Don’t Go to Google, TripAdvisor, or OurFaves for Restaurant Reviews

If you come to my town, where's the best place to go to look for that perfect restaurant, or opinions about a place you're considering dining at? Me. Seriously.

And maybe a few of my friends.

We know the truth.

You can sometimes get some of that truth from Toronto Life and Zagat.

Once in awhile, we'll maybe go write a review on Yelp. But probably not. If you ate at 50 places a year and 30 of them are really good, you'd tire of writing it up.

So, if you go to Google's review aggregation that includes results from TripAdvisor, OurFaves, and Google itself... you'll see a bunch of inexplicable one-star reviews for some of the best restaurants in the city.

"I could barely see my food..." Have you heard of ambience?

"The staff was unfriendly..." ... after you put ketchup on the filet of sole.

"Cramped..." Sorry it isn't the Rainforest Cafe.

Don't get me wrong: I'm a big fan of user-generated content and recommendations. Unfortunately, when you go to TripAdvisor, you often have to wade through the most inexplicable, knuckle-dragger "reviews" of some of the best hotels and restaurants known to mankind.

It's also an interface issue with Google's review aggregation, though. The Harbord Room actually averages four stars on Yelp... but you wouldn't know this from Google's tally, which makes it look like there are a lot of dissatisfied visitors, and the average looks like it's just over one star.

You know what? Just let me make the reservation, and if the place is no good, I'll take full responsibility. :)

In the meantime though you could trust good reviewers like this guy.

Can Search Engines Sniff Out "Remarkable"?

I never tire of listening to experts like Mike Grehan speaking about the new signals search engines are beginning to look at, because it's so important to bust the myths about how search engines work.

To hear many people talk, today's major engines are faced with little more than a slightly-beefed- up, slightly larger, version of a closed database search. Need the medical records for your patient Johnny Jones, from your closed database of 500 medical records, just type in johnny or jones or johnny jones, and you're good to go. Isn't that search, in a nutshell? It is: if you can guarantee that you're referring to a nutshell like that. But with web search, it's nothing like that.

The World Wide Web now has a trillion pages or page-like entities... that Google knows about. (They don't know what to do with all of them, but they'll admit to the trillion.) Some observers estimate that there will soon be five trillion of these in total, too many to index or handle. Who knows, maybe 10% of that could be useful to a user or worthy of indexing. But until some signal tells the search engine to index them in earnest, they'll just sit there, invisible. That's out of necessity: there's just too much.

The difference isn't only quantitative, it's also qualitative. User queries have all sorts of intents, and search engines aren't just trying to show you "all the pages that match". There are too many pages that match, in one way or another. The task of measuring relevancy, quality, and intent is far more complex than it looks at first.

And on top of that, people are trying to game the algorithm. Millions of people. This is known as "adversarial" information retrieval in an "open" system where anyone can post information or spam. The complexity of rank ordering results on a particular keyword query therefore rises exponentially.

In light of all this, search engines have done a pretty good job of looking at off-page signals to tell what's useful, relevant, and interesting. The major push began with the linking structure of the web, and now the effort has vastly expanded to many other emerging signals; especially, user behavior (consumption of the content; clickstreams; user trails) and new types of sharing and linking behavior in social media.

This is a must, because any mechanical counting and measuring exercise is bound to disappoint users if it isn't incredibly sophisticated and subtle. Think links. Thousands of SEO experts are still teaching you tricks for how to get "authoritative" inbound links to your sites & pages. But do users want to see truly remarkable content, or content that scored highly in part because someone followed an SEO to-do list? And how, then, do we measure what is truly remarkable?

Now that Twitter is a key source of evidence for the remarkability of content, let's consider it as an interesting behavioral lab. Look at two kinds of signal. The first is where you ask a few friends to retweet your article or observation, and they do. A prickly variation of that is where you have a much larger circle of friends, or you orchestrate semi-fake friends to do your bidding, with significant automation involved.

But another type of remarkable happens when your contribution truly makes non-confidantes want to retweet and otherwise mention you. When your article or insight achieves "breakout" beyond your circle of confidantes, and further confirming signals of user satisfaction later on when people stumble on it.

Telling the difference is an incredible challenge for search engines. Garden variety tactical optimization will work to a degree, mainly because some signals of interest will tend to dwarf the many instances of "zero effort or interest". But we should all hope that search engines get better and better at sniffing out the difference between truly remarkable (or remarkably relevant to you the end user) and these counterfeit signals that can be manufactured by tacticians simply going through the motions.

Can Search Engines Sniff Out "Remarkable"?

I never tire of listening to experts like Mike Grehan speaking about the new signals search engines are beginning to look at, because it's so important to bust the myths about how search engines work.

To hear many people talk, today's major engines are faced with little more than a slightly-beefed- up, slightly larger, version of a closed database search. Need the medical records for your patient Johnny Jones, from your closed database of 500 medical records, just type in johnny or jones or johnny jones, and you're good to go. Isn't that search, in a nutshell? It is: if you can guarantee that you're referring to a nutshell like that. But with web search, it's nothing like that.

The World Wide Web now has a trillion pages or page-like entities... that Google knows about. (They don't know what to do with all of them, but they'll admit to the trillion.) Some observers estimate that there will soon be five trillion of these in total, too many to index or handle. Who knows, maybe 10% of that could be useful to a user or worthy of indexing. But until some signal tells the search engine to index them in earnest, they'll just sit there, invisible. That's out of necessity: there's just too much.

The difference isn't only quantitative, it's also qualitative. User queries have all sorts of intents, and search engines aren't just trying to show you "all the pages that match". There are too many pages that match, in one way or another. The task of measuring relevancy, quality, and intent is far more complex than it looks at first.

And on top of that, people are trying to game the algorithm. Millions of people. This is known as "adversarial" information retrieval in an "open" system where anyone can post information or spam. The complexity of rank ordering results on a particular keyword query therefore rises exponentially.

In light of all this, search engines have done a pretty good job of looking at off-page signals to tell what's useful, relevant, and interesting. The major push began with the linking structure of the web, and now the effort has vastly expanded to many other emerging signals; especially, user behavior (consumption of the content; clickstreams; user trails) and new types of sharing and linking behavior in social media.

This is a must, because any mechanical counting and measuring exercise is bound to disappoint users if it isn't incredibly sophisticated and subtle. Think links. Thousands of SEO experts are still teaching you tricks for how to get "authoritative" inbound links to your sites & pages. But do users want to see truly remarkable content, or content that scored highly in part because someone followed an SEO to-do list? And how, then, do we measure what is truly remarkable?

Now that Twitter is a key source of evidence for the remarkability of content, let's consider it as an interesting behavioral lab. Look at two kinds of signal. The first is where you ask a few friends to retweet your article or observation, and they do. A prickly variation of that is where you have a much larger circle of friends, or you orchestrate semi-fake friends to do your bidding, with significant automation involved.

But another type of remarkable happens when your contribution truly makes non-confidantes want to retweet and otherwise mention you. When your article or insight achieves "breakout" beyond your circle of confidantes, and further confirming signals of user satisfaction later on when people stumble on it.

Telling the difference is an incredible challenge for search engines. Garden variety tactical optimization will work to a degree, mainly because some signals of interest will tend to dwarf the many instances of "zero effort or interest". But we should all hope that search engines get better and better at sniffing out the difference between truly remarkable (or remarkably relevant to you the end user) and these counterfeit signals that can be manufactured by tacticians simply going through the motions.

Yusuf Mehdi’s Too-Candid Comments About Abandoning the Long Tail

Credit Yusuf Mehdi for honesty: in his remarks at SES New York last week, as reported by eWeek, he noted that Microsoft fell well behind Google in search because it focused on doing well for popular queries, when it should have known that search is "all about the long tail."

It is bizarre, because every notable failure in search since 1994 has basically been in the realm of curated results and chances are, that trend will continue. Whether they're hand-edited search results or partially "produced" variations of web index search focusing on improving the treatment of head terms using the efforts of channel producers, the market kept coming back with the same response: this approach doesn't scale. A website with opinions about what people should focus on is not a search engine, it's just a website. And that creates a serious positioning problem when you're competing in the "search engine" space, which needs to scale to help people find hard-to-find information. Forget the long tail: channel producers and editors even do a poor job of producing information around the "torso". As information and customer demands evolve, it becomes difficult to keep up, and many of the real world uses of the search engine begin to look like a "demo" of "well, this is how it works over here, on this query, in theory, and eventually we'll get back to extending the technology so it works for the stuff you're looking for, with partners who provide information in a way that you prefer, which changed in the past year."

Here's a list of some of the search engines that haven't caught on precisely because they failed to understand and gear up for the massive scale required in the search engine business, focusing instead on curating results for a limited set of popular queries or categories:

  • Yahoo Directory
  • Open Directory
  • LookSmart
  • Ask Jeeves
  • Mahalo
The list could probably be much longer.

Others have fared a bit better because they didn't claim to be search engines. These include:
  • About.com
  • Squidoo
Obviously, many of these properties are of limited use in the real world of finding info.

The bizarreness doesn't stop there, however. A significant aspect of the PR rollout of Bing was focused on the fact that Microsoft knew it would be most effective -- again -- at doing better for users in the realm of more popular types of searches, ceding long tail excellence to Google. In terms of positioning, that's like saying Microsoft is good at negotiating partnerships, designing interfaces, and subscribing to web services. That's like saying Microsoft is building a portal. That's like saying Microsoft is Yahoo.

Google itself is no saint when it comes to long tail accomplishments and relevance. On many counts, all search engine companies have waved white flags on truly scaling to address all potential content, because there is just too much of it (and too much spam). Dialing back on the ambitions of comprehensiveness, to devote more screen real estate to trusted brands and search experiences that are tantamount to paid inclusion, is Google's current trend, much as it was for companies like Inktomi and Yahoo in the past.

The industry consensus is that search is far from solved. But a prerequisite to solving any problem is trying. Microsoft is signaling that they will continue to dip a toe in the water and essentially "wimp out" when it comes to addressing scale and complexity issues. This is in line with what they've done all along, and the positioning for Bing. The question is: if Google's wimping out too, wouldn't you rather use the relatively less wimpy search company that has committed a massive budget to R&D, probably 30X Microsoft's? By sending these signals, Microsoft is not exactly giving users good reasons to use their products. It's reminiscent of the trajectory taken by companies like AOL and Yahoo, who didn't feel that search was a problem that could or should be solved by them, so they contented themselves with staying hands-off and creating a workable project largely driven by feeds, partnerships, and ideas external to their own company.

To SEO's, Mehdi's ruminations on the long tail must be heartening. It says, in essence, "spam away."

Yusuf Mehdi’s Too-Candid Comments About Abandoning the Long Tail

Credit Yusuf Mehdi for honesty: in his remarks at SES New York last week, as reported by eWeek, he noted that Microsoft fell well behind Google in search because it focused on doing well for popular queries, when it should have known that search is "all about the long tail."

It is bizarre, because every notable failure in search since 1994 has basically been in the realm of curated results and chances are, that trend will continue. Whether they're hand-edited search results or partially "produced" variations of web index search focusing on improving the treatment of head terms using the efforts of channel producers, the market kept coming back with the same response: this approach doesn't scale. A website with opinions about what people should focus on is not a search engine, it's just a website. And that creates a serious positioning problem when you're competing in the "search engine" space, which needs to scale to help people find hard-to-find information. Forget the long tail: channel producers and editors even do a poor job of producing information around the "torso". As information and customer demands evolve, it becomes difficult to keep up, and many of the real world uses of the search engine begin to look like a "demo" of "well, this is how it works over here, on this query, in theory, and eventually we'll get back to extending the technology so it works for the stuff you're looking for, with partners who provide information in a way that you prefer, which changed in the past year."

Here's a list of some of the search engines that haven't caught on precisely because they failed to understand and gear up for the massive scale required in the search engine business, focusing instead on curating results for a limited set of popular queries or categories:

  • Yahoo Directory
  • Open Directory
  • LookSmart
  • Ask Jeeves
  • Mahalo
The list could probably be much longer.

Others have fared a bit better because they didn't claim to be search engines. These include:
  • About.com
  • Squidoo
Obviously, many of these properties are of limited use in the real world of finding info.

The bizarreness doesn't stop there, however. A significant aspect of the PR rollout of Bing was focused on the fact that Microsoft knew it would be most effective -- again -- at doing better for users in the realm of more popular types of searches, ceding long tail excellence to Google. In terms of positioning, that's like saying Microsoft is good at negotiating partnerships, designing interfaces, and subscribing to web services. That's like saying Microsoft is building a portal. That's like saying Microsoft is Yahoo.

Google itself is no saint when it comes to long tail accomplishments and relevance. On many counts, all search engine companies have waved white flags on truly scaling to address all potential content, because there is just too much of it (and too much spam). Dialing back on the ambitions of comprehensiveness, to devote more screen real estate to trusted brands and search experiences that are tantamount to paid inclusion, is Google's current trend, much as it was for companies like Inktomi and Yahoo in the past.

The industry consensus is that search is far from solved. But a prerequisite to solving any problem is trying. Microsoft is signaling that they will continue to dip a toe in the water and essentially "wimp out" when it comes to addressing scale and complexity issues. This is in line with what they've done all along, and the positioning for Bing. The question is: if Google's wimping out too, wouldn't you rather use the relatively less wimpy search company that has committed a massive budget to R&D, probably 30X Microsoft's? By sending these signals, Microsoft is not exactly giving users good reasons to use their products. It's reminiscent of the trajectory taken by companies like AOL and Yahoo, who didn't feel that search was a problem that could or should be solved by them, so they contented themselves with staying hands-off and creating a workable project largely driven by feeds, partnerships, and ideas external to their own company.

To SEO's, Mehdi's ruminations on the long tail must be heartening. It says, in essence, "spam away."

Google AdWords: No More Last-Click-Attribution Blues

Getting credit for an online conversion - and giving due credit to all recent influences - has been one of the hottest topics in digital marketing over the past couple of years. The urgency of the matter has grown as media costs -- especially click prices on paid search keywords -- have risen.

Marketers have been so hungry for better attribution of "keyword assists" (or simply, the non-overriding of the first click in the sequence towards purchase, whether that's over a matter of hours or many months), they've been willing to explore cumbersome customizations in a variety of analytics platforms, including Google Analytics.

But if you're looking to simply analyze the contribution of paid keyword searches on Google Search that preceded the keywords that led directly to a sales conversion (aka "assists"), you'd prefer to see all that data rolled up conveniently within Google AdWords itself, showing the data in handy formats that might make it easy to change your bidding patterns. In particular, earlier stage keywords (typically, before a last-click brand search) would now be revalued in your model; you'd bid them higher in cases where they made assists.

Happily all of this is now rolling out in AdWords as part of a reporting initiative called Search Funnels. A variety of reporting options help you tap into the power of this new information.

Earlier, when I defended the "last click"'s merits as an attribution method, I pointed to some data by Marin Software showing 74% of etail conversions only have one associated click - even counting assists. Moreover, Marin's approach bucketed prior clicks categorically, arguing that if a prior click was very similar in intent or style to the last click, then the extra information wouldn't be enough to cause you to alter bidding patterns anyway. That knocked the number of truly "assist-powered" conversions (that you could actually attribute properly) down to 10% or less.

This is where Google's new reporting needs to be scrutinized closely. In your individual case it could be quite valuable, but in current individual case studies Google may have on hand, anywhere from 70-95% of conversions only have one click to speak of. If Marin's logic above is even close to sensible, then it does underscore the limits to assist data. There will be some value attributable to assist keywords in around 10% of conversions, give or take. That's actionable but not earth-shattering. Of course, this is going to be most valuable to advertisers who have a lot of prior influencer clicks hiding behind a high number of clicks that are currently attributed to a last-click on the brand name.

To pump up the role of prior keywords, it might be fair to also point to assist impressions - views of the ad on Google Search where the ad wasn't clicked, but shown. But in those cases was the ad really seen? Perhaps not, but there may be some value in knowing what search keywords got the searcher's research motor running. Perhaps they clicked on a competitor's ad. Google is offering impression assist data as well with this release, which will be sure to delight trivia buffs, AdWords junkies, and Google's accountants alike.

Remember, we're not just talking about multiple searches all done in a single day, or in one session. Google is logging the time and date of every search by that user prior to a purchase/lead, and when a conversion happens, full funnel information is available as to the time lag between clicks and before the conversion.

Adding in impression assists to the mix, we may see past search query information for up to 20-25% of conversions in some advertiser accounts. Again, while not stupendous, this at least counts as extremely important and material to how you approach keyword value.

The ease of sorting in order of frequency of conversion by assist keyword helps not only to see the keywords in question, but with the "keyword transition path" view, you can see what last click converters they preceded, to better understand the consumer mindset. The screen shot below is a canned Google example while the program is still in beta. In my briefing I saw a more typical and valuable case example that showed the frequency (fictitious example to replace the one I saw) paths like "almond milk calories" > planethealthnut or "milk alternative" > planethealthnut. Whereas the brand might have got disproportionate credit for this conversion in the past, now, keywords like [milk alternative] or [almond milk calories] might attract higher bids, even more so if you experiment over time, allowing for more repetitions of your "research stage keywords" over many months.

In my opinion, "paths" work fairly well as a metaphor here and are not too misleading because the "funnel" steps tend to be relatively coherent and causal in practice. They aren't necessarily so, however. The reason these reports can look sensible is because they're drawn from a narrow universe of high-intent keywords that advertisers are avidly bidding on. You're not going to see a paid search keyword funnel path like "drawbridge in mexico" > james mcbleckr phone 415 > nike > air jordans used > nike.com largely because Nike doesn't have most of the keywords in that path in their paid search account. Truly generating causal paths out of all the things someone does online prior to a conversion is likely to be incredibly messy, but that's a much longer story.

Long story short: life is indeed a lot simpler when viewed through the prism of an AdWords account. And today, advertisers are getting what they desperately seek: easy-to-use information about paid keyword search attribution so that the last click doesn't override all other attribution data.

Google AdWords: No More Last-Click-Attribution Blues

Getting credit for an online conversion - and giving due credit to all recent influences - has been one of the hottest topics in digital marketing over the past couple of years. The urgency of the matter has grown as media costs -- especially click prices on paid search keywords -- have risen.

Marketers have been so hungry for better attribution of "keyword assists" (or simply, the non-overriding of the first click in the sequence towards purchase, whether that's over a matter of hours or many months), they've been willing to explore cumbersome customizations in a variety of analytics platforms, including Google Analytics.

But if you're looking to simply analyze the contribution of paid keyword searches on Google Search that preceded the keywords that led directly to a sales conversion (aka "assists"), you'd prefer to see all that data rolled up conveniently within Google AdWords itself, showing the data in handy formats that might make it easy to change your bidding patterns. In particular, earlier stage keywords (typically, before a last-click brand search) would now be revalued in your model; you'd bid them higher in cases where they made assists.

Happily all of this is now rolling out in AdWords as part of a reporting initiative called Search Funnels. A variety of reporting options help you tap into the power of this new information.

Earlier, when I defended the "last click"'s merits as an attribution method, I pointed to some data by Marin Software showing 74% of etail conversions only have one associated click - even counting assists. Moreover, Marin's approach bucketed prior clicks categorically, arguing that if a prior click was very similar in intent or style to the last click, then the extra information wouldn't be enough to cause you to alter bidding patterns anyway. That knocked the number of truly "assist-powered" conversions (that you could actually attribute properly) down to 10% or less.

This is where Google's new reporting needs to be scrutinized closely. In your individual case it could be quite valuable, but in current individual case studies Google may have on hand, anywhere from 70-95% of conversions only have one click to speak of. If Marin's logic above is even close to sensible, then it does underscore the limits to assist data. There will be some value attributable to assist keywords in around 10% of conversions, give or take. That's actionable but not earth-shattering. Of course, this is going to be most valuable to advertisers who have a lot of prior influencer clicks hiding behind a high number of clicks that are currently attributed to a last-click on the brand name.

To pump up the role of prior keywords, it might be fair to also point to assist impressions - views of the ad on Google Search where the ad wasn't clicked, but shown. But in those cases was the ad really seen? Perhaps not, but there may be some value in knowing what search keywords got the searcher's research motor running. Perhaps they clicked on a competitor's ad. Google is offering impression assist data as well with this release, which will be sure to delight trivia buffs, AdWords junkies, and Google's accountants alike.

Remember, we're not just talking about multiple searches all done in a single day, or in one session. Google is logging the time and date of every search by that user prior to a purchase/lead, and when a conversion happens, full funnel information is available as to the time lag between clicks and before the conversion.

Adding in impression assists to the mix, we may see past search query information for up to 20-25% of conversions in some advertiser accounts. Again, while not stupendous, this at least counts as extremely important and material to how you approach keyword value.

The ease of sorting in order of frequency of conversion by assist keyword helps not only to see the keywords in question, but with the "keyword transition path" view, you can see what last click converters they preceded, to better understand the consumer mindset. The screen shot below is a canned Google example while the program is still in beta. In my briefing I saw a more typical and valuable case example that showed the frequency (fictitious example to replace the one I saw) paths like "almond milk calories" > planethealthnut or "milk alternative" > planethealthnut. Whereas the brand might have got disproportionate credit for this conversion in the past, now, keywords like [milk alternative] or [almond milk calories] might attract higher bids, even more so if you experiment over time, allowing for more repetitions of your "research stage keywords" over many months.

In my opinion, "paths" work fairly well as a metaphor here and are not too misleading because the "funnel" steps tend to be relatively coherent and causal in practice. They aren't necessarily so, however. The reason these reports can look sensible is because they're drawn from a narrow universe of high-intent keywords that advertisers are avidly bidding on. You're not going to see a paid search keyword funnel path like "drawbridge in mexico" > james mcbleckr phone 415 > nike > air jordans used > nike.com largely because Nike doesn't have most of the keywords in that path in their paid search account. Truly generating causal paths out of all the things someone does online prior to a conversion is likely to be incredibly messy, but that's a much longer story.

Long story short: life is indeed a lot simpler when viewed through the prism of an AdWords account. And today, advertisers are getting what they desperately seek: easy-to-use information about paid keyword search attribution so that the last click doesn't override all other attribution data.