In the year 2000, one in 22 new US trademark filings ended in .com. Not one in 22 technology filings. One in 22 of everything: the restaurants, the plumbers, the shampoo, the tractor parts. Nothing else in the trademark record comes close.
The federal trademark register is one of the few naming records that runs continuously from the 1800s to now, so we loaded the full US Patent and Trademark Office case file dataset, 11.2 million marks with text, and counted how often each naming convention appears. Filing volumes grew enormously over that period, so every figure here is a rate rather than a count: marks per 10,000 filed in that year.
The whole curve
In 1993 the rate was 0.10. Four years later it was 46.8, which already looks like a trend. Then 1998 closed at 83.1 and 1999 at 458.7, a 5.5x jump in a single year, and 2000 held the same altitude at 461.6.
For scale: “& Co”, the most recognisable naming convention in the older half of the register, peaks at 68.0 per 10,000 in the 1900s. The .com peak is 6.8 times higher than that. It is the largest single naming event in 150 years of the register, and it lasted two years.
The crash, and the part nobody remembers
2001 came in at 120.1, a fall of 74% from the peak in twelve months. That much is the familiar story.
The unfamiliar part is what happened next. .com does not keep falling. It climbs back for most of the decade and reaches 113.0 by 2009, within striking distance of the 1998 rate, before beginning the long decline that takes it to 22.8 in 2023. The bubble popped in 2001. The naming convention took another twenty years to fade, and on this evidence it faded because something replaced it rather than because it failed.
Cyber ran first, not alongside
Ask most people to place “Cyber” naming and they will put it with .com, in the late nineties. The register disagrees. Cyber peaks in 1995 at 40.2 per 10,000, and it is already falling while .com is still climbing. By 2000, the year .com tops out, Cyber is down to 18.7.
These were two different waves with a five year gap, and the earlier one was over before the famous one began. It is the sort of thing that is obvious in the data and invisible in memory.
Each wave hands over to the next
Widen out and the same shape repeats. Every one of these terms rises, peaks, and gives way, and they do it in order.
| Term | Peak | At peak | 2020s |
|---|---|---|---|
| Systems | 1980s | 43.3 | 7.2 |
| Technologies | 2000s | 46.2 | 19.7 |
| Solutions | 2000s | 51.2 | 27.8 |
| Smart | 2000s | 48.8 | 26.5 |
| Cloud | 2010s | 15.9 | 11.0 |
| AI | 2020s | 11.9 | 11.9 |
Five of the six have already peaked. Only AI has not.
Where AI sits
AI is flat at well under one per 10,000 for twenty years. It reaches 1.1 in 2015, 2.2 in 2016, then sits near 8.4 through 2022 before jumping to 25.7 in 2023, a 3.1x rise in the first full filing year after ChatGPT.
On the .com curve, 25.7 per 10,000 is roughly where .com stood in 1995: past the point where you could call it noise, and two years before the part everyone remembers.
That is where the data stops. The case files we hold run through the 2023 filing year, so we can say where AI sits on the arc and we cannot say whether it has peaked. Anyone telling you which of those it is does not have the filings to know yet, and neither do we.
How this was counted
Two of these terms are close to untestable as bare text searches, and it is worth showing why, because it is the reason this analysis is rarely done.
Searching marks for “AI” anywhere in the text returns 380.7 per 10,000 at peak. The true figure is 11.9. The difference is AIR, HAIR, CHAIR, MAIL and every other word with those two letters inside it: 96.9% of the matches are wrong. “Net” is 93.2% wrong for the same reason, collecting INTERNET, KENNETH, BONNET and MAGNET. Every pattern behind the charts above requires a word separator, so “AI” matches a word and not a fragment.
One further check matters for a dataset like this. A single mark filed across seven product classes is seven rows, so one company protecting a name broadly can look like seven companies choosing it. Every headline above was therefore computed twice, once per filing and once per distinct mark text per year. The two agree within a few percent throughout, with a single exception we excluded by name: Cloud in 2000, where 402 filings turn out to be 97 distinct marks belonging to one applicant’s defensive portfolio.
Rates are per 10,000 marks filed in that year, dated by filing date where the register records one and registration date otherwise. Source is the USPTO Trademark Case Files dataset, 2023 case files, to March 2024.
If you are naming something now, the register will tell you whether it is already taken. You can check a name against all fifty states, or look at formation and trademark activity in a single state such as Delaware.