Sentiment analysis has become a product feature, which is a shame, because the manual version takes an afternoon and produces something no dashboard can: a manager who has actually read what guests wrote.
The star rating is not sentiment. It is a summary the guest chose, and it compresses "the food was extraordinary and the room was too loud to speak in" into a three.
What follows is a method for finding the themes in a quarter of reviews. It is dull, it works, and you only have to do it four times a year.
Why the average rating tells you nothing you can act on
Your rating moves slowly and lags everything. By the time it has shifted enough to be noticed in a meeting, the cause is two months old and the manager involved has moved to another site.
It also hides the useful part. A quarter of fours can be four different restaurants: good food with slow service, good service in a tired room, or the one where everybody mentions the same server by name and nobody has told him.
What you want is themes and their direction. That means reading, and reading in a structured way so that what you find survives your own mood on the day you read it.
The method
One person, one afternoon, one sitting. The sitting matters more than it sounds like it should.
- Export or copy every review from one quarter, across every platform you monitor. All of them, including the fives.
- One row per review: date, platform, rating, location, and the text itself.
- Read them in one go if you possibly can. Themes announce themselves when you read forty in a row and stay hidden when you read four a day between other jobs.
- Tag each review with up to three themes from a fixed list, and give each tag a direction: positive, negative, or mixed.
- Tally the tags. Then go back and read the ten reviews behind the largest theme again, slowly.
The theme list
Fix the list before you start. Inventing categories as you go means the first twenty reviews were coded under different rules from the last twenty, and the tally becomes an artefact of the order you read them in.
Below is a starting list. Adapt it once, for your own room, and then leave it alone for a year so the comparison holds.
THEMES — tag up to three per review, each + / − / mixed FOOD-QUALITY the cooking itself FOOD-VALUE price against expectation PACE waiting, gaps between courses, rushed WARMTH was the team glad to see them COMPETENCE did the team know things ROOM noise, seating, temperature, the loos BOOKING everything before arrival RECOVERY something went wrong and was handled NAMED-STAFF a person named, well or badly TALLY theme | positive | negative | mixed | reviews mentioning Also record: how many reviews you decided were not about the restaurant at all. Watch that number.
The rules that keep it honest
Coding your own restaurant is an exercise in self-deception unless you set the rules first and then follow them when they are inconvenient.
- Code what is said, not what you know. "Slow" gets a PACE tag even when you remember that Saturday and can explain exactly why it happened.
- A review can carry two directions. Praise for the food and a complaint about the room is a positive tag and a negative tag, not a neutral review.
- Have two people code the same fifty reviews separately, then compare. You will disagree on some of them, and the disagreements are the most interesting part of the afternoon.
- Count the ones you dismissed. Everybody discards a few reviews as unreasonable. Write down how many. If the number keeps growing, you are not coding, you are defending.
- Keep the sheet. Next quarter's entire value is in the comparison, and a sheet nobody can find is an afternoon spent for nothing.
What to do with the tally
Three themes. Not nine. The largest negative, the largest positive, and whichever theme moved most since last quarter.
The largest negative gets a named owner and one specific change. The largest positive gets protected — this is the step people forget, and it is how restaurants accidentally renovate away the exact thing guests kept mentioning.
The mover is your early warning. A theme that has doubled in a quarter is describing something that happened recently, which means it is still fixable and somebody still remembers doing it.
Then close the loop: when you reply to the reviews inside that theme, say what changed. A quarter of careful coding that produces no visible change is a filing exercise with better handwriting.
When a tool earns its place
At one or two sites, do this by hand. The reading is the product. Handing it to software means nobody in your company has read what your guests wrote, and you will notice the difference in how your managers talk about the room.
At nine sites and several hundred reviews a quarter, hand coding stops being possible and something has to group them for you. What to insist on is that the tool shows you the reviews behind a theme, so a manager can read the ten that matter rather than accept a number and a coloured arrow.
Either way the tally is not the output. The output is one sentence somebody can say at the next managers' meeting, and a person has to have read enough to write it.