Personalizing Core Experiences

Outcome: Increased average 7-day retention from 2 to 4 days for a global product with 40 million DAU

Grammarly is a one-size-fits-all product. All users get the same experience regardless of context. Grammarly’s Daily Active Users were steadily declining. To improve engagement and differentiate the core product in the competitive AI landscape, we introduced personalization. Grammarly learns and adapts to users’ preferences and context.

I rallied executive support to bring ML partners on board and teamed up with product managers across the core product to push strategic changes forward. With a war unfolding in the background and our engineering team impacted, I stepped up—reworking the roadmap and finding scrappy, creative ways to keep delivering without losing momentum.

Team

Services

Strategy
FACILITATION
1 → 2
Interaction Design

DATE

2022 - 2023

ROLE

LEAD PRODUCT DESIGNER

Audience & Problem

Grammarly has helped millions of students polish essays, emails and stories with suggestions around tone, vocabulary and structure. But as Gen Z pushed for authenticity and a cultural reckoning highlighted the need for inclusivity, it became clear: helping students succeed also means honoring their unique voices.

The relevance model for suggestions prioritized accuracy over nuance—leaving users to sift through suggestions that didn’t always match their context. Sometimes, it missed the mark entirely.

Like Abigail, who lost a parent to cancer. Grammarly suggested she write “parents” instead of “parent”—a small change that felt deeply insensitive especially when she was writing her personal statement for college.

It was time to build a product that not only corrects, but connects.

Problem

  • Reduce friction during editing
  • Strike a balance between the users’ natural voices and “objectively” good writing

Goals

  • Learn and adapt to users’  preferences and context
  • Shift perception that Grammarly is a trusted AI writing assistant
  • Metrics to keep an eye on: DAU, 7-Day Retention, interactions per session

Approach

Strategy & Kickoff

I dug into our core users' pain points to spot where personalization could really shine—and landed on revision as the sweet spot. It was key for user value and measurable business impact.

Once I had the context, I ran workshops to align stakeholders around a clear north star. I also set guiding principles to keep our team grounded—especially when navigating data collection discussions—so we stayed respectful of user privacy.

Creating Effective Permissions Requests

Grammarly never stored user’s text. But with a personalized relevance model, our team needed to store user’s text to learn their voice and progress overtime.

I worked on educating our new and existing users the value of a personalized revision experience and the impact it can have on their writing to get their permission.

I presented these requests at the appropriate time and gave them the option to reverse their initial decisions to reinforce trust.

Implicit & Explicit Feedback

We knew users weighed audience, context, and personal style when reviewing Grammarly’s suggestions—but we wanted to dig deeper.

I led in-product surveys to uncover what really influenced their choices and improved feedback mechanisms within the core product to capture those insights. That data helped shape the personalized relevance model.

Personalized Revision

As Grammarly’s personalization model got smarter—learning when users meant to break rules or lean into their unique voice—I explored how to make personalized suggestions easier to engage with.

The existing interface funneled users into a rigid, linear edit flow, putting grammar first and leaving structure, flow and style in the dust.

I proposed a more flexible assistant experience where users could tailor suggestions to what they cared about. Due to team capacity shifts during the war in Ukraine, I re-scoped the idea to fit within the existing UI—still delivering a personalized revision experience without a full overhaul.

  • 🧪 Experiment 1—Inline Editing: I prioritize clarity suggestions directly within the text.
  • 🧪 Experiment 2—Contextual Options: I introduced a range of rewrite options that aligned with each user’s voice and writing context, giving them more control without sacrificing authenticity.
  • 🧪 Experiment 3—Style Bundling: I explored stylistic editing by automatically grouping suggestions; those that matched the user’s voice were bundled for quick acceptance, while mismatched ones were separated for easy review or dismissal.

Experiment 1 - Before

Experiment 1 - After

Experiment 2 - Before

Experiment 2 - After

Experiment 3 - Before

Experiment 3 - After

Experiments 2 and 3 were a win—users checked their text more often and spent more time engaging with suggestions, so I shipped them to GA.

Experiment 1? Not so much. Inline suggestions disrupted the writing flow, so I rolled it back and proposed an expanded Grammarly Assistant UI when users were reviewing longer pieces of text to make advanced suggestions for text structure more approachable without getting in the way.

This alternative surfaced needed edits by structure, flow, and style—helping users prioritize. The expanded UI increased text checks and encouraged deeper review, boosting overall engagement.

Sharing fun feedback to build loyalty

As Grammarly’s personalized relevance model got smarter about each writer’s unique voice, I turned those insights into playful, Gen Z–savvy moments that felt personal, motivating, and shareable—keeping users engaged and coming back for more.

Results

Impact

  • Average 7-day retention increased from 2 to 4 days
  • 14-day premium cancellations decreased by 9%
  • Uninstalled rate decreased by 17%
  • CSAT score changed from dissatisfied to satisfied by 18%
  • Interactions per user per session:
    • frequency of text checks: 5 text checks (target = 3 text checks)
    • time spent reviewing suggestions: 2.6 minutes (target = 1 minute)

learnings

Design
  • For personalization to be successful, AI  must be in sync with the user’s context and preferences often (who I am today is different than who I was yesterday)
  • Not all parts of the core product need to be personalized, users don’t want AI to be a “clutch” that keeps them dependent
  • A risk management plan for changing data collection policies needs principles and visual diagrams to ensure all XFN partners have a shared language to discuss the implications to the business, product and user
Leadership
  • Rescoped and renegotiated with stakeholders the critical components needed to make this experience successful for both users and Grammarly
  • Balanced short-term design and long-term design priorities while also hiring
  • Collaborated with the growth team to deliver parts of this experience due to the war  (Engineering team was based in Kiev, Ukraine)

Next Steps

  • Revamp Goals to also ensure that the personalized relevance model aligns with users writing aspirations (grow vocabulary)
  • Reinforce throughout the user experience why users should personalize and the impact it can have in real world outcomes