Even though my view may indeed be biased, this pull request is still relevant in my opinion:

• I don't deny that Google Custom Search is a great tool but it is more focused on searching for web pages. The way Docsearch works is by indexing hierarchically your documentation: it's thus more faithful to its structure
• The UX provided by Google Custom Search is not the best one available, and as-you-type search is somehow nicer to use than an intermediate page showing the results
• Algolia indeed has less tried and tested products than Google, but a lot of efforts have been put in Docsearch and it is already live in several "big" open source projects: https://facebook.github.io/react/, http://docs.scala-lang.org/, https://babeljs.algolia.com/, etc.

(I am here only talking about the search within the documentation provided by Google Custom Search and in no way about the search http://www.google.com is able to deliver)

Not convinced:
model selection => does not show the chapter 3 of the user guide ("Model selection and evaluation")
BaseEstimator => no relevant results
projected gradient nmf => no relevant results
check_x_y => no relevant results
Fabian Pedregosa => no results

For each of these examples, the first link in Google Custom Search is very relevant.

Hi @TomDLT !

The url targeted (or excluded) by the scraper are configurable. I indeed excluded the "summary" pages, but I can add them as well as the "about" pages. May I change the configuration so can you give it another try?

We rarely accept a PR without demanding changes; sure, tweak it. We absolutely need API reference included, and it does not seem to be there.

Are those changes made in our HTML/JS invocation, or in an API client configuration interface?

The changes I am currently making are done to a JSON configuration file used by our crawler.
The JSON file in itself is public and can be found here: https://github.com/algolia/docsearch-configs/blob/master/sklearn.json
The crawler actually using this file is not public however

The configuration has been updated and the demo is up to date here
Tell me what you think!

Thanks for the update, this is better.
However I personally don't like that:

• If I choose a wrong link and go back to previous page, I have to type again the search.
• In the same way, I can't open several pages in several tabs from one unique search.

Also:

• perhaps our documentation isn't as well-structured as it could be - or as hierarchical as Algolia insists it should be - but most queries I type give the same heading and page name, i.e. the title is triplicated in the search results UI
• entering classifier in Google gives me the example "Classifier comparison" as the top result, and narrative "Supervised learning" next, better than all top-5 results returned by Algolia.

I am closing this this PR as "not yet", to let Algolia improve UI, ranking, handling of less-structured documentations, etc.