Now Google is using their RankBrain machine learning for search, artificial intelligence is affecting search results. Previous best practice is dead and the old rule book can be torn up. Machine learning does things which are counter-intuitive to humans. Consequently A/B split testing is the best way forward for optimising performance in Google search.

Google search now uses machine learning

The Google algorithm was previously well understood.  That was before Amit Singhal, head of Google’s search team and master of the ranking algorithm left in 2016 under something of a cloud.  Singhal apparently preferred that the rationale for search results would be clear. However, those days are gone.  Singhal’s replacement, John Giannandrea, was previously the head of Google’s artificial intelligence. The search engine now uses a machine learning technology called RankBrain to help deliver search results.

Machine learning has side effects

Machine learning can bring dramatic improvements where applied, but it also has side effects.  Machine learning optimises things in ways which are not intuitive to humans.  Consequently, it is difficult for humans to understand what is going on.

Testing reported by SEO agency Distilled at SMX Munich in March 2017 found that SEO professionals were worse than a coin flip at predicting which pages would rank best in Google when given a choice of two.  This suggests that predicting performance in Google search is becoming increasingly difficult. Not only does machine learning act in ways that are opaque to humans, some sites seem to respond differently to others for the same change. So it seems that there is a new rule book and it cannot be understood by humans.

The way forward is A/B split testing

The good news is that the introduction of machine learning may mean that digital marketers can optimise even more than before.  However, if there is no longer a best practice and every site is different, it follows that the way forward is to make extensive use of A/B split testing.

Split testing for SEO is not the same as split testing for CRO

Split testing for Conversion Rate Optimisation can demonstrate which version of a page performs best by alternating between the two. For SEO it doesn’t work like that. You cannot have two versions of the same thing for the Googlebots or you are likely to get in trouble.  Therefore you need to divide multiple pages into separate batches.

  • Create two batches of pages which are based on the same template
  • Make a change to all pages in one batch
  • Analyse which batch performs better over time with a statistically robust quantity of traffic

Then you can use Google Analytics or your API to create two separate reports, one for each batch, and observe.

…or maybe you should just give Gingerbolt a call…