← Craft

Two Levels and Tags: Taxonomy Under a URL Constraint

The new site’s design allowed two levels in the URL: a category and a subcategory. That was it.

What I had to fit into those two levels was thousands of support articles from several sources, built up over years, with categories nested inside categories nested inside more categories.

This is how we got it done, and what I’d tell anyone staring down the same constraint.

The setup

The product is a direct-to-consumer one. It’s expensive, it’s technical, and most owners eventually need a specialist to work on it with them. The client had two main sources of information about it, each with heavily overlapping subject matter. One was an expert-answered Q&A archive: customers asked questions, a specialist answered, and the vetted answers were rolled into a tagged, searchable list. The other was a resource library of how-tos and technical documentation, closer to the customer-facing experience. A couple of months in, a third source turned up, and it had to be reviewed before it was retired, too.

The goal was a single source of truth. My job was to lead the sorting, keep everyone moving, and deliver solid information architecture for a new system and new templates, all on a deadline.

Deep trees hide duplicates

Here’s the problem with a category tree that goes four or five levels down: you can’t see the duplicates from the top. Two answers to the same question can sit in completely different branches, and no amount of staring at the category list will tell you.

So we went in through titles. We pulled blocks of content that shared keywords, then read them one at a time and made a call: keep it, combine it with something else, rewrite it, or retire it. A lot of it was tied to products that no longer existed or were no longer supported, and that went quickly. The rest did not. Thousands of articles, one at a time, for months.

Get out of the water

At first, I let the client teams lead. We were trying to organize an ocean of content, and I was sitting in a boat floating over the top of it, trying to help the client see how massive the undertaking was. They managed their book of content from deep in the water. It was hard for them to see just how much work really needed to be done, and I mainly just offered perspective while we got started. Once they finally accepted how complicated and messy sorting was one small cup at a time, they were ready to get on board with me and start sorting the ocean of content into buckets instead of into tiny little cups. From there, the sorting exercise started taking off very quickly. SEO strategy gave us the order of work. I was able to sort through spreadsheets utilizing keywords and categorizing articles into categories that way. I did a lot of the work in a massive spreadsheet that I’d bring to client check-ins for discussion. They would walk away with the spreadsheet and validate individually each article in the categories I suggested. Customers and the professionals who work on these products search by product and product year, because the product changes every year and so do the answers. That was an easy place to start. From there, we worked backward to the broader categories, questions about the brand, the product overall, and the class of product.

Break ties on purpose

Product and year were our category and subcategory. But plenty of articles didn’t belong to just one product or one year, so we needed a tie-break order:

  1. Product, if it’s still an option.
  2. If not, year.
  3. If not, the area of the product, the specific part, or a technology. Sometimes a subject was a technology available across multiple products and years, and that could count as a product.

Everything, product and year included, also went in as a tag. We tagged as much as the system would allow, to give site search better metadata to work with, and the search bar worked better because of it.

Name it for the person in a hurry

Titles did a lot of the work. The naming convention led with the thing someone would be looking for: the exact part, then the issue, then the viewpoint of the product. The “viewpoint” was the side or viewpoint of the product that someone would be looking at when working on their issue. We’d seen some evidence in site search that some folks who didn’t know what part they were looking at would search in that way. We knew that users would be searching while in varying degrees of distress and would need the issue they were researching to be easy to see in an article, easy to find on a page, and easy to search.

Inside the articles, we used clear FAQ-style headers, with step-by-step instructions where they were needed. I modeled the structure on a typical customer’s journey while solving an issue: they glance, find what they need, and get out of there quick.

An exercise in faith

Toward the end, it turned into an exercise in faith. We knew there were still duplicates out there. The goal was to get everything categorized into one category and one subcategory, with the leftover duplicates cleaned up as we found them.

That’s easier to say than to live with. A project like this can feel like it will never end. People worry they’ll finish and realize they missed a whole pile, and that their peers will judge the work as messy. The stakes weren’t imaginary, either. Missing content on an expensive, technical product frustrates customers fast.

The guidelines nobody asked for

The editorial guidelines weren’t an original ask. My team and I recommended them, because there were so many stakeholders and collaborators that people needed something to turn to when they questioned whether a piece of content was valid or how it was put together. I structured the document like a content strategy: the goal and the North Star first, then how that works in everyone’s day-to-day.

The first months after launch were rough

It shipped, and the client’s teams were very happy with it. As expected, customers took a little longer. For the first three to six months, support requests were up, and so was the use of site search.

A few things helped. We made some careful adjustments to the experience, like adding more calls to action for the resource search. We put messages on the search page for people who were lost, explaining what had changed, why, and how to use the new repository. And we deprecated the old links to the old repositories and redirected them to the new experience. Once customers got their bearings, organic search picked up the slack in a big way, and people found the resource center much faster.

How I’d run it now, with Claude as my analyst

I did this project by hand: category keywords pulled into one tab, subcategory keywords into the next, and so on down the tree, for months. If another project like it landed on my plate today, I wouldn’t do it that way. I’d make Claude my analyst partner. It does the grinding and offers suggestions, and I’d make the calls.

1. Combine the lists.

I’d join a crawl of the site with whatever CMS export of URLs I can get. If I can, I’d pull the article text into a column too, so Claude can see what each piece is for, not just what it’s called. I’d upload the combined file and ask Claude to sort the list and group the content that’s exactly alike.

2. Group the near-duplicates.

Claude doesn’t make embeddings itself, so I’d have it write a script. I wouldn’t tell the script which words to look for. I’d let it score how similar each article is to every other one, then cluster everything above a cutoff. I’d also ask Claude to explain how it scored. Most of the time, it comes down to the actual words in the content, especially the distinctive ones.

3. Gut-check the clusters.

AI reads very literally, and that can lead it astray fast. Content written in “insider language,” the way professionals talk about a product and owners don’t, can land in some odd clusters. I’d eyeball every cluster, because the next steps take me straight into categorizing and migration planning. If clusters look off, I’d go back with more context, like the editorial guidelines and a few examples of articles I’ve categorized by hand.

4. Decide what matters.

Before anything gets scored, I’d get the list of supported products from the client. Then I’d pick a few high-level signs of priority. Publish date is the obvious one. Newer content is likelier to be accurate than older content on the same subject, so the date has to be in the crawl. But newer isn’t automatically right, so I’d treat date as one factor, not the answer. I’d also pull organic traffic and backlink data into the sheet, so nothing with real traffic or links gets retired without its own redirect, and I’d add any keywords I need prioritized. Then I’d give Claude a goal: migrate only content that isn’t duplicated or that scores well on these factors, and that’s relevant to the products we’ve defined. I’d also ask it to list its own scoring factors before it scores anything. It’ll probably come up with some I haven’t thought of.

5. Flag every article.

Then I’d let Claude score the clusters and flag each article: keep, retire as irrelevant, retire as duplicate, modify a duplicate, modify bad content, and so on. I’d settle the flag names before it starts. The client’s experts would still validate every flag, especially on technical instructions. And before I trust any of it, I’d run the whole process on one small batch first and spot-check a sample.

6. Sort the obvious stuff first.

I’d start with the keep list and look for quick buckets I can see with the naked eye. Obvious clusters are the fastest way to get a lot done. If one feels too obvious, I’d ask whether it’s only obvious to me because I work inside this content every day. A customer may search exactly that way and need exactly that category. I’d also ask whether the buckets run too deep. Here’s my gut check: would someone Google their exact problem and land right on this page, or would they start on a category page and look around? Then I’d work backward from the big categories while trying to keep it simple.

7. Fix it with the experience.

Most importantly, I’d try not to overthink it. Good customer experience can fix a lot more than you’d expect. Adding modals or tooltips to category pages, re-editorializing content that already exists, or just really solid linking strategy can solve problems that content categorization and search can’t. Not everything has to be solved with information architecture.

8. Map the redirects.

Last, I’d have Claude draft the redirect map, old URL to new, and I’d check it with a crawl.

What I’d tell you first

  • When trying to empty an ocean, you get farther with buckets than cups.
  • Pick a tie-break order and use it every time.
  • Well-written titles and headers can solve problems IA can’t.
  • Meet post-launch challenges with creativity and good user experience.

More from the blog