As advanced personalized marketing technologies become more accessible, online retailers are better equipped to target individual users with content that resonates and compels, in a familiar voice. Personalized product pages with user-tailored words, phrases, sentences, and sections will convert more buyers and create more engaged and loyal customer relationships.
Instead of using a human brain to analyze how users interact with language, we use a linguistic neural network that communicates performance and scoring data to unique Natural Language Generation Personas. Each NLG Persona is programmed to write in a unique author/brand voice, using machine-learning to decide what to write and who to write it to.
Much like the brain of a human writer, Writesof uses a vast linguistic neural network. Language is broken down into segments, or strings, each containing a set of values. Language decisions occur, firstly, on conditions of who the language is communicated to, and secondly, the intended communication objective. Like humans, Writesof’s systems learn based on knowledge of other human communicators – writers and respondents. If a product page is of the writer, the message sender, customers are message receivers and respondents.
When a customer clicks on an ad or visits a product page, their responses (on-page engagement and actions) reveal user feedback to the message sender. In our system, we can interpret and analyze this type of user feedback much better than a human can. In this sense, our system is far more informed about product communications and customer feedback than even large content writing teams. And our computational linguistics systems also analyze unconventional message senders and respondents, such as search bots, ad impressions, and internal feedback sources. Feedback is the only way for a person, or a machine, to learn the precise manner in which a message is received, understood, and responded to. By harnessing the power of customer data,
Primarily, Writesof analyzes human language inputs sourced from product page text and metadata (brand and user-generated), social media, and product ads. Language is segmented by site, page, brand, category, authors, and by user-consumer. Along with social media and product ads, product and category pages represent the majority of an online retailer’s communications fingerprint and footprint.
To operate with optimal predictive capabilities, the editing system requires high volumes of on-page user engagement data. If your organization sells thousands of products, with, at minimum, average conversion rates, you’re probably a good match. As we create new product page versions and begin split and multivariate testing them, even a site with fewer than 10,000 products can begin testing millions of words of fresh content within a matter of weeks.
Of course, our system does not automatically write and edit content by itself without input from human writers. It first requires a lot of training and tuning, with respect to brand voice, author style(s), as well as message mechanics, objectives, and grammar structures. Over time, with each new page-visit, the system learns how different types of users interact with and respond to different types of language.
As an illustration, if you’re an online retailer with 100,000 product pages and healthy traffic volume, that’s roughly 200-500 million words of text that users interact with – all high-value data. Writesof’s machine learning capabilities function better with larger volumes of user engagement and conversion data. We use this data to calculate correlations between language features and customer behaviors. With long and diversified product assortments, language is more diversified, enabling our system to learn how certain subjective and objective features and forms of language influence different types of customers. User data updates in near-real-time, which means our NLG systems can make editing decisions on a daily or hourly basis. With a simple adaptor API, it can even deliver refreshed and revised product page versions to your Product Information Management system, just as soon as our system processes the updated user data. This will enable you to rotate through product page versions by time of day, as well as periodic events, promotions, and seasons. It can even automatically rotate landing page versions, testing a wider variety of language with different types of users. This enables our systems to perform A/B and multivariate analysis on a greater variety of message features and forms. In most cases, we have found that this type of analysis does not necessarily improve our system’s predictive abilities, except in certain cases with highly diversified product offerings of fewer than ten thousand product pages. This is true because our system does not analyze product pages – rather, it analyzes words and word usage, with respect to audience interactions. Our systems are constantly analyzing every piece of language for its perofrmance values as that piece relates to every other piece of language on a page, in a category, site-wide, even measuring off-site usage and performance. particular product categories are offered with with minimally diversified product offerings of fewer than five thousand SKUs. with one exception being highly diverse product offering, comprised of fewer than 10,000 SKUs. For especially long product lines, this is unnecessary, as our systems already operate by deciding which version performs best based on performance data during that time period. Seasdaily and is even capable of automatically publishing edited content without human oversight.
Our automated editing solution also prefers to learn from several human writers, as it can extract a wider range of language features and functions data, assigned to a machine persona – a program in our NLG that enables us to restrict content generation to certain words and grammars used by a specific author. Machine personas are built by the very text that an author publishes, given that the client has tagged an author’s works and that you have access to at least hundreds of pages of author-specific content.
parse out compositions by author. highly diverse language – several human authors with different styles. interpret larger volumes of user engagement and conversion data, networked to a more diversified language set. Generally speaking, our NLG systems learn by analyzing user engagement with specific words and phrases in their respective locations within the content. The system then processes the engagement-language data relationships to make probabilistic decisions on how to write and edit content – for specific products, to specific users or user groups.
Our machine learning model utilizes logic that we call programmatic incentivization. When chunks of content perform well with certain users in a certain product category, the system is rewarded with additional editing flexibility. When chunks of content perform below existing or prior benchmarks, the system restricts editing decisions for that user group and category. If an automatically edited chunk of content performs poorly again, that particular chunk of content is sent, in a prioritized queue, to a specified editor for supervision. In most cases, human editors will see a list of edit suggestions, prioritized by the predicted value of a language chunk’s usage, in context with the page’s content.
Authorized human editors can override any machine-generated content, on any page, at any time. Language is never removed from the system. Even poorly performing language is valuable, as we use its correlative data to influence writing and editing decisions for current and future projects. We actually encourage human editor input, as each time an expert makes an edit, high-value language chunks are parsed and classified into our systems, where they are analyzed for engagement performance just like every other word on the site. All edits, whether performed by our NLG or by a human, are logged and tagged with valuable metadata that helps our systems understand how editing decisions are influenced by various conditions and input sources.
Our systems essentially transform human edits into high-value data that influences future writing and editing decisions. Please note that high-quality Natural Language Generation must have a source of intelligence. For this reason, in order for an organization to qualify for our automated content editing solutions, it must be able to demonstrate healthy traffic volume and conversion rates on thousands of product pages. At this time, we simply don’t have the resources to provide value to companies that are struggling with conversion metrics. If your conversion rates are around averagehealthy, we can substantially improve them
Note: High quality, fully enriched product page content in. , also making better decisions, over time, about how to write content and writing new product pagecontent that uniquely targets users and user groups. and conversion data. the best results with different users. , firstly, based on who you are writing it to, and secondly, what you are writing about. The who part of the forbased on their understand a message’s mechanics and attributes, and you understand the reader’s mechanics and attributes, the mechanics and subjective attributes of a particular message, all of its parts, functions, and reader perception, andcan reveal immensely valuable data can reveareveals challenges and opportunities at all points of the customer journey. It helps writers make better decisions in product page user and search optimization. It provides valuable insight into user interests, attitudes, and behaviors that help online retailers refine their targeting tactics.
If your organization is actively focused on off-site and on-site personalized marketing, Writesof can help you unlock your data’s true potential with a personalized messaging solution that generates full product descriptions, from title to meta tags. Our Natural Language Generation systems are uniquely designed to write about product features and benefits, in rich detail and with persuasive artistry.
We deliver personalized content that is more effective at converting buyers. With our product descriptions, customers will remember your brand. They will remember how informative your product pages are. If you diversify product page messaging by user, customers will find your content more relatable, creating new opportunities for after-sale engagement, user-generated content, reviews, and return purchases.
