Why Google Just Can’t Suggest

Part 2 – How Google Suggest works

In this part I’ll be explaining the internal structure of how Google Suggest is implemented (free hint; the answer is “badly”) and why Google did it like they did it. First, let me indulge you with a short history lesson. In August 2008, The Google engineering team in Haifa, Israel announced the launch of Google Suggest in the US after a 4 year beta run in Google Labs. Rejoice oh (American) searchers of the world wide web, Google will now help you search like never before. Oh joy.

Global roll out

Around this time Google started gathering search data in order to roll out this feature to every single country where Google search is present. This included the UK, Ireland, France, Japan, Brazil and every other kingdom where Google operates. Soon, everybody would be able to indulge in the same Google awesomeness that our brothers and sisters in the US had been enjoying.

Now, this is a lot of data to gather – so how should Google manage search suggestions for all these countries to provide the best user experience? Well, a structure like this would have been pretty perfect:

In summary, this shows a structure of having a separate search suggestion list for every Google search website. So, you search on Google.es and you get suggestions of queries other Google.es users have been searching for; you search on Google.co.uk and get UK relevant suggestions. Suggestions are weighted based on popularity (that’s what the “pop” arrow is for) and updated periodically to take account of searching trends. Pretty straight forward eh? Indeed, if you’ve not looked at Google Suggest too closely, you might be under the illusion that it works like this. The advantages of a structure like this (as opposed to say, grouping all search suggestions in one big list) are:

1. Lookup efficiency (only searching through words relevant to the target users’ country)
2. No “suggestion bloat” (no possibility of irrelevant search suggestions)
3. Maximise the number of relevant search suggestions available for each country

Now there are downsides to this method; notably in development and maintenance resources. We’ve often heard that Google (that’s the largest Internet brand ever) has limited engineering resources. And yeah, if you consider that Google has a lot of search engine pages in English (including offering a non-default English option in several countries), and quite a few in Spanish and so on, then there would be a lot of search suggestions being stored multiple times internally.

But so what? This is Google people! When you read their core “ten things we know to be true” values, the very 1st sentence on the very 1st bullet point states clearly:

Since the beginning, we’ve focused on providing the best user experience possible.

Phew. I was worried then for a second. I thought the team for Google Suggest might have cut some corners to save resources, but that would definitely not offer the best user experience.

OK, I know sarcasm sometimes doesn’t translate well into the written word, so let me shatter any possible illusions. They didn’t choose a structure that offered the best user experience (at least not for people outside the US). I hope you’re sat down, as what they actually implemented was this:

Yep, to save a bit of development time and maintenance resource, as the Google Suggest team already had a big list of words searched for by Americans on Google.com, they decided it would be a good idea to group suggestions by language rather than by location.

There’s a technical definition for this very structure; it’s called half-assed.

Now, if you’re living in Germany or France for example, the consequences of this structure will be of little concern to you as you’re in by far the dominant country for search volume in your language. However, for internet users in English speaking countries (outside the US), the ramifications of having every English search suggestion available to every English Google search engine are huge. With a badly thought out structure, there is always a likelihood of bias towards the country that provides the most search volume (for English language suggestions, that’s the US of course).

Obviously, the Google Suggest team couldn’t just use global search popularity before launching “localised” Google Suggest (although it wouldn’t surprise me if they considered it) as there would have been an almighty fuss, and rightly so. However, the crappy “one language fits all” structure did leave them with a bit of a quandary of how to provide relevant suggestions in each relevant country. So, how did they get around this?

Predictably, with another half-assed solution. When Google Suggest launched on engines outside the US in March 2009, the accompanying blog post was dominated with the fact that search suggestions would be relevant depending on which country engine you were using to search. Sounds terrific, except all they had done was include an additional weighting system for each country. So, every English language suggestion (of sufficient popularity) had a weighting for Google USA, Google UK, Google Ireland, Google Australia and so on, based on the volume of queries in that particular country. Not so impressive after all.

But does it actually work? Well, let me demonstrate a couple of test searches so you can judge for yourself

The H Test

Let’s try a search for h in various English language Google engines, and see what the results are…

We can see that hotmail is the most popular h search in all countries (go Microsoft) and the rest of the results seem to make sense; home depot shows up for the USA and Canada, halifax in the UK and home and away in Australia (*snigger*). But then, there are also some clearly visible flaws. The term how to have sex’ in Ireland for example. Ask yourself this:

1. Are people in Ireland really searching on Google for guidance on how to have sex?
2. Even if they are, are they really adding an apostrophe to the end of the search term???

The answer is of course not. They’re simply typing h or ho, seeing that result and clicking out of curiosity (curious bunch the Irish). This then maintains the popularity of the search on Google.ie and it remains in the top suggestions indefinitely. Then thanks to the structure I’ve alluded to already, the same search suggestion (complete with apostrophe) appears in every other English language Google search. Quite a flawed system really when you think about it, and quite easy to game as many people have found.

The Hudghton Test

For this test, I’m going to search for my surname. This isn’t some crazy ego trip I promise; it’s actually a handy way to demonstrate the search term popularity weighting effect…

With my surname being quite uncommon, in most countries there is no weighting for any term containing hudghton and they are just listed in a global popularity order. There are various family members of mine on there; my dad is John Hudghton and so on. For this experiment, don’t worry about the top result (as Google will first try and match in the order you type), concentrate on the names in suggestion 2 onwards.

You’ll notice that I’ve not included the Australian results yet and there’s a good reason for that. Max Hudghton is a recently retired Aussie rules football player. Apparently he was better than at least one other player; magic. The reason he makes a good example is this; no one outside of Australia has a “bleedin’ clue struth mate” what Aussie rules football is (me included). If my structure analysis above is correct, Max should be the most highly weighted Hudghton name and thus appear first in the Google Suggest list on Google Australia. Well, let’s see:

Well whad’ya know, there Max is as the first name; what a hero. Incidentally, if you’re wondering if country by country search suggestions vary based on your IP, I can assure you they don’t. I tested with various proxy servers just to make sure. The new even more localised Google Suggest obviously may factor in your IP for different areas of the US, but not top level country versus country suggestions. I’m not going to cover the new localised US searches in this post; suffice to say it’s just a case of building more crap on crap foundations, instead of fixing the core issues. Every suggestion they show in that blog post you can still access from any English language Google engine.

So what’s the problem?

If the weighting system works (to a reasonable extent), then why do I think Google Suggest is so badly implemented? Well, consider the second example in particular:

• No one in Australia (or indeed anywhere outside the UK) is searching for my dad for goodness sake, yet Google Suggest is still showing his name as a valid suggestion. That’s not a good user experience. I’ve imaginatively dubbed this suggestion bloat and it affects a huge volume of searches.
• Google Suggest can only store a finite number of suggestions. The poor implementation means that there may be useful suggestions for Australian or English speaking Hong Kong users for example, that don’t make the cut for the global English suggestion list. Obviously, they would have done if the suggestions lists had been grouped by location instead of language.

But the big problem is neither of those two things, oh no. By far the greatest issue with Google Suggest is what I call suggestion merging.

Eh? What’s Suggestion Merging?

That’s a great question, thanks for asking. Suggestion merging is without doubt the biggest mistake the Google Suggest team have made thus far. And why was it brought about? Through the poor initial implementation of the system as I’ve detailed above.

When you’re creating a system of any size, if you start off with crap foundations, you can only ever build more crap on top of it. If you try and build something good on top of bad foundations, the system will invariably collapse. I mean seriously, this advice is so old it’s in the bible (Matthew 7:24-27 FTW). And yeah, you don’t get many SEO bloggers quoting the bible as a reference I know

The problem the Google Suggest team had, purely because of grouping by language instead of location, is that they now had a huge list of English suggestions. Serving up AJAX suggestions from a list that big is always going to be server intensive, however you split and manage your data. So naturally, they wanted to optimise this list in any way they could, without removing suggestions.

Now that’s all well and good in principle; but how could they do this? Well, I imagine the conversation at Google Haifa went a bit like this (B for boss, E for employee):

B: We have a huge list of English words; how can we optimise them before we launch “localised” Google Suggest?
E: Well, I was reading online and there’s lots of words that are interchangeable
B: Go on..
E: According to xxx website, words containing ise can also be spelt using ize; both forms are correct!
B: Really, that could work..
E: So how about we automatically merge where the volume difference between ise and ize versions is large?
B: Great idea! One question, how do we determine which version to use?
E: By search volume.
B: Of course, but since we group by language, what volume measure should we use to determine this change?
E: ..
E: Well, it would have to be global volume.
B: ..
B: OK, make it so.

I mean seriously, you have GOT to be kidding me. They used global search volume to determine how to merge suggestion terms?

Now look, I’m not saying the conversation was exactly like that, but at some point a ridiculous design decision was made to allow this to happen. But what does it actually mean for you and me, my fellow UK Google searchers? Well you’ll be glad to know this part is almost over and I’m going to explore the consequences of this change in part 3. Just to give you a quick taster, an example that caused a lot of fuss recently is the term search engine optimisation:

This is a term that has had the “suggestion merging” treatment. Google have looked at terms containing search engine optimisation or search engine optimization and just kept the most popular one, based on global search volume. The example above is actually very clear; most of the generic terms have more volume for the “iz” versions (thanks to US search volume). For the UK targeted terms, the “is” version was kept as this is what we search for in the UK and hence had more global volume. Just in case I’ve lost you slightly, take a look at this example too:

And there you can see it again; the most popular global variant is the one that is kept during the merging process. In this case, only UK users are searching for SEO in Manchester, so they’ll naturally use the “English” spelling and this will therefore be the most popular variant. Strangely, in a limited number of cases both variants are kept (search engine optimisation company for example is available as a suggestion in both the ”is” and “iz” forms). A possible reason might be that there wasn’t sufficient global volume difference to merge the terms.

Regardless, the whole thing is bloody ridiculous really. Look at is this way; are more people searching for my dad than search engine optimisation? He might like to think so, but it’s definitely not the case (sorry dad). Yet john hudghton has a search suggestion and search engine optimisation does not. This change has major ramifications for SEOs and ordinary users alike…. I’d suggest you take a deep breath, have a cup of tea and then move on to part 3 where I have plenty of examples

1. Why Super Brands Don’t Care About SEO

Be Sociable, Share!

•