Case study · Business listings

Moving listings
to a new vendor

The platform had to move its business listings to a new vendor by a fixed deadline. Choosing the vendor wasn’t my call, but how we’d switch was. I set the usability bar every vendor had to meet, got an AI requirement into the scorecard, specified a layer that keeps the experience the same whichever one won, and built a live prototype that tested all four vendors.

Watch the tour

CompanyThryv
RoleDesign & research lead
Team4 PMs · 2 designers · 2 development teams
Timeframe2026
FocusSystems · Research · AI · Mobile
TL;DR

The platform had to move business listings for 160K locations to a new vendor by a fixed deadline, and no one had defined what “don’t make it worse” meant. Choosing the vendor wasn’t my call, but how we’d switch was: one usability bar every vendor had to meet, an AI requirement in the scorecard, and a layer that keeps the experience the same whichever vendor won. I built a live prototype that tested 4 vendors’ APIs and AI. The vendor is chosen and launches in 2028, so there’s no adoption data yet.

At a glance
4
Vendors’ APIs and AI tested in one live prototype
3→1
Studies combined into one usability bar
160K
Locations in scope
01 · Context

Two years to switch vendors

The platform had to replace its listings vendor on a fixed deadline, with the new partner fully live by 2028. I led design and research, and my job was to make sure the switch didn’t quietly make the experience worse.

The decision had four parts: which vendor and what kind of integration, how the change would affect existing users, what a new backend would make possible, and how to get PMs, dev leads, marketing and support to agree.

Who we designed for

Our persona was a roofer who runs her own business. We scored every option on whether it would help her, rather than on what the platform could technically do.

What businesses are paying for

They want to show up when someone asks an AI search tool, not only in directory listings.

A gap in the research

None of the earlier studies covered reconnecting listings to a new vendor, because that flow didn’t exist yet.

One lens in the middle, a roofer running her own business, with four evaluation epics around it. Traditional citations: show up with the right info without updating 50 or more directories by hand. AI search citations: be in the answer when a homeowner asks an assistant for the best roofer. Reviews: replies go out without her writing them. Insights: tell her when she is falling behind competitors and what to do about it.
The four evaluation epics, each scored against what helps a roofer run her business. That’s why the vendor with the most features didn’t rank first on fit.

The scorecard rated each vendor on three things: how much our UI would have to change, API performance, and AI capability through an MCP server. The final vendor decision wasn’t mine.

02 · My role

Design and research lead on the evaluation

I owned design and research on the proof of concept, and worked with one other designer so research and prototyping could run at the same time.

01

Combined the usability baseline

Our listings screens scored an A in three earlier usability studies. I combined those three studies into one bar any new vendor had to meet. Before that, “don’t make it worse” had no number behind it.

02

Specified an abstraction layer

I specified a layer between our interface and the vendor’s API, so the frontend would work with whichever backend was chosen.

03

Put an AI criterion in the scorecard

I got the AI interaction model added to the evaluation. Does the AI give her a draft to approve, or an empty box to type into? That became a requirement instead of a nice to have.

04

Built a live prototype

Instead of mock data, it connected to all four vendors’ sandbox APIs, and through MCP to their AI. Every status and score on screen was live, so we judged each vendor, AI included, on how it actually behaved. The prototype called each vendor’s live sandbox API, and the sandboxes were seeded with demo businesses. That’s why the captions say demo data.

The AI assistant screen with suggested prompts including What should I prioritize fixing right now, Draft a reply to my worst recent review and Which listings have issues
The AI requirement in practice. Each entry point gives the user something to approve or decline instead of an empty box. Approving a drafted review reply takes a few seconds. Writing one from scratch often never happens.
Thryv AI comparing Apple Maps with the business profile, drafting three corrections for phone, Saturday hours and category, then the fields moving from Needs attention to Sent, waiting for Apple Maps
Concept, in the interactive prototype. Thryv AI drafts the corrections for a listing that doesn’t match, and nothing is sent until the owner approves. Sent fixes wait for the site to update before they count as matching.
The research and design plan in six steps. Goals and criteria, synthesizing prior research, vendor testing with users, designing the experience and building a live prototype on real sandbox data are complete. The last step, validating and prioritizing for score, feature priority and MVP scope, is in progress.
The research and design plan. Steps 0 to 4 were done before the vendor decision. The last step, testing with real users, sets the score, feature priority and MVP scope.
03 · The abstraction layer

Keeping our UI separate from the vendor’s API

The integration raised three design problems. The vendor API returned fields that didn’t map cleanly to our patterns. What it could return limited what I could show, including partial states, delays while content generated, and AI output of varying confidence. And small business owners had to trust that AI written content described their business accurately.

SMB user reviews & approves Abstraction layer · my design surface Vendor A backend Vendor B backend
Two of the four vendors shown. The same layer works for any vendor.

The abstraction layer was my call. It meant that whichever vendor won, the frontend wouldn’t need to be rewritten.

Switching the prototype backend from Vendor A to Vendor B: the listing rows sync, and a note maps each vendor’s own status values to the same four statuses the owner sees
The layer in the interactive prototype. Switching the backend from Vendor A to Vendor B changes the status values the vendor sends, and the layer maps both to the same four, so the screen stays the same. Values from the vendors’ sandboxes, on demo data.
Listings Management: a table of directory platforms with Connected, Processing, Issue Found and Not Connected statuses, beside a business profile completion score
The listings screen the layer protects, captured from a demo account.
04 · Research findings

Three findings that shaped the design

Hours didn’t match how people think

People expect to mark a closed period inside their open hours, not split the day into two blocks. This came up in two separate studies.

Wording already confused people

Our existing terms already made people hesitate. Adding the vendor’s terms on top would make it worse.

People couldn’t tell what would happen

Before clicking, people couldn’t tell the difference between disabling a listing and disconnecting it.

Hours editor in the prototype: Monday open 7 AM to 5 PM with a break from 12 to 1 PM shown as a gap on the day’s timeline, and an Add a break link on every other day
Hours with a break inside the day, in the interactive prototype, so a lunch hour is one closed period instead of two separate blocks.
Confirmation dialog explaining that the two actions differ. Disable keeps the listing connected with its data preserved, and it can be turned back on any time. Disconnect removes the listing, stops reviews syncing, needs verifying again and cannot be undone.
Each option now says what will happen and whether it can be undone.
05 · Mobile

Reviews and posts on a phone

Business owners reply to reviews and check how their posts did on their phones, so the reviews and social pages had to work on a phone too, with the same Thryv AI drafts as on desktop.

Recorded from the interactive prototype, on demo data. These Thryv AI flows are a concept.

The worst review first

Thryv AI opens with what needs attention: the negative reviews still waiting, and which one to answer first.

A reply in one tap

The suggested reply is ready when the review opens. Switch the tone or regenerate it, and Post reply stays pinned within reach.

Posts at a glance

Social leads with what failed, then how posts did, then the comments still waiting, each with a drafted answer.

The reviews inbox on a phone: a 4.4 average from 86 reviews, a Thryv AI brief about two negative reviews, and the reviews that need a response
A one star review from Alex M. with a suggested reply, tone choices for Friendly, Professional and Empathetic, and a Post reply to Google button
The social page on a phone: an alert that Instagram didn’t publish a post, reach, likes and comments for the last 30 days, a customer question to answer, and upcoming posts
A published post on a phone with its reach, likes and comments, and a drafted reply to a customer’s question about free estimates
The owner’s phone, in the interactive prototype on demo data: the reviews inbox, a suggested reply with three tones, the social page, and a post with a drafted answer to a customer’s question. These Thryv AI flows are a concept.
06 · Where it stands

No launch numbers yet

The vendor has been chosen, but the new partner doesn’t go live until 2028, so there’s no adoption data yet. This is what I can show so far.

CompleteThe usability bar every vendor had to meet, and the AI requirement in the scorecard. Both were mine.
CompleteProof of concept execution, integration setup and testing across all four vendors.
CompleteScoring, internal alignment and the vendor decision gate.
NextMigration planning and rollout, with the new partner live by 2028.

Let’s talk