Autoreplies with articles allow you to automatically suggest relevant help center articles via email and web form in response to a customer support ticket.
There are a number of ways you can write and format your articles to work best with autoreplies. You may find that you can't or don't want to use all of the suggestions in this article, but applying just one or two can greatly improve your autoreply results. Presenting customers with better recommendations can increase the autoreply resolution rate, which is the percentage of enquiries resolved by autoreply from the total enquiries where it offered suggestions.
This article contains the following topics:
Structuring and organizing your information
One area where you can make a difference is in the structure and organization of your articles. The suggestions in this section include both customer-facing and behind-the-scenes decisions.
Creating titles for your articles
Article titles are considered when determining whether an article is a good candidate for solving a customer's problem. If an article's title closely matches the text of a support request, it's more likely to come up as a suggestion in an autoreply.
To take advantage of this, you should give your articles titles that are similar to how your customer might phrase their support request.
Try formatting your article titles in one of the following ways:
- As a question: "How do I reset my password?"
- As a simple, active phrase: "Resetting a password"
Narrowing an article's focus
Wherever possible, an article should cover a single, specific topic. How this impacts your articles depends on how you present information to your customers:
- Issue/resolution articles: If you prefer to write your articles around issues and resolutions, try to focus each on a very specific issue with a single resolution. For instance, rather than having a single article covering multiple, related billing issues – payment options, requesting a refund, and challenging a charge – write three articles, one for each issue. If possible, offer a single resolution for each issue as well. This way, you can guide your customers to the resolution you prefer, and if necessary, link to other articles that may offer alternative resolutions.
- Process articles: The same general principle applies to articles built around processes and procedures. Say, for example, you need an article walking a customer through an account set-up process, which involves two main tasks: Creating an account and selecting account options. You could have a single article covering both tasks, since they are both part of the account setup process. However, breaking it up into two articles, one for each task, will help the bot identify the article that best matches your customer's needs, and help your customer locate the applicable information without having to wade through unnecessary text.
Writing article introductions
The first 75 words in an article are weighed more heavily than the subsequent text. By including as many related keywords and context as possible in an article's introduction, you will boost the accuracy of your suggested articles.
For example, if you're writing an article about your company's return policies, don't include a preamble about how you understand that sometimes a shirt is the wrong size or that everyone makes mistakes. As charming as that might be, it gets in the way of evaluating the article’s content. Start your articles cleanly and precisely: "We accept returns for a full refund within 30 days of purchase. After 30 days, you can request a store credit," then get into the details.
Using alternative text
What you write behind the scenes can impact autoreply results as well. Use alternative text on your embedded images and videos to add descriptive keywords.
Using segmentation and labels
The power of article labels really comes through when you have a situation where you can segment your customers when tickets are created. A good example of this is a game developer supporting iOS and Android platforms. The developer has two channels for incoming tickets: one for iOS issues and one for Android. Using the trigger conditions, you can create two autoreply with articles triggers (one for each channel) and then use article labels to serve up only iOS or Android articles where they make sense. So, stop for a second and really think about how you could best segment your customers and create useful conditions around them, then serve them the best articles possible! This is described in more detail in the best practices for using labels with autoreplies guide. For information on working with triggers for autoreplies, see Creating new triggers for autoreplies.
Even if you are not using customer segmentation, labels have a huge impact on autoreply results. Use them thoughtfully. For example, if an article is about a topic that does not tend to generate customer questions, you might omit the label you've added. This helps create a smaller, more curated pool of articles when making suggestions.
Thanks for the heads up Rafael! I've updated the link to the correct article :)
This is helpful! I have a question which I think can best be explained by an example.
There are some things that our users cannot get an answer to in an article, no matter how hard we try. For example, if they've forgotten their username, no article can tell them what it is.
We have an article that gives them tips to remember what their username is, but let's say it doesn't help and they mark the article "unhelpful". How does Answer Bot interpret that?
It seems like if enough people mark it unhelpful, then AB is going to start thinking that's not a good article to serve to people who forgot their username...when in fact, it actually IS the best article.
Can we "reverse" those unhelpful marks, or is it even necessary?
Thank you for leaving a comment here. Currently, there's no way of deleting these votes. However, this is still possible, but by using Votes API.
First, you would have to find the ID of the Votes that you want to delete. One way to do this would be to use the List Votes endpoint to show the votes in a specific article. The result would be something like this:
The value property will show you the 'upvote' (1) or 'downvote' (-1). You can get the Vote ID from here and then use the Delete Vote endpoint.
The result after deleting the first vote:
There's also no direct consequence to having many unhelpful votes in an article. The votes are mainly used for reporting only.
You can also consider Disable "Was this article helpful?" in Zendesk Guide.
I hope this helps clear things.
Hi Plo Mangsat, thanks for the reply. I could use a little more guidance so I can decide how to proceed.
I do think that there is a consequence to having downvotes: it makes the article seem less helpful to end users when they see a lot of downvotes, as if we are not keeping our articles up to date. That is a serious consequence in terms of brand trust.
But before I disable it, I want to understand exactly how this works with Answer Bot.
It seems like the downvote (and upvote) count has been used a lot more since we started using Answer Bot. Is that just because people happen to be going to the articles more often, or is the up/downvote actually tied into the AB functionality?
For example, in a support ticket, I see that someone has been shown 3 articles by AB. The person clicked one of them "not helpful". Did that article get a "downvote" from this interaction?
Let's say that the person is shown 3 articles and one of them is the answer they need, so they click "Yes, solve my ticket" (or whatever it says). Now did that article get an upvote?
If AB and voting are interconnected, I need to know how disabling it will affect AB.
Article votes won't affect the functionality of your AB's article recommendation. Whenever a user upvotes or downvote an article, the article gets a vote for that interaction. I tested this as well that a user can only vote once in an article regardless of how many times they visited the article.
I hope this helps.
I have enabled answer recommendation but the bot fails to suggest any relative published article. Can anyone help?
We also suggest that you check Best practices: Helping Answer Bot find the right articles more easily to improve article recommendations
Is there any way to add alternative text for Answer Bot to pick up? For example, we had a customer who didn't get our password reset article as a suggestion because her email said she needed a "pass word re-set". We don't want to write "pass word" in the article text, but would like it to get picked up for suggestions.
Alternate text can only be used if you have embedded photos and videos in your articles. Once you have opened the article in Edit mode, click Source Code. You can modify the value in alt= to have the additional keywords your prefer.
Is there any difference between optimizing articles for the widget and optimizing to be recommended in support tickets or will optimizing articles for the widget also optimize them for recommended articles in a support ticket?
The reason for asking is because the widget is suggesting the right articles but when I send a support ticket with exactly the same phrase it suggests wrong articles.
They serve the same purpose.
When you write about these 75 first words, which weigh more heavily, do these include the words from bullet lists?
For example, a client asked "Where can I buy your product? I live in Sweden". We have a help center article which lists all the countries in which our product is available (which includes Sweden and is in total less than 75 words). But the answer bot didn't suggest this article. Do you know why?
Why does the widget suggest correct articles but the support ticket suggests totally incorrect articles when exactly the same phrase is used?
I have created a ticket for you to investigate further. Please wait for my update via email.
How are you defining the Resolution Rate metric? Can that be added to the top of the article, as per the Suggestion Rate, please?
Resolution rate is defined as "The percentage of enquiries resolved by Answer Bot from the total enquiries where it offered suggestions." -- see Metrics and attributes for Zendesk Answer Bot. But I'll see if I can get it mentioned here :)
Like Denise, I am also wondering about what is counted as the first 75 words. We like to include a "subtitle" (heading 4) at the top of the article, and then a table of contents (linking to other sections of our article) before we start in on the content. Will those be included in the first 75 words? If so, that should feed the answer bot more precise information, but if not we'll need to find a different solution. Thank you!
I'd like to get an answer to Jana's question too. We have a lot of articles that start with a TOC before the content. Ideally, there should be a tagging option to select which part of the article text is most important, not just the first X words. Wrapping the text in a tag such as <ab>...</ab> (via a button, for our less code-y article writers) would allow us to control the text that AnswerBot is using, and include important things like a bullet list as noted by someone above. But on the other hand, if a bullet list is not being counted, then maybe that's a good way to format our TOC so it gets ignored? Would love to get some clarity on this. Thanks!
Hi Jana and Marci, the first 75 words of the article body are weighted regardless of any formatting or styling. So a TOC or summary can be a good thing, but keep in mind that the model is going to perform best with natural language so if you can phrase your TOC as questions, issue symptoms, or some other format that aligns with how your users ask about a topic, you'll see the best results.
Unfortunately, there's not a way for you to tell Answer Bot to ignore something in the first 75 words. We've run into this limitation at Zendesk ourselves - we used to include a disclaimer at the top of some articles acknowledging that translation was done by a machine, but this was confusing to Answer Bot. We have since experimented with including this disclaimer as an image instead, as well as just putting the disclaimer text at the bottom of the article.
Please sign in to leave a comment.