Article recommendations in chat

When a visitor sends their first chat message, the window can show up to three of your help center articles that match it, under the heading "These articles might answer your question:". The visitor then presses This answered my question or I still need help.

• This answered my question closes the conversation, tags it resolved-by-article and adds an internal note naming the article. If the visitor writes again, the ticket reopens like any chat.

• I still need help leaves the conversation open for your team, who see it as usual.

The conversation reaches your inbox the moment the visitor sends that first message, before they answer, so nobody waits on the suggestion.

Matching is a word search of your published articles. No AI reads the message or picks the articles, and the list of suggestions is the same every time for the same words.

Before you start#

• Create a chat (Settings > Chat) and a live help center with published articles. Until there is a chat, Workflows > Article recommendations shows only Create a chat.

• Admins change these settings. Other roles can read them.

Set it up#

Workflows > Article recommendations has one card for each chat.

1. Switch on Suggest articles in for the chat.

2. Choose the Help center. By default the chat suggests from its own help center. A help center marked "(not live)" suggests nothing until you publish it.

3. Choose How closely an article must match (below).

4. Under Never suggest, add articles that match many messages but rarely answer them, such as your terms of service. Remove one from the list to allow it again.

5. Save the card.

The same switch appears on the chat's own Automate tab, under Article suggestions.

How closely an article must match#

Words are compared in the article's language, so "shipping" also finds "shipped". Titles count more than summaries, and summaries more than the body, so a clear title does most of the work.

Try a message before you save#

Each card has Try a first message. Type something a customer might write, such as "How long does delivery to Canada take?", and press Try. You see every article that shares a word with it, its score, and one of:

• Suggested: it reaches the score this setting needs and would be shown (three at most);

• Excluded: it is on your Never suggest list;

• Too weak: it matched, but not closely enough.

The test uses your unsaved choices, so you can compare Broad, Balanced and Strict before saving. If nothing shares a word with the message, the card says so, and that conversation would go to your team with no suggestion.

Which articles can be suggested#

• Only published articles in a live help center.

• Only in the conversation's language. When an article has no published version in that language, the help center's default language is used.

• Never an article on the Never suggest list.

To get better suggestions, improve the articles: put the words customers use in the title and summary. See Write and organize articles.

When no suggestion appears#

• Suggestions are made for the first message of a conversation only, not for later messages.

• Conversations that start from a flow's hand-over or an order request get none, because they already carry the visitor's answers.

• A visitor who already has a conversation open keeps writing in that conversation.

See the results#

Analytics > Automation counts first messages that got suggestions, how many visitors said an article answered their question, how many still needed help and how many did not answer. Conversations closed by an article count as resolved by automation. See the automation report.

There is no screen yet that lists each past suggestion for you to mark right or wrong. The Try a first message box and the Never suggest list are how you tune what is shown.