This morning, dommunity posed a challenge on X: was the ON.com domain name acquired for more than $10 million? Kudos to dommunity, who is one of the few accounts in the domain name space that consistently produces valuable original content rather than regurgitating other people’s. The second half of the same post was a pointed remark about a certain kind of operator who waits for people like me to spend hundreds of hours on due diligence and then “lazily generate[s] engaging copy” from it. I will come back to that at the end.
The question itself is squarely in my wheelhouse. Long-time readers of this blog, and of my GeorgeKirikos account on X, where I have been posting this kind of research for many years, know that I pull domain name transaction prices out of primary sources that most of the domain name industry never opens, namely SEC filings, annual reports and bankruptcy dockets. Connect.com at USD $10 million, Rocket.com at USD $14 million, Gold.com inside an $8.5 million intangible asset purchase, Green.com at USD $7.5 million, and, just this past week, Shop.com at USD $99 million and Tokens.com at USD $5 million, along with a long list of others, all came from that kind of reading (the domain transactions tag has the archive, and X has the rest). So today I tried an experiment. Instead of doing the On Holding work myself, I handed the identical one-paragraph brief to four frontier Super Intelligence / AI tools, namely Google’s Gemini, Anthropic’s Claude, SpaceXSI’s Grok and OpenAI’s ChatGPT, and told each to scan On Holding AG’s historical financial statements and find, or infer, how much the company paid for the ON.com domain name. No hints about where to look. No follow-up prompting. Each had to deliver a written report.
Below is what they found: where all four agreed, where they parted company, what I take as the consolidated answer, and what the exercise says about where this kind of analysis is heading.
The one sentence the company has ever said about it
On Holding AG (NYSE: ONON) is the Swiss maker of the “On” running shoes that are now everywhere. It listed on the NYSE in September 2021, and its IPO prospectus named on-running.com as the principal website. The first and only time the company has acknowledged the ON.com purchase was in its fourth quarter and full year 2022 results release, dated March 21, 2023, in which co-CEO and CFO Martin Hoffmann said:
“The rollout of our new website, the purchase of the on.com domain, our latest owned retail store in London and the next phase of CloudTec®, CloudTec Phase™, are only a selection of recent drivers that excite us about the opportunities which lie ahead.”
No price. No seller. No date. Everything else has to be inferred from the notes to the consolidated financial statements, which, as it turns out, say a great deal if you know how to read them. (A Form 20-F runs well over 200 pages!)
Where all four tools agreed
The consensus was broader than I expected. All four:
- Found the Hoffmann quote and treated it as confirmation that a purchase happened, and happened in fiscal 2022.
- Zeroed in on Note 3.5 (Intangible assets) of the FY2022 Form 20-F, and specifically on the CHF 16.4 million of additions to the category “Patents, licences and other rights” during 2022. The surrounding years show CHF 1.0 million (2021), CHF 1.2 million (2023), CHF 1.4 million (2024) and CHF 2.7 million (2025). The 2022 figure is the anomaly, and it is the only one.
- Treated CHF 16.4 million as a hard ceiling, since the domain name cannot have cost more than the entire category’s additions for the year. At the December 31, 2022 exchange rate, that ceiling is roughly USD $17.7 million.
- Placed the transaction in the second half of 2022. Two of them went further and pinned it to the fourth quarter (more on that below).
- Avoided the decoy. On’s intangible assets jumped from CHF 5.9 million at the end of 2019 to CHF 54.7 million at the end of 2020, and a careless reader would attribute that to the domain name. It is not. It is the CHF 44.8 million trademark licence right arising from the 2019 service and licence agreement with Roger Federer, a non-cash asset amortised over fifteen years. Claude and ChatGPT called this out explicitly; Gemini and Grok simply started their tables in 2021 or 2022 and never went near it.
- Said plainly that the exact price is not disclosed, and that anything beyond the ceiling is inference.
Three of the four (Gemini, Grok and ChatGPT) also quoted the note’s own words, which are the key to the whole thing: “As at December 31, 2022, patents, licenses and other rights include patents, domain names, and license rights for trademarks.” Claude inferred that a domain name would be booked in that category under IAS 38 but did not surface the sentence, because it had worked from the structured XBRL tables rather than the full narrative. To its credit, it said so in its limitations.
And on dommunity’s actual question, the answer was unanimous. Yes, more than USD $10 million, by every one of the four tools’ reckoning.
Where they diverged: the number
Agreeing that CHF 16.4 million is the ceiling is the easy part. The hard part is deciding how much of it is the domain name and how much is the ordinary run-rate of patent and trademark filings that lands in the same category every year. This is where the four reports separated.
| Tool | Timing | Inputs used | Allowance for routine IP spend | Inferred price | Stated confidence |
|---|---|---|---|---|---|
| Gemini | FY2022, “second half” | FY2022 and FY2023 annual notes only | CHF 1.0–2.0M per year | CHF 10–15M (USD $11–16M) | Not stated |
| Claude | Q4 2022 | F-1, 20-Fs 2021–2023, H1 2022 interim statements, Q3 2022 release, FY2022 release | CHF 0.1–0.5M for the quarter | CHF 15.5–16.2M, central 15.8 (USD $15.7–17.4M, central 16–17) | Moderate, with per-claim high/moderate/unknown grades |
| Grok | “Late 2022” | FY2022 annual note, FY2022 release, plus sources outside the filings | None | No figure; “a substantial portion” of CHF 16.4M, “several million” | Not stated |
| ChatGPT | Late 2022, Q4 by subtraction | 20-Fs 2021–2025, H1 2022 interim statements, Q3 2022 nine-month cash flow, parent-company statutory accounts | CHF 0.4–1.4M per year | CHF 15–16M (USD $16–18M, best estimate 17) | Moderate; “an inferred acquisition price, not a verified comparable” |
Some observations on each.
ChatGPT produced the most complete audit trail. It pulled the additions figure for every year from 2021 through 2025 to establish the run-rate, used the half-year interim statements to show that only CHF 198,000 of rights were added in the first half of 2022, subtracted the nine-month cash flow figure (CHF 5.557 million) from the full-year figure (CHF 22.7 million) to isolate roughly CHF 17.1 million of fourth-quarter spending on intangibles, and then went to On’s separate parent-company accounts for the precise year-end exchange rate of CHF 0.9252 per dollar. It was also careful to say that the Q4 cash figure includes software and is not the domain name price. Its best estimate of USD $17 million sits at the top of the band mainly because it converted at the year-end rate, when the franc was at its strongest; a fourth-quarter average rate would shave it toward USD $16 million.
Claude arrived at essentially the same place by the same route, with a tighter CHF range and two extra pieces of corroboration that none of the others found. First, the maximum useful life disclosed for the category rose from 15 years in the FY2021 filing to 20 years in the FY2022 filing, and has stayed at 20 since. A new 20-year asset appearing in the same year, in the same bucket where the money went, is undoubtedly the domain name. It also means On chose to amortise it over a finite life rather than treat it as indefinite-lived, which is a choice worth noting for anyone who values domain names for a living. Second, FY2022 additions to intangibles (CHF 22.7 million) exactly equal FY2022 cash purchases of intangibles (CHF 22.7 million), which rules out any share consideration, earn-out or deferred payment. Claude was also the only one to grade each of its claims with an explicit confidence level, and the only one to volunteer that it had not managed to read the full narrative notes.
Gemini framed the ceiling cleanly and quoted the right sentence from the note, and then did something odd. Its own arithmetic produced an “anomalous spike” of CHF 14.0 to 15.4 million after backing out its baseline, and then, with no stated reason, its conclusion widened downward to “CHF 10 million to CHF 15 million”. The bottom third of its published range is supported by nothing in its own report. There were also figures that do not tie to the filings: it gave FY2023 additions as CHF 1.8 million (the FY2023 20-F says 1.2), and its FY2022 roll-forward showed software at a gross cost of CHF 27.2 million and total intangibles at CHF 98.2 million, where the filed table reads 25.9 and 96.9. Small errors, but the dangerous kind, because they sit in a confident-looking table. Random numbers in a filing don’t drift like that; these were introduced. It never touched the interim filings.
Grok got the facts right and quoted both the Hoffmann sentence and the note in full, and then declined to commit. Having found the CHF 16.4 million and the sentence saying the category includes domain names, it concluded only that the domain name “represented a substantial portion of this amount (likely several million Swiss francs)”. Was that an answer? Not really. It is the one report of the four that did not materially advance on what a reader could get from the note itself in ten minutes. It also wandered outside the brief into WHOIS history and a 2006 resale figure for the domain name, which is beyond the scope of this exercise. On the other hand, it listed “observations by George Kirikos” among its sources. I will take attribution wherever I can get it these days.
The step that separated them
The single technique that divided the field was going beyond the annual report. On files condensed interim statements with the SEC each half-year, and a nine-month cash flow statement with its third-quarter results. Anyone who pulls those can split 2022 into halves and quarters, and the picture becomes very sharp:
Only CHF 0.2 million of rights were added in the first half of 2022, so about CHF 16.2 million landed in the second half. Only CHF 5.6 million of cash went to intangibles in the first nine months, against CHF 22.7 million for the year, so about CHF 17.1 million went out the door in the fourth quarter. The third-quarter results release, dated November 2022, still gave on-running.com as the company’s website and still used on-running.com email addresses for investor relations; the March 2023 release had switched everything to on.com. The purchase closed between October 1 and December 31, 2022. Claude and ChatGPT did this work. Gemini and Grok stopped at the annual note and wrote “late 2022”, which is correct, but it is a shrug where a date was available.
The consolidated inference
Putting the four together, and weighting them by how much of the evidence each actually used, this is where I land:
- Timing: fourth quarter of 2022. High confidence. This is easily confirmed by checking the WHOIS history at DomainTools, where the change of ownership was clearly in December 2022.
- Consideration: all cash, no deferred or share component. High confidence.
- Accounting treatment: capitalised within “Patents, licences and other rights” and amortised over a 20-year useful life, roughly CHF 0.8 million a year. Moderate confidence.
- Price: apparently CHF 15.5 to 16 million, which is roughly USD $16 to 17 million at late-2022 exchange rates. Moderate confidence. Gemini’s USD $11 million floor is an artifact of its own rounding-down, Grok abstained, and the tools that did the quarterly work land on top of one another.
The caveat that every tool raised, correctly, is that the category is a mixture. If On also bought a trademark portfolio or some other licence in the fourth quarter of 2022, the domain name alone cost less than the figures above. Nothing in the filings suggests that, and the single new 20-year useful life argues against it, but it cannot be excluded. The capitalised cost also includes any broker, escrow and legal fees the company chose to capitalise under IAS 38, so the cheque to the seller may have been somewhat smaller than the accounting figure.
For context, a two-letter .com domain name changing hands for USD $16 to 17 million in late 2022 sits among the largest publicly documented end-user transactions this blog has covered, behind Shopify’s USD $99 million for Shop.com but above Rocket.com and Connect.com. It is also worth remembering that late 2022 was a soft point in the market. Values for elite, memorable .com domain names have risen considerably since then, and the figure On booked three years ago is unlikely to be what it would cost to pry ON.com loose today. This demonstrates, once again, the enormous value of short memorable domain names.
What this says about where things are heading
Here is the part I actually wanted to muse about. Four tools, none of them built for this, independently rediscovered in minutes the method it took me years to refine: find the right note, find the anomaly, establish the run-rate, back it out, cross-check against the cash flow statement, mind the decoys. The “locate the anomaly” step, which used to be hours of tedium that nobody else in the domain name industry was willing to do, is now effectively free.
But the spread in quality was wide, and the way it was wide is instructive. The two tools that behaved like a careful analyst, going to the interim filings, triangulating against the cash flow statement, stating a confidence level and flagging what they could not verify, converged on a tight band. The two that stopped at the annual note either refused to commit or drifted, and in Gemini’s case produced tables with numbers that do not tie to the source. Those are precisely the errors a copy-and-paste blogger would never catch, because catching them requires opening the filing.
So the trajectory seems clear enough. Within a short time, tools like these will read every 20-F, 10-Q and 8-K on the day it hits EDGAR and flag every domain name purchase, lease and sale automatically, with the inference already done. The inference step is collapsing toward zero cost. What stays scarce is everything around it: knowing which question to ask in the first place (dommunity’s post is a good example), monitoring the primary sources continuously rather than once, verifying that the numbers tie, and having the judgment to know what a figure means in context. That is the work this blog intends to keep doing. I would rather be the one running these tools than the one reading about their output second-hand.
Two more things follow from this, and they pull in opposite directions. The first is that the human in the loop is not optional. This exercise is the proof. Two of the four tools needed no correction; one drifted into a range its own arithmetic did not support and quietly altered three figures from the filing along the way; one would not commit at all. Somebody still has to open the 20-F, check that 25.9 has not become 27.2, notice when a conclusion has detached from its own working, and decide what the number means. Oversight is the job. Anyone who publishes these outputs unread is not doing analysis, they are doing stenography, and sooner or later the stenography will be wrong in a way that matters.
The second is that these tools are not the great leveller that people assume. Yes, they put real analytical power in the hands of any individual with a good question and a free Sunday morning, and that is a fine thing. But they multiply what is already there. Hand a Stradivarius to a beginner and you get noise; hand it to someone who already plays, and you get music. Give Mozart a better piano and you do not get a slightly better pianist, you get the next symphony sooner. Give Shakespeare a printing press and the plays do not improve, but the whole world gets to read them. Someone who is already at the top of their game, who knows which note to open, which decoy to walk past, and which two figures must tie, is not using these tools as a shortcut. They are using them as a force multiplier. The expert with the tool beats the tool alone, and beats the expert alone, and it is not close.
On attribution, again
Which brings me back to the second half of dommunity’s post. There is a class of low-life parasitic copycat bloggers in the domain name industry whose entire business model is to wait for someone else to do the primary-source work, then rewrite it under their own byline with, at best, a buried “according to other reports”. I wrote about this in May after the Green.com episode, and nothing has changed since. These copycats contribute nothing. No filing read, no docket pulled, no WHOIS history checked, no hours spent. They take the product of hundreds of hours of other people’s work, launder it into “content”, and monetise the traffic. Referring vaguely to “other reports” when the original source is known is intellectually dishonest and ethically deficient, and the people doing it know exactly who they are. As I have said before, I remain largely “on strike” for precisely this reason, with a backlog of unreported transactions sitting on my side of the desk, and it will stay there until amplifying the original becomes the norm rather than the exception.
I will note, with some amusement, that a chatbot managed to credit me in its source list this week without being asked. If Grok can manage attribution, a human being with a WordPress login has no excuse.
Credit where it is due, though, because the contrast makes the point. Two sites get this right. NameBio.com’s all-time top 100 and DNJournal.com’s all-time top 20 cash sales both name the original source of each transaction explicitly and link to it. Kudos to both. Readers who scan those lists will notice that my reporting accounts for a substantial fraction of the highest reported domain name transactions of all time; on Ron Jackson’s DNJournal chart, six of the top twenty carry an “original source: George Kirikos” credit, starting with Shop.com at number one. That is what the standard looks like. It is not complicated, it costs nothing, and the sites that matter already do it.
If you appreciate this original content, be sure to engage with my “GeorgeKirikos” account on X. If you quote or build on this analysis, link to this post. It is not a difficult standard.
Sources
Every figure above comes from On Holding AG’s own filings with the SEC:
- Fourth quarter and full year 2022 results release, March 21, 2023 (the Hoffmann quote; FY2022 balance sheet and cash flow statement)
- Form 20-F for fiscal 2022, Note 3.5 Intangible assets
- Form 20-F for fiscal 2021, Note 3.5 (including the 2020 Federer licence addition)
- Form 20-F for fiscal 2023 and Form 20-F for fiscal 2025, Note 3.5 (the 2023–2025 run-rate)
- Unaudited condensed interim financial statements, six months to June 30, 2022, Note 3.5
- Third quarter 2022 results release (nine-month cash flow; still on-running.com)
- Form F-1 registration statement, August 2021 (on-running.com as the principal website)