Spatial Stack with Matt Forrest

Why do AI assistants recommend only about 1.2% of local businesses?

In this episode, Matt Forrest sits down with Albi Wiedersberg, VP of Product Management at the Overture Maps Foundation, to unpack what it takes to build a places layer that AI and every other product can actually trust.

Albi pulls back the curtain on how Overture is tackling one of the hardest problems in mapping: keeping a global record of places complete and current. Discover how GERS (Global Entity Reference System) gives every place, building, road, and boundary on Earth a permanent ID, how the travel industry is using it to stop reconciling hundreds of internal IDs, and how Overture reset every operating status to blank in May and rebuilt it from pooled signals from members like Uber, Meta, TomTom, and Tripadvisor.

Whether you're building AI assistants, working with POI data, or trying to understand why your favorite local shop never shows up in a chatbot answer, this conversation breaks down the data layer underneath it all.

In this episode, we cover:

- What GERS (Global Entity Reference System) is and why Albi compares it to the ISBN for books
- Why AI assistants miss more than 98% of real places
- The "viable places data set" and why coverage is the gate to adoption
- What grounding means for location data, and a study by Voygr finding one in five LLM place answers were hallucinated or out of date
- Spatial knowledge graphs and pre-baked spatial joins
- Why open or closed is the hardest simple problem in POI data
- Where the places data set is headed in the next year

LINKS:

Albi Wiedersberg on LinkedIn: https://www.linkedin.com/in/wiedersberg/

LEARN MORE

Overture Maps Foundation: https://overturemaps.org
Overture Explorer: https://explore.overturemaps.org
Overture Docs: https://docs.overturemaps.org
GERS Bridge Files: https://docs.overturemaps.org/gers/bridge-files/
Voygr Local Search & Discovery Performance report: https://voygr.tech/report/

ORATOR, built by Wherobots, led by Daniel Smith:
- From Concept to Prototype: Grounding AI & LLMs with Overture's Cross-Theme Knowledge Graph: https://overturemaps.org/blog/2026/from-concept-to-prototype-grounding-ai-llms-with-overtures-cross-theme-knowledge-graph/
- Wherobots' Spatial Graph RAG for the Physical World: https://wherobots.com/blog/spatial-graph-rag/

CHAPTERS:

00:00:00 – Intro
00:02:13 – Albi's path from finance data to Google Maps
00:04:06 – Automotive maps and the self-healing map
00:05:28 – Why a complete global data set is so hard
00:08:24 – What is GERS? The ISBN for places
00:10:47 – A new hotel gets a GERS ID
00:13:51 – AI recommends 1.2% of local businesses
00:16:09 – The jewelry store in Corfu and the viable places data set
00:21:15 – What grounding means for places data
00:24:09 – Spatial knowledge graphs and pre-baked joins
00:30:38 – Operating status: the hardest simple problem
00:32:36 – Signals, coverage, and the next 12 months
00:36:22 – How to get started and contribute signals

---

📊 FREE: The Modern GIS Skill Map
The 5 skills that actually matter in modern GIS (and what you can stop learning). Based on a survey of 1,400+ geospatial professionals.
➡ Get the free training + PDF guide: https://forrest.nyc/go/training/

📰 Weekly modern GIS insights: https://forrest.nyc

CONNECT WITH ME
📸 Instagram: https://www.instagram.com/matt_forrest/
💼 LinkedIn: https://www.linkedin.com/in/mbforr/
📧 Newsletter: https://forrest.nyc
🌐 Website: https://forrest.nyc

What is Spatial Stack with Matt Forrest?

Welcome to The Spatial Stack, where modern geospatial technology takes center stage. Our episodes feature interviews with leading experts, insightful discussions on the integration of AI and big data in spatial tech, and case studies on groundbreaking projects worldwide. Tune in to stay ahead in the rapidly evolving world of geospatial technology!

If you ask an AI for a good bakery near you,

there's a real chance it never mentions the

one that's been on your corner for 20 years.

There's one number that stood out to me before

I had this conversation.

AI assistants recommend about 1.2%

of local businesses.

The other 98%

just aren't showing up.

I work with Places data all the time.

I'm in the Overture Explorer almost every week,

and I still don't have a clean answer for why

that happens.

So I ask the person whose job it is to fix it,

LB Weitersberg,

VP of product at Overture Maps Foundation,

but before that,

more than seven years at Google Maps working

on the Places API and automotive maps.

And he doesn't think that it's the model's fault.

And so in my mind,

it's it's really a data problem,

right?

Not primarily a model problem.

And even though I guess I would have an interest

in it being a model problem,

because then we don't have to fix it.

So he told me about a jewelry store in Corfu.

It started as a home business during COVID.

Now it has a front door on a street with tourists

walking in.

The consumer AI assistants couldn't find it.

Overture could though,

because the owner kept her Facebook page up to date.

The fix Overture is betting on is called GERS,

a permanent ID for every place,

building,

road,

and boundary on Earth.

His analogy is the ISBN.

And then the

ISBN was introduced and it didn't really make

the books better,

but it made everything you could do and build

on top of the books possible.

A few of the things I didn't expect,

a single hotel can sit in more than 100 systems,

each with its own ID for the same building.

One in four businesses opens,

closes,

or moves every year,

and almost nobody reports a closure.

So in May,

Overture wiped every operating status in the

place's dataset back to blank and started rebuilding

it from signals.

Uber drop offs,

meta,

TomTom,

TripAdvisor,

about 8.5 million US places were refreshed

in the June release alone,

because getting opened or closed wrong can

cut both ways.

Wrongly open,

it hits a user who drives there,

wrongly closed,

the merchant suffers,

and so there's no safe default.

This is the spatial stack,

and as always,

the future is spatial,

so let's get into it.

All right,

Albi,

welcome uh onto the show.

Excited to have you here today.

Um,

I'll let you give a quick intro about yourself.

Tell us who you are,

a little bit about your background,

and yeah,

a little bit about what you're working on at Overture.

Absolutely.

Yeah.

So my name is Albi Weidersberg.

I'm from Berlin and Germany,

working with the Overture Maps Foundation,

where I am the vice president of product management,

the official title.

You know,

sometimes uh people ask,

like,

why would a data focused open uh open data

kind of foundation effort have a product manager?

Which is a great question and one that is very

exciting for me because we're actually it means

we're approaching it any type of a product

where you're trying to understand the user

and you're trying to plan what you're gonna build.

So yeah,

how I I got here I think is is kind of interesting

because

I entered maps a little sideways,

let's say.

I have a background economics,

finance,

accounting,

and and then consulting.

Um but everything I did across those jobs had

always had to do with with data.

I built financial uh data systems at Siemens,

I built kind of or work

deeply in procurement data systems and consulting,

and that was always the threat.

And you know,

I went to grad school in Berkeley and after

that 2015 I joined Google,

not in

the maps role first,

but followed the manager to maps uh shortly after.

And that's honestly how I ended up in maps,

and I was there for more than seven years working

on a lot of mapping focused products.

Started with,

I think,

Places API at Google,

and then the relaunch of Google Maps platform,

and then a product that I'm still very proud of,

uh,

eco-friendly routes.

I'm not sure if you if you've used it in the

Google Maps app.

But then spent a couple of years on actually

automotive maps at Google.

And that is the part that I would say in hindsight

really got me here with Overture.

Observed very early on,

of course,

with the phones,

right?

Smartphones,

how the world filled with uh connected sensors,

but then it increasingly became the car.

And the the really

fascinating aspect is how these sensors can

help compare the map to the real world and

then flag any type of change where the two diverge.

So

you can create basically a map that heals itself.

And that idea

really stuck with me.

But I also realized kind of being in a company

building this,

that there is a type of a structural limit

that a company on its own can build because

you need you need data from sensors,

and that means you need to set up partnerships

and deals,

and you have to do it like everyone in the

industry does that bilaterally.

And you know,

that's not a criticism of anyone's business model,

but it's really what bilateral deals do is

that they don't compound.

And so I had watched Overture and open data

get really good.

And then the question and when Overture started

in late 2022 was kind of what if a group of

companies comes together and does what I've

been what I had been working on,

automotive and self-healing maps.

And what if companies build this self-healing

map together and they they set up these these deals,

these trades and these data sharing agreements

once and then everyone benefits?

Yeah,

no,

I that that's I I think it's a really interesting

background taking into account everything that

I think you've been able to do from you know

working in the spatial space,

kind of almost,

you know,

finding your way into maps and understanding

the complexity of it,

even going back to the,

you know,

your economic background of the economics of

global mapping,

I think is is really complicated.

And I don't think a lot of people necessarily

understand how hard it is to create that global

data layer.

So I think I want to start there is to figure out,

okay,

you know,

just from a perspective of what does it take

to create a complete global data set?

Like what,

why is that so complicated?

Because I think a lot of people,

when they interact with the data,

um,

you know,

they might not necessarily know it's overture

right away or things like that.

But you know,

they're looking at a map,

they can see it,

they can quickly find and search for anything

on the globe.

Why is that so complex?

And,

you know,

how is overture starting to solve that?

I know it's a really big question on pack,

as like right off the bat,

but um,

I think you're in a unique position to at least

start to,

you know,

kind of figure out that question a little bit.

So let's start there and then we can kind of

dive into some specifics.

Sure,

absolutely.

At a high level,

the world is complex and the map is uh is has

the ambition to represent the world.

Obviously,

a map is always just a subset,

like you have to kind of figure out what which

part of the real world do you want to actually

model and you know showcase on a map.

With Overture,

we have we have picked what we call a few data

themes that we f that we care about.

It's still hard because even if you were able

to perfectly capture those those themes and

the real world as as it kind of can be modeled

along them in a perfect way,

it would also continue to outdate because the

world changes.

And so I think,

you know,

we spend a lot of effort,

particularly on places,

uh meaning point of interest data,

because that's one of them the themes that

outdates the most,

right?

So one in four or five businesses every year change.

And yeah,

so the ever since maps were started,

being it in like a paper form or then in digital form,

the problem of properly measuring the world

and capturing it and then refreshing it was

a very hard one.

Now,

I talked about the the sensors and the opportunities

of using,

you know,

new ways of capturing how the world,

what the world looks like and how it's changing.

And you know,

that is also a quite a messy process where

there's a lot of noise and a little bit of

signal and figuring that out,

what is actually that we care about in in that

that world of uh out there in the world of

signal that you can get from from sensors and

other sources of data.

That's still a hard one.

And it's it's uh even even for things that

you would believe are are fairly straightforward,

there it takes a lot to align on what the what

the map and what the kind of digital version

of the real world truth looks like.

Yeah.

And I think one way you've started to,

or guess overture started to solve that is with GERS.

And I think that is a term that more people

should know and be aware of.

For someone that's hearing that term for the first time,

can you talk to me a little bit about what that is?

Like what's the one-sentence version and why

is it such a critical backbone to what you're

doing at Overture and the data there?

Yeah,

GERS,

one of the most one of the of the most elegant acronyms.

It's uh it's a permanent ID for every place,

every building,

every road boundary on earth.

I mentioned the themes before,

right?

So every theme in Overture carries a GERS ID

on every entity that is modeled in the themes.

That would be a technical answer,

I guess.

Why it's important is that basically two different

systems or people or analysts or whoever it is,

they can actually agree that they're talking

about the same restaurant,

for example,

without having to guess.

And the the analogy that I really like is with

ISBN for books,

where every shop,

distributor,

library used to have their own way of referring

to the same book.

And each of them spent a lot of time and effort

reconciling those,

right?

And then ISBN ISBN was introduced and it didn't

really make the books better,

but it made everything you could do and build

on top of the books possible.

Yeah,

it I I I I honestly use the library reference

a lot myself,

to be honest,

when I'm talking about data and systems and

everything like that,

because it is um I've never used it in that

way to kind of give like that global reference system,

right?

Of like,

here's how all things are interconnected.

So,

um,

but I use it a lot when I when I talk about

anything with data because it is one of those

systems that has,

you know,

very specific ways about you know interacting

with systems and stuff like that.

So let's just imagine a scenario where,

you know,

a new hotel opens up,

you know,

in a city,

and uh maybe that's it a at a specific point,

you know,

it's like a point of interest,

and you know,

but there's other things associated with it,

and and there are other things inside a hotel

that usually exist,

things like restaurants or there's a building,

stuff like that that's associated with it.

That gets a new GERS ID.

What happens after that takes place and and

what can someone immediately start to do with

that when that data comes in?

Yeah.

Obviously,

the hotel owner works out wakes up and they

have been Gersified.

So that's great on its own.

Yeah,

I mean,

what it enables,

I mean,

hotels are part of the travel industry,

and then that's uh that's really a huge supply

chain on its own.

And so I would say by by receiving that GURS ID,

the whole travel supply chain has one ID they

can align on.

And uh,

you know,

that can touch whatever a hotel is involved with,

like a trip booking,

but also all the property systems for the hotel brands,

loyalty programs,

uh the maps out there that consumers use,

ride sharing and tour operators.

So there's a whole stack of different stakeholders

who are involved in this.

What we're talking to the hotel uh folks a bunch,

and is they have more than 100 systems and

each have their own ID for the same hotel.

And so that means every single link between

systems is a custom integration in the kind

of legacy world,

right?

And uh,

and what that what that causes is that actually

bookings fail because they can't agree on that

they're talking about the same place or the

same building,

or even subsets of the place and building,

as you mentioned.

But GERS can give them a common language,

we call it like a universal plug for travel.

It's basically an open ID.

And it doesn't mean that all these systems

need to be replaced.

There's actually what we call bridge tables,

the translation between an ID that is being

used internally by someone and GERS.

And so the idea is that the integration has

to be done once,

and then the industry has something to align around.

So instead of building a hundred different

adapters for a hundred different systems,

you can use GERS.

And uh it's it's very similar to,

I think,

a payment network.

Nobody gets excited about payment networks,

I guess,

but uh it's it nothing works without it.

And it goes further,

right?

Because what the hotel can then do is AI assistants,

they might not may not know that something

exists or they recommend something that doesn't exist.

Uh but the agent or the LLM they can look up

the ID and they can verify that what they're

talking about,

the hotel is real and it's not hallucinated,

it's not a scam.

So we have members in Overture.

Uh I wrote a blog post about it with like in

hotel and Lynx Rest in Hilton who who build

um discovery agents on top of natively that

understand GERS.

And so,

you know,

this is only possible because two or more sides

agree which hotel they're talking about.

Yeah,

I was gonna ask you the the the cost of that system,

but I but I think you answered it here where

there's multiple different systems,

multiple different IDs.

They're not agreeing on each other.

I mean,

you know,

it seems simple when a name should equal a name,

a place should equal a place,

but it's not as clear as that.

And and it's it's almost strange that we there

wasn't some sort of common system to say that

that first of all,

this thing is real.

Here's what it is,

and here's how two systems talk about that.

It's it's it's pretty interesting that that

has taken so long to get to that point.

But um you're you're definitely leading the

way about with that at Overture.

And I know you mentioned sort of AI assistance.

I I found an article that that you you were

writing or you wrote,

and it basically said,

and correct me if I'm if I'm wrong in this quote,

but AI assistants recommend just about 1.2%

of local businesses when you sort of talk to them.

That could be hotels,

that could be anything.

Sort of how is how is that between the work

you're doing now at Overture and and or even

GERS or things like that,

how is that helping bridge that gap so AI can

not only recommend the correct places,

but also share more about what's actually there

and do that correctly?

Yeah.

First off,

I I found that number,

it's not my number,

but I found that number quite astonishing.

And and it is particular because it means there

is like more than 98%

of places out there that the AI does not recommend

and it's the long tail.

I I imagine these are actual existing places

in the real world,

like a Turkish bakery around the corner,

it's been there for 20 years.

But it's either that nobody has a commercial

interest to describe this business in the right

way that you know the models want to use it,

or the information might exist,

but the model doesn't have a reliable layer

of data that is organized in a way it can understand.

And so basically the the 1.2% I was referring to,

that's the sliver that is actually really well

documented.

And then the rest kind of falls off the edge.

And that's a market failure,

and that's exciting because in shared infrastructure

has always been used to address market failures

from an economic perspective.

So what

we have then is basically

a business that it

either it doesn't exist

and the model recommends it,

so someone will actually drive across town

and then come across a shuttered storefront.

I think that's the worst end user-consumer

type experience.

Um and ultimately it becomes the AI company's problem.

Or alternatively,

the the business exists,

it doesn't get recommended,

right?

It's part of these 98%.

And the business owner has no idea why,

because there isn't clearly established rules

like there were with SEO,

where it's like,

uh,

okay,

you can call someone or uh work with a company

who optimizes that you show up.

And so the business basically stops being findable.

And that is,

and that is a real problem for for many businesses

uh out there.

And the way that that manifests in overture

is in the places theme.

It's you know,

places listings,

and that's where businesses and all these different

things can go.

Why,

why is the places layer so critically important?

And what are some of the things that are happening

with that to sort of really make it,

I guess I guess as as you described it,

a really viable places data set in terms of uh updated,

you know,

complete,

searchable,

findable.

Why is that piece so critical?

Yeah,

it's super critical.

Maybe,

maybe to answer that,

uh let me just give you an example real quick

because you know I do this when I'm on vacation

and my my family finds it ridiculous.

Uh,

but I actually look up the businesses I come

across uh when I'm when we're traveling and

I look them up on the Overture Explorer,

which is like a map we have on the Overture

website where you can uh explore our data.

And so a year about a year ago,

we were in Corfu and there was a jewelry store

that started actually as a home run business

during COVID,

like I think many people did,

but now it's a real shop,

right?

So it invites tourists in and it has a it's

at the like street,

has a front door,

all the things.

And so I was looking it up um or trying to

find it on the like the typical consumer AI assistants,

and it couldn't find it because I suspect because

it used to be a home run small business,

and there's a language cutoff and so uh sorry,

a knowledge cutoff and so forth.

But on Overture and in our Explore feature,

we actually knew about this business,

and it was because the owner had actually done

what was available to her,

the Facebook page.

She kept it updated.

And so what happened then is it got ingested

into Overture,

we assigned the ID and it became part of our

open data set.

And so the information was was basically out

there and it was reachable,

but the failure was that that it didn't exist

in these in these products.

And in my mind,

it it's because there

is currently no layer of data that is complete

enough so that these these tools can actually

bet their answer on,

right?

And then if you can only depend on a source

some of the time,

then product teams don't wire it into the product.

And then it falls back on what the model remembers,

and the model might not remember like the jewelry

store in Corfu.

And so in my mind,

it's it's really a data problem,

right?

Not primarily a model problem.

And even though I guess I would have uh an

interest in it being a model problem because

then we don't have to fix it.

But yeah,

that's uh that's that's kind of what's going on.

And then to your question,

right?

I like kind of based on this problem.

A year ago,

we we had uh an audit with uh with our group

of leadership and overture,

we call it the steering committee.

It's basically our board.

Uh and we looked at adoption and how we know

just the state of affairs and our data.

And the pattern was really clear that uh that

it always came back to the places data set

as the as a very important asset.

And the logic is kind of like this,

right?

If if we want people to build on overture,

then we have to enable them so they can stop

stitching together all these data sources that

they're currently stitching together and and

instead rely on our entities,

like the jewelry store.

And and also kind of including entities that

only this company knows about and that they

might want to contribute to overture.

That's also an important point.

Because ultimately we need an ID on anything

on everything,

and you can't put an ID on something that you

don't know about.

And so coverage,

meaning knowing about all the places in that example,

is the gate.

And below a certain threshold,

someone building an AI assistant or any product,

they won't,

they

will,

they they won't adopt and they will keep their

own pipeline and they will keep maintaining

their own pipeline and all these costs,

and you know,

then they haven't fully adopted overture and GERS.

And that's and that's why we decided that um

we're gonna put a lot of effort into creating

what we call the viable places data set um

so that anyone can build their product on top

of it and it becomes load bearing for them.

It's not just like,

you know,

a data set uh a sherry-picking data set,

but actually something they can bet on.

Yeah.

And I I will say that you are not the first

person working in geospatial to test things

while on vacation.

Uh,

and you are definitely not the first person

on this podcast to have talked about testing

things on vacation.

So um,

it is a very common thing for people that work

with this data to test it in places that are

not their own backyard and seeing how it works

in the real world.

So,

but I think besides that,

what what what you've highlighted here is I

think one of the important pieces about how

places can really start to create an underlying

data infrastructure,

right?

It's not just having listings that are up to date,

but that's one piece.

It's where all the data is being collected from,

um,

ensuring that you know it's flowing through

in a consistent format and then creating it

for any system,

not just an AI,

but but able for any system to be able to read,

and like you said,

ultimately for that system to bet on the accuracy,

the validity,

the fact that it's up to date on all those

different pieces.

I think what that all comes down to is sort

of having a a grounding layer for places.

And a lot of times we're tying that into,

you know,

AI systems.

And I don't want to get to,

you know,

specifically how we bring that in.

But for someone who's,

you know,

maybe approaching the term grounding,

you know,

uh,

you know,

from for the first time,

or someone who's like,

you know,

building a product or wants to use this,

what what does that mean to overture to create

a grounded places layer?

You said,

you know,

sort of that that complete viable places layer.

What does that look like?

And when someone's coming to use this data,

what can they expect from it,

both now and then,

you know,

from the vision going forward?

Yeah.

And,

you know,

fairly uh just to kind of admit it up front,

I mean,

we are still figuring that out in detail as well.

I think it's still early days,

even though many people are working on AI day

in and day out these days.

I think it's still early day to kind of figure

out what spatial and location grounding actually

will look like at scale.

Um now,

in in terms of what does grounding mean and

like how how we define it.

I mean,

uh,

someone who doesn't hasn't been in contact

with uh with AI grounding,

I always describe a model without grounding

as kind of like an analyst working off of something

they remember,

like a map they remember from a while ago.

So they it's still it's still useful to some degree.

There they can work with it in a confident

and articulate way,

but it's also a little bit out of date and

there might be some inaccuracies.

And so what we're handing the industry ideally

is somewhat like an authoritative layer that

you can cite in the work.

And uh so you can basically on the one hand

make sure in terms of recall that there aren't 98%

of entities missing from a response,

and on the other hand,

that the kind of responses are that that you

give are grounded in reality.

I think so.

Our member Voyager ran a study where they tested

all the leading uh LLMs about with like location

and geospatial questions,

and I believe one in five responses about places

were still somewhat either hallucinated or out of date.

Yeah.

I I think that our goal is that we will provide

this kind of layer of authoritative layer or

grounding layer,

if you will,

that people can rely on where they can get

the the recall correct and the hell where they

can control hallucinations.

And and that is so complete that they don't

have to use anything else.

And at least when it comes to kind of understanding

is this a real-world entity or not.

Not necessarily talking about rich data,

that's a different topic.

If you want to bring in rich data,

that's possible.

And the a a good grounding layer will actually

help you identify which rich data you want

to bring in efficiently so that the context

window doesn't explode from the start,

which is also a problem that many people in

the in the AI industry are are trying to solve.

Yeah,

I when I explain to people,

you know,

that maybe don't you know interact with the stuff often,

you know,

kind of how you know AI interacts with location,

where it's good,

where it's missing.

I I I think it,

you know,

a lot of what I try to bring forward is that

you need to,

first of all,

geography is changing and moving all of the time,

that the real world is in constant,

you know,

motion,

literally.

And and POIs is a key part of that business,

is change their hours.

They,

they,

they might move,

they are opening and closing,

and and all these different things are happening

all the time.

Whereas the models,

you know,

the the language models themselves are sort

of capped in time.

They might know big things that don't move,

like they'll they'll know where you know the

Statue of Liberty is or the Eiffel Tower,

and they they know things about you know,

names maybe that don't change often,

like neighborhoods or stuff like that.

But that underlying layer has to,

you know,

it has to reach out to that somehow.

It has to connect to that and with without

getting too much in the specifics of of all

those different things.

That that's one key component of that.

But there's also the sort of thematic interaction

with that data,

right?

And and I'm gonna kind of move into the the

knowledge graph kind of concept of where things

are going.

You know,

the

places have,

you know,

there's certain component about where they are,

latitude,

longitude,

streets,

neighborhoods,

all that different stuff.

But then there's also the how they relate to each other.

And I think you talked about context window.

Are they how does a veterinary office relate

to a doctor's office or a pet food,

a pet supply store or things like that?

There's these other relations thematically

and otherwise that I think are also important.

And you can kind of embed that information

with spatial context and making it very fast

and searchable.

What,

what,

what are the when I talk about a spatial knowledge

graph,

what does that mean to you and how does overture

fit into something like that?

Yeah,

this is an important question about our data

model in a certain way.

The way I would describe the

data model so far is basically we have these themes,

as I mentioned,

and there are flat lists telling you,

you know,

what a place is.

If you contrast that with a with a graph,

it would actually tell you not just what it is,

but also what it's part of.

And almost every real question from from people

out there has a component of uh of this of

the second part.

So when someone asks,

uh,

is you know,

is there somewhere near the station where I

can work for an hour before my train?

You kind of have every component in this,

you know,

other than kind of the station and the the

maybe cafe that we're gonna bring back as a response.

Every component is uh is a relational component

that I would expect to be in a graph.

And yeah,

there's a lot of prototyping work going on

in in Overture,

and we are we're trying to figure out what

the what the graph looks like that brings together

basically our themes.

Kind of how have have you or Overture members

started to kind of explore this a little bit

or start to experiment on that?

How what have they sort of done so far in this space?

Yeah,

absolutely.

Yeah,

we have had a bunch of exploration.

I mean,

we've been quite frankly talking about this

with our members for more than a year,

probably almost two years now.

And I would highlight in a prototype called

Orator that uh was built by our member Wearobots.

And actually,

I should say Daniel Smith uh led it.

And Overture doesn't build this necessarily,

right?

Which is kind of an important point of our

model and how we work,

how we work together,

how we enable the stack and people working

with Overture data.

The way it the way it's constructed is that

a geometry acts as the tissue,

the kind of tissue across that graph concept,

somewhat like a foreign key,

even though like it's not a relational database

concept in this case.

But yeah,

for example,

if a place point sits inside of a building polygon,

then this graph would already know that these

two entities,

building and place,

are related,

which is very helpful,

including for an LLM who that might not just

want to do this type of join on the fly.

And nobody writes that kind of link by hand.

It's actually uh a spatial join that is uh pre-baked,

so to speak,

in the graph.

And and that matters because uh I what I've

seen from from the prototype is how quickly

this can be done for I think it was the San

Francisco Bay Area and then Manhattan,

and we also kind of blocked about that,

which are very different areas of the world

in terms of density,

and it was one pipeline,

and there was no manual curation or manual

work from from what I've seen.

But the knowledge graph can be uh reproduced

and for every release so that it actually stays fresh,

stays fresh and and can re-establish these

relationships and the graph format based on

every release which we do monthly.

Yeah.

And and I think what people

why they what where

else the graph is important is that it models

not just the relationship between the entities,

but also a bunch of the metadata and context around it.

So you kind of how important every edge of the graph is,

what the confidence is that was derived,

if it's inside,

or if it's likely the same,

or if there's the different attributes that

connect entities.

And without that,

you don't have a knowledge graph,

basically.

You without that,

you have back to the to the start,

you have a list of

assertions or facts,

and that's really the difference.

Yeah,

it's it's funny because I think it's you're

pre-baking in some spatial awareness into the graph.

And like if it's just a pure knowledge graph,

you know,

it knows that you know these thematic things

or or these terms are are close to each other.

But then you can also bake in,

you know,

spatial proximity or distance or neighbors.

And and I think it gives it a really easy to

access layer.

Because I I think if you were to ask me a few

months ago or even a year ago,

how how would I solve that problem?

Well,

SQL can solve it.

You can query it on demand and you know,

you can build a whole MCP to do that.

I think that's you know,

that that was one way that I was thinking about that.

But to really scale it up,

you know,

to make it as as most efficient as possible,

that's what you can do.

And I think it's creating data products to fit.

And I think that's,

you know,

from what we've been talking about,

you can you're you're giving people the ability

to build their products on top of this and

scale it and have consistency in all those

different pieces.

And I think solving that very first layer of,

you know,

making sure that it's like you said,

a viable places data set that people can build

off of is,

you know,

enabling those types of innovations,

which is really,

really,

really cool to see.

To to go back down to that viable piece again.

What like what is one like data quality problem

that's really,

really hard to solve that maybe sounds like it's simple,

but it is is in in actuality a really challenging

problem from a POI perspective.

Yeah,

it's uh it's it's interesting.

I mean,

I would pick operating status,

which is modeled as open or closed,

couldn't be any simpler.

So it looks very trivial,

but in the schema,

right?

But then it's not.

As I said like at the beginning,

uh I think one in four businesses opens,

closes,

or or relocates in a given year.

And the problem really is that nobody has a

strong incentive to report a closure.

Uh so it doesn't,

you know,

like as a business,

if I have to close,

I wouldn't have,

wouldn't be worried to go in and file something.

It typically just stops and there is no notification

or fee.

And then all these failures happen that I described.

They're not symmetrical.

It makes it really,

really nasty.

Wrongly open,

it hits a user who drives there,

wrongly closed.

The the merchant suffers.

And so there's no safe default.

You can't really be conservative about the problem.

And then you can't also you cannot solve this

by looking harder at any single existing data

set out there in the open.

And so,

yeah,

what's kind of a bit of a uh pivot we made

in May with our places data set,

we reset all the operating status values to blank.

And then in June,

we started filling them back up from something

we called pool signals.

That's uh that's a really cool example of a

kind of collaborative effort around data where

we're actually able to infer um what the operating

status of a business is so we can start to

solve this problem at scale.

Yeah,

I I I had no idea that you set all the uh statuses

and started solving it from from ground up.

So it's it's cool to hear how this is actually

working in very real time and building it out.

And and you know,

it's exciting to hear how this is all progressing

so fast,

which is which is really awesome.

If you if you look like a year,

let's just say like a year from now,

what what what do you think the place's data

set's gonna look like?

What are people gonna be building on top of it?

What are your plans for it?

Maybe this is a fancy,

uh,

you know,

fancy way to ask the question of what's next

and and kind of what's coming.

Where do you see it going and and how do you

see people using it?

Yeah,

actually,

let me talk about the last uh kind of 12 months

a little bit because that actually sets up

nicely how we're thinking about it going forward.

So,

you know,

we so first off,

the places data set is already very important.

Like some of our key members,

Meta and others,

are using it in products that reach billions of people.

These are these aren't like pilots,

it's actual at scale products.

But you know,

one of the key questions or key challenges

is is the coverage,

and that you know,

we're we're addressing.

And the other key problem is the freshness

that we just talked about.

And so when you look at our June release,

we actually updated uh for eight and a half,

I believe,

million places in the US.

We updated the operating status,

and that was based on what we call signals.

So,

as a as an example,

Uber,

who's a member in Overture,

they send us aggregated anonymized ride and

forred place visits,

and can imagine it as like a ride end somewhere,

and that is like a small piece of evidence

that a place is operating because it was visited.

Standalone,

the the signal isn't super crazy meaningful

and always correct,

but aggregate it with other similar but different

signals from other members,

and you'll get what we call a heartbeat for a place.

And that's particularly helpful for the volatile

categories like restaurants and bars and venues

and thereof.

And before someone complains,

I'll just say there is no privacy problems

in this because we don't like there's no rider

data or no route information or anything in

the data that we receive.

It's really just uh was there a ride activity

at a given place,

not a trace about anything like a person or a ride,

you know,

and uh it's it's the only version of this that

should exist,

uh frankly.

You can also not totally derive that a place

has closed just because the ride activity stopped.

But that's where other signals come in,

right?

So we have other members,

we have Meta,

we have TomTom,

we have TripAdvisor.

They all see the world from different angles

and uh and have different slices of signals.

And I don't think any kind of commercial data

vendors can can kind of contribute or or sell

this at the level that we can we can get access

to all these different perspectives.

And that's that kind of indicates where we're

going because yes,

we have been bootstrapping the coverage of

the places data set,

and we have focused on the US and and key categories

in the US,

such as food and drink and and shopping and

landmarks and lodging.

And we're already building it out further now

for Western Europe.

So we're moving region by region to establish

viable uh coverage of our data.

At the same time,

we're scaling up this type of signal collaboration

with more providers joining and with more regions

also like coming online to update the existence

and operating status of places.

And so,

yeah,

I can imagine in a year we don't have just

the kind of Western world,

but we are actually

expanding to other regions.

I know our members have mentioned how they

would like to do something similar in uh in

different regions across Asia or even uh Latin

America and so on.

And um,

this is definitely something that is uh very,

very important for us and kind of high up the

priorities list.

Yeah.

It's gonna be interesting to see how it all

kind of comes together here and how all the

pieces start to kind of stack up because it's

it's been cool to watch everything from Overture

kind of develop over time,

but I'm I'm really curious to see how places is gonna,

you know,

kind of use all these different pieces from

you know all the data streams and the signals

that you're getting,

how you can kind of create those those real

pieces and and go from there.

Yeah,

is there anything else uh people should know?

How can they get involved,

use the data,

things you have coming up?

Yeah,

what where where can they kind of you know get started?

I'd say it's two things.

One benefit of having an open data product,

check it out,

right?

There's not there's not a lot uh to get access.

Uh you can actually right away head over to our website,

overture maps.org,

where you'll find the explorer to investigate

our data and see it.

Um,

but also the excellent docs.overture.org for

everything you want to know about kind of the

sources and licenses,

we get a lot of questions and where is the

data available,

what type of data,

what does your schema look like?

That's all there.

You don't need any permission,

that's the whole point,

right?

So you can get started.

And then the one thing I really care about,

and a little bit of an ask to anyone kind of

listening to this,

if your organization business sees a slice

of the physical world similar to what I described

about the place

signals.

Um,

for example,

you know you have some indication of when a

place is open,

closed,

or there's a new place,

or any other thing that uh we track in our

data themes byproduct of your operations,

we sometimes call it.

Then talk to us about kind of learning more and

eventually contributing that signal because

most likely that signal is very unique and

you're the only one who's able to actually

contribute and help us with that.

Yeah.

Very cool.

Yeah,

it's it's

I definitely check out the Explore app.

I will make sure to link that in the show notes

or description below because I I use that on

quite a regular basis just to see and and be

able to explore without kind of you know fully querying.

But check the data out and use it.

I think it it's it's like one of the easiest

ways to just get access and start seeing what

you can do.

And then yeah,

definitely make sure to,

you know,

kind of sign up and stay,

you know,

stay aware of all the news and things that

are coming out.

There's a lot of great ways to engage with

overture data and make sure you're using it.

So yeah,

Albi,

thanks for taking the time to chat with me today.

I'm excited to see where things are gonna go

because I think it's like,

as you pointed out,

it's a critical layer for so many different

things from operations to you know,

the all the way through to like stuff like

we talked about with hotel bookings that I

would have never thought about.

Having that viable data layer is gonna be really

critical.

So I'm excited to see where this goes.

And uh thanks for chatting.

I'm sure we'll uh be talking again in the near future.

Thanks so much,

Matt.

Thanks for having me.

All right.

Thanks.