7 lessons worth learning about early social networks statistics

What to take away

Lesson 1: Nobody was counting

An operator running a service for a few hundred people had no reason to measure them. The service was a hobby, a club, a side project or a department's experiment. Its scale was known the way a village's is: by recognition, not by tally.

Where money changed hands, counting had a purpose. The WELL, opened in Sausalito in 1985, charged roughly $8 a month plus about $2 an hour of connect time, so its billing records held a number. That number covered paid accounts, not the people using them.

On Usenet, started in 1980 by Tom Truscott and Jim Ellis at Duke and the University of North Carolina, the countable thing was the number of sites taking a news feed. BITNET, running from 1981 between the City University of New York and Yale, counted member institutions.

FidoNet, which Tom Jennings began in San Francisco in 1984, counted nodes, meaning the sysops who ran them. The server log, where one existed, was kept for debugging and rotated when the disk filled.

It was not an audience record, and nobody treated it as one. The idea that it might one day be historical evidence did not occur to anyone in a position to preserve it.

Where operators did report figures, they reported them to attract members or investors. The figures were chosen for effect, and there was no independent check. A number in a press release from the period is a number the operator wanted printed.

Lesson 2: The units were not people

Even a kept count would count something other than what the reader wants to know. Accounts were one unit, and one person could hold several, which were never cleaned. Logins counted sessions, not people.

Connections counted a stored relationship without saying whether either party remembered it. Messages counted activity by the most active. Classmates.com, launched in 1995, counted registrations, including the free ones, and put contact with former classmates behind a paid tier that typically ran a few dollars a month.

None of these converts to people without a rule, and the rule was never written down, because nobody at the time needed to convert. GeoCities, which Yahoo bought in 1999 for about $3.6 billion in stock, counted homesteaders, its word for accounts, not visitors.

A modern figure that says how many people used an early service has performed that conversion silently. It rests on an assumption the reader is not shown, applied to a count the reader cannot inspect. LiveJournal, started by Brad Fitzpatrick in 1999, labeled accounts active by a login window the site set itself.

The same unit problem persists on every current service, and the counts described under online status and identity are defined by their issuers in exactly this way. The difference is that the current issuers exist and can be asked. The early ones do not.

Lesson 3: The data went with the operator

Services closed, and their databases closed with them. GEnie, run by General Electric Information Services from 1985, shut down in 1999. Prodigy's classic service ended the same year, and AOL bought CompuServe in 1998 and folded its consumer service in over the following years.

The account tables, the connection lists and the logs were on machines that were sold, wiped or thrown out. No institution was collecting them. What escaped was whatever an individual member had saved for private reasons, a record of that member's corner and nothing more.

The archive kinds described under internet culture archives hold public pages, which can be counted in a limited sense. The operational data that would support an audience figure was never public and is in none of them. A count assembled from captures is a count of captures.

Lesson 4: What survived is not a sample

Everything that survived from the period survived because it was popular, public, textual or someone's favorite. That is survivorship bias in its purest form: a figure built from the remains describes the remains. Those remains were selected by the accidents that preserved them.

Usenet is the clearest case. Its text survived through archives Deja News began building in 1995, which Google acquired in 2001, so the surviving corpus counts posts. A bulletin board running on one phone line in a bedroom left nothing of the kind.

The small, private and unremarkable services, which were most of the services, are absent. A total that omits them is not a lower bound on anything, because the omitted part is of unknown size.

No method repairs this. A biased sample of known size can be corrected; a biased sample of unknown size and unknown composition cannot, and that is what the record of this period is.

Lesson 5: Retrospective estimate dressed as measurement

Most circulating figures were made years later. People needed a number for a chart, so they used partial records and assumptions to close the gaps. SixDegrees.com, launched in 1997 and shut down in 2001, is usually given in round millions of registered profiles, and that figure traces back to the company's own marketing.

The estimate was cited. The citation was cited, and the chain looked like measurement. Trace any early network scale figure back through its sources, and it usually ends at one estimate with no stated method, or at nothing.

The platform timelines page describes the same process for dates: agreement among sources that share an ancestor is not confirmation. For figures it is worse, because a date can at least be checked against a document, and there is no document behind most of these numbers.

Lesson 6: What the record can carry

The record supports structure and sequence claims on the early social networks page: which capabilities existed, in what combination, and roughly in what order. CompuServe carried forums and email before the web.

Friendster, launched in 2002, put the friend-of-a-friend model on a profile page, and MySpace followed in 2003. It supports relative statements, made carefully: one form was more widespread than another in surviving material, with the survival bias acknowledged. It supports participant testimony about what their corner felt like, marked as testimony.

It does not support a total, a growth rate, a share or a numeric comparison between services, and this site states none. A page called statistics that gives no statistics is the honest version.

Lesson 7: Reading a figure when you meet one

Common questions

Are there really no reliable figures at all?

There are figures with stated methods, for specific services, at specific moments, where an operator or a researcher happened to count and the count happened to survive. CompuServe's own subscriber announcements are one type. They describe that service at that moment. Nothing generalizes from them.

Why do so many sources give the same number?

Because they are copying the same estimate. Agreement is evidence of a shared source, not of accuracy.

Would archived pages let someone count now?

They would let someone count captured pages, which measures what was public, popular and crawled. The count would be of the archive, not of the period.