Posted by rjonesx.
Alright, so here's the situation. You have a million-product website. Your competitors have a lot of the same products. You need unique content. What do you do? The same thing everyone does — you turn to user-generated content. Problem solved, right?
User-generated content (UGC) can be an incredibly valuable source of content and organization, helping you build natural language descriptions and human-driven organization of site content. One common feature used by sites to take advantage of user-created content are tags, found everywhere from e-commerce sites to blogs. Webmasters can leverage tags to power site search, create taxonomies and categories of products for browsing, and to provide rich descriptions of site content.
This is a logical and practical approach, but can cause intractable SEO problems if left unchecked. For mega-sites, manually moderating millions of user-submitted tags can be cumbersome (if not wholly impossible). Leaving tags unchecked, though, can create massive problems with thin content, duplicate content, and general content sprawl. In our case study below, three technical SEOs from different companies joined forces to solve a massive tag sprawl problem. The project was led by Jacob Bohall, VP of Marketing at Hive Digital, while computational statistics services were provided by J.R. Oakes of Adapt Partners and Russ Jones of Moz. Let's dive in.
What is tag sprawl?
We define tag sprawl as the unchecked growth of unique, user-contributed tags resulting in a large amount of near-duplicate pages and unnecessary crawl space. Tag sprawl generates URLs likely to be classified as doorway pages, pages appearing to exist only for the purpose of building an index across an exhaustive array of keywords. You’ve probably seen this in its most basic form in the tagging of posts across blogs, which is why most SEOs recommend a blanket “noindex, follow” across tag pages in Wordpress sites. This simple approach can be an effective solution for small blog sites, but is not often the solution for major e-commerce sites that rely more heavily on tags for categorizing products.
The three following tag clouds represent a list of user-generated terms associated with different stock photos. Note: User behavior is generally to place as many tags as possible in an attempt to ensure maximum exposure for their products.
- USS Yorktown, Yorktown, cv, cvs-10, bonhomme richard, revolutionary war-ships, war-ships, naval ship, military ship, attack carriers, patriots point, landmarks, historic boats, essex class aircraft carrier, water, ocean
- ship, ships, Yorktown, war boats, Patriot pointe, old war ship, historic landmarks, aircraft carrier, war ship, naval ship, navy ship, see, ocean
- Yorktown ship, Warships and aircraft carriers, historic military vessels, the USS Yorktown aircraft carrier
As you can see, each user has generated valuable information for the photos, which we would want to use as a basis for creating indexable taxonomies for related stock images. However, at any type of scale, we have immediate threats of:
- Thin content: Only a handful of products share the user-generated tag when a user creates a more specific/defining tag, e.g. "cvs-10"
- Duplicate and similar content: Many of these tags will overlap, e.g. "USS Yorktown" vs. "Yorktown," "ship" vs. "ships," "cv" vs. "cvs-10," etc.
- Bad content: Created by improper formatting, misspellings, verbose tags, hyphenation, and similar mistakes made by users.
Now that you understand what tag sprawl is and how it negatively effects your site, how can we address this issue at scale?
The proposed solution
In correcting tag sprawl, we have some basic (at the surface) problems to solve. We need to effectively review each tag in our database and place them in groups so further action can be taken. First, we determine the quality of a tag (how likely is someone to search for this tag, is it spelled correctly, is it commercial, is it used for many products) and second, we determine if there is another tag very similar to it that has a higher quality.
- Identify good tags: We defined a good tag as term capable of contributing meaning, and easily justifiable as an indexed page in search results. This also entailed identifying a "master" tag to represent groups of similar terms.
- Identify bad tags: We wanted to isolate tags that should not appear in our database due to misspellings, duplicates, poor format, high ambiguity, or likely to cause a low-quality page.
- Relate bad tags to good tags: We assumed many of our initial "bad tags" could be a range of duplicates, i.e. plural/singular, technical/slang, hyphenated/non-hyphenated, conjugations, and other stems. There could also be two phrases which refer to the same thing, like "Yorktown ship" vs. "USS Yorktown." We need to identify these relationships for every "bad" tag.
For the project inspiring this post, our sample tag database comprised over 2,000,000 "unique" tags, making this a nearly impossible feat to accomplish manually. While theoretically we could have leveraged Mechanical Turk or similar platform to get "manual" review, early tests of this method proved to be unsuccessful. We would need a programmatic method (several methods, in fact) that we could later reproduce when adding new tags.
Keeping the goal in mind of identifying good tags, labeling bad tags, and relating bad tags to good tags, we employed more than a dozen methods, including: spell correction, bid value, tag search volume, unique visitors, tag count, Porter stemming, lemmatization, Jaccard index, Jaro-Winkler distance, Keyword Planner grouping, Wikipedia disambiguation, and K-Means clustering with word vectors. Each method either helped us determine whether the tag was valuable and, if not, helped us identify an alternate tag that was valuable.
- Method: One of the obvious issues with user-generated content is the occurrence of misspellings. We would regularly find misspellings where semicolons are transposed for the letter “L” or words have unintended characters at the beginning or end. Luckily, Linux has an excellent built-in spell checker called Aspell which we were able to use to fix a large volume of issues.
- Benefits: This offered a quick, early win in that it was fairly easy to identify bad tags when they were composed of words that weren’t included in the dictionary or included characters that were simply inexplicable (like a semicolon in the middle of a word). Moreover, if the corrected word or phrase occurred in the tag list, we could trust the corrected phrase as a potentially good tag, and relate the misspelled term to the good tag. Thus, this method help us both filter bad tags (misspelled terms) and find good tags (the spell-corrected term)
- Limitations: The biggest limitation with this methodology was that combinations of correctly spelled words or phrases aren’t necessarily useful for users or the search engine. For example, many of the tags in the database were concatenations of multiple tags where the user space-delimited rather than comma-delimited their submitted tags. Thus, a tag might consist of correctly spelled terms but still be useless in terms of search value. Moreover, there were substantial dictionary limitations, especially with domain names, brand names, and Internet slang. In order to accommodate this, we added a personal dictionary that included a list of the top 10,000 domains according to Quantcast, several thousand brands, and a slang dictionary. While this was helpful, there were still several false recommendations that needed to be handled. For example, we saw "purfect" correct to "perfect," despite being a pop-culture reference for cat images. We also noticed some users reference this saying as "purrfect," "purrrfect," "purrrrfect," "purrfeck," etc. Ultimately, we had to rely on other metrics to determine whether we trusted the misspelling recommendations.
- Method: While a tag might be good in the sense that it is descriptive, we wanted tags that were commercially relevant. Using the estimated cost-per-click of the tag or tag phrase proved useful in making sure that the term could attract buyers, not just visitors.
- Benefits: One of the great features of this methodology is that it tends to have a high signal-to-noise ratio. Most tags that have high CPCs tend to be commercially relevant and searched frequently enough to warrant inclusion as "good tags." In many cases we could feel confident that a tag was good just on this metric alone.
- Limitations: However, the bid value metric comes with some pretty big limitations, too. For starters, Google Keyword Planner’s disambiguation problem is readily apparent. Google combines related keywords together when reporting search volume and CPC data, which means a tag like “facbook” would return the same data as “facebook.” Obviously, we would prefer to map “facbook” to “facebook” rather than keep both tags, so in some cases the CPC metric wasn’t sufficient to identify good tags. A further limitation of the bid value was the difficulty of acquiring CPC data. Google now requires running active Adwords campaigns to get access to CPC value. It is no simple feat to look up 5,000,000 keywords in Google Keyword Planner, even if you have a sufficient account. Luckily, we felt comfortable that historical data would be trustworthy enough, so we didn't need to acquire fresh data.
- Method: Similar to CPC, we could use search volume to determine the potential value of a tag. We had to be careful not to rely on the tag itself, though, since the tag could be so generic that it earns traffic unrelated to the product itself. For example, the tag “USS Yorktown” might get a few hundred searches a month, but “USS Yorktown T-shirt” gets 0. For all of the tags in our index, we tracked down the search volume for the tag plus the product name, in order to make sure we had good estimates of potential product traffic.
- Benefits: Like CPC, this metric did a very good job of consolidating our tag data set to just keywords that were likely to deliver traffic. In the vast majority of cases, if “tag + product” had search volume, we could feel confident that it is a good term.
- Limitations: Unfortunately, this method fell victim to the same disambiguation problem that CPC presents. Because Google groups terms together, it is possible that on some occasions two tags will be given the same metrics. For example: “pontoons boat,” “pontoonboat,” “pontoon boats,” “pontoon boat,” “pontoon boating,” and “pontoons boats” were in the same traffic volume group which also included tags like “yacht” and “yachts.” Moreover, there is no accounting for keyword difficulty in this metric. Some tags, when combined with product types, produce keywords that receive substantial traffic but will always be out of reach for a templated tag page.
- Method: This method was a no-brainer: protect the tags that already receive traffic from Google. We exported all of the tags from Google Analytics that had received search traffic from Google in the last 12 months. Generally speaking, this should be a fairly safe list of terms.
- Benefits: When doing experimental work with a client, it is always nice to be able to give them a scenario that almost guarantees improvement. Because we were able to protect tags that already receive traffic by labeling them as good (in the vast majority of cases), we could ensure that the client had a high probability of profiting from the changes we made and minimal risk of any traffic loss.
- Limitations: Unfortunately, even this method wasn’t perfect. If a product (or set of products) with high enough authority included a poor variation of a tag, then the bad variant would rank and receive traffic. We had to use other strategies to verify our selections from this method and devise a method to encourage a tag swap in the index for the correct version of a term.
- Description: The frequency with which a tag was used on the site was often a strong signal that we could trust the tag, especially when compared with other similar tags. By counting the number of times each tag was used on the site, we could bias our final set of trusted tags in favor of these more popular terms.
- Benefits: This was a great tie-breaker metric when we had two tags that were very similar but needed to choose just one. For example, sometimes two variants of a phrase were completely acceptable (such as a version with and without a hyphen). We could simply defer to the one with a higher tag count.
- Limitations: The clear limitation of tag frequency is that many of the most frequent tags were too generic to be useful. The tag “blue” isn’t particularly useful when it just helps people find “blue t-shirts.” The term is too generic and too competitive to warrant inclusion. Additionally, the inclusion of too broad of a tag would simply create a very large crawl vs. traffic-potential ratio. A common tag will have hundreds if not thousands of matching products, creating many pages of products for the single tag. If a tag produces 50 paginated product listings, but only has the potential to drive 10 visitors a year, it might not be worth it.
- Method: Stemming is a method used to identify the root word from a tag by scanning the word right to left and using various pattern matching rules to remove characters (suffixes) until you arrive at the word’s stem. There are a couple of popular stemmers available, but we found Porter stemming to be more accurate as a tool for seeing alternative word forms. You can geek out by looking at the Porter stemming algorithm in Snowball here, or you can play with a JS version here.
- Benefits: Plural and possessive terms can be grouped by their stem for further analysis. Running Porter stemming on the terms “pony” and “ponies” will return “poni” as the stem, which can then be used to group terms for further analysis. You can also run Porter stemming on phrases. For example, “boating accident,” “boat accidents,” “boating accidents,” etc. share the stem “boat accid.” This can be a crude and quick method for grouping variations. Porter stemming also is able to clean text more kindly, where others stemmers can be too aggressive for our efforts; e.g., Lancaster stemmer reduces "woman" to "wom," while Porter stemmer leaves it as "woman."
- Limitations: Stemming is intended for finding a common root for terms and phrases, and does not create any type of indication as to the proper form of a term. The Porter stemming method applies a fixed set of rules to the English language by blanket removing trailing “s,” “e,” “ance,” “ing,” and similar word endings to try and find the stem. For this to work well, you have to have all of the correct rules (and exceptions) in place to get the correct stems in all cases. This can be particularly problematic with words that end in S but are not plural, like "billiards" or "Brussels." Additionally, this method does not help with mapping related terms such as “boat crash,” “crashed boat,” “boat accident,” etc. which would stem to “boat crash,” “crash boat,” and “boat acci."
- Method: Lemmatization works similarly to stemming. However, instead of using a rule set for editing words by removing letters to arrive at a stem, lemmatization attempts to map the term to its most simple dictionary form, such as WordNet, and return a canonical “lemma” of the word. A crude way to think about lemmatization is just simplifying a word. Here’s an API to check out.
- Benefits: This method often works better than stemming. Terms like “ship,” “shipped,” and “ships” are all mapped to “ship” by this method, while “shipping” or “shipper,” which are terms that have distinct meaning despite the same stem, are retained. You can create an array of “lemma” from phrases which can be compared to other phrases resolving word order issues. This proved to be a more reliable method for grouping variations than stemming.
- Limitations: As with many of the methods, context for mapping related terms can be difficult. Lemmatization can provide better filters for context, but to do so generally relies on identifying the word form (noun, adjective, etc) to appropriately map to a root term. Given the inconsistency of the user-generated content, it is inaccurate to assume all words are in adjective form (describing a product), or noun form (the product itself). This inconsistency can present wild results. For example, “strip socks” could be intended as as a tag for socks with a strip of color on them, such as as “striped socks,” or it could be “stripper socks” or some other leggings that would be a match only found if there other products and tags to compare for context. Additionally, it doesn’t create associations between all related words, just textual derivatives, so you are still seeking out a canonical between mailman, courier, shipper, etc.
- Method: The Jaccard index is a similarity coefficient measured by Intersection over Union. Now, don’t run off just yet, it is actually quite straightforward.
Imagine you had two piles with 3 marbles in each: Red, Green, and Blue in the first, Red, Green and Yellow in the second. The “Intersection” of these two piles would be Red and Green, since both piles have those two colors. The “Union” would be Red, Green, Blue and Yellow, since that is the complete list of all the colors. The Jaccard index would be 2 (Red and Green) divided by 4 (Red, Green, Blue, and Yellow). Thus, the Jaccard index of these two piles would be .5. The higher the Jaccard index, the more similar the two sets.
So what does this have to do with tags? Well, imagine we have two tags: “ocean” and “sea.” We can get a list of all of the products that have the tag “ocean” and “sea.” Finally, we get the Jaccard index of those two sets. The higher the score, the more related they are. Perhaps we find that 70% of the products with the tag “ocean” also have the tag “sea”; we now know that the two are fairly well-related. However, when we run the same measurement to compare “basement” or “casement,” we find that they only have a Jaccard index of .02. Even though they are very similar in terms of characters, they mean quite different things. We can rule out mapping the two terms together.
- Benefits: The greatest benefit of using the Jaccard index is that it allows us to find highly related tags which may have absolutely no textual characteristics in common, and are more likely to have an overly similar or duplicate results set. While most of the the metrics we have considered so far help us find “good” or “bad” tags, the Jaccard index helps us find “related” tags without having to do any complex machine learning.
- Limitations: While certainly useful, the Jaccard index methodology has its own problems. The biggest issue we ran into had to do with tags that were used together nearly all the time but weren’t substitutes of one another. For example, consider the tags “babe ruth” and his nickname, “sultan of swat.” The latter tag only occurred on products which also had the “babe ruth” tag (since this was one of his nicknames), so they had quite a high Jaccard index. However, Google doesn't map these two terms together in search, so we would prefer to keep the nickname and not simply redirect it to "babe ruth." We needed to dig deeper if we were to determine when we should keep both tags or when we should redirect one to another. As a standalone, this method also was not sufficient at identifying cases where a user consistently misspelled tags or used incorrect syntax, as their products would essentially be orphans without “union.”
- Method: There are several edit distance and string similarity metrics that we used throughout this process. Edit Distance is simply some measurement of how difficult it is to change one word to another. For example, the most basic edit distance metric, Levenshtein distance, between "Russ Jones" and "Russell Jones" is 3 (you have to add "E","L", and "L" to transform Russ to Russell). This can be used to help us find similar words and phrases. In our case, we used a particular edit distance measure called "Jaro-Winkler distance" which gives higher precedence to words and phrases that are similar at the beginning. For example, "Baseball" would be closer to "Baseballer" than to “Basketball” because the differences are at the very end of the term.
- Benefits: Edit distance metrics helped us find many very similar variants of tags, especially when the variants were not necessarily misspellings. This was particularly valuable when used in conjunction with the Jaccard index metrics, because we could apply a character-level metric on top of a character-agnostic metric (i.e. one that cares about the letters in the tag and one that doesn’t).
- Limitations: Edit distance metrics can be kind of stupid. According to Jaro-Winkler distance, "Baseball” and "Basketball" are far more related to one another than "Baseball" and "Pitcher" or "Catcher." "Round" and "Circle" have a horrible edit distance metric, while "Round" and "Pound" look very similar. Edit distance simply cannot be used in isolation to find similar tags.
- Method: While Google's choice to combine similar keywords in Keyword Planner has been problematic for predicting traffic, it has actually offered us a new method to identify highly related terms. Whenever two tags share identical metrics from Google Keyword Planner (average monthly traffic, historical traffic, CPC, and competition), we can conclude that there is an increased chance the two are related to one another.
- Benefits: This method is extremely useful for acronyms (which are particularly difficult to detect). While Google groups together COO and Chief Operating Officer, you can imagine that standard methods like those outlined above might have problems detecting the relationship.
- Limitations: The greatest drawback for this methodology was that it created numerous false positives among less popular terms. There are just too many keywords which have an annual search volume average of 10, are searched 10 times monthly, and have a CPC and competition of 0. Thus, we had to limit the use of this methodology to more popular terms where there were only a handful of matches.
- Benefits: When a tag could be mapped to a Wikipedia entry, this method proved to be a highly effective at providing validation that a tag had potential value, or creating a point of reference for related tags. If the Wikipedia community felt a tag or tag phrase was important enough to have an article dedicated to it, then the tag was more likely to be a valuable term vs. random entry or keyword stuffing by the user. Further, the methodology allows for grouping related terms without any bias on word order. Doing a search on Wikipedia creates a search results page (“pontoon boats”), or redirects you to a correction of the article (“disneyworld” becomes “Walt Disney World”). Wikipedia also tends to have entries for some pop culture references, so things that would get flagged as a misspelling, such as "lolcats," can be vindicated by the existence of a matching Wikipedia article.
- Limitations: While Wikipedia is effective at delivering a consistent formal tag for disambiguation, it can at times be more sterile than user-friendly. This can run counter to other signals such as CPC or traffic volume methods. For example, "pontoon boats" becomes "Pontoon (Boat)", or “Lily” becomes "lilium." All signals indicate the former case as the most popular, but Wikipedia disambiguation suggests the latter to be the correct usage. Wikipedia also contains entries for very broad terms, like each number, year, letter, etc. so simply applying a rule that any Wikipedia article is an allowed tag would continue to contribute to tag sprawl problems.
- Method: Finally, we attempted to transform the tags into a subset of more meaningful tags using word embeddings and k-means clustering. Generally, the process involved transforming the tags into tokens (individual words), then refining by part-of-speech (noun, verb, adjective), and finally lemmatizing the tokens ("blue shirts" becomes "blue shirt"). From there, we transformed all the tokens into a custom Word2Vec embedding model based on adding the vectors of each resulting token array. We created a label array and a vector array of each tag in the dataset, then ran k-means with 10 percent of the total count of the tags as the value for number of centroids. At first we tested on 30,000 tags and obtained reasonable results.
Once k-means had completed, we pulled all of the centroids and obtained their nearest relative from the custom Word2Vec model, then we assigned the tags to their centroid category in the main dataset.
[('beach', 'NN'), ('photographs', 'NN')]
[('seaside', 'NN'), ('photographs', 'NN')]
[('coastal', 'JJ'), ('photographs', 'NN')]
[('seaside', 'NN'), ('photographs', 'NN')]
[('seaside', 'NN'), ('posters', 'NNS')]
[('coast', 'NN'), ('photographs', 'NN')]
[('beach', 'NN'), ('photos', 'NNS')]
The Categorization column above was the centroid selected by Kmeans. Notice how it handled the matching of "seaside" to "beach" and "coastal" to "beach."
- Benefits: This method seemed to do a good job of finding associations between the tags and their categories that were more semantic than character-driven. "Blue shirt" might be matched to "clothing." This was obviously not possible without the semantic relationships found within the vector space.
- Limitations: Ultimately, the chief limitation that we encountered was trying to run k-means on the full two million tags while ending up with 200,000 categories (centroids). Sklearn for Python allows for multiple concurrent jobs, but only across the initialization of the centroids, which in this case was 11 — meaning that even if you ran on a 60-core processor, the number of concurrent jobs was limited by the number of initialization, which in this case, was again 11. We tried PCA (principal component analysis) to reduce the vector sizes (300 to 10) but the results were overall poor. Finally, because embeddings are generally built based on probabilistic closeness of terms in the corpus on which they were trained, there were matches that you could understand why they matched, but would obviously not have been the correct category (eg "19th century art" was picked as a category for "18th century art"). Finally, context matters and the word embeddings obviously suffer from understanding the difference between "duck" (the animal) and "duck" (the action).
Bringing it all together
Using a combination of the methods above, we were able to develop a series of methodology confidence scores that could be applied to any tag in our dataset, generating a heuristic for how to consider each tag going forward. These were case-level strategies to determine the appropriate methodology. We denoted these as follows:
- Good Tags: This mostly started as our “do not touch” list of terms which already received traffic from Google. After some confirmation exercises, the list was expanded to include unique terms with rankings potential, commercial appeal, and unique product sets to deliver to customers. For example, a heuristic for this category might look like this:
- If tag is identical to Wikipedia entry and
- Tag + product has estimated search traffic and
- Tag has CPC value then
- Mark as "Good Tag"
- Okay Tags: This represents terms that we would like to retain associated with products and their descriptions, as they could be used within the site to add context to a page, but do not warrant their own indexable space. These tags were mapped to be redirected or canonicaled to a “master,” but still included on a page for topical relevancy, natural language queries, long-tail searches, etc. For example, a heuristic for this category might look like this:
- If tag is identical to Wikipedia entry but
- Tag + product has no search volume
- Vector tag matches a "Good Tag"
- Mark as "Okay Tag" and redirect to "Good Tag"
- Bad Tags to Remap: This grouping represents bad tags that were mapped to a replacement. These tags would literally be deleted and replaced with a corrected version. These were most often misspellings or terms discovered through stemming/lemmatization/etc. where a dominant replacement was identified. For example, a heuristic for this category might look like this:
- If tag is not identical to either Wikipedia or vector space and
- Tag + product has no search volume
- Tag has no volume
- Tag Wikipedia entry matches a "Good Tag"
- Mark as "Bad Tag to Remap"
- Bad Tags to Remove: These are tags that were flagged as bad tags that could not be related to a good tag. Essentially, these needed to be removed from our database completely. This final group represented the worst of the worst in the sense that the existence of the tag would likely be considered a negative indicator of site quality. Considerations were made for character length of tags, lack of Wikipedia entries, inability to map to word vectors, no previous traffic, no predicted traffic or CPC value, etc. In many cases, these were nonsense phrases.
All together, we were able to reduce the number of tags by 87.5%, consolidating the site down to a reasonable, targeted, and useful set of tags which properly organized the corpus without wasting either crawl budget or limiting user engagement.
Conclusions: Advanced white hat SEO
It was nearly nine years ago that a well-known black hat SEO called out white hat SEO as being simple, stale, and bereft of innovation. He claimed that "advanced white hat SEO" was an oxymoron — it simply did not exist. I was proud at the time to respond to his claims with a technique Hive Digital was using which I called "Second Page Poaching." It was a great technique, but it paled in comparison to the sophistication of methods we now see today. I never envisioned either the depth or breadth of technical proficiency which would develop within the white hat SEO community for dealing with unique but persistent problems facing webmasters.
I sincerely doubt most of the readers here will have the specific tag sprawl problem described above. I'd be lucky if even a few of you have run into it. What I hope is that this post might disabuse us of any caricatures of white hat SEO as facile or stagnant and inspire those in our space to their best work.
Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don't have time to hunt down but want to read!