01

Context

Flickr was a photo and video hosting service and an online community for photographers. Flickr’s main revenue source was a subscription service called Flickr Pro. We wanted to improve the service to attract new subscribers and reduce churn.

“Magic View” was an image recognition and object categorization feature launched in 2015 on Flickr. It was deprecated a few years later for cost reasons. We had a strong intuition that reintroducing image recognition as a Pro feature would help us achieve our goal of making the subscription service more attractive.

Note: Confidential strategy from this project has been obfuscated. Charts, documentation, and product direction shown here are not the actual ones from the project.

02

Research

Our first step was to explore the field of opportunities that could make Pro more attractive. This would ensure we would prioritize the strongest ideas, which may or may not have included image recognition. We used our personas to drive discussion with the team. We discussed what each persona valued, what their needs were, and what they might be interested in paying for as part of a subscription service.

The top ideas from the brainstorm covered a wide scope, from feature redesigns to partnerships and integrations. They all centered around “selling the dream” of what the Pro subscription could include for each persona.

Each idea was mocked up in high fidelity as a believable new feature or product change. We ran these ideas by our free users in a survey to gauge their interest. The survey results made it clear that better organizational tools were indeed a critical need that our free users were willing to pay for.

We did a deep-dive into understanding organization and search behaviors via another survey, as well as started a discussion with our Alpha Group, an audience of enthusiastic users who are consulted for early feedback.

This research revealed that integrating automation of certain organization tasks via AI/ML tools would be an impactful improvement to alleviate the biggest user pain points.

To prove willingness to pay, we ran a fake door test on the website with our target audience. The data showed that even with our lowest confidence conversion rate, we would easily reach our ROI targets for the feature. It also gave us a list of early adopters who could help us test the new system.

03

AI/ML Models

We looked into pricing for a few popular image recognition services, as well as weighed the pros and cons for using the third-party services versus building our own.

We discovered that building our own service kept costs down, had favorable data privacy, and maximized ROI for the feature.

We used a combination of three premade non-fine-tuned models to craft our custom image recognition system.

04

Integration

The new image recognition service would allow us to re-introduce a “Magic View” to the user’s photo library. We would also highlight this feature in search, so empty results would inform users that our system could find objects and styles in their untagged photos if they were subscribed to Flickr Pro.

05

Next steps

The system was built to scale and improve many aspects of the product experience. Depending on the business need, we could:

This project was paused for business critical work.