My intent was to write a post about practical application of the theory of sets in PPC. Instead, it turned out to be an essay. That is what you usually get when you let a [former, but who cares?] linguist into search engine marketing. In this essay, I give an overview of match types, how keywords with different match types correlate with different sets of search queries, best practices and how they evolved as match types changed, answers to some practical questions about campaign structure and analysis, and – a typical “I don’t know how to do it right” disclaimer at the end. But I sincerely don’t. The objective of the essay is to make you view keywords and match types from a slightly different prospective. Constructive feedback and suggestions are welcome.
Please note that I am using examples from low-volume high-margin campaigns. Thus, this approach is not universal and can be changed significantly depending on the behavior of a specific PPC campaign.
* * *
After more than 5 years in paid search (that’s one heck of an old-timer, right? But it is something considering how fast PPC platforms evolve and change) –I have finally realized that the beauty of my work is that it makes you think. It’s all about thinking. It’s all about the way you think. You cannot do it mechanically, there are no rules of thumb; best practices become obsolete faster than you learn how to apply them. I am not saying that paid search is the only area that gives your brain a little bit of exercise, but that’s what I do, isn’t it? PPC is about thinking and a hint of math.
As far as the math part is concerned, this post was partially inspired by Mike Nelson’s article, “Applying the Theory of Sets in Match Types,” (Search Engine Land, Oct. 22, 2012). But it was not some kind of Eureka! moment for me. For quite some time before that I could not help imagining the logic behind how search queries are matched to keywords with different match types as some kind of Venn diagrams.
Begin at the Beginning: Match Types Evolution and Best Practices
Every self-respecting PPC marketer can give you a definition of various match types in their sleep. But it was not like that just several years ago. Many PPC marketers would simply use broad match, and only the most sophisticated ones would try phrase and exact.
Initially, broad match, besides the actual words compiling keywords (or “key-phrases”) from your account in any order with any other words before, after, or between them, included the following: singular/plural forms, spelling variations, various verb forms, synonyms. In 2007, Google introduced an improvement to broad match – expanded broad match – combining queries from two successive searches, thus adding previous user search behavior to the mix.
Phrase match would be triggered by the phrase you had in your account, excluding spelling variations, singular/plural, etc. (and account managers would use misspellings generators to compile huge lists for phrase match). Exact… well, it is pretty obvious.
Search queries using matching phrase and exact keywords were considered more relevant, indicating higher intent and thus more valuable, so one of the best practices was to use variations of core keywords in phrase match plus broad match to catch more impressions. After expanded broad had been introduced, some advertisers started to rely solely on phrase and exact to exclude irrelevant searches (read “AdWords Expanded Broad Match: How to Combat Google’s Cash Grab”).
Phrase match for misspells was used in the hope of getting less expensive clicks for keywords that competitors would probably not include in their lists. The more sophisticated the spelling variations lists, the higher the probability of having fewer competitors entering the auction for a given keyword and the lower CPC. Then it was recommended to use all 3 match types allowing Google to match a search query with the closest keyword – that is how it was supposed to work, at least theoretically. The problem is that broad match still includes everything else.
One of the popular ideas was then to replicate a campaign for various match types – so that you would end up with 3 campaigns instead of 1 – and bid higher on more precise match types. The latter one seems a little odd, because replicating campaigns would not give you much… and this is when the theory of sets comes into play for the first time.
Query Sets and Recent “Improvements” of Match Types
Broad match keywords are matched to a set of queries. Suppose the number of queries is not infinite (which, unfortunately, it is – I have seen 29 variations of spelling an 8-character word). In order to “replicate” a broad match campaign, you will have to include more phrase match keywords than you have in broad match to cover the same query set. And even more ones in exact match. Otherwise you are not exactly “replicating” a campaign: instead, you create three different campaigns that cover different query sets.
However, since the number of queries that can trigger a broad match keyword is infinite at the end of the day, it would make more sense to come up with phrase and exact matches for only “core” keywords. It is not even necessary to separate them into different campaigns. Keyword-level bidding might solve the problem.
In 2010, modified broad match type was introduced to combine the flexibility and a pretty broad reach of – sorry for the tautology –broad, with the relative precision of phrase. Modified broad keyword would be matched to a set of queries containing close variations of whatever you have, in any order, plus any other words. This was definitely a more limited set because it did not include synonyms. You can use as many plusses as you want to limit the set of search queries and make your targeting rather fine-grain.
The most recent change to match types was announced in April 2012: now phrase and exact match include close variants and different grammatical forms – so-called “near match”. Which, if you think in terms of sets of queries that they can be matched to, changes the latter ones.
Going back to best practices. The last best practice we discussed is having all match types in hopes that Google will use the most precise match to map search queries to your keywords. The problem is that it does not always work this way. If a search query can be matched to a phrase match keyword, it can also trigger a broad match one. Most likely, Google will choose the keyword with the highest bid. I personally feel hesitant about paying more when I could have paid less.
Besides, by running different match types simultaneously we end up with overlapping query sets. Some queries can be matched to any match type: if an exact match is triggered, why wouldn’t the same keyword in broad be? It is pretty random (if we cannot track the way Google does it and can only guess). And if we still get everything we could have gotten with just broad or broad modified – do we even need phrase and exact?
Overlapping Query Sets: What to Do
One of the ways to go about overlapping query sets is to select only one match type. The top performing one. The problems I see here are the following.
Problem 1. The experiment might not be valid, especially if you bid differently on different match types. Besides, with overlapping query sets what if your phrase match keyword gets only half of what it could have got while the other half is mapped to broad match?
Problem 2. The definition of “top performing” is unclear. Is it the amount of conversions (whatever you consider that to be: a lead, a quality lead, or a signed case)? Conversion rate? Cost per conversion? CTR? CPC? Cost per conversion seems to be the sanest criteria, but it is virtually impossible to use it in the sparse data sets that we usually have in our low-volume, high-margin accounts.
Problem 3. Keyword-to-phone call tracking that we have does not capture match type, which in our case means we lose over half of the data. The first problem can be solved by isolating different match types with the help of negative keywords: add phrase and exact to a broad match ad group as negatives, add exact as a negative to a phrase match ad group.
My issue with this approach is that it may artificially decrease average account CTR (and quality score) by not allowing broad match keywords that generate the most impressions to be triggered by relevant search queries. Quality score is supposed to be calculated based on impressions where search query is identical to a keyword, and you are going to completely eliminate the chance of them being identical with phrase and exact negatives. Besides, in the case of single-word keywords, what would they be matched to in broad match if we add a phrase negative? Suppose we have an ad group that contains +injury and -“injury” – doesn’t it create some kind of conflict?
The second problem… I am not exactly sure how to approach KPIs here. With insufficient conversion data, I am personally inclined to look at CPC hoping to decrease overall cost and come up with some other ways to improve conversion rates. No one cancelled ad text and landing page testing.
As for the third problem, it can theoretically be solved by using different phone numbers for different match types to track them separately.
But even if we manage to set up a perfect experiment there is still one more problem: are the results actionable? What if you have – again – a low-volume, high-margin campaign where every conversion counts? Are you still going to get rid of whatever brings more expensive conversions (but does bring some)? If broad match that performs worse in terms of CPA than, let’s say, phrase brings most conversions – will you only be running on phrase?
At this point I am almost ready to start with modified broad and a lot of negative keywords, add phrase match keywords based on search queries, compare CPC, and pause whatever turns out to be more expensive. Or begin with phrase and exact match for whatever we know performs well historically in terms of the quality of traffic and brings conversions — and use modified broad match for everything else while avoiding overlapping sets. Frankly, I do not believe there is only one best way to set up a campaign: too much depends on historical data, potential amount of traffic, goals, budgets, etc.
Confused? Let’s Look at the Numbers
Let me share the results of a little experiment. We have campaigns with all three match types and overlapping sets of search terms. I decided to look at what overlaps what to see if we can eliminate at least some of the overlapping sets.
The first campaign I looked at was a branded one, with relatively high avg. quality score. The chart below shows how most of impressions come from search terms matched to all 3 match types; at the same time, a significant amount of impressions (over 12%) comes from exact match keywords. Phrase match contribution is less significant. In this case (and look at CPCs!) it might make sense to use exact for core keywords (probably even opt out of the near match) and modified broad to capture everything else.
The second chart is for a non-branded campaign, with rather low avg. quality score.
It is interesting that there are no “non-duplicate” impressions for exact here at all. It does not mean the latter one did not generate any; it only means that the same keywords triggered different match types. Randomly. In this case, I am almost inclined to think that we could start with just modified broad, gradually adding exact (and corresponding exact negatives) and carefully tracking CPC and other available KPIs.
Needless to say, this chart does not give a complete picture because the campaign analyzed here is literally built around one single-word keyword. The relationship between sets is, in fact, more complex. The reason is that, for example, single-word search queries can be matched to a whole bunch of single-word keywords and longer ones containing the same component. In this case, it looks like phrase match of a single-word keyword is the worst thing you can do because it “blocks” more specific keywords. (E.g. injury lawyer can be matched both to “injury lawyer”– a more specific keyword and “injury”– a phrase match keyword that is too broad in this case). Besides, CPC for phrase looks pretty discouraging.
Here is a distribution of unique terms for the campaigns mentioned above.
The chart above is for a branded campaign, the chart below is for non-branded one. Search query distribution is pretty interesting (both campaigns are set not to use near match).
Did I manage to make it all even more confusing? Here’s an example of suggested sample structure. Instead of an ad group containing +injury and [injury] (overlapping sets) it suggests +injurywith exact negatives for “core keywords” and separate ad groups for the latter ones in exact match. I am not saying it is a good (or bad) structure; it just helps to avoid overlaps.
Conclusion: What’s the Plan?
Actually, forget about your keywords for a minute. Think about what people type in and what you want to capture in that ocean of queries. Your keywords are just criteria you set for the traffic that might be relevant to what you are advertising, and not exactly what will be typed into a search box. Keeping in mind that every keyword can be matched to a set of queries (with near match it can), think of cutting off the unnecessary and limiting these sets without missing the traffic you really want.
People will spell any way they want; they will come up with weird search queries and tell a search engine what they want to find in illogical and unpredictable ways. According to Google (“Are Search Queries Becoming Even More Unique? Statistics from Google”) 20% of search queries are unique and 70% do not have exact match keywords. This data is over 2 years old, but I do not believe that there have been dramatic changes since then. (I will be grateful for fresh statistics, though). What I want to say is this: you are not dictating anything. You are just carefully choosing what you let Google use to trigger your ads. You can control the overall theme and, to some extent, the degree of precision.
The last thing I want to touch upon is keyword performance analysis. If we make numerous changes to our lists of keywords and match types, how do we analyze the performance? For example, is keyword data before the changes still valid and useful? Or should we consider the campaign we end up with after our “optimization efforts” different? It may have a new structure, different approach to match types and thus different list of keywords, plus sometimes new ads, maybe even new landing pages. Will data for keyword analysis before changes be as relevant as after?
I believe that it will. First, the ads will still be triggered by much the same set of core queries – unless we change the theme of the campaign completely. Second, we do not have the luxury of slicing data too much. On the other hand, there is no reason why a keyword that converted well several months ago would stop doing so now, regardless of campaign structure. It is still matched to the same set of queries, isn’t it?
So, what do we do? What’s the plan? I do not know. I have just tried to share my ideas and provoke some thinking to shift the focus from keywords as… well, keywords, to them being merely criteria limiting sets of search queries. Please do not take what I have written for granted. Actually, do not take anything for granted, especially in PPC. As a friend of mine – a mathematician, by the way – says, “It is always true; except for when it isn’t”.
