This document describes all places where AI features are available in Webtrends Optimize. Note that the content of this will rapidly change as we continue to build on the product, with AI being a core focus of future developments.
Introduction
Webtrends Optimize uses AI features with some core principles for every single feature that we build.
First is the inability for our providers (Claude/OpenAI/Cloudflare/etc.) to train their models on yours and our data. This is a core agreement tier built into every subscription we have.
Second is to self-host wherever possible, which puts a physical space between the rest of the world and your data in what we refer to as Sovereign AI. For many features, we have our own models, run on our own hardware, and so your data is not shipped out to large API providers which is very common with other platforms. This is easy to do for features that are not time-critical, like processing your Survey responses into sentiment scores, and generating Predictive Heatmaps.
To summarise - Q: Is customer data used to train WTO models, or those of services WTO uses? A: No. Not for a single service we provide.
Regex Checker
What is it?: Customers can write regular expressions in our UI to target pages. E.g. /p/\d+ to match pages like mysite.com/p/123456. Many customers struggle to write these expressions. An AI regex checker allows us to provide some guidance to you, ironing out bad practices and hoping for easier success.
What data is used to power this? We have our own ruleset of things to look for. Your specific regular expressions are fed into the model we use along with our rules as a list of things to look for.
Predictive Heatmaps
What is it? People are likely familiar with heatmaps for click behaviour and scrollmaps for scroll behaviour. These look at data collected and paint them onto a page. Predictive Heatmaps in WTO takes predictions of a user's visual attention (similar to if you ran eye-tracking studies in a lab) to predict where people will look on your web page.
What data is used to power this? We have our own model. It's self-hosted. Your input screenshots are "painted" on with our heatmaps and returned to you. We do not use your screenshots to train our models.
Synthetic User Testing
What is it? This is an assistant to help you run UX heuristic teardowns of web pages, and look for new test ideas. Specifically, it focuses on personalisation, and so all teardowns and ideas are based on user personas.
What data is used to power this? All data used for this feature is currently only using what the AI can publicly see from your web pages, and what it might have been trained on as general knowledge. E.g. it knows that you're a bank/airline/retailer/etc, what market you're in (budget, premium, etc.), what competitors you have etc. We do not use any of your data to power the generations found in this feature.
In the future - users have asked for a way to add their own personas, provide their own customer research from which we can interpret personas, etc., and so this feature will in the future allow you to input your own internal data to generate personas. We will address security as a primary concern when building that next phase of this feature.
AI Coding Copilot
What is it? In our coding editor, much like Github Copilot - as you type, we will offer suggestions of what the next block of code could/should be as an autocomplete-style interface.
What data is used to power this? We collect the code you're working on to generate what next line/s of code we suggest. These are only used at the time of generation and are not stored. No customer data is collected for training purposes.
AI Visual Editor Copilot
What is this? In our UI, we will soon have the ability for users to be on their web page (via our WYSIWYG editor) and enter a prompt to transform the page instead of clicking and manually changing things themselves. E.g. "change this heading to say X" "create a new batch of quick-links here" etc.
What data is used to power this? Your prompt, the page and screenshots of it, will be fed into a series of composed AI prompts. But not retained for training purposes.
Visually-similar Product Recommendations
What is this? We typically generate product recommendation carousels based on product feeds, where we recieve attributes like name, price, brand, colour, etc. Many customers have poorly-tagged product feeds, and so this feature scans your product photography, infers many attributes about each product such as brand, colour, style, shape etc. and uses that as the basis for "Similar products to this" recommendation carousels.
What data is used to power this? We need you product photography, and the embeddings we generate from this are all pooled into a large Vector Database. The sharpness of our recommendations is based on our ability to store and retain access to your photography. No other data is required to power this feature.
AI Reporting Insights
What is this? When in our reporting, many insights are buried behind filters e.g. Mobile Safari is performing badly, users in Manchester and London really like Variation 2 etc. The effort to find these insights is high manually, but easy with AI. Our feature applies a large number of dimensions to your dataset, applies our usual AB Testing statistics, and then sends the data off to an AI to summarise where we can look for winners and losers of subsegments. Data is only sent for tests where you click an AI Insights button in the UI.
What data is used to power this? Your test performance data, specifically views and metrics captured. We do not send any test metadata (what you are testing, why, where, the code used, etc.) as part of this process. And no data is used to train our models or those of our suppliers.
AI Survey Analysis
What is this? We allow users to run surveys in our platform, but analysis of the captured data is typically very manual. This feature allows you to:
Understand and interpret sentiment scores of your feedback, build averages etc.
Extract common themes (poor ux, excessive popups, etc.)
Find all responses that match a given theme, e.g. "show me all responses related to excessive popups".
Note that this feature will be available in late 2026, and will not be enabled for any survey by default.
What data is used to power this? Your survey feedback is captured already in WTO, and this data is ran through several models to run sentiment scores, extract themes, etc. This is run on sovereign hardware and no data is retained for training purposes.