Velocity - URL to AI Videos in One Click

About Velocity

Velocity is an AI powered URL to video generator for e-Commerce products. We had a soft launch for the product to test the waters and to gather valuable insights from potential users and early adopters.

The tool targeted at improving product discoverability through creating product videos and faster iterations for marketing ads to boost sales, all from a single click.

It was meant to be a simple tool for people with minimal tech or creation skills.

Userflow (high-level)

User registration

Video Generation

The beginning of the problem

Post the first month of the soft launch, we saw that, with our targeted marketing and through managed sales, we were able to get users to our platform but the churn was significant.

96% of the users drop off entirely within 28 days
Over 60% dropping off after the first use

We were aggressive in our definition off drop off since it was a new product testing the waters so the drop off consideration was for users who did not use more than 1 session in the first week of their use and 0 sessions in the consecutive weeks.

We really needed to get to the bottom of this. So we had to define a strategy to tackle this.

The strategy to understand the problem better.

1. Check existing data

   1. Check CTR from landing page to sign up

   2. Check number of videos generated per person on average

   3. Total number of videos downloaded

   4. Check full story recordings for chosen profiles (atleast 10)

   5. Like per videos generated
      
      

2. Identify known users who churned or prospective users to try with and observe user’s behaviour during the session and listen to what they were talking about

   1. 10 users recruited for an interview with an incentive
      
      

3. Solve for each problem individually to improve overall usability strategically

   
   

Understanding where we went wrong

Observations:

Key insights from the data:

• 40% of users did not create a video after sign up
  
  

• Many created a video, but never viewed it due to the long generation time
  
  

• People who generated videos clicked “Regenerate” a few times, but never downloaded a video
  
  

• More dislikes on generated videos than likes

From user interviews :

1. Users who landed on the page or signed up fell mostly in the category of:

   1. Performance or E-commerce marketers (>50%).

   2. Small business owners selling a niche product line (>30%).
      
      

2. They were attracted by the ease-of use and single click generation USP, meaning that they did not have to depend on Marketing agencies or UGC creators for video creation.
   
   

3. The text generated by the videos were not impressive or catered to their product’s language.
   
   

4. The colours used in the generated videos did not match their brand language and there was no way to do it.
   
   

5. Users who did not sign up did not completely understand what they were getting from the landing page.
   
   

6. Both groups seemed to want easier ways to generate videos as per the Ad management tools they were using, like Meta Ads, etc
   
   

7. We received an initial SUS score of 48 which was extremely bad.

   
   

Diving deeper and diverging into solutions:

1. There is a drop off at the landing page due to not knowing what users were going to get, also users did not understand what link to provide or what data to provide.

   1. Maybe a try feature before login

   2. send the generated video via mail

   3. Try sample links
      
      

2. Users wanted better copy generated for their videos

   1. We needed to rethink the Agents used for phrase generation and template selection

   2. Images that were used in videos were of lower quality
      
      

3. Even with good copy, users wanted to make minimal tweaks on the video, without going into a complex editing interface
   
   

4. Wanted easier integrations with their ad management tools
   
   

5. There was a high failure rate for many links from random marketplaces

   
   

Prioritizing what to solve for:

After understanding the key problems we needed to prioritize what to fix and how to tackle them quickly and efficiently.

I prioritized 3 of the problems to fix first, which had the most efficient chances to cater to improve the user metrics of the product.

1. Problem: Discovery and pre-user sign up issues

   Solution: Allow users to create 1 free video before signup
   
   

2. Problem: The videos generated by our Agentic AI pipeline did not cater to the needs of the target audience.

   Solution: Improve the text generation, image generation and template selection pipeline in the Agentic AI work flow. Update rules to cater to the target audience better.
   
   

3. Problem: Even when the videos looked good, users wanted to make minor edits for faster A/B marketing and minor edits.
   Solution: Create a simplified edit feature to allow users to customize text, update images and choose options for audio and voiceovers, easy enough for a marketer or small business owner to be able to edit videos to their liking.

1. Fixing Discovery and pre-user sign up issues

• I had a discussion with stakeholders and persuaded them on the need to allow one free video generation, which could act as a hook for users to understand what our tool does, we were already giving users 50 free credits post sign up, and the stakeholders were convinced that it was worth trying.
  
  

• We needed to keep the generation flow simple enough for low-skilled and explorative users.

  
  

Updated User flow:

Made quick wireframes and did some quick static testing to see hoe the flow would work. And then moved on to the visual designs

User flow update with screens:

Additional considerations:

From a business perspective we had to make sure we restrict generations and downloads, since each generation costs $0.2 to the company. The considerations and additional tech implementation included :

1. Block download for videos, and allow download only after signup.

2. Lock video generation to one video per device, to restrict bad actors from misusing the "one free video" offering.

Success metrics that were considered:

1. Lead conversion rate for Discovery to video creation

2. Lead conversion rate for Discovery to sign up

2. Making our Agentic AI pipeline cater to user needs

The AI pipeline needed to cater better towards the type of content that the user needed, without hallucinating, and making sure that the pipeline would make the right decisions in choosing the right templates, colors and text for the product it is generating it for.

One thing that was certain was that we needed more templates for the AI agent to choose from, we had an internal team of Motion designers working on creating more relevant templates as per the needs of the target users.

To test the best possible phrase generation, we created a prototype that generated phrases from scraped data of PDP links submitted to it. And the LLM will score each phrase generated across 8 parameters. Then we asked 10 relevant (internal and enterprise) users to score it according to them, of a fixed set of diverse PDPs.

1. Created a quick prototype to A/B test 2 different models across several parameters, tested with 8-10 target users

[Screens to be added]

2. Annotate our templates to cater to different types of products

   1. Which meant we needed a better Annotation tool - This required a longer process - so parked for later ( but was done - will create a separate case study for this )

   2. Categories needed to be marked by experts

3. Update the colour logic used for the video updation as per the product colours

3. Users wanted to make quick edits to the generated videos

The next thing to try was to add a minimal edit interface:

We looked into the type of users we had, who were not creatives but users who had an eye for quality content. They would not want to get their hands dirty with video editing in detail, they wanted quick results.

So we Introduce a minimal edit feature to cater to the needs of the users we were targeting.

The key features we needed were

1. Basic with text edit and image edits

2. Simple to control audio genre and voiceover themes

The outcomes

1. The new flow for a pre-logged in generation improved user’s initial adoption rate

   1. 70% of the users landing on the page created a video pre-login

   2. 40% of users landing on the page logged in as opposed to 6% before

   3. There was still room for improvement

2. We found out that Claude 3.7 made more consistent marketing copy though Gemini 2.0 was more creative, it halucinated and came up with unwanted text

3. The videos created were of better quality and performance marketers appreciated it

   1. But they needed more control - more like a brand kit - parked for later (separate case study)

[ 2+3 increased the likes to dislike ratio on the videos ]

1. The edit feature was used by over 20% of the users before downloading

2. The SUS score rose to 64, which meant that we were moving in the right direction with most features

3. The landing page still caused a little confusion, but it performed better

More enhancements

1. A confidence score

2. A brand kit for consistent video content in the brand language

3. Thumbnails to identify videos better

4. Better version history

5. More integrations with Ad managers

6. Text better suited for marketing insights

7. Improve the template creation and annotation pipeline for templates (since we needed to increase the options that the AI agent was able to choose from)