Summary

  • Packet Clearing House's Internet Exchange Directory prints eleven column labels and, in the static subset I retrieved, one LINX LON1 row: 930 participants, 2,218,258 prefixes, 4.31T peak, 3.65T average, established 8 Nov 1994.
  • The dataset documentation defines a "Participants list url" field and a "Date IXP last updated" field, and in the text I read it states no sourcing rule, no refresh cadence, no directory-level date, and no unit for the traffic figures.

Eleven columns, one row I can quote

The page is a table, and the table has a header. As retrieved, it reads Region, Country, City, IXP Name, Participants, Peak, Avg, IPv6, Prefixes, Established, URL. One row follows: Country United Kingdom, City London, IXP Name LINX LON1, 930 participants, 4.31T peak, 3.65T average, a blank IPv6 cell, 2,218,258 prefixes, established 8 Nov 1994 (directory page).

The same page tells you not to trust it as a whole, in its own words: "JavaScript Required - This page requires JavaScript to work. You are viewing a static copy which is only small subset of the entire IXP Directory." That sentence is the most useful thing on it. What I have is a fragment of a larger instrument, printed by the publisher, and I will treat it as exactly that: one row, quoted as printed, not a census of the world's exchanges and not a full export.

For a reader deciding whether two networks share a building, that row is the ordinary first stop. So the question worth asking is what the row is made of.

The schema has a place for the exchange's own list

The directory's dataset documentation is where the row becomes legible. It opens with a sentence I will quote in full, because it is the publisher's own description of its product: "Packet Clearing House maintains the global directory of Internet exchange points. The directory reports on past and current IXPs, peering subnets used by the members to exchange traffic, locations, equipment and membership."

Then it lists fields. Twenty-three of them: id, ctry, cit, reg, name, url, stat, date, prfs, lat, lon, prts, traf, avg, trgh, ipv6_avg, pch, updt, ranks, prts_url, q9, iata, regct. Two of those fields carry definitions in the retrieved text, and both are the ones that matter here. prts_url is defined as "Participants list url". updt is defined as "Date IXP last updated".

The schema, in other words, has a slot for an exchange's own participant list and a slot for the date the row was last touched. It documents JSON by default, CSV by appending ?format=csv, and a Subnet Member Details endpoint returning ip, fqdn, ping, asn, org and prefixes, documented as JSON-only.

That is a real instrument. It is also, at this level of abstraction, a filing cabinet with labelled drawers and no registrar's signature.

What the retrieved text leaves out

The documentation I read says nothing about where the values come from. It does not say who fills prts_url, whether exchanges submit it, whether anyone checks it, or whether the field is usually empty. It does not say how often a row is refreshed, or on what trigger. It prints no last-updated date for the directory as a whole. Not every listed field gets a definition, and the definition of prts — the participants figure behind the 930 — is not among the ones I retrieved.

I want to be exact about the shape of that sentence, because it is the article. I am not telling you these rules do not exist. I am telling you that the pages a stranger lands on, fetched on 19 November 2026, do not print them. A methodology note could sit behind a login, on a page whose navigation I did not follow, or in a document written for customers. A directory is not obliged to publish its intake rules. But a row is only evidence if a reader can find out what it counts.

The only provenance I found is dated, and it is not the publisher's current page

Two statements about where these rows come from do exist in public. Neither is a PCH methodology page, and neither is current.

The first is a mailing-list post of 9 July 2014, attributed to Bill Woodcock, in the NANOG archive: "The PCH directory is twenty years old, and is independently verified by our staff. So what’s there isn’t always up-to-date, but we do differentiate between rumor and something that’s been verified by someone going and laying eyes on it." Twelve years old, and candid in a way no corporate page would be: staff verify, and freshness is a matter of degree.

The second is a 2016 paper, A Comparative Look into Public IXP Datasets, reporting a crawl dated 19 September 2014. It says that, according to direct communication with PCH staff, about 70 per cent of listed IXPs were compiled by PCH staff, 25 per cent contributed by the Internet community and about 5 per cent added by the IXP operators themselves; that PCH peers at many exchanges itself and uses the BGP information from those peerings to derive participant lists; and that participant data is counted per port, so one ASN can appear more than once as a member of an exchange.

Read those two sentences together and the 930 changes character. If a participant is a port, then 930 is a count of entries, not of distinct networks, and the reader comparing it to their own member roll is comparing two different objects. I did not open the full paper, so I report that as a secondary source's finding about a 2014 crawl, not as PCH's rule today.

No unit, no window, so the row cannot be aged

The row prints peak 4.31T and average 3.65T. It does not print what T is, on that page or in the field list I retrieved. Terabits per second is the obvious reading and I am not going to adopt it on obviousness; the honest statement is that the directory's own static table gives a magnitude and leaves the unit to a definition I could not find. The IPv6 column is blank for this row, which is a blank cell and not a zero.

Nor is there a measurement window. A traffic figure without a period is a number missing its denominator: 4.31T over what interval, sampled when, averaged how? The per-row updt field would answer the second half of that, if a row carried it. The static table I retrieved does not.

The detail page agrees on identity and shows me nothing else

LINX LON1 has its own page. It confirms the row's identity: Short Name LINX LON1, City London, Country United Kingdom, Organization Type Association, Status Active, Media Type Ethernet, Managed By LINX, Organization LINX, Established 8 Nov 1994. That agreement is worth something. The aggregator's index and the aggregator's per-exchange record tell the same story about what the thing is.

Then the page runs out. The Subnets table reads "No subnets". The sections that would carry member detail read "MRTG Loading...". I am not going to convert a placeholder into a finding. A dynamically rendered member table that did not populate in my fetch is a fact about my fetch. It is not evidence that PCH's database holds no subnets for LINX LON1, and it is not a contradiction of the 930 in the row. It is simply the point where the instrument stopped being readable to me.

The quick read that stands in for the argument

LINX publishes the rules that decide who may be a member, and the port price to the pound; the register of members itself returned a cookie-consent shell to an unauthenticated read that same day, in my own reading published on 19 November 2026. No member name, ASN or count came back. So the one number in circulation about that exchange's membership, in front of a stranger, is the aggregator's: 930.

That is not an accusation. It is the structure of the thing. The exchange keeps its roll; the directory prints a count of it; the reader gets a number with no stated supplier, no stated counting rule and no stated age, and the only place to check it is the party that chose not to publish.

Where the publisher does print a date

One page in the set does carry recency, and it is not the directory. The IXP locations summary prints "Last updated Wednesday 18th of November." No year in the retrieved text. It also prints "This report is free for reuse and adaptation with attribution." A report offered for reuse, refreshed on a schedule the reader can see, sitting beside a directory that offers no such line is a distinction worth holding onto. The summary is not the table, and I will not use its date to age the row.

Two chrome pages, recorded as what they returned

The address I first asked for, pch.net/ixpdir, returned site frame: a maintenance notice reading "Website maintenance planned from Mar 1st 05:00UTC until Mar 4th 05:00UTC. Portions of our web site may be unavailable.", plus navigation labels "Download datasets", "Submit a change" and "Feature/Bug Report", and no rows. pch.net/resources/data returned HTTP 404. The working table was at /ixp/dir and the working documentation at /ixp/data. That is navigation learned by fetching, and I record it because a reader trying to check my work will hit the same doors.

"Submit a change" is the interesting label, and it is only a label. It tells you the publisher anticipates corrections arriving from outside. It does not tell you who may submit, what evidence is required, or what happens to the row afterwards.

What the row is good for

Treat it as a lead with a provenance gap, and it stays useful. Treat it as a certificate and you are citing a number whose counted object, supplier and age you cannot state. Between those two uses sits every peering deck, procurement memo and filing that has ever quoted an exchange directory without saying which column it read.

BTW has read Packet Clearing House's corporate record before — a 4 November 2026 Research article put a California company number, a tax classification, a masthead sentence and one public number-resource record against a single directory label. This piece is the other half: not what the organisation is, but what its operational instrument does with an exchange's data. The answer, on the pages I could read, is that it builds a labelled slot for the exchange's own list and then declines to say who is expected to fill it.

The fourteen limits this reading carries

  1. Everything here is a reading of pages retrieved on 19 November 2026. A page that returned frame, an error or a partial rendering is a fact about that retrieval, not a statement about what the publisher holds.
  2. The static table says of itself that it is only a small subset of the directory. The LINX LON1 row is one row, not a census, and no AMS-IX row was captured in what I retrieved.
  3. The table prints 4.31T and 3.65T without printing the unit, and defines none of its columns in its own text. Those figures are quoted as printed and are not converted into a measured bandwidth.
  4. The field list names fields and defines prts_url and updt. It does not define every field, and no definition of the counted object behind the participants figure was retrieved.
  5. The documentation prints no sourcing rule, no refresh cadence and no directory-level date in the text retrieved. That is an absence from the pages I read, not proof that no methodology statement exists anywhere.
  6. "No subnets" and "MRTG Loading..." are what the rendering showed. They are not evidence about the contents of the database and not a contradiction of the 930.
  7. The 930 is PCH's row value, not a verified count of unique member networks. The per-port counting in the 2016 paper would make a participant count a count of entries; that report was not verified in the full paper.
  8. The 9 July 2014 list post and the 2016 paper are not PCH's current published methodology. Their dates and attributions travel with every use here, and nothing in them is presented as describing 2026 practice.
  9. The paper was read as landing and abstract text in a retained snapshot; the full paper was not opened.
  10. The IXP locations summary is a different page from the directory, its update line carries no year in the retrieved text, and it is not used as the directory's own cadence.
  11. The requested /ixpdir returned frame and navigation with no rows; /resources/data returned HTTP 404. Both are recorded as retrieval results.
  12. My LINX reading is my own prior work: the member portal returned a cookie-consent shell on 19 November 2026 and no member name, ASN or count was obtained. Nothing here infers a count from any other figure.
  13. No BGP session, looking glass, collector, RPKI repository, IRR database or directory API query was run, and no live routing measurement is reported.
  14. Nothing here asserts that a row is wrong, that an exchange was omitted or favoured, that any status was bought, or anything about the conduct, staffing, funding or storage of Packet Clearing House or any exchange that a cited document does not itself state.