Re: Geofeeds are adversarial — was Re: Geofeeds are good

[Tom Beecher via NANOG <nanog () lists nanog org>]

> Large-scale ground truth evaluations require access to data that is
> typically only available to enterprises internally. Our NANOG 96
> presentation (https://nanog.org/events/nanog-96/content/5678/) and our
> peer-reviewed research (https://dl.acm.org/doi/10.1145/3676869) present
> what we can share publicly about our methodology and accuracy. We would
> welcome an independent academic benchmark if one were to emerge.


The information presented in the paper and preso is nothing about the
accuracy of IPInfo's data. What you present is a comparison between the
results from your probe network, and published geofeeds. If your test and
the geofeed AGREE on the location, you classify that geofeed as 'correct' ,
if they do not agree, you classify it as 'incorrect'.  This is nonsensical.

Your multilateration method of taking the straight line distance and speed
of light in fiber to create distance polygons is flawed. This works in
radio when waves do propagate in straight lines, but terrestrial fiber
networks don't work this way. The physical network infra between two points
is never a great circle route. There are locations all over the world where
two endpoints are in the same city, but would show RTTs in the hundreds of
ms simply because the two networks involved don't connect on the same
continent.

Any statements about the 'accuracy of geofeeds' therefore mean nothing. For
example, you claim that 92% of geofeeds are accurate to the country level.
But what that really means is 92% of your tests agree with geofeeds. It
tells us nothing about if your test was accurate or not.

It is also important to point out that even if we assert that your test are
correct , you point out that geofeeds only cover 1.5% of IPv4, and 0.7% of
IPv6 addresses are covered by geofeeds.  This means that for basically 99%
of all addresses possible, your only reference is your ping/multilateration
method, which as stated previously will have issues accurately placing an
address, meaning for almost the entire internet, your statement of accuracy
is 'just trust me bro'.

On Mon, Feb 16, 2026 at 12:48 AM Abdullah DevRel of IPinfo via NANOG <
nanog () lists nanog org> wrote:

> Hi Hank,
>
> Large-scale ground truth evaluations require access to data that is
> typically only available to enterprises internally. Our NANOG 96
> presentation (https://nanog.org/events/nanog-96/content/5678/) and our
> peer-reviewed research (https://dl.acm.org/doi/10.1145/3676869) present
> what we can share publicly about our methodology and accuracy. We would
> welcome an independent academic benchmark if one were to emerge.
>
> — Abdullah | DevRel, IPinfo
> _______________________________________________
> NANOG mailing list
>
> https://lists.nanog.org/archives/list/nanog () lists nanog org/message/EE5AB4UEX6KHLKGNC27OJKPNDA2RYFHO/
_______________________________________________
NANOG mailing list 
https://lists.nanog.org/archives/list/nanog () lists nanog org/message/CIV6EDFJHRSA6JIGASJKYG33M53HYTM2/