Event Feedback Survey: Write It and Summarise the Answers With AI

Keep an event feedback survey to ten questions: six ratings, two comment boxes, one "would you come again" and one permission to quote. Send it the next morning, export the replies, strip names, and have AI sort the comments into a fixed theme list. Count the themes in your spreadsheet, not the chat, then put the results on one page.

The AI step is the reading, not the counting. Language models read comments well but miscount them: ask for a tally and the totals drift. So the AI does the tagging, the sheet does the sums, and you spot-check both. Ten questions should take most attendees about two minutes on a phone, short enough to finish in one sitting.

Follow me on Instagram@sagnikteaches

The worked example is an illustrative one-day conference: 220 attendees, 96 replies, three session tracks and a drinks reception. Every figure and comment below is invented to show the method, so do not read any of it as a benchmark for your own event.

Connect on LinkedInSagnik Bhattacharya

The ten questions, with the decision each one feeds

Every question earns its place by changing something you will do. If you cannot name the decision, cut the question. Questions 1 to 6 are ratings, 7 and 8 are open comments, and 9 and 10 are single choices.

Subscribe on YouTube@sagnikteaches

Wording does most of the damage in a first draft. Three fixes I would make:

• "How were the speakers and the venue?" becomes two questions, 3 and 4. A single score for two things cannot tell you which one to fix.
• "Don't you agree the lunch was excellent?" becomes "How would you rate the food and drink?" Leading wording inflates scores and ruins any comparison with next year.
• "Any other comments?" becomes "What is the one thing we should change first?" A blank invitation gets "no" or "thanks". A specific one gets something you can code and act on.

Label the ends of every scale (1 is poor, 5 is excellent) and keep the direction the same throughout. Question 5 joins food and drink because one caterer usually owns both; split them if two suppliers do. If your event has several session tracks, add an optional question listing the sessions and ask which one mattered most. It costs the reader ten seconds.

Pick a form tool that exports cleanly

Any form tool will do if it passes four tests: it exports every reply to a spreadsheet or CSV, it works on a phone without an account, it gives you a separate link for each audience, and its free limits fit your reply count. As of October 2026 the vendors' own pages say this.

The reply cap matters more than it looks. A 220-person event with a 45% reply rate produces about 99 replies, one under Jotform's free limit, and a 300-person event would not fit. For step-by-step builds, the Tally survey walk-through and the tutorial on Gemini in Google Forms cover two of these tools in detail.

Send it the morning after, and work out your reply rate

Send the survey the next morning, while the day is fresh and attendees still have your email open. Put the link in the thank-you email and a QR code on the closing slide and the exit signs. Send one reminder after about three days, worded so that people who have replied can ignore it, and close the survey after a week so the numbers stop moving while you write.

Use one link per audience: attendees, speakers, exhibitors or sponsors, and volunteers. Their experiences differ, and a blended average of a speaker's green-room comfort and an attendee's lunch queue tells you nothing. Open the form with a short, honest introduction:

Thanks for joining us on [date]. Ten questions, about two minutes.
Your answers are anonymous unless you tell us who you are. We use them to improve the
next event, and we may quote comments without your name if you say yes on the last question.
Please leave out names of individual people; we remove any we find.

Your reply rate is replies divided by people who attended, not by people who registered. In the example, 96 replies from 220 attendees is 43.6%. Treat every percentage you report afterwards as "of those who replied". Below roughly 20% I would present results as pointers and read every comment myself, because the people who answer are rarely a fair slice of the room.

Strip names from the survey export before the AI reads it

Export to CSV or send the replies to a sheet, then take out the name, email and phone columns, and add an ID column (1, 2, 3 and so on) so you can trace a coded comment back to its row. Keep the audience or ticket type, because you will want to compare groups. Then read down the comment columns for names of staff or other attendees, and replace each with a role. Treat the cleaned file as something you could show a supplier, because whatever is in it is what you are handing to the AI. If the export arrives untidy, with stray spaces or merged cells, the tutorial on cleaning messy data in Excel covers the usual fixes.

Where you paste the data matters. Business plans such as ChatGPT Business, ChatGPT Enterprise, Claude Team and Claude Enterprise do not train on your content by default, and consumer plans have a model-training switch in their privacy settings. Setting names change, so check the current screen rather than a screenshot. If you can use neither, code a sample by hand and skip the AI for the rest.

Draft a theme codebook from a sample, then freeze it

Coding means tagging each comment with a theme label. The codebook is the fixed list of labels and what each one covers, and it must not change halfway through, or counts from the first batch stop matching the last. Build it from a sample first: pull 40 comments from the "change first" column, mixed across ticket types, and ask the AI to propose the list.

Below are 40 comments from an event feedback survey, one per line, answering
"What is the one thing we should change first?"
Read them all, then propose 8 to 10 theme codes for an event organiser.
For each code give: a short capitalised label, a one-sentence definition that says what is
included and what is not, and one example comment copied word for word.
Do not count anything yet. If a comment fits two codes, say which would be primary.
List any comment that fits none of your codes.

A typical first answer, abbreviated and illustrative, offers labels such as QUEUES, CATERING, ROOM TEMPERATURE, SOUND, WI-FI, SESSION LENGTH and SIGNAGE. It is too fine-grained, and two codes overlap. I would merge QUEUES into the area where the queue happened (food, registration), fold ROOM TEMPERATURE into VENUE, combine SOUND and WI-FI into TECH, and stop at nine codes. That edit is your judgement, and it is the most valuable five minutes of the whole process.

Freeze it. Save the table in a document, name it with the event and date, and paste it into every coding prompt unchanged.

Code every comment against the frozen codebook

Now give the AI the whole list in batches of about 100 rows, with the codebook and strict instructions about the output. Batches keep it from skipping rows, and the final line lets you catch it when it does.

You are coding survey comments for an event organiser. Use ONLY this codebook:
[paste the nine-code table, with the Covers and Does not cover columns]
Input: a table with columns ID and Comment (100 rows).
Output: a table with exactly 100 rows and these columns: ID, Primary code, Second code
(or blank), Tone (positive, negative or mixed), Quote-worthy (yes or no).
Rules: use the labels exactly as written. Treat sarcasm as negative. If a comment is too
short to code, use OTHER. Do not summarise, count or comment on the data. Keep every ID unchanged.
After the table write one line: "Rows returned: N".

A realistic failure: you send 100 rows and the reply says "Rows returned: 97". The model has dropped three rows, perhaps long or oddly formatted ones, and has not said which. Never accept a batch whose count differs. Paste the IDs of the returned rows back with the instruction "list the IDs missing from this output", or re-run the batch smaller. Two minutes of checking here saves a report that quietly ignores three people.

Spot-check twenty rows by hand

Pick 20 coded rows at random: sort by a column of =RAND() values and take the top twenty. Read each comment beside its code and tone. My rule of thumb is that more than two disagreements out of twenty means the definitions are too loose, so tighten the "does not cover" column and re-run that batch. Three failures turn up again and again.

• Sarcasm. "Fantastic, another 40 minutes in the lunch queue" can come back as positive when the prompt has no rule for it. Adding "treat sarcasm as negative" helps, and the spot-check is how you find out whether the rule is working.
• Mixed comments. "Great keynote, shame about the lunch queue" in the "change first" column should be FOOD. Tell the AI which question each batch answers, and say that for "change first" the primary code is the thing the person wants changed.
• Local shorthand. Room names, acronyms and nicknames for a session mean nothing to the model. Add a line to the prompt such as "Room 2 is the side room with the panel sessions" and re-run the affected rows.

Also scan the OTHER pile. If it holds more than about one comment in ten, a theme is missing from the codebook. Add a code, then re-run all batches so every count uses the same list.

Count in the sheet, not in the chat

Paste each batch's coded table beside the original comments, matching on ID, so every comment has a primary code in your sheet. Then count with ordinary formulas. With the codes in column D and the question's comments in the same rows:

Comments coded FOOD:        =COUNTIF(D2:D80,"FOOD")
Share of coded comments:    =COUNTIF(D2:D80,"FOOD")/COUNTA(D2:D80)
Average overall rating:     =AVERAGE(B2:B97)
Promoters (9 or 10):        =COUNTIF(C2:C97,">=9")
Detractors (0 to 6):        =COUNTIF(C2:C97,"<=6")

A pivot table does the same job if you prefer clicking. In the example, 79 people wrote a "change first" comment and 91 wrote a "best part" comment, so these are the counts by primary code.

The rating questions come out the same way. Of 96 replies, 41 scored 9 or 10 on question 2 (42.7%) and 17 scored 0 to 6 (17.7%), so the recommend score, promoters minus detractors, is +25. On question 9, 61 said yes, 29 maybe and 6 no. The tutorial on AI analysis of recommend-score comments goes deeper on the 0 to 10 question if it matters to your sponsors.

Here is why the sheet does the counting. Suppose you ask the chat for a tally and it says "about 30 comments mention food". The sheet, counting FOOD as the primary code in the "change first" column, says 24. The chat has swept in the six food comments from the "best part" column as well. Both numbers are defensible, and only one answers the question you asked. The sheet gives you the same number every time and shows its working.

The one-page report, filled in

The report is a summary of verified numbers, so give the AI the numbers and forbid it from making any. The prompt below carries them in.

Write a one-page post-event report for the organiser using ONLY the figures below.
Do not calculate or estimate anything yourself.
Figures: attendees 220, replies 96 (43.6%), overall rating 4.2, recommend score +25,
would come again: yes 61, maybe 29, no 6. Ratings: programme 4.3, venue 3.7, food 3.4, registration 4.4.
"Change first" counts (79 comments): FOOD 24, VENUE 19, FLOW 11, TECH 9, SESSIONS 8, NETWORK 5, COMMS 2, VALUE 1.
"Best part" counts (91 comments): SESSIONS 47, NETWORK 28, FOOD 6, VENUE 4, FLOW 2, OTHER 4.
Quotes (permission given): [paste five quotes].
Structure: one-sentence headline; what worked (3 bullets with counts); what to fix first
(3 bullets, each with one suggested action); the five quotes; one decision for the next event.
Say "of 96 replies", never "of attendees". Under 300 words.

An illustrative reply, trimmed, looks like this:

Headline. Attendees rated the day 4.2 out of 5, and 61 of 96 replies would come again, but food and room comfort drew most of the criticism.
What worked. The programme was the best part in 47 of 91 comments and rated 4.3. Networking came next at 28 of 91. Registration scored 4.4, the highest area.
Fix first. Food and drink, 24 of 79 comments, rated 3.4: add a second serving point and send the caterer a dietary count five days out. Venue, 19 of 79, rated 3.7: set the heating earlier and add a speaker in the back half of Room 2. Flow, 11 of 79: build a 15-minute buffer into the afternoon.
Decision. Keep the programme format; move the saved budget into catering capacity and room sound.

Before it goes anywhere, check every figure against the sheet and every claim against the comments. The AI will sometimes add a flourish, such as saying that networking comments were "mostly about the drinks reception", which you can only keep if you have read enough comments to know it is true. Cut anything you cannot trace. Plan on about an hour for a hundred replies, from export to finished page, and longer the first time because the codebook has to be built.

What to do with the answers in the week after

A survey nobody acts on teaches attendees not to answer next time. Within a week, do three things. First, turn each "fix first" item into a task with an owner and a date on the next event's plan; the checklist tutorial in this series shows how to feed lessons back into a task library. Second, send the relevant excerpts to the people who can act on them: the caterer's food scores, the venue's room comments, each sponsor's mentions. The tutorial on sponsor reports and team debriefs covers that hand-off. Third, tell attendees what changed. A short note costs little and shows people their answers were read, which gives them a reason to reply next time.

Subject: What you told us, and what we are changing

Thank you to everyone who answered the survey after [event]. Three things came up most:
lunch queues, temperature and sound in Room 2, and the afternoon running late.
For [next event] we are adding a second serving point, setting the heating earlier
and building a buffer into the afternoon. What you loved about the keynote and the drinks
reception stays exactly as it is.

Keep the codebook, the counts and the one-page report in the event's folder. Next year's survey is only comparable if it uses the same ten questions and the same nine codes, and the earlier file is the proof of what you measured. Where you do change a question, note it, because a score that moved after a rewording has not necessarily moved in the room.

Keeping attendee comments private when you use AI

Free-text comments carry more personal detail than their authors expect: a colleague's name, a health condition behind a dietary request, a complaint about a named member of staff. Remove those before pasting. Use a business plan or the model-training switch in the privacy settings, keep the raw export in your own storage, and delete the working copy once the report is signed off. Quote only comments from people who said yes on question 10, and strip anything that identifies them. Sponsors and venues need excerpts and counts, not the whole file.

If you have attendees in the EU, data-protection rules may apply to the survey as well, so ask your data-protection adviser before storing or sharing the raw replies. For most small events the practical answer is the same as the habit above: collect little, say why, and keep it short-lived.

Questions about running an event survey

What if I only get 20 replies?

Read all of them yourself; at that size the AI adds little. Use it to draft the one-page report from your own notes, and quote counts rather than percentages, such as 7 of 20 replies mentioned lunch. A small self-selected sample leans towards the keenest and the angriest attendees, so present it as a pointer, not a measurement.

Can I use the AI inside my survey tool instead?

You can, and some form tools now offer summaries. The checks do not change: freeze a codebook, count in a sheet, spot-check twenty rows and keep names out of the data. Tool features and limits change often, so test the built-in summary on a copy of your own replies before you rely on it.

Should an event survey be anonymous?

Usually yes, because people are franker when they are not named. If you want to follow up a serious complaint, add an optional contact box as the last question and keep that column out of any file you give to an AI. Say plainly on the form how answers are stored and who sees them.

When should I send the survey, and how many reminders?

Send it the morning after, while the day is fresh, with the link in a thank-you email and a QR code on a closing slide or exit sign. One reminder after about three days is enough. Close the survey after a week so your numbers stop moving while you write the report.

Further reads

• Hotel Guest Feedback: Turn Post-Stay Surveys Into Fixes With AI — Another hospitality case: comments turned into a ranked fix list.
• Museums and AI: Turn Visitor Feedback Into Next Season's Plan — Visitor comments into a short list of changes for next season.
• AI for Marketing an Event: A Six-Week Promotion Plan for Venues — A week-by-week promotion calendar for a one-off 120-guest event, with prompts, partner emails, tagged links and a weekly check on ticket sales.
• AI in Live Events: Postponed-Show Messages and Refund Replies — A facts sheet, a first-notice prompt and refund-reply wording for small promoters and venues, with a person approving every send.
• Bookshop Event Ideas: Author Nights, Book Clubs and Signings With AI — A filled-in event plan, a publisher outreach email, ticket-page copy and a promotion run for author nights, signings and book clubs.
• Event Sponsorship Proposal: Tiers and Pitch Emails Drafted With AI — Price four sponsor tiers from your budget gap, then have AI draft the proposal and a four-email pitch sequence. One worked event, with the sums.
• Yoga Workshop Ideas: Plan, Price and Promote a Weekend With AI — A worked weekend workshop: the break-even ticket count, three price tiers, a fact sheet, three emails, five posts and a weekly sales check for studio owners.
• AI Entertainment for Events: What Party Hire Firms Can Add to a Package — Five AI add-ons a party hire firm can sell, a costed caricature station for 150 guests, and the venue and consent points to settle before listing any.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Tally's and Jotform's pricing pages (free plan limits), Microsoft Support's Forms guide (question types) and Google's Forms pages (export to Sheets), all read in October 2026. Example figures and comments are illustrative.