Copied to clipboard
← All posts
From the logs · 9–15 November 2026

One Week, 1.5 Million Requests: Who Actually Asks an IP Lookup API

We took seven days of our own request log apart. Most of the traffic never saw our website, a couple of hundred machines did a fifth of the asking, and somebody out there is very keen to find a file called .env.

· 10 min read· 238,725 addresses

At 23:20 UTC on Wednesday, November 11, our request log, which had been ticking along at about 150 lines a minute, started writing close to 1,000. Nothing was wrong on our side. An online store we have never spoken to, one of the sites that calls our lookup API from its own pages, was suddenly being visited from 17,145 different addresses in about 40 minutes, and every one of those browsers asked us where it was. By midnight it had tailed off.

We only noticed while preparing this article, which says something about how often we read our own logs. So we read a whole week of them, Monday, November 9 through Sunday, November 15, and counted. This is what an IP lookup service looks like from the inside. Several of the numbers were news to us.

1,530,820requests logged in seven days
238,725distinct addresses that sent them
197countries, across 9,425 networks
3.0%of requests were someone opening a page on this site
What the log holds

One row per request: which endpoint was called, the value that was looked up, the address that asked, the page that referred it (if the browser sent one) and the time. Nothing else. Addresses are blanked after 90 days, as the privacy policy describes. Everything below is a count or a share. No address, and no site that uses the API, is named anywhere in this piece. We removed our own uptime monitor and a sister service of ours first, which together accounted for 85,277 requests and would otherwise have topped every chart.

Most of our traffic never sees our website

We think of ipaddress.to as a website with an API attached. The log disagrees. Page views on this site, every tool page and blog post and the homepage, came to 45,274 requests, or 3.0% of the week. The other 97.0% were API calls, and 92.7% of everything was a single endpoint, the IP lookup.

Who calls it turned out to be the interesting part. 83.7% of all requests came from browsers sitting on somebody else's website. A store or a landing page includes a line of JavaScript that calls the API with no argument, the API answers with the caller's own country, city and network, and the page uses that to pick a currency or a shipping notice. We can tell because the browser sends the referring page along, and because 99.8% of these lookups are an address asking about itself.

142 sites did this during the week. They are not evenly matched: the busiest one sent 48.2% of the embedded calls and the top ten sent 97.7%. So a fair description of our audience is "the customers of about ten web shops", which is not the sentence we would have written on the About page.

Fig. 1 · every dot is about 500 requests

Where a week of requests came from

Regroup the same stream by who was asking, by the kind of network they were on, or by IP version.

Hover or tap a lane for its numbers.
Lane heights follow each group's share of requests, with a floor so the smallest ones stay visible. "Unknown" networks are those our classification table has no entry for, mostly transit carriers and mixed-use operators.
Show the numbers
GroupAddressesRequestsShare of requestsMedian per addressSeen once
Embedded on other sites211,8631,281,57783.7%242.9%
Direct API calls15,878144,4099.4%177.9%
Our own pages10,82790,8855.9%229.1%
Scanners15713,9490.9%180.0%

The other three crowds

The remaining sixth of the traffic splits into three groups that behave nothing like each other.

Direct API calls, meaning requests with no referring page, came from 15,878 addresses and made 144,409 requests. That average is useless, because three machines made 104,400 of them: 55,884, 31,571 and 16,945 lookups respectively, each working through a long list of other people's addresses. That is a back-end enriching its own logs, and it is exactly what the API is for. Take those three away and what is left looks different again: 40,009 requests from thousands of addresses, 77.9% of which appeared exactly once, and 59.4% of which were in India. One call each, no referrer, mostly asking for the caller's own location. We cannot see what software sends them. It has the shape of a phone app checking where it is on first launch.

Our own pages drew 10,827 addresses. We would love to call them readers. But 42.2% of them sit in hosting networks, against 15.4% on consumer ISPs, and the country list goes United States, Singapore, China, which is a map of cloud regions more than of people. Google's crawler alone, from 143 addresses, produced 25.5% of this group's requests. A good share of the audience for our tool pages is other programs.

Scanners are the smallest group and the easiest to recognize: 157 addresses whose requests were mostly for pages that do not exist. They get their own section further down.

The shape of the week

A normal hour brought in about 8,800 requests. The quietest of the week was Wednesday at 09:00 UTC with 5,283, and embedded traffic bottoms out around 09:00 UTC every day, when the United States (41.8% of embedded addresses) is asleep. The curve climbs through the American afternoon and evening. Weekends were not a rest: Sunday was the busiest day with 238,202 requests and Thursday the slowest with 188,475. When most of your traffic is shoppers on other people's stores, your week looks like retail.

Fig. 2 · requests per hour, UTC

168 hours, and one that does not belong

  • Embedded on other sites
  • Direct API calls
  • Our own pages
  • Scanners
Move across the chart to read any hour.
Stacked by group. The spike at Wednesday at 23:00 is 38,698 requests, more than four times a typical hour.
Show the numbers
DayRequestsBusiest hour (UTC)Requests in that hour
Monday, November 9226,14314:0011,627
Tuesday, November 10196,95423:009,744
Wednesday, November 11235,23123:0038,698
Thursday, November 12188,47500:0014,024
Friday, November 13212,82320:0015,706
Saturday, November 14232,99223:0014,745
Sunday, November 15238,20200:0016,143

Then there is Wednesday night. In the hour from 23:00, embedded traffic went from its usual 7,500 or so to 34,399 requests. One site accounted for 25,843 of them. Sales do that, and so do newsletters. But of the 18,250 embedded addresses we saw that hour, 3,719 belonged to just three cloud providers in Asia, the wave began inside a single five-minute window, and it ended as abruptly as it started. People shopping do not arrive from rented servers in formation. Our guess is bought traffic or a bot run against that store. We cannot see their checkout, so a guess is all it is.

We saw somebody else's bot problem only because their pages ask us where each visitor is.

A few addresses do most of the asking

Divide 1,530,820 requests by 238,725 addresses and you get 6.4 each. The median address made 2. Nearly half of them, 106,409 or 44.6%, appear in the log exactly once all week, and 89,732 of those are the same event: a person opened one page on one shop, their browser asked us one question, and we never heard from that address again.

At the other end, the busiest 238 addresses (0.1% of the total) made 18.2% of all requests, and the top 1% made 46.7%. The bottom half of the address list, 119,363 of them, produced 8.6%.

Fig. 3 · 1,000 squares, each 0.1% of addresses

Count the addresses, then count what they sent

The squares are sorted from heaviest to lightest. Switch views and each one takes the room its requests deserve.

  • Top 0.1%238 addresses
    18.2% of requests
  • Next 0.9%2,149 addresses
    28.6% of requests
  • Next 9%21,485 addresses
    26.8% of requests
  • Next 40%95,490 addresses
    17.8% of requests
  • Bottom 50%119,363 addresses
    8.6% of requests
The single red square is 238 addresses that each made at least 593 requests. The darkest block is the half of all addresses that made one or two.
Show the numbers
Slice of addressesAddressesRequests eachShare of requests
Top 0.1%238593–55,88418.2%
Next 0.9%2,14965–59328.6%
Next 9%21,4858–6526.8%
Next 40%95,4902–817.8%
Bottom 50%119,3631–28.6%

This matters for anyone setting a rate limit. A per-address limit of, say, sixty requests a minute sounds like a policy for all users. In practice it is a policy for a few hundred machines, and the other 99% will never come within sight of it. The same goes for anything keyed on "returning visitors". Most addresses do not return.

Whose networks, and how much IPv6

We sorted every address by the kind of network that announces it. 40.8% were on consumer ISPs, 23.5% in hosting and cloud networks, and 34.9% in networks our classification table has no label for, which is a gap in our data we would rather admit than paper over. Flip Figure 1 to "Network type" and the imbalance shows: hosting is 23.5% of addresses but 45.1% of requests. Servers ask more often than people do.

The number that made us stop was inside the embedded group. These are, in principle, shoppers. Yet 22.4% of the addresses that loaded a store page and ran its JavaScript were in a datacenter. Some are people behind a VPN. A fair number are headless browsers, price scrapers and ad-verification bots, all of which execute the page like a customer would and none of which buy anything.

IPv6 gives a second, cruder reading of the same thing. Overall 36.0% of addresses were IPv6, but the groups are far apart: 38.9% for embedded visitors, 15.6% on our own pages, 10.2% for direct API callers and 6.4% for scanners. Phones and home broadband have moved to IPv6. Rented servers and attack tooling mostly have not bothered. A crowd's IPv6 share is a rough measure of how much of it is people.

Fig. 4 · share of addresses that were IPv6

IPv6 depends on where you are, and on what you are

By country

France84.1%
Singapore72.6%
Germany58.1%
United States41.0%
Mexico36.4%
Brazil33.4%
Vietnam32.4%
United Kingdom28.7%
China27.8%
Australia19.9%
Canada19.3%
India12.4%
Belgium11.0%
Colombia10.2%
Argentina9.2%
Hong Kong0.5%

By kind of visitor

Embedded on other sites38.9%
Direct API calls10.2%
Our own pages15.6%
Scanners6.4%
36.0%of all 238,725 addresses were IPv6. The countries on the left are the sixteen that sent the most addresses, and they span 83.6 points from top to bottom.
Pink is IPv6, the dim remainder IPv4. These are shares of the addresses that reached us, which lean toward shoppers on a small number of sites. They are not national adoption figures.

By country the range is enormous. France leads at 84.1%, Singapore and Germany follow, the United States sits near the overall figure, and Hong Kong comes last at 0.5%. If your application still stores addresses in a column sized for IPv4, a third of this week would not have fit.

What the scorer thought of them

Network labels only go so far, so we drew a random sample from each group, 4,656 addresses in all, and ran each one through the same IP score anyone can query. (If you have not met it: it estimates how likely an address is to be something other than an ordinary home or phone connection. This post explains the mechanics.)

Fig. 5 · scored sample, 4,656 addresses

Risk bands, group by group

  • Low
  • Medium
  • High
  • Critical
Embedded on other sitesn = 1,500 · median score 12
78%16%
Direct API callsn = 1,500 · median score 13
73%26%
Our own pagesn = 1,499 · median score 13
52%34%
Scannersn = 157 · median score 68
24%19%22%35%
Up to 1,500 addresses per group, picked at random and scored on November 17. The scanner group is small enough that all of it was scored.
Show the numbers
GroupConsumer lineHostingVPN, proxy or TorOn a threat listVerified crawler
Embedded on other sites64.7%24.7%9.7%2.1%0.5%
Direct API calls71.1%26.4%0.7%0.9%0.4%
Our own pages28.0%43.2%1.3%14.3%26.2%
Scanners34.4%59.9%3.2%59.9%3.8%

The embedded crowd is mostly what the shops think it is. 77.5% of the sample scored low and 64.7% sat on consumer lines. Another 9.7% were on a VPN and 3.5% behind a privacy relay of the kind now built into phones and browsers, and those are still people. The 6.3% in the top band are a different matter. Most of them (5.4% of the whole sample) sit in networks known for tolerating abuse, and there they were on a product page, running the store's JavaScript like any customer.

Direct callers look more like households than datacenters: 71.1% on consumer lines and next to no VPNs, which fits the phone-app reading from earlier. Our own pages come out worst of the three groups that are supposed to be legitimate. Barely half the sample scored low, 26.2% were verified crawlers, and 14.3% appear on at least one threat list. People do visit. They are outnumbered.

The scanners are where a score earns its keep, and also where it shows its limit. 59.9% of them were already on a threat list and 56.7% scored high or critical. But 24.2% scored low. A quarter of the addresses that spent the week rattling doorknobs on our server have a clean record everywhere we know to look, and 34.4% of the group were on consumer lines, which usually means somebody's router or camera has a second job. A reputation score reports what an address has been caught doing. It has nothing to say about a scanner in its first week.

What the scanners were looking for

Across the week, 459 addresses asked for 3,344 different paths that do not exist here, 11,265 requests in total. Most of those addresses tried a path or two and left; the 157 in our scanner group are the ones that kept going. None of it is aimed at us in particular. Every public web server gets this, all day, and the list of what they ask for doubles as a survey of the mistakes people make when deploying software.

Fig. 6 · requests for pages that do not exist

A week of wrong doors, sorted by what was behind them

887 paths · 3,661 requests
GET /.env404×45from 23 addresses
GET /.git/config404×42from 20 addresses
GET /.env.local404×32from 18 addresses
GET /.git/HEAD404×32from 14 addresses
GET /.env.backup404×30from 17 addresses
GET /.env.bak404×30from 17 addresses
GET /.env.production404×30from 17 addresses
GET /.env.old404×29from 16 addresses
GET /.env.dev404×28from 15 addresses
GET /.aws/credentials404×27from 13 addresses
GET /.env.example404×26from 13 addresses
GET /.env.prod404×26from 13 addresses
GET /.env.save404×26from 13 addresses
The most requested paths in each family, with how many times and from how many addresses.
All of these returned 404. We do not run WordPress, we do not keep secrets in the web root, and there is no PHP shell called this_is_a_new_hello_world.php.
Show the numbers
FamilyDistinct pathsRequests
Secrets8873,661
PHP shells1,0202,436
WordPress3211,273
Logo hunting124717
AI tool configs1987
Everything else9733,091

The biggest family is the hunt for secrets. We counted 417 different spellings of a path containing .env: .env.bak, .env.old, .env.save, /backend/.env, /laravel/.env and so on, 1,829 requests between them. Each is somebody's guess at where a tired developer left a copy of the production credentials. Alongside those come .git/config, AWS credential files, Terraform state and CI pipeline definitions.

Two smaller families were new to us. One is aimed at AI coding tools: 19 paths, 87 requests, asking for things like /.claude/settings.json and /.cursor/mcp.json, plus the spend log of an LLM proxy. The volume is small. But a year ago these paths were on nobody's wordlist, and the files they are after tend to hold API keys with a credit card behind them.

The other is odd. 11 image paths, among them /static/icon/bjch.png and /assets/images/logo-cn-vpnweb.png, were each requested from 31 to 33 different addresses, about once per address. Nothing breaks when a logo is missing, so this is not an attack. It is a census. Several of those file names come from the stock themes of VPN and proxy control panels, and whoever runs the scan is mapping where such panels are installed by checking for their artwork. We are a site about IP addresses, so perhaps we looked promising.

Scanners keep hours of their own. Their requests bunch into bursts around 02:00 and 18:00 UTC, with no relation to anybody's working day. As a group they also made 2,169 calls to our score API during the week, nearly all from one address, in between probing for files. We do not know what it was checking. We like to think it was looking itself up.


If you run a site, look at your own

None of this needed special tooling. It needed a log with a referrer column and an afternoon. A few things we would check first on anyone's server, having now done it on ours.

Count addresses and requests as two separate things, because they tell different stories and the average of one over the other describes nobody. Look at where your API is being called from. Ours is in production on 142 sites we did not know by name, and the busiest of them moves our traffic more than anything we do ourselves. Compare the IPv6 share of your signed-in users with the share among the requests you suspect are automated. And sort your 404s by path, once. If one of those 417 spellings of .env would have returned a 200 on your server, that is worth knowing before a stranger with a wordlist finds out.

Method: one week of request-log rows, 2026-11-09 00:00 to 2026-11-16 00:00 UTC, aggregated per address and per hour. Groups are assigned per address in this order: mostly 404s (scanner), any page view on this site (own pages), any API call carrying a referring page (embedded), otherwise direct. Network type comes from the autonomous system announcing the address. Shares may not sum to 100 because of rounding.