app access and applicable permissions determine what is available. For this task, a pasted table is sufficient. If you later connect a service, review the scope and identify the exact source you intend to use; a larger set of accessible files is not automatically better evidence.
Make blanks and labels unambiguous
The prompt below names the columns, supplies the records and explains the blank. Without that instruction, the assistant might count Thursday as zero or silently discard it without telling you. Neither is an acceptable interpretation when the centre needs to know that a record is incomplete. The wording also distinguishes total attendance from unique learners: one person could have attended several sessions.
Paste the complete prompt into a fresh Gemini chat. Ask for the arithmetic, not just an answer. The useful calculation is 12 + 18 + 0 = 30 across three recorded counts, followed by 30 ÷ 3 = 10. There are four session rows, but using four as the denominator would adopt an extra assumption about the missing Thursday value.
Analyse these fictional session records only.
session,attendees
Tuesday,12
Wednesday,18
Thursday,
Friday,0
The blank Thursday value is missing, not zero.
Zero on Friday is a known recorded count.
Report total recorded attendance, number of known counts,
mean among known counts, and missing session labels.
Show the arithmetic. Do not infer unique learners or why counts changed.
Finish with one sentence explaining what the missing value prevents
us from concluding. Do not access connected apps or change any file.This is a chat prompt containing CSV-style text, not executable code. The independent expected results are total 30, three known counts, mean 10 and Thursday missing.
Check what the file actually contributed
Gemini Apps can accept supported uploaded files and help analyse their contents. For a real spreadsheet, identify the relevant sheet, columns, unit and reporting period. Ask the assistant to show a few source rows and its treatment of blanks before summarising the whole file. Check those rows against the workbook, including whether a displayed zero comes from a formula or a recorded value.
The official file documentation warns that large inputs can lead to missed details or connections. Upload success does not establish complete coverage. Reduce the input to the relevant rows when possible, keep the column definitions, and ask the assistant to state any unreadable parts. A shorter extract is useful only if it retains the context needed to interpret the values.
For pictures or scanned tables, verify the recognised numbers and labels first. A mistaken 18 read as 13 changes the mean while still producing plausible arithmetic. When a chart is generated, compare its plotted values, labels and missing-value treatment with the original records. A blank should not become a zero-height bar without a clear explanation of that choice.
Separate source links from proof
Gemini may show sources or related links, including references to supplied files or connected material when those sources are available. Not every response includes links. A link can help you investigate, but you must open it and check whether it supports the specific statement. A general page about learning centres cannot explain why this centre's Wednesday attendance was higher.
If you ask an additional current question, such as the centre's published opening hours, keep it separate from the attendance calculation. Identify the official source, the relevant date and the exact claim it supports. Search information about the venue does not fill a missing internal attendance record. If evidence is unavailable, report the gap rather than asking for a more confident answer.
A generated explanation such as “Friday was quiet because of rain” is a hypothesis unless you supplied relevant evidence. Even a real weather record would not by itself establish why people stayed away. Describe what the data shows first, then state what additional information would be needed to test an explanation.
Use independent checks and small revisions
Verify the total and mean with a calculator or spreadsheet you control. Count the known rows explicitly, including Friday's zero. Then replace Thursday's blank with 6 and predict the new result before asking Gemini again: the total becomes 36, the number of known counts becomes four, and the mean becomes 9. This checks that the assistant responds to the data change rather than repeating its first calculation.
Ask for a short report only after the calculations are settled. A faithful report says there were 30 recorded attendances across three sessions with known counts, while Thursday is missing. It should not say that 30 different people attended, that Friday's class was cancelled, or that the centre had an average of ten attendees across all four sessions. Those are different claims.
If an answer is wrong, point to the precise row or calculation rather than merely saying “try again”. Keep the original records beside the result so revisions remain anchored to evidence. This method transfers to budgets, survey summaries and inventories, provided you define units and missing values appropriate to each task.
Keep analysis and changes distinct
A summary in chat is not an edit to the centre's source workbook. If a connected tool can create or modify a file, review the proposed destination and exact change before authorising it. Verify the resulting file and preserve the original when you need an audit trail. A generated report should state its source, reporting period and missing-data treatment so another person can reproduce the calculation.
The finished exercise demonstrates faithful summarisation of a small invented dataset. It does not establish the accuracy of real attendance collection, a causal explanation for a trend or complete analysis of a large workbook. Those require additional evidence and checks beyond the assistant's response.
Try it yourself
Use the fictional attendance prompt, check its four expected results, then replace the missing Thursday value with 6 and predict the revised total and mean before running it again. Write a two-sentence report with no invented explanation.
Show hint
For the first run, expect total 30, three known counts and mean 10; Thursday stays missing. For the second, expect total 36 and mean 9 across four counts. A good report distinguishes attendances from unique people and a missing value from zero. It does not claim that a source workbook was edited.
Check your understanding
4 questions to revisit the ideas. No score is saved, and you can retry. This is separate from Skill Quiz; it does not award XP or unlock guides.
Sources & further reading
Primary references for the concepts in this guide. Examples and exercises are illustrative; verify provider-specific behaviour against the version you use.