Put every comment in one sheet, strip out names, read thirty yourself to draft ten or so themes, then have an AI assistant tag each comment against those themes in batches. Count the tags in a spreadsheet rather than in the chat, and spot-check a tenth of them. The most frequent fixable themes become a short list of changes for next season.
AI speeds up the sorting; it does not choose what to fix. A small museum with a couple of hundred comments can finish the whole job in a day of volunteer time, and most of that is typing up paper cards rather than anything clever. Sector-wide surveys and policy papers tell you what visitors in general expect, which helps with strategy, but they cannot tell you that the stairs to your upstairs gallery have no sign. Your own comments can.
The museum and its 212 comments
The figures here are invented. A small local-history museum has a part-time manager and about 30 volunteers. It is open from April to October and counted 6,800 visits this season. Visitors left feedback in four places:
| Source | Comments | Notes |
|---|---|---|
| Comment cards at the desk | 118 | Handwritten; the card asks "What one thing should we improve?" |
| Comments book by the exit | 46 | Handwritten; mixed praise and complaints |
| Online form after booking (Google Forms) | 31 | Typed; already in a sheet |
| Review sites | 17 | Public reviews copied across |
| Total | 212 | Over about 6,800 visits |
It is the end of October and the museum is closing for the winter. The manager wants a short, costed list of changes for next April to put to the trustees. The steps below follow what that manager does, in order, with the time each takes.
Step 1: one sheet, one comment per row, identifiers removed
Create a spreadsheet in Excel or Google Sheets with these columns: ID, date, source, visitor type (family, adult, school group, unknown), comment, primary theme, secondary theme, problem or praise, and a note. Give each comment an ID such as C001 as you enter it. Typing up the 118 cards and 46 book entries takes a volunteer about a minute each, so a little under three hours; the online and review-site comments are pasted in minutes. If you would like to try photos of the cards, test ten first and compare each result with the original, because a misread word changes the theme. The tutorial on what breaks when ChatGPT reads photos and files shows what to look for.
Before any text goes to an AI tool, remove personal information. Visitors often write names, ages or phone numbers, and children's cards can carry a first name and a school. An example before and after (the details are invented). The original reads: "Thanks to [first name] on the desk, my daughter [first name] (7) loved the trail. Please call me on [phone number] about school visits." After clean-up it reads: "Thanks to the volunteer on the desk, my daughter (7) loved the trail. Wants to discuss school visits." The phone number goes into a separate list for the education officer to follow up, never into the chat.
Choose the account carefully. A work account on a plan that does not train on your content by default is the better choice, or else a personal plan with the model-training switch turned off in its privacy settings. Either way, keep the master sheet with identifiers on your own drive and paste only the cleaned comment column. The museum AI policy checklist covers the wider rules for visitor data.
Step 2: read thirty comments before the tool sees any
This step is easy to skip, and skipping it tends to produce a generic list of themes that could belong to any business. Read about thirty comments, a mix from every source, and note the themes you see. Doing this first means the themes belong to the museum, not to the model. Choose them on purpose: ten from each of the three biggest sources, including some very short and some very long ones, so that you are not reading only the first thirty cards of one busy weekend. After thirty comments the manager drafts nine themes and one catch-all, and writes a rule for each edge case that came up while reading. The list below is the one the manager ends up with after the spot-check in Step 4.
| Theme | Counts as | Edge rule |
|---|---|---|
| Volunteers and staff | Comments about people who welcomed, guided or helped | Ignore names; praise and complaints both count |
| Toilets and baby-changing | Toilets, baby-changing, accessible toilet | Café toilets count here |
| Signs and wayfinding | Finding the museum, galleries, stairs, exits | Car park signs count here |
| Café and refreshments | Food, drink, café service and café prices | Café prices go here, not under Opening hours and price |
| Children's activities | Trails, quizzes, hands-on objects, activity sheets | School-group comments go here if about activities |
| Seating and step-free access | Places to sit, ramps, lifts, steps, wheelchair routes | Mentions of a visitor's own disability are not a theme; tag the barrier |
| Labels and text size | How easy the written information is to read | Content errors go to Other |
| Lighting | Light levels in galleries | Kept separate because conservation limits what can change |
| Cold galleries | Temperature and heating | Includes "too hot" if it appears |
| Opening hours and price | Admission price, opening days and times | Not café prices |
| Other or unclear | Anything else, or general praise with no topic | Comments about other places go here |
Step 3: the tagging prompt, and what came back
The prompt below is the final version. The first run used a merged "Labels and lighting" theme and had no rule about irony, and Step 4 shows why both changed. Paste the theme list and rules into the prompt, then add the comments in batches of about 40, which keeps each reply short enough to check. If you would rather upload the sheet, Claude accepts CSV and XLSX files and ChatGPT and Gemini also accept uploaded files, but check your plan's file limits first. For a few hundred short comments, pasting in batches is simpler and every batch can be checked on the screen.
You are helping a small museum sort visitor comments into themes.
Use ONLY these themes, named exactly as written:
Volunteers and staff; Toilets and baby-changing; Signs and wayfinding;
Café and refreshments; Children's activities; Seating and step-free access;
Labels and text size; Lighting; Cold galleries; Opening hours and price;
Other or unclear.
For each comment return one table row: ID | Primary theme | Secondary theme
(or none) | Problem, praise or unclear | Reason in under 12 words.
Rules: the primary theme is the one the visitor spends most words on.
Café prices go under Café, not Opening hours and price.
If the tone is ironic or sarcastic, mark it as a problem and say so.
If you are unsure, use "Other or unclear" and say why.
Do not count, summarise or comment on the batch.
COMMENTS
C014: ...
C015: ...The last rule matters more than it looks. Assistants are good at classifying a comment and unreliable at adding up, so the instruction keeps the arithmetic where it belongs, in the sheet. A few rows of the output look like this (illustrative):
| ID | Primary | Secondary | Type | Reason |
|---|---|---|---|---|
| C014 | Signs and wayfinding | none | Problem | Could not find the stairs to upstairs gallery |
| C015 | Volunteers and staff | Children's activities | Praise | Volunteer helped child finish the trail |
| C016 | Toilets and baby-changing | none | Problem | No baby-changing in the downstairs toilet |
| C017 | Cold galleries | Labels and text size | Problem | Freezing upstairs and labels too small |
| C018 | Café and refreshments | none | Problem | Café shut before 3 pm; scones dear |
The second column is what keeps multi-topic comments from losing half their meaning. C017 says it is freezing upstairs and the labels are tiny, which is two problems with two different fixes, and the secondary column lets you count both.
Check each batch as it arrives. Assistants like to invent a theme of their own, such as "Customer service" or "Facilities", even when told to use only your list. When a row uses a theme that is not on the list, paste the batch back with a line such as "Rows C031 and C038 used themes that are not on my list; redo those rows using only the listed themes", and save the corrected rows into the sheet as you go.
Step 4: the spot-check that decides whether to trust the tags
Pick 20 comments at random, just under a tenth of 212, and have a volunteer tag them without seeing the AI's answers. Compare. On the first pass the volunteer and the tool agree on 17 of 20, which is 85%. The three disagreements teach the manager more than the 17 matches: one comment about a dark gallery was filed under Labels, one sarcastic remark was filed as praise, and one comment about the café toilets had gone under Café.
Those disagreements change the rules. Lighting and labels become separate themes, because the museum can reprint a label but its conservation needs limit how bright a gallery can be. The prompt gains a rule that irony is a problem, and the edge rule that café toilets count as toilets. The manager reruns the batches that were affected, then checks a fresh random 20, not the same ones, since tuning rules to a sample and then testing on it flatters the result. This time 19 of 20 agree, which is 95%, and the manager is satisfied that the tags are good enough for counting, not for judging individual comments.
A note on what "good enough" means: people disagree with each other too. If you can spare a second volunteer, have them tag the same 20 comments and see how often the two people agree. If the humans match each other only 17 or 18 times out of 20, a tool that matches a careful volunteer 19 times is doing as well as a second person would.
Now count in the sheet, not in the chat. With the primary theme in column F, type of comment in column H and the data in rows 2 to 213 (212 comments plus the header), these formulas work in both Excel and Google Sheets:
=COUNTIF(F2:F213,"Toilets and baby-changing")gives the number of comments on a theme.=COUNTIFS(F2:F213,"Toilets and baby-changing",H2:H213,"Problem")gives how many of those raise a problem.- A pivot table with the theme as rows and the type as columns produces the whole table below in a minute.
Step 5: counts, then a score for what to fix first
The counts, on the final theme list, add up to 212:
| Theme | Comments | Raise a problem |
|---|---|---|
| Volunteers and staff | 42 | 2 |
| Toilets and baby-changing | 31 | 29 |
| Signs and wayfinding | 28 | 25 |
| Café and refreshments | 24 | 13 |
| Children's activities | 22 | 8 |
| Seating and step-free access | 19 | 17 |
| Cold galleries | 14 | 14 |
| Labels and text size | 11 | 10 |
| Opening hours and price | 10 | 5 |
| Lighting | 6 | 6 |
| Other or unclear | 5 | 1 |
| Total | 212 | 130 |
So 130 of the 212 comments raise a problem, about 61%. That is expected when the main card asks what to improve, and it is a reminder that the question wording shapes the data; next season the card could also ask what visitors enjoyed. Volunteers and staff are the most-mentioned theme, and almost all of it is praise.
Do not let the 40 praise comments about volunteers disappear into a count. They are worth reading aloud at the volunteers' end-of-season meeting, and any named thanks should come from the original cards, because the cleaned sheet that went to the AI no longer has the names. Praise is also the cheapest thing a museum can give its volunteers.
Counts alone mislead, because the most-mentioned problem is not always the best one to fix. Score each problem theme with a simple rule: problems × fixability ÷ cost. Fixability is 3 if volunteers could do it in a weekend, 2 if it needs a supplier or a small budget, and 1 if it needs building work or a committee decision. Cost is 1 for low (under about $300), 2 for medium ($300 to $1,500) and 3 for high (over $1,500). These bands are illustrative; use your own.
| Theme | Problems | Fix | Fixability | Cost | Score |
|---|---|---|---|---|---|
| Signs and wayfinding | 25 | Six directional signs and a floor line from the entrance | 3 | 1 | 75 |
| Seating and step-free access | 17 | Six folding seats and a sign showing the step-free route | 3 | 1 | 51 |
| Café and refreshments | 13 | Price board and closing time on the door | 3 | 1 | 39 |
| Toilets and baby-changing | 29 | Fold-down changing shelf; sign to the second toilet | 2 | 2 | 29 |
| Children's activities | 8 | Reprint trail sheet with pencils at the desk | 3 | 1 | 24 |
| Labels and text size | 10 | Larger print for the 12 most-read labels | 2 | 2 | 10 |
| Cold galleries | 14 | Note on the website and desk now; heating review for trustees | 1 | 3 | 4.7 |
| Lighting | 6 | Explain conservation limits on a label; no change to levels | 1 | 2 | 3 |
| Opening hours and price | 5 | Trustees to review | 1 | 2 | 2.5 |
The top five scores go into the plan: signs, seating and step-free access, café, toilets and children's activities. The manager adds larger-print labels as a stretch item, because it is cheap and the 12 labels are easy to identify, and sends cold galleries, lighting and hours to the trustees with the interim notes. The scores are for discussion, not a verdict: the toilets have the most problems and rank fourth because the cheap fix only partly solves it, which is a useful thing to say out loud to the trustees.
Step 6: writing the trustee note, briefing and "you said" panel
With the table done, the AI is useful again for the writing. Give it the table (not the raw comments) and ask for three things: a one-page note for the trustees, a short briefing for volunteers, and the wording for a "you said, we did" panel at the desk.
Using only the table below, write (1) a one-page note for trustees with the
five committed changes, cost bands and a timetable for April,
(2) a five-bullet briefing for volunteers on what is changing,
(3) a "you said, we did" panel of under 80 words.
Quote the comment counts exactly. Do not describe praise as "overwhelming"
or add any claim that is not in the table.A first draft is likely to include a line such as "visitors overwhelmingly praised our volunteers". The actual figure is 42 of 212 comments, about 20%, and 40 of those 42 were praise, so the line becomes "volunteers and staff were the most-mentioned theme, in 42 of 212 comments, and 40 of those were praise". The panel might read: "You said: the upstairs gallery is hard to find. We did: added six new signs and a floor line from the entrance. You said: there is not enough seating. We did: added six folding seats; ask a volunteer if you would like one. Still to come: a baby-changing shelf in spring." Whatever the tool writes, check each number against the table and each promise against what the trustees have approved before anything is printed. The same table can also support accreditation paperwork; the accreditation evidence checklist shows where visitor feedback fits.
Where the AI went wrong in this run
Four mistakes showed up, and each is typical of comment sorting.
- Two themes merged into one. "Too dark" and "labels too small" both landed under Labels. Only a person who knows about conservation limits would think to split them, which is why the manager reads the sample before accepting the themes.
- Sarcasm read as praise. "Lovely that the toilets are still 'being refurbished' after three years" was tagged as praise. The fix was a rule about irony and a human read of every comment the tool marks as praise but which contains words like "still" or "again".
- Comments about somewhere else. A review-site comment praised a castle's views and complained about its car park. The tool filed it under Signs and wayfinding. It belongs in Other, and the edge rule now says so.
- An over-confident summary. The trustee note's "overwhelmingly praised" line, above, was the tool filling a gap with a flattering adjective. Giving it counts rather than raw comments, and telling it not to add claims, removes most of this.
None of these errors would have shown up in a quick skim of the output; they show up only because of the 20-comment check and the rule of giving the writing step only the checked table. For the general version of the method, the tutorial on analysing customer feedback with AI covers survey data in any business, and the hotel guest feedback tutorial shows the same loop in another setting.
Make next season's cards easier to sort
The data is only as good as the question. Three small changes to the card make next year's sorting quicker. Ask two questions rather than one: "What did you enjoy most?" and "What one thing should we improve?", so that praise and problems are not mixed in one box. Add a tick-box for who the visitor came with (family, adults, school group, on my own), which gives you a visitor-type column without guessing. And add a line such as "May we quote you, without your name?" with yes and no boxes, so that the "you said, we did" panel can use real words with permission. Put the same questions, with a QR code, on the online form so that every source produces the same columns. If you would rather not type up cards at all, a short form on a tablet by the exit does the typing for you, at the cost of fewer handwritten comments.
Checking next season that the fixes worked
Keep the same theme list next year and add a season column to the sheet. The point of the exercise is the comparison: if signs were 28 comments this season, a good result next April to October is a handful, and if the number has not moved, the fix did not work and the sign is in the wrong place. Counts need to be read against visits, since a busier season produces more comments. Dividing by visits gives comments per 1,000 visits; this season's 28 sign comments over 6,800 visits is about 4.1 per 1,000.
The whole job, in this example, comes to about seven hours: roughly three for typing up the paper comments, half an hour reading the first thirty, an hour for the tagging batches, 45 minutes for the two spot-checks, half an hour for the counts and scoring, and an hour for the three written outputs and their checks. Nearly half is typing, which no AI removes. If you would rather build the theme list and the spot-check routine with someone before handing it to volunteers, bring last season's comments to a 1:1 call.
Questions museum managers ask about analysing visitor comments
How many comments do we need before this is worth doing?
Around 50 is enough to see patterns, and below that you can read them all yourself in half an hour. The method earns its time from a couple of hundred comments upwards, or when several people collect feedback in different places. The spot-check and the count in a spreadsheet matter more than the number of comments.
Can AI read handwritten comment cards from photos?
Some assistants can read handwriting from a photo, but accuracy varies with the writing and the lighting, and a misread word changes the theme. Test ten cards, compare every result with the original, and decide from that whether typing the cards up by hand is quicker than correcting the output.
Is it acceptable to paste visitor comments into an AI chat?
Only after removing names, contact details and anything that identifies a person, especially children, and only in an account whose settings you have checked for model training. Keep the original sheet with identifiers outside the chat. Check what your visitor privacy notice says about how comments are used.
How often should we repeat the exercise?
Once per season is enough for most small museums, with a lighter check at the midpoint if something big changes, such as a new gallery. Keep the same theme list from year to year so that you can compare counts, and only add a theme when comments clearly need it.
Further reads
- Event Feedback Survey: Write It and Summarise the Answers With AI — Writing the survey itself, then summarising the answers.
- AI Policy for a Small Library: One Page, Patron-Data Rules — Handling visitor and patron data when AI is involved.
- Can an AI Guide for Museums Stick to Your Own Collection Notes? — A visitor-facing AI guide that sticks to your own collection notes.
- How to Clean Messy Data in Excel: Step-by-Step Guide — Tidying a messy comments sheet before you analyse it.
- 60 AI Prompts for Google Sheets That Actually Work (Copy, Paste, Automate) — Prompts for working with a comments sheet in Google Sheets.
- 8 Examples of AI in Tourism From Guided Tours to Group Bookings — Eight practical AI uses for tour guides and attractions, each with the tool type, set-up time, a worked example and a check that it worked.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the worked example uses invented figures for illustration; statements about file uploads and model-training settings follow OpenAI, Anthropic and Google help pages as recorded in October 2026.