Optimising articles for AI search and providing search guidance

[Leri Andrews]

Bear with me - this is a long one!! 

With ZING it was possible to understand why some articles floated to the top of the search results and to address issues by:

1. Adjusting the weighting to prioritise meta, title and description 

2. Tell authors to write quality articles and include useful meta

3. Train users with 'keyword' searching

Result: they can type in 'baby', confident that the meta in the 'Maternity Leave Policy' article will bring that up even though it's not in the title/description and only mentioned 3 times in the policy.

Now I have AI search which is a black box in comparison, training itself based on user behaviour.  Documentation suggests the search behaviour should follow:

1 word: It does a search. it does not do a second search using OR

2-7 words: it does a search using AND....

• ...if "insufficient results" are returned, it does a second search using OR and will return results that match at least half (rounded up) of the terms
• ...if sufficient results are returned, it does not do a second search using OR

8 words: it does a search using AND...

• ...if 1+ results are returned, it does not do a second search using OR
• ...if 0 results are returned, it does a second search using OR and will will return results that match at least half (rounded up) of the terms

The problem

I've been telling my users to not type in just one word, but use it as you would any other search engine. You wouldn't, for example, type 'holiday' into google and expect to be shown best-value last-minute holidays in Spain, half-board.  You'd type in 'last-minute holiday Spain half-board' or even assume a natural language capability and type 'I really need holiday right now in Spain'.

What I am actually finding is that typing 'HR guidance on compassionate leave in UK for all employees' (10 words) gives my users irrelevant content (think all sorts of other leave coming up, filled with the word 'leave'). Typing 'UK Compassionate Leave' a la google, gives similarly irrelevant results.  Only typing in 'compassionate' results in the actual policy appearing at the top. This is completely counter-intuitive to me.

Is anyone finding the same?  Any solutions?

[Paul Curwen · Accepted solution]

I know what you mean by black box 🙂 Have you tried using result improvements rules. Boost actions? These influence which results appear first see:

https://docs.servicenow.com/bundle/tokyo-platform-administration/page/administer/ai-search/task/boos...

You can also block content you don't want to appear: 

https://docs.servicenow.com/bundle/tokyo-platform-administration/page/administer/ai-search/task/bloc...