When you perform a search on Google, you’ll often see “sponsored links” at the top and to the right of your search results. Advertisers can bid to have their advertisements appear with search results through Google’s Adwords program.
When an advertiser creates an ad for the Adwords program, they attempt to choose the keywords that their ads may appear beside in search results. The premise behind this approach is that allowing advertisers to target keywords in searches that are relevant to what those advertisers offer means that the ads searchers see will be relevant for what searchers are looking for.
Usually, those ads will show up in response to a current query that a searcher has typed into the search engine, but if you’ve been performing a number of searches, Google has sometimes looked at your earlier queries in addition to your current one to determine which advertisments to show you.
For example, you search for the word [golf] and received a set of search results, along with some sponsored links. If you then search for the word [shoes], you may have seen ads (in the past) on the search results page for “golfing shoes.”
Google recently published a patent application that describes how they might use a previous query to influence which advertisements show up during a current search.
I believe the first time that I saw this happen was a couple of summers ago, but I’m not seeing ads that appear to have been influenced by more than one query the past couple of days. Or at least I didn’t until I was writing this post, and came across a more subtle variation, described below.
Using Previous User Search Query To Target Advertisements
Invented by Mayur Datar and Roberto J. Bayardo
Assigned to Google
US Patent Application 20090006207
Published January 1, 2009
Filed June 25, 2008
Abstract
A system and method to target advertisements to a user search query using a previously entered user search query as well as the current search query. The previously entered search query can accompanying the current search query in a query referrer, or can be determined separately from the current search query.
The relevance of the previous search query is determined concerning the current search query, and if found relevant a combination of the two query terms is used to find a set of advertising keywords from which candidate advertisements to be targeted and served to the user with search results are found. The process can be extended to additional search queries.
The patent application goes into a fair amount of detail on how they might decide whether a current query and a previous query are related, and whether or not they should serve ads related to both queries.
It also describes how they might expand a current search query to match up with potential advertising keywords when there isn’t an earlier query from a searcher. In that instance, they might look at “related words,” or perform something known as “query broadening.”
Related words
Ads could be shown from words that are related to a query, and are found through mining information from term clustering tools, synonym dictionaries, and machine learning systems.
For example, the words “cars” and “autos” might be considered to be related to the word “car.” If a searcher types in a search for the term “ford car,” advertising keywords for the equivalent queries “ford cars” and “ford autos” might be used to decide which ads to show searchers.
The patent application tells us that they might use the process described in an earlier Google patent filing to find these related terms: Increasing the number of relevant advertisements using a relaxed match.
Query broadening
In addition to finding related words, and substituting those words into a query phrase to find keywords to target advertising, related queries might also be used.
For example, the query term “Disneyworld Florida” might be considered to be a related query to “theme parks,” and if someone searches for “theme parks,” or a slight variation of that term, the advertising system might use the related query “Disneyworld Florida” as an advertising keyword for targeting ads.
This kind of query broadening is described in another Google patent application: Identifying Inadequate Search Content.
These “related words” and “query broadening” methods gather information about the relationships between query terms and between whole queries from places like the search query log files of the search engines, and other data sources. But, looking at a current query, and the previous query used by a searcher during an immediate search session might add some additional context regarding what someone is searching for.
Techniques for targeting ads and/or finding ad keywords for targeting ads based on a previous query
The patent application describes additional techniques that it might use separately, or together, to find keywords that it might use to target ads shown for a search when looking at a current query and a previous query.
Overlapping words
When there are words that are common to previous query and current query, it might be helpful to keep those common overlapping words in at least some of the query broadening and related word sets.
For example, someone searches for “hotels in Manhattan” and then searches for “cheap hotels.” The term “hotels” is an overlapping term common to both of those queries, and it might make sense to include the word “hotels” when broadening a query to determine which ads to show a searcher.
The advertising system might keep the word “hotels” as one of the advertising keywords used to determine which ads to show since it appears to be important to what the searcher is looking for.
Query term summation
Words might be added from a previous query to come up with ads for a current query.
I followed up my searches for “hotels in Manhattan” and then “cheap hotels” with a search for “cheap manhattan.” and received one sponsored link – for:
Manhattan Hotels from $80
Get Cheap Manhattan Hotel Rates.
Amazing Discounts. Book Now!
It appears that Google added “hotels” to the keywords used to decide which advertisement to show me alongside my search for “cheap manhattan.”
Conclusion
While I was reading this patent filing, I couldn’t help but wonder if the “overlapping words” and “query term summation” processes might also be applied to the results that we see for normal web searches in addition to being a method used to come up with advertisements shown to searchers.
The query broadening patent filing, Identifying Inadequate Search Content examines how search results might sometimes be expanded during a Web search when there aren’t many results for a query.
Might Google also sometimes look at overlapping words in queries used during the same search session to show results that are more relevant for words that appear in more than one query from the same searcher? Or add words from previous searches in a search session, as it does in query terms summations, to determine which search results you see?
It looks like something that they could do.
Added 2009-1-9 – Great post from Jeremy Chatfield at the Merjis Internet Marketing Blog from last March which approaches this topic from a paid search perspective, and is highly recommended – Is AdWords Search History Permutation Fraudulent?
Hi Jeremy,
I like your phrase “search history permutation” to describe how previous queries in a search session might influence which sponsored links might be shown to searchers, and your blog post raises some great points about the practice. I’d definitely recommend that anyone who reads this post visit your post on the topic (and I’m going to add a link to your post at the bottom of mine, above.)
I’ve seen something else also, which might influence which advertising keywords might be used to generate sponsored links for a set of results, and I’m not going to give it away here until you’ve written up the results of your investigation. I’m wondering if we have both have the same suspicion.
I’ve been hoping that Google would publish a patent filing on this topic.
HI Bill – thanks. The issue that I found was related to how Google decides which keywords receive impressions. I suspect that whatever you’ve found is nothing to do with Neural Network Training Sets 🙂
Go ahead and publish your insights. Please. I’m sure you’ll give me something new to think about!
Hi Jeremy,
I like a good Neural Network Training Set myself sometimes. 🙂
I noticed that when I see a message in the top right of the search results page along the lines of:
that one or more of the sponsored links shown may focus upon that location. So my search for [convention center] included this sponsored link result:
Convention Center
Check Out Local.com To Find
Convention Center In Your Area!
Local.com
Philadelphia, PA
Did that sponsored link appear because of search history permutation, since my previous queries included “philadelphia,” or was it because of a Results based personalization of advertisements in a search engine?
I wasn’t signed in to Google personalized search, but I’ve written recently about How Google Might Personalize Search Results Outside of Personalized Search.
So, did I see that sponsored result based upon directly upon the use of a previous query, or based upon a personalized profile delivering me an advertisement?
Hi Bill, Google’ve already been using search history permutation. I was wondering why some of my clients were getting some strange search results, about this time last year. See link no longer available for some details. After more investigation, the *main* cause of problems was something else, which I’ve yet to write up… Search History Permutation, at least as applied to paid search, can produce some dreadful advert matches.
Hadn’t come across the patent, though. Thanks!
Hi Dave,
Thank you. Quiet night at home on a Friday night here, too.
Interesting to consider how previous queries might also impact organic searches. The examples are helpful. Kind of fun to try out a number of different queries, in a couple of different browsers, and see the differences in the sponsored links shown as well.
Hi Bill… just me out doing some reading on a Friday night (yea, wild man… I know). I was hesitant at first as an organic search geek, but the query analysis is always interesting and drew me in. There’s always signals to be mined there as you hinted, for organic as well as ad serving. I also like when the provide the examples, somehow they seem to help paint the picture.
Well, off to read some more… just wanted to see what U’ve been up to.
Thanks for the post… talk to U soon.
Hi Bill – I to do an approximately annual set of experiments on geolocational targeting in paid search. I’ve not done the latest round yet. I do pay some attention to organic results – we do a little SEO as well as paid search – but I can’t control organic results as much as I can control paid search; the research is harder to do, as it involves placing specific search queries from and to specific locations to look for evidence of specific results. I haven’t done all the research for this years’ posting on geolocation. That said, here’s some observations, mostly drawn from PPC.
Google definitely – as I think you’ve said before – use Search Query Parsing to break the search query into up into fragments that allow them to specialise the search. Geolocation being a key component. Location seems to be economically important – most economists still show that most discretionary spend is still mostly local in most economies, despite the rise of the internet. Respecting location is probably a good move by Google. 🙂
When I investigated, a year ago, Google used only a single previous query to affect the current paid search response. That is, if the previous query was “dentist philadelphia” and the current search was “tax refund”, then Google would synthesise “dentist refund”, “dentist tax”, “tax philadelphia”, “philadelphia refund”. However, if you repeat “tax refund”, or add a third search that doesn’t feature “philadelphia” or “dentist”, then those synthetic searches would not be used – with new synthetic searches based on the current and a single previous search, again.
Google definitely does use identified locations, and the behaviour appears to have changed over the last year. About 18 months ago, they were not identifiably using local geotargeting over, for example, national borders. That is, if you advertised holidays only to New Yorkers (NY geotarget), then it used to be that your adverts ran only in the USA. They will now show to searchers in the UK, if they type a query like “vacation in new york”. Location is becoming increasingly important to paid search. I don’t, yet, see large changes in organic search results, when I’m *not* signed in to personalized results.
However, that may be partially because most of the locations for internet connectivity that I use in the UK do not have a defined location other than “being in the UK”. When I do use a location that Google have identified down to a city level, I notice that I see more local results.
I think I’m seeing, in the preparatory experiment series, indications that when I use a location that Google’s IP location database has identified to city level, paid search with localisation. If I’m *also* signed in to Google for personalised results, then I think the results are suggesting that localisation is also happening. However, my current experiments are focused on transnational advertising – that looks like a new behaviour.
You like your Neural Network Training Sets well fried, or raw? 😉
Last year I idly wondered about whether or not status messages on facebook, or in recent tweets, might also be used to inform search recommendations:
link no longer available
(and I guess this one also – http://blog.ouseful.info/2008/12/12/arise-ye-databases-of-intention/ )
For example, you could imagine an overlap in search terms and recent tweets allowing additional info from a tweet to augment or inform the search?
never thought about this aspect!
now when i search i hav to keep an eye on it 🙂
Hi Jeremy,
I’ll be looking forward to your post on geolocation when you’re finished with it. Thank you for sharing some of your preliminary observations.
Google does seem to use search query parsing in searches. The inclusion of a specific location does appear to be a pretty strong signal of a searcher’s intent, and the search engines do seem to place a strong reliance on its usage.
The claims section of this patent application does raise the possibility of the use of a third query, even though the description section doesn’t really address that possibility. So, it might make sense to pay attention to more than just the previous query if you are looking at this behavior in more depth.
It makes sense to show geographically related results to people outside of a specific geographical region if their query includes that location. If Google is now doing that, it seems like a step in the right direction.
I’m a big fan of machine training systems regardless of whether they call themselves neural net or not. 🙂
Hi Tony,
Very interesting suggestions. Would Google inform its search system and advertising system of actions taken on social systems like Twitter? I would say that it’s a possibility.
We know that Google keeps an eye on your browsing history, and if you’re signed into personalized search, it gathers a fair amount of information about what you are doing.
One of the patent applications that I ran across over the past year from one of the major search engines (can’t recall which one at this moment) mentioned that they would watch your searches and selections of search results at other search engines. If something like that is being done, then paying attention to status updates on twitter or Facebook isn’t too much of a larger step to take…
Hi aniroy1986,
Sounds good. Let us know if you see something interesting. 🙂
About the “search history permutation”. Does google use all data registered at browsers search history? or do they use the last 10 search query? query within 24 hours? or anything else?
Hi Charles,
The patent application focuses upon the the query immediately before the current query in its description, but raises the possibility that they might use the previous two queries in the claims section of the patent.
This is different than using a person’s search history, or all searches over a period of time, or a string of many queries over a search session. It doesn’t require being logged into a Google Account, or need to call up a searcher’s browsing or search history.
Hi,
I don’t think that this will help to users in all cases.
Bcoz, if user search for query1 and he tries for the second query. As per this patent if the result given result related to the previous query and user is looking for something looking for a different result i.e if he get the source using the first query1 and he is looking for a different source using the quesy2 which is irrelevant to the query1. At this time if user get the same kind of result, he will be frustrated. So this has to be considered for some cases only and not for all the cases of searches.
Hi cdsseo,
You’re right. The patent application does discuss the point you make. If the query terms don’t appear to be related, then it’s quite likely that they won’t let the earlier query influence the sponsored links that are shown with the present (or current) query.
As a consumer, this type of methodology during searches is really annoying. As a marketer, it tends to work to my advantage. For instance, anyone searching for something like “dental continuing education” will probably get shown lots of ads for dentists’ offices, which would have nothing to do with the search. In a way, this just wastes resources but still increases exposure for the business.
I read a while ago that every single site that uses the google adwords program needs a privacy policy and if this becomes regular across the content network then this will be even more the case.
Hi Jimmy,
In Google’s information about Quality Scores, they do mention that they would like to see a privacy policy for an advertiser’s site if it asks for personal information from visitors. If they don’t ask for that kind of information, then one doesn’t seem to be required.