Designing conversational AI map search for PayByPhone
A search bar that understands sentences like "cheap parking near English Bay, 15 minutes' walk, for 2 hours", so drivers can plan parking without fiddling with filters.
Role
· Senior Product Designer
Partner
· Product Management
Company
· PayByPhone
Focus
· Conversational AI
Mobile screen
Standard search
Mobile screen
AI Search
Mobile screen
Results
66.5%
of people who open the map leave without picking a spot
12.5%
of monthly users open the map tab at all
1,300+
cities, each with its own parking rules
Overview
Finding parking on the map was hard
PayByPhone works in over 1,300 cities across North America and Europe, and handles millions of parking sessions.
Most people already know where they parked. They see a sign, type in the zone code and pay. But drivers who want to plan ahead or look around got a map that made it hard to compare options, and most of them gave up.
The question
Could drivers just say what they need (price, walk time, how long they'll stay, permits) and get a good answer, without cluttering the mobile map?
The problem
Most people never get through the map
PayByPhone was built for people who arrive, enter a zone code and pay.
Amplitude data showed that only 33.5% of people who open the map end up selecting a spot. The other 66.5%, about 584,158 people, leave without one. Only 12.5% of monthly users open the map tab in the first place. The other 87.5% skip it and type in a zone code.
People who do browse decide quickly. Within 30 to 60 seconds they either spot something that works or they close the map.
Why filters fall short
Parking prices and rules don't fit into a filter
01
Street prices keep changing
On-street rates change block by block, by time of day and during events. Many are tiered, like $1.02/hr for the first two hours and $1.71/hr after that. A single flat price on a pin is misleading. And building filters for 1,300+ cities with changing rules would be heavy to build and hard to keep up to date.
02
Too much to set at once
Drivers think about where they're going, their budget, how far they'll walk, how long they'll stay, overnight rules and permits, all at once. Asking them to set that many dropdowns on a phone screen is too much.
Research
What others already do well
Before designing anything, I looked at where typing or talking already makes searching easier.
Google Maps
You type or speak into one search box. Voice works as a quick way to enter text, not as a separate hands-free mode. I took that as a sign that voice doesn't need its own screen.
Parky.AI and Parking Base
Both turn confusing city parking rules into plain language. That tells me people really want help understanding what a sign or a zone means.
Why AI
Two things a filter can't do
A simple filter is fine for one question. AI earns its place in two specific ways.
1. One sentence instead of many taps
"Cheap parking near English Bay Beach, within 15 minutes' walk, for 2 hours" is a whole search in a single action.
2. Comparing what you'll really pay
A price filter only reads the rate on the sign, and that breaks with tiered pricing. In testing, a pin that showed a lower hourly rate ended up costing more for a 3-hour stay than a tiered pin with a higher headline rate. Sending the search through an AI layer lets us sort results by the total cost of the stay.
Flat-rate pin
$2.97/hr
looks cheaper, but costs $8.91 for 3 hours
Tiered pin
$3.75
total for the same 3 hours, despite the higher headline rate
Exploration
Four ideas I tried first
I stress-tested each of these against real search examples before choosing a direction.
Exploration image
Idea 1 · filter sheet
Idea 1
A big filter sheet
A bottom sheet full of dropdowns, sliders and checkboxes.
Why I dropped it
City rules differ so much that it would break across 1,300+ cities, and it asked for too many taps on a small screen.
Exploration image
Idea 2 · chatbot
Idea 2
A chatbot
A full-screen chat window for open-ended conversation.
Why I dropped it
It invited questions about billing, refunds and accounts, which we didn't want to take on. It was also too slow for a quick utility app where drivers need to see the map right away.
Exploration image
Idea 3 · one bar
Idea 3
One bar for everything
A single search field with an AI sparkle, "Search or ask AI...", handling both addresses and questions.
Why I dropped it
The app couldn't tell what people meant. If someone typed "Denman", should it look up an address quickly or ask the AI? The layout jumped around depending on the guess.
Exploration image
Idea 4 · pull-up drawer
Idea 4
A pull-up drawer
An AI-first drawer you pull up to see recent searches and template cards.
Why I dropped it
It pushed normal address search out of sight. That would slow down the 87.5% of people who just want a quick, direct lookup.
The solution
Two search modes, and you pick before you type
A simple switch sits right above the search bar: Standard or ✨ AI Search.
Because people choose how they want to search before typing a word, the app never has to guess. Everyday address lookups stay fast, and the AI is there when someone has a more complicated request.
Mobile screen
Mode switch
Mobile screen
Typing a request
Mobile screen
Results on the map
Prototype video
Add the video here
Key details
How it works
Four choices that make AI search feel safe and easy to trust.
01
Your sentence turns into pills you can edit
What you type becomes small tags: Near "English Bay", Within 15 min walk, 2 hr, ≤ $5/hr. If the AI gets one part wrong, tap the X on that tag. No need to retype the whole sentence.
Mobile screen
Search pills
Mobile screen
Walk time and duration
02
Walking time and parking time are two different things
The app keeps how far you'll walk separate from how long you'll park. Walking distance follows real routes, not a straight circle on the map, which can be misleading.
03
Permit-only spots stay hidden unless you say you have a permit
Say "I have a residential permit" and those spots appear. Until then they're left out. That's safer than a small warning tag that's easy to miss.
Mobile screen
Permit handling
Mobile screen
Map pins with total cost
04
Pins show what you'll really pay
Once you give a duration, map pins show the total cost of your stay, like ~$2.28 · 2 hr, and match the list. No mental math.
Usability testing
What testing showed, and what I changed
I ran moderated walkthroughs of the interactive prototype. Three problems came up, and I fixed each one.
Problem 1
"15 min walk" was read as parking time
What happened
The parser took "15 min walk" as a 15-minute parking stay. It worked out the cost of a short visit instead of finding spots near the beach.
What I changed
Walking distance and parking time are now always read separately. The pills show what the AI understood, so people can check it at a glance.
Problem 2
Map pins and the list showed different prices
What happened
The list showed the total cost of the stay, but the map pins still showed hourly rates. People had to do maths to compare them.
What I changed
Pins and list cards now update together. Both show the total cost and time, like ~$2.28 · 2 hr.
Problem 3
Permit spots were easy to pick by mistake
What happened
Permit-only zones were mixed in with normal results and marked with small text. Some people picked spots they weren't allowed to use.
What I changed
Permit-only spots are hidden unless the person says they have a permit. Otherwise they sit in a collapsed section at the bottom of the results.
Success metrics
How we'll know it's working
Four things to watch after launch.
Search to session
How often AI search leads to a parking session, compared with standard search.
Correction rate
How often people edit or remove a pill. It's a direct sign of how well the AI understands them.
Time to start a session
Whether searches with several requirements are faster than the old filter flow.
Price accuracy
How often we show a "price updated" notice, so we can improve the cost estimates.
What's next
What I left out on purpose
A standalone filter bar
This is a separate project running alongside AI search.
A voice-first, hands-free mode
This is planned for Apple CarPlay and Android Auto, where talking is already the main way people interact.