Introducing the Content Helper: From RAG to Riches
![](https://staging.wpvip.com/wp-content/uploads/sites/3/2024/03/image2.png?w=1024)
We are excited to announce the release of major updates to the Content Helper.
The Content Helper helps organizations use their content data and their performance analytics to get bespoke, performance-driven suggestions from large language models—and does it all from the WordPress editor where you already work. These new Content Helper features accelerate content production, improve results from SEO, improve audience engagement, and also increase the data fluency of your team.
We are using Retrieval Augmented Generation (RAG), which is a technique that organizations use to enhance the accuracy and reliability of generative AI models by specifying exactly the data they want the AI to reference when generating text or other outputs.
Our new Content Helper employs a RAG-based workflow that leverages both your existing content library and the performance metrics (page views, search traffic, etc.) connected to that content. It’s the difference between “Oh, another software with GPT” and “Wow, this understands our business and is driving search traffic!”
Before we begin, this feature is turned off by default and requires you to opt-in.
Enhance your content with AI to make it perform like you wish it could
Too much AI hype has focused on “magic” demos of writing entire articles from scratch. It produces content that is coherent and occasionally correct—but is it actually good?
The real opportunity is in using data to analyze content performance, ascertain best practices, and making it easy to apply them to content production. The Content Helper combines intuitive tools built for humans with AI-driven analysis no human could possibly achieve on their own.
Businesses today need to publish better content, faster. They have large teams of journalists or marketers, they have more content—and more data about the performance of that content—than they know what to do with. They think AI might solve their problems but they struggle to connect all of these things together effectively.
With the Content Helper, you can use your full content repository and the performance data for that content to get LLM-generated suggestions that are bespoke to your business and grounded in driving results. This is not model innovation—we’ll leave that to OpenAI, Anthropic, Google, and the like—instead, it’s data innovation.
An example use of the Content Helper
Our Content Helper is not asking GPT “Hey, suggest 5 titles for the text in this content.” There are dozens if not hundreds of tools for that workflow.
Instead, we—internally to our systems—have analyzed your content library to understand what every piece of content is about, and we’ve stored for you the performance data that your content generated (page views, search referrals, conversions, etc.).
We combine all of this to be able to say: “Find 20 of our articles that are similar to this current topic, take the 5 that generated the most search results, and send their titles to OpenAI along with the text of the article currently being worked. Ask OpenAI to create title suggestions that follow those examples that drove performance. Then, take the suggestions and make them easily available in the WordPress Editor for creators to either use or discard.”
It’s generative AI—but it’s focused, data-driven, contextualized to your business, and integrated in the space where you already work.
Key capabilities of the Content Helper
Today, the Content Helper generates Smart Links, performance-driven titles and excerpts, and brings analytics to your editing experience. Make sure to request access and read a full walk-through of its capabilities. In the meantime, here are some of the highlights:
- Direct integration with the CMS. Make data and AI-driven optimization available with a click of a button right in the CMS editor, resulting in faster and more effective content workflows.
- Automated Smart Linking for SEO performance. Click a button and the Content Helper analyzes all your published content, inserting links in your draft to your most relevant, top performing content—improving search results. Choose how many links you want to insert, and whether you want to insert them just in the block you’ve selected or across the entire article.
- Accelerated title suggestions. Suggest title options that will generate the most clicks via analyzing every piece of content a company has ever written and its performance.
- Automated excerpt generation. Analyzing all previous content, generate a content summary for new content that will draw the most attention and clicks on search engines.
- Integrated performance analytics. Bring data about referrers and other post performance to the editing interface, and make it easy to click into the full dashboard for deeper analysis.
How it works: a semi-technical walk-through
In this section we’ll be walking through the diagram below, to give some understanding of what’s happening behind the scenes with the Content Helper.
![A diagram showing the workflow of the Content Helper. Content Library and Content performance goes into the Editorial interface. The next step in the diagram has four parts: Smart Linking, Title Suggestions, Excerpt Generator, and New Workflows. Next are content ranking, content trimming, and Packet created for LLM interaction. At this step, the process interacts with OpenAI for functional data, reference data, prompt guidance, and prompt requests. Lastly, results are piped into WP Admin for discard or use. There are two prerequisites noted: the Parse.ly tracker installed on your site, and that Content Helper and AI features are enabled in WP Admin.](https://staging.wpvip.com/wp-content/uploads/sites/3/2024/03/image.png?w=1032)
The Content Helper uses content similarity and content performance to interact with AI to prompt titles, smart links, excerpts, and other results for the piece of content currently being created.
Data is not shared before a user prompts the features within the WordPress editor. Then, data specific to selected content is sent in the interaction with Open AI.
Prerequisites, your content library and content performance
Customers using WordPress VIP benefit from the platform both storing the things they have written, and automatically analyzing that content. This analysis is powered by our internal natural-language processing, which is built on top of open-source models.
This process is run internally on our infrastructure. It’s fundamental to our providing features such as Smart Tags, which suggests tags for your content, and our Content API, which powers recommendations on your site.
What this system does is calculate content similarity. It uses transformer models to create vectors for every piece of content. Think of it as mathematically calculating how similar various pieces of content are. A simplified example is in the diagram below:
![A diagram showing two documents: Document 1 says "Obama speaks to the media in Illinois." Document 2 says "The President greets the press in Chicago." Between these two sentences is a chart showing how similar keywords from these sentences are calculated.](https://staging.wpvip.com/wp-content/uploads/sites/3/2024/03/workflow-1.png?w=1032)
“Obama speaks to the media in Illinois” and its alternate, “The President greets the press in Chicago” are two sentences where none of the main words are repeated—but they are effectively communicating the same thing. Vectorizing your content enables us to calculate content similarity in this way.
We also provide measurement of all the results of user interactions with your website and apps. Page views, engaged time, search referrals, conversions, and so on. These metrics are also categorized not just by the URL they occur on, but also by the metadata attached to them: title, publish date, author, tag, section, word count, and so on.
When you combine these two systems, you can say that our platform effectively knows what your content is about and also knows which of your content produces better results.
This is what we think is so powerful, and this is what we are bringing to up-level prompt engineering with large language models. We can better prompt models like OpenAI, to produce performance-driven results instead of generic ones, and we can democratize access to those interactions via our WordPress VIP editorial interface, so your team doesn’t need to be trained on another tool in another system. It’s already where your team works.
User interaction in the editorial interface
Let’s move on to the second piece of the Workflow Diagram above, where the box says “Editorial Interface: User interaction with Content Helper AI features in WP Admin.” It’s important to note the order that this occurs in.
It is user interaction that prompts the process of sharing data with OpenAI. Additionally, our contract with OpenAI states that what we share with them cannot be used for training purposes. We are not preemptively sharing data with OpenAI, and we are not sending your entire library over to them for analysis. Instead, when a user is working on a specific piece of text in the WordPress editor, they then prompt our Content Helper AI features and begin the whole process.
Today, those features are Smart Linking, Title Suggestions, and the Excerpt Generator. More are on the way. When a user clicks the button to prompt those features, we prepare a packet of data to send to OpenAI.
Interaction with OpenAI
The exact details of what’s in those packets change based on which feature is prompted, and which text the user has prompted them for (the entire article versus just one block of text, for example)—but at a high level the process is the same:
- We take the selected text that is being worked on. This is your “functional data” in the diagram.
- We reference the database of your content and your performance data to determine the content that is both most similar and generated the most results. We trim that content—sometimes to just the first 300 words, sometimes to just the title, it depends on which Content Helper feature is being used. This is your “reference data” in the diagram.
- We send these two to OpenAI with specific prompts we’ve engineered to ensure consistent, usable results (things like following style, length of output, etc).
- OpenAI returns results for the prompt we’ve sent it, and we pipe those into the WordPress editorial interface.
- The user either decides they like them, or they don’t.
Of the features we just walked through, we are particularly excited about the Smart Linking feature, which analyzes the text you are working on against the performance of your historical content in order to automatically insert links into your content that will increase SEO performance. There are lots of companies working on helping people generate Titles and Excerpts, but we haven’t seen others working on Smart Linking yet.
See it in action—watch the webinar
We hosted a webinar with Penske Media Corp and 10up to discuss various experiments implementing AI into content publishing workflows—both what worked, and what didn’t.
You’ll hear best practices and get a quick demo, all in one spot.