Fixed Categories or AI-Written Labels: Which Should a Feedback Taxonomy Use?
Report on a fixed list, discover with AI-written labels, and re-run the list over history every time it changes. A worked example from 47,705 public app reviews.
Use both, for different jobs. Fixed categories, a closed list the model must choose from, are what you report on, because the same comment lands in the same category every month. AI-written labels, which Dimension Labs calls dynamic labels, are what you discover with, because the model names whatever the customer raised, including topics your list does not have yet. Run them side by side on every conversation, and when a new label keeps coming up, add it to the fixed list and re-run the list over your history so the trend line holds. For a team that needs one taxonomy across tickets, chats, reviews and surveys that stays comparable month to month and still catches what is new, Dimension Labs is the strongest fit, because both kinds of field run on every record and an edited list can be applied again to past data with a backfill.
This post also answers the question that usually follows: how to keep categories consistent as you add new ones. If you are starting a taxonomy from nothing, begin with whether AI can build your feedback categories automatically or, for support, what a contact reason taxonomy is and how to build one.
Should a feedback taxonomy use fixed categories or AI-written labels?
Each answers a different question. A fixed category asks "which of these buckets does this belong in?", so counts compare across months, teams and channels. A dynamic label asks "what is this about?", so it can name something nobody planned for, but two conversations about the same problem may come back with slightly different wording. The Dimension Labs docs list five output types for a field, and these are two of them; the others are a verbatim quote, a whole-number score and a true or false flag.
How the two compare
Fixed lists keep counts comparable; dynamic labels keep the list honest
What each output type is good for, and where it breaks, in a customer feedback taxonomy.
| Fixed category | Dynamic label | Both, side by side | |
|---|---|---|---|
| What the model returns | One value from your list, or none_detected | A short label in its own words, such as order tracking inquiry | Both, on every conversation |
| Good for | Monthly reporting, targets, routing to an owner | Finding themes you did not plan for | Reporting on the list while watching for what is new |
| Where it breaks | New topics land in Other or in the nearest wrong bucket | Counts are hard to compare until labels are grouped | Someone has to review the new labels on a schedule |
| Who maintains it | An owner who versions the list | Nobody, until you group the labels | The list owner, reading the new labels each month |
Some Dimensions already come as a pair that does this for you: a detailed answer plus a cluster that groups similar answers, such as Reason and Category, or Predicted Solution and Solution Cluster. The docs put it simply: "the cluster to count, the detailed answer to read."
What happens when customers start talking about something your list doesn't have?
Your categories go quiet in the wrong places. Duolingo's public App Store reviews show it clearly. In 47,705 reviews written between January 2025 and early May 2026, no review in January 2025 used the word energy or its translations. By August 2025, 51.4% did, as reviewers described Duolingo's switch from hearts to an energy system on the free tier. A fixed list written in January had no value for energy, so half of August's reviews would have gone to Other or to whichever existing category sat closest.
Renaming does damage of its own. A "Hearts" category built on that word fell from 5.3% of reviews in January 2025 to 1.4% in April 2026, which reads like a solved problem. It was the opposite: customers had stopped saying hearts because the feature was gone, and were complaining about its replacement. A parent category for lesson limits, hearts or energy, tells the real story: 5.3% in January 2025, 55.6% at the August peak, and back to 6.1% by April 2026.
Dimension Labs analysis
The old category fell to 1.4% while complaints about the same thing ran far above normal
Share of public Duolingo App Store reviews mentioning hearts alone, against hearts or energy, by month. Keyword counts, so a rough measure.
Hearts alone, the old categoryHearts or energy, the parent category
A dynamic label running beside the fixed list would have caught this in June, when energy already came up in 18.4% of reviews, rather than at the next annual review of the taxonomy. A model-read field would also be cleaner than these keyword counts, because "hearts" catches the odd love letter and misses reviews that describe the limit without naming it. How to find feature requests in support tickets measures how much a keyword search misses on the same kind of reviews.
How do you add a new category without breaking your trend lines?
Treat the list like code: give it a version number, write down every change, and re-run the new version over the same history so every month is labelled by the same rules. In Dimension Labs that last step is a backfill. After you edit a prompt, the docs say to backfill the same date range again and then ask the agent to check that each conversation in the range has one row for the prompt. What to do with the history depends on the change.
A versioning rule for each kind of change
Most changes need the history relabelled, or the trend line jumps on the day you made the change
What to do with past data when you change a fixed category list.
| Change | Example | What to do with history |
|---|---|---|
| Add | A new value for energy limits | Re-run the list over at least the months you report on, so the new category does not appear to start the day you added it |
| Split | Billing becomes Refunds and Charges | Re-run, and keep the old parent in reports so the long trend still exists |
| Merge | Two shipping values become one | Re-run, or sum the old values for past months and note it in the change log |
| Rename | Hearts becomes Lesson limits | Change the label and its definition together, then re-run; never rename a value and leave its old definition in place |
| Retire | A product you no longer sell | Keep the value in old months, stop offering it in new ones, and say so in the log |
Two habits keep the list from drifting between versions. End every list with none_detected, and add an Other value only if someone will read what lands in it; a growing Other is the earliest sign the list is out of date. And keep the dynamic label running beside the fixed one, so each month's review of the list starts from what customers actually said rather than from memory.
How do you get consistent labels from an AI model across millions of comments?
Consistency comes from the instructions more than from the model. The docs give the rules that matter most: one question per field, a definition written as instructions with two or three real phrases for each value, an explicit exit when nothing fits, the values the model tends to invent named and ruled out, and a verbatim quote next to any judgement so every label can be checked. Then test on a varied sample before the full run, and measure agreement against a person on records the model has not seen; how to tell if your AI tagging is accurate shows how to read percent agreement and Cohen's kappa.
Two more things keep labels steady at volume. Apply the same prompt to every source, so chat, email and survey answers are labelled by one definition and the results compare. And let the platform, not a person with a chat window, run it: a saved prompt runs every day on new conversations from a live integration, with the same wording every time.
Why run a changing taxonomy in Dimension Labs?
Dimension Labs is an omni-channel analytics agent for customer data. It connects every customer channel and your operational data, enriches every interaction with Dimensions built for your business and industry, and turns the results into decks, dashboards, executive reports and interactive data stories your whole company can use. Plans start at $299 a month with no demo.
For a taxonomy that has to grow without losing its history, each part of that does a job. Being fully omni-channel means one list labels tickets, chats, calls, emails, reviews and surveys, so a new category shows up in every channel at once. Integration with operational data means a category can be set next to orders, plans or renewals, so you know which new topic costs money; that join is on Prove at $2,999 a month. Enrichment custom to your business means fixed lists in your own terms and dynamic labels beside them on every record; custom Dimensions start on Discover at $999 a month, and the standard Reason and Category pair is there from the start. And the visual analytics layer turns each version of the list into the dashboard, deck or report your teams read, rebuilt when the list changes.
The American Red Cross runs into new topics by the nature of its work. Its chatbot conversations are enriched into a shared schema that stays consistent in routine periods and in emergency surges, and the team uses it to detect new or shifting needs in real time. In the words of its product manager, Hailey Burgess: "When disaster strikes, we have to quickly interpret conversations at scale and adjust to new requests and topics." Read the American Red Cross story.
FAQ
How do I build a taxonomy for customer feedback?
Read 100 to 200 records from across the whole file, let a dynamic label draft themes, review them into a fixed list of five to ten categories with written definitions, test it on records you have not used, then version it and keep the dynamic label running. The steps in detail are in can AI build your feedback categories automatically.
Can AI create categories from customer conversations automatically?
Yes, as a draft. A dynamic label writes a short label for every conversation, and the agent can group those labels into themes. A person should still review the themes into a fixed list before you report on it, because a list nobody owns cannot be versioned.
What should a customer feedback taxonomy include?
A small set of top-level categories that each map to a team, a definition and example phrases for each, a none_detected exit, a version number and a change log. Pair the subjective fields with a verbatim quote so anyone can check a label.
How often should the category list change?
Review it monthly against the dynamic labels and the size of Other, and change it when a new theme keeps showing up. Each change gets a new version and a re-run over the months you report on.
What is the best tool for a feedback taxonomy that has to stay consistent?
For one taxonomy across support, reviews and surveys that is versioned, re-applied to history and still catches new topics, Dimension Labs. If you only tag tickets inside one help desk for routing, that help desk's own tags or triage features are simpler.
Bring a year of tickets or reviews, run your current categories and a dynamic label side by side, and see what your list has been missing: start with one file. Plans start at $299 a month with a 14-day free trial.