Issue Description
I use cognitive services TranslateArray to translate my users comments. One of the advantages of this service is that we can use ProfanityAction to mark every profane words in the destination language. I also make use of the automatic language detection, so that I do not have to identify the content before sending it in.
When I get my translation back for a destination language which match the source language, the profanity is not marked. Is there another endpoint I could/should hit, or a parameter I do not know about, or is there a possible improvement of the service ?
Corresponding Documentation
Follow the cognitive service protocol to hit the TranslateArray endpoint, with an english sentence containing profanities, with the ProfanityAction: Marked behavior: http://docs.microsofttranslator.com/text-translate.html#!/default/post_TranslateArray
Reproduction Steps
Send an English sentence with profanities
Translate to fr, notice correctly marked profanities
Translate to en, notice the missing profanities tag
Expected Behavior
Profanities should be marked even if no translation occured.
Actual Results
I obtained the unmodified sentence back.
There is nothing in the documentation that specifies what happens if the source and target language are the same. My guess is that if it sees that they match it will do nothing.
However, there is a specific API that detects profanity for any given language: Content Moderation for Text. The API docs are here.
The Text - Screen function does it all – scans the incoming text (maximum 1024 characters) for profanity, autocorrects text, and extracts Personally Identifiable Information (PII), all while matching against custom lists of terms.
Your observation that Translator API does nothing if source and target languages are the same, is correct. Not an answer, just clarification.
I need to set up A/B testing for my homepage using Google Experiments. Thing is, my page serves 2 languages - English and German.
The path to homepage itself is simply / for EN, and /de for DE.
Then, I have the following variations of the homepage prepared:
/experiment-A and /de/experiment-A
/experiment-B and /de/experiment-B
The goal for the experiment is user's registration. I don't care which language the user has set. The question is: Do I need to set up 2 separate experiments for the two language versions?
Yeah, I'd set up two experiments. The plus side is you'll also be able to get some insights from it. I find Germans behave quite differently than our North American visitors.
google has an API for downloading search suggestions:
https://www.google.com/support/enterprise/static/gsa/docs/admin/70/gsa_doc_set/xml_reference/query_suggestion.html
unfortunately, as far as i can tell, these results are specific to your location. for an analysis, i would like to be able to define the city/location that google thinks it is making the suggestion to. here's what happens when i scrape from dar es salaam, tanzania:
http://suggestqueries.google.com/complete/search?client=firefox&q=insurance
["insurance",["insurance","insurance companies in tanzania","insurance group of tanzania","insurance principles","insurance act","insurance policy","insurance act tanzania","insurance act 2009","insurance definition","insurance industry in tanzania"]]
i understand that a vpn would partially solve this issue, but only by giving me a different location and not lots of locations. is there a reasonable way to replicate this sort of thing quickly and easily from, say, the 100 largest cities in the united states?
confirmation that results differ within the usa-
thanks!
Google will use your IP and your location history (if turned on) to determine your location.
To be able to go around it, you can spoof your IP while logged off your google account (but I don't know if google will consider it a trial of hacking no matter what your intentions are).
Another way is to use Tor browser (even though it is not it's original purpose). You can configure tor to exit from a certain country using the Exitnode parameter in the torrc config file
As found in the docs:
ExitNodes node,node,…
A list of identity fingerprints, country codes, and address patterns of nodes to use as exit node
But if you want a fast way to do it, I don't think that's possible since google wants to know the real location of the users and have put a lot of effort into making such tricks fail.
The hl param for interface language changes the search results, but I can't tell if it's actually changing the location. For example:
http://suggestqueries.google.com/complete/search?client=chrome&q=why&hl=FR
Here's an example with 5 different values of hl:
http://jsbin.com/tusacufaza/edit?js,output
We have an interesting 'bug' in our JS code that only fires when a user agent has a specific combination of parameters -- specifically when IE8 sends both Trident/4.0 and MSIE 6.0;
We have checked the GA export data; it appears to export only the pre-digested browser information:
http://code.google.com/apis/analytics/docs/gdata/gdataReferenceDimensionsMetrics.html#browser.
Forum post 87919 on forums.digitalpoint.com (link removed since I'm a n00b)
refers to pulling a cross-segment report for more detail; however, that no longer appears to be on the GA front end interface.
Why do we need this instead of just fixing it? If it turns out it only impacts a few users, we can schedule the fix later in the cycle; if it's impacting 20% of our base, it becomes far sooner to fix.
So, the question - how can we pull a specific User Agent string from GA; pull all UA strings from GA or run a Regex against GA to get a count of a matching UA string?
We're also working with the SA team to enable UA logging on the apache level (very high volume website; logging is turned way down).
The best you're going to get out of the Data Export API (or the Google Analytics interface) is processed information: browser / browser version (Internet Explorer / 8.0). There is no way to get the original user-agent string.
Also note that if you have a high volume website you're going to run into data sampling especially if you're looking at a date range beyond larger than a day or two (more on sampling).
Now, I realise the initial response to this is likely to be "you can't" or "use analytics", but I'll continue in the hope that someone has more insight than that.
Google adwords with "autotagging" appends a "gclid" (presumably "google click id") to link that sends you to the advertised site. It appears in the web log since it's a query parameter, and it's used by analytics to tie that visit to the ad/campaign.
What I would like to do is to extract any useful information from the gclid in order to do our own analysis on our traffic. The reasons for this are:
Stats are imperfect, but if we are collating them, we know exactly what assumptions we have made, and how they were calculated.
We can tie the data to the rest of our data and produce far more accurate stats wrt conversion rate.
We don't have to rely on javascript for conversions.
Now it is clear that the gclid is base64 encoded (or some close variant), and some parts of it vary more than others. Beyond that, I haven't been able to determine what any of it relates to.
Does anybody have any insight into how I might approach decoding this, or has anybody already related gclids back to compaigns or even accounts?
I have spoken to a couple of people at google, and despite their "don't be evil" motto, they were completely unwilling to discuss the possibility of divulging this information, even under an NDA. It seems they like the monopoly they have over our web stats.
By far the easiest solution is to manually tag your links with Google Analytics campaign tracking parameters (utm_source, utm_campaign, utm_medium, etc.) and then pull out that data.
The gclid is dependent on more than just the adwords account/campaign/etc. If you click on the same adwords ad twice, it could give you different gclids, because there's all sorts of session and cost data associated with that particular click as well.
Gclid is probably not 100% random, true, but I'd be very surprised and concerned if it were possible to extract all your Adwords data from that number. That would be a HUGE security flaw (i.e. an arbitrary user could view your Adwords data). More likely, a pseudo-random gclid is generated with every impression, and if that ad is clicked on, the gclid is logged in Adwords (otherwise it's thrown out). Analytics then uses that number to reconcile the data with Adwords after the fact. Other than that, there's no intrinsic value in the gclid number itself.
In regards to your last point, attempting to crack or reverse-engineer this information is explicitly forbidden in both the Google Analytics and Google Adwords Terms of Service, and is grounds for a permanent ban. Additionally, the TOS that you agreed to when signing up for these services says that it is not your data to use in any way you feel like. Google is providing a free service, so there are strings attached. If you don't like not having complete control over your data, then there are plenty of other solutions out there. However, you will pay a premium for that kind of control.
Google makes nearly all their money from selling ads. Adwords is their biggest money-making product. They're not going to give you confidential information about how it works. They don't know who you are, or what you're going to do with that information. It doesn't matter if you sign an NDA and they have legal recourse to sue you; if you give away that information to a competitor, your life isn't worth enough to pay back the money you will have lost them.
Sorry to break it to you, but "Don't be Evil" or not, Google is a business, not a charity. They didn't become one of the most successful companies in the world by giving away their search algorithm to the first guy who asked for it.
The gclid parameter is encoded in Protocol Buffers, and then in a variant of Base64.
See this guide to decoding the gclid and interpreting it, including an (Apache-licensed) PHP function you can use.
There are basically 3 parameters encoded inside it, one of which is a timestamp. The other 2 as yet are not known.
As far as understanding what these other parameters mean—it may be helpful to compare it to the ei parameter, which is encoded in an extremely similar way (basically Protocol Buffers with the keys stripped out). The ei parameter also has a timestamp, with what seem to be microseconds, and 2 other integers.
FYI, I just posted a quick analysis of some glcid data from my sites on this post. There definitely is some structure to the gclid, but it is difficult to decipher.
I think you can get all the goodies linked to the gclid via google's adword api. Specifically, you can query the click performance report.
https://developers.google.com/adwords/api/docs/appendix/reports#click
I've been working on this problem at our company as well. We'd like to be able to get a better sense of what our AdWords are doing but we're frustrated with limitations in Analytics.
Our current solution is to look in the Apache access logs for GET requests using the regex:
.*[?&]gclid=([^$&]*)
If that exists, then we look at the referer string to get the keyword:
.*[?&]q=([^$&]*).*
An alternative option is to change your Apache web log to start logging the __utmz cookie that google sets, which should have a piece for the keyword in utmctr. Google __utmz cookie and you should be able to find plenty of information.
How accurate is the referer string? Not 100%. Firewalls and security appliances will strip it out. But parsing it out yourself does give you more flexibility than Google Analytics. It would be a great feature to send the gclid to AdWords and get data back, but that feature does not look like it's available.
EDIT: Since I wrote this we've also created our own tags that are appended to each destination url as a request parameter. Each tag is just an md5 hash of the text, ad group, and campaign name. We grab it using regex from the access log and look it up in a SQL database.
This is a non-programmatic way to decode the GCLID parameter. Chances are you are simply trying to figure out the campaign, ad group, keyword, placement, ad that drove the click and conversion. To do this, you can upload the GCLID into AdWords as a separate conversion type and then segment by conversion type to drill down to the criteria that triggered the conversion. These steps:
In AdWords UI, go to Tools->Conversions->Add conversion with source "Import from clicks"
Visit the AdWords help topic about importing conversions https://support.google.com/adwords/answer/7014069 and create a bulk load file with your GCLID values, assigning the conversions to you new "Import from clicks" conversion type
Upload the conversions into AdWords in Tools->Conversions->Conversion actions (Uploads) on left navigation
Go to campaigns tab, Segment->Conversions->Conversion name
Find your new conversion name in the segment list, this is where the conversion came from. Continue this same process on the ad groups and keywords tab until you know the GCLID originating criteria
Well, this is no answer, but the approach is similar to how you'd tackle any cryptography problem.
Possibility 1: They're just random, in which case, you're screwed. This is analogous to a one-time pad.
Possibility 2: They "mean" something. In that case, you have to control the environment.
Get a good database of them. Find gclids for your site, and others. Record all times that all clicks occur, and any other potentially useful data
Get cracking! As you have started already, start regressing your collected data against your known, and see if you can find patterns used decrypting techniques
Start scraping random gclid's, and see where they take you.
I wouldn't hold high hope for this to be successful though, but I do wish you luck!
Looks like my rep is weak, so I'll just post another answer rather than a comment.
This is not an answer, clearly. Just voicing some thoughts.
When you enable auto tagging in Adwords, the gclid params are not added to the destination URLs. Rather they are appended to the destination URLs at run time by the Google click tracking servers. So, one of two things is happening:
The click servers are storing the gclid along with Adwords entity identifiers so that Analytics can later look them up.
The gclid has the entity identifiers encoded in some way so that Analytics can decode them.
From a performance perspective it seems unlikely that Google would implement anything like option 1. Forcing Analytics to "join" the gclid to Adwords IDs seems exceptionally inefficient at scale.
A different approach is to simply look at the referrer data which will at least provide the keyword which was searched.
Here's a thought: Is there a chance the gclid is simply a crytographic hash, a la bit.ly or some other URL shortener?
In which case the contents of the hashed text would be written to a database, and replaced with a unique id.
Afterall, the gclid is shortening a bunch of otherwise long text.
Takes this example:
www.example.com?utm_source=google&utm_medium=cpc
Is converted to this:
www.example.com?gclid=XDF
just like a URL shortener.
One would need a substitution cipher in order to reverse engineer the cryptographic hash... not as easy task: https://crypto.stackexchange.com/questions/300/reverse-engineering-a-hash
Maybe some deep digging into logs, looking for patterns, etc...
I agree with Ophir and Chris. My feeling is that it is purely a serial number / unique click ID, which only opens up its secrets when the Analytics and Adwords systems talk to each other behind the scenes.
Knowing this, I'd recommend looking at the referring URL and pulling as much as possible from this to use in your back end click tracking setup.
For example, I live in NZ, and am using Firefox. This is a search from the Firefox Google toolbar for "stack overflow":
http://www.google.co.nz/search?q=stack+overflow&ie=utf-8&oe=utf-8&aq=t&client=firefox-a&rlz=1R1GGLL_en-GB
You can see that: a) im using .NZ domain, b) my keyword "stack+overflow", c) im running firefox.
Finally, if you also stash the full landing page URL, you can store the GCLID, which will tell you the visitor came from paid, whereas if it doesn't have a GCLID, then the user must have come from natural search (if URL tagging is enabled of course).
This would theoretically allow you to then search for the keyword in your campaign, and figure out which adgroup them came from. Knowing the creative would probably be impossible though, unless you split test your landing URLs or tag them somehow.