Integrate Queries from external similar webshop & keyword planner
A
Anton Siegordner
This was a discussion point with Markus K. today. In order to optimize not only spelling or various similar wordings inside a cluster is there a possibility to get suggestions beyond the obvious missspelling etc. Goal for us would be to integrate ALL possible search queries into the existing cluster. example:
Montageprofil -> Clustered already in various spelling types. In addition we need to add: bauprofil, 4kant profil, 6kant profil, profilschiene etc. (infos I do get when I e.g. look at other DIY shops or ask the google keyword planner...
Andreas Wagner
Hi Anton,
while I agree that this idea sounds promising in theory, our past experience has clearly shown that it doesn't perform well in production. The core issue lies in the nature of the related terms — they are related, but not true synonyms. As a result, they often fail to convey the same user intent or accurately represent the same and complete set of products over time. This leads to two critical risks: misinterpreting the user's intent or unintentionally excluding relevant products when the intent is correctly understood.
This challenge is precisely what led us to develop our new capability, NeuralInfusion, which will soon be generally available.
A
Anton Siegordner
Andreas Wagner thanks for the response and your perspective on this. While I get your point and I guess this is how you developed your business model. I still believe that my idea could work - for us but also for many other businesses for specific queries. I also admit, this can be something to work on inside the own search engine, not on external side with searchhub.
Andreas Wagner
Anton Siegordner: thx for your reply.
I think we need some clarification first to have some common ground as I agree that this idea is useful:
First things first: This has nothing to do with our business model. The only two reasons why we don't want to support or add such a feature are explainability and consistency.
What you are essentially asking for is called query expansion: using the terms provided this would mean that we would have to build a union of results from all these queries.
As of today the independent queries would result in the following results sizes and confidences:
-montageprofil: 2 results - 1,2%
-bauprofil: 4 results - 0%
-4kant profil: 4552 results - 0.1%
-6kant profil: 383 results - 0%
-profilschiene: 97 results - 8%
If we would add all those queries into a single cluster you would independent of the user query always show the result for profilschiene!
But the products shown have nothing to do with what the user was actually searching for using one of the other queries - as the intent is different.
Since we are NOT a search engine we can only pick the best available result from the underlying search engine for a given intent.
Without NeuralInfusion (which means that we are running a search system that actually understands the intent and the product assortment in the background on our end) we can NOT manipulate a given search result and instead re-route the user to a different better one.
Only by NeuralInfusion we can add, remove, pin or rerank a searchresult and achieve a result that actually helps the searcher to find what he is looking for without leaving relevant product from your assortment aside.
Please let me know if this explanation helps.