Leadership and readiness
First week startup traction analysis: what 7 days can and cannot prove
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
First week startup traction analysis can show that a launch reached people and that a pipeline has begun; it cannot show a growth rate, conversion, retention or repeat purchase, which need later periods. A replay of the Model Context Protocol's public server list makes the point: week one added 2 third-party listings, the median week through week 13 added 15, and the busiest week added 189.
License terms
First week startup traction
Key takeaways
- Week one is a level measured once; a growth rate needs at least two periods after the launch effects fade.
- For a data licensing business, sign-ups, inquiries and first deals each settle only at a later event: a passed review, a written scope, an accepted delivery, a second purchase.
- On the MCP server list, week one added 2 listings and ranked 46th of 66 weeks; the median week from weeks 2 to 13 added 15.
- The 10 busiest listing days held 43.2 percent of 1,439 entries: spikes recorded review batches, not demand.
- Repeat evidence arrived months later: of 66 launch-month entries, 58 were still listed in February 2026.
What can first week startup traction analysis actually prove?
It can prove that a launch reached people and that a pipeline has started to form. It cannot prove a growth rate, a conversion rate, retention or repeat purchase, because a rate needs at least two periods and the later steps have not had time to happen. Week one is a level, measured once, under conditions that will not repeat.
That distinction is old in startup practice. Paul Graham's essay Startup = Growth sets a good growth rate during Y Combinator at 5 to 7 percent a week, a figure that only exists as a comparison between one week and the next. A week-one count has nothing to compare with, and the next few weeks are shaped by the launch itself: announcement traffic, partners lined up in advance and a backlog of people who were waiting for the doors to open.
For a business that licenses data, the gap is wider than for a consumer app. A sign-up from a company offering records is not licensable supply until its rights and records have been reviewed; an inquiry from a model developer is not demand until a written scope exists; a first signed deal is not repeatable revenue until a second buyer, or the same buyer again, pays for more. The ledger below separates what each early signal can carry.
The week-one evidence ledger for a data licensing business
Read each early signal for the claim it can support and the event that would settle the claim it cannot. The settling events are stated as events, not days, because their timing differs by company and is itself something to ask for:
| Early signal | What it can establish | What it cannot establish yet | Evidence that settles it |
|---|---|---|---|
| Supplier sign-ups | The offer reached companies willing to apply | That they hold records they may license | Applicants that passed rights and records review, with rejection reasons |
| Buyer inquiries | A model developer found the company | Budget, task fit or willingness to pay | Inquiries that became a written scope, counted by buyer |
| First signed license | One counterparty accepted terms once | Price durability or delivery capability | Signed terms, open conditions and acceptance status |
| First accepted delivery | The company can prepare data that passes one test | Repeatability across suppliers and buyers | A second accepted delivery from a different supplier |
| First cash received | Money moved for a delivered dataset | That it recurs | Receipts at the buyer's next purchase or refresh |
| Second purchase by a buyer | Repeat demand from that buyer | Breadth across buyers | Repeat purchases spread over several buyers |
Replaying a real launch: the Model Context Protocol server list
A public record lets the exercise be checked, so we replayed one. When Anthropic announced the Model Context Protocol on November 25, 2024, it shipped an open-source repository of servers. Its README soon held a public list of third-party servers, in two sections: Official Integrations, which the README describes as maintained by companies building production-ready servers for their platforms, and Community Servers. Every addition is a dated commit, so the list's first week can be compared with what followed.
We walked the main branch of the repository at commit ef9808f, read every version of the README and dated each entry by the first version that listed it. Week one, November 25 to December 1, 2024, added 2 third-party listings. The Official Integrations section did not exist until December 3. The list then grew for more than a year, was marked as no longer maintained on November 26, 2025, and was removed in April 2026 in favor of the MCP Registry, after 1,439 distinct entries.
A reader at day seven would have seen one of the quietest weeks in the list's life. Week one ranks 46th of 66 weeks (tied weeks share a rank), and it holds 0.1 percent of everything ever listed:
| Window after announcement | Third-party listings | Median per week | Share of all 1,439 |
|---|---|---|---|
| Week 1 (Nov 25 to Dec 1, 2024), all a day-seven reader saw | 2 | Not applicable | 0.1% |
| Weeks 2 to 4 | 64 | 21 | 4.4% |
| Weeks 2 to 13 | 158 | 15 | 11.0% |
| Weeks 14 to 26 | 340 | 22 | 23.6% |
| Weeks 27 to 52 | 893 | 22.5 | 62.1% |
| Busiest week (week 29, June 9 to 15, 2025) | 189 | Not applicable | 13.1% |
Why the busiest weeks measured the gatekeeper, not demand
Because a listing appeared when a maintainer merged it, not when someone asked. The 1,439 entries landed on 151 distinct days, and the 10 busiest days hold 621 of them, 43.2 percent. On June 9, 2025 alone, 128 entries were listed, most of the busiest week's 189. Those spikes record review batches clearing a queue, which says more about the gatekeeper's calendar than about demand that week.
Data licensing has the same structure. A supplier is counted only after a rights and records review, a buyer only after a scope is written and terms are signed, and each of those gates is staffed. Signings that bunch at quarter end or after a hiring round are throughput, not demand, and a week-one count taken before the first gate has cleared anything measures nothing but interest.
Launch weeks add two more distortions. Anthropic's announcement named early adopters already integrated at launch, such as Block and Apollo, and development-tool companies working with the protocol, so part of any launch count is business development done beforehand. And interest itself moves: Sadeghi and colleagues, studying experiments at Microsoft, describe novelty effects that fade over time and primacy effects that grow as users adopt a change. A first week can overstate or understate what follows; the MCP list did the second.
How wide is a conversion rate built on one week?
Too wide to rank companies on, in almost every real case. Brown, Cai and DasGupta's review of binomial intervals in Statistical Science (2001) found the textbook Wald interval erratic and recommended the Wilson or Jeffreys interval for small samples, so the figures below use Wilson 95 percent intervals. The same 25 percent conversion rate means very different things at 12 inquiries and at 120.
The zero-repeat row matters most for first-week claims. When no buyer has had time to come back, the zero is not a measurement at all. Even once buyers have had the chance, zero repeats from 12 is compatible with a true repeat rate anywhere up to about 22 percent at 95 percent confidence, or 25 percent by the rule of three. The inputs are illustrative and describe no company:
- Illustrative: 3 conversions from 12 inquiries is 25.0 percent, with an interval of 8.9 to 53.2 percent.
- Illustrative: 30 conversions from 120 inquiries is also 25.0 percent, with an interval of 18.1 to 33.4 percent.
- Illustrative: 2 of 2 first buyers renewing is 100 percent, with an interval of 34.2 to 100 percent.
- Illustrative: 1 accepted delivery from 5 attempts is 20.0 percent, with an interval of 3.6 to 62.4 percent.
- Illustrative: 0 repeat purchases from 12 first buyers has an upper bound of 22.1 percent (exact) or 25 percent (rule of three).
Repeat purchase cannot show up in week one
Repeat purchase is the evidence a data licensing business most needs and the one a launch week cannot contain. The standard customer-base models make the dependency explicit: Schmittlein, Morrison and Colombo's Counting Your Customers model in Management Science (1987) works from the number and timing of each customer's previous transactions. With one transaction per customer there is nothing to model.
The MCP replay shows how late such evidence arrives. Of the 66 entries listed in the first four weeks, 58 (87.9 percent) were still listed in the last version of the list in February 2026, and for 14 of them (21.2 percent) the same owner listed another server in a later week. Neither number existed at day seven; for the two week-one entries, neither owner listed again. In licensing, the equivalent is a buyer returning for a refresh, a second dataset or a renewal, and it appears only at that buyer's next purchase cycle.
Where the week-one ledger could mislead an investor
Five things could make this framework, or the public example behind it, wrong for a particular company. Some would understate an early signal and some would overstate it, so weigh each one before applying the ledger:
- Pre-sold contracts. A license signed in launch week may rest on work done before launch, as the MCP announcement's early adopters had already integrated the protocol; it is real evidence of the pipeline, but not of the launch.
- Free listings are not paid demand. Adding a server to a public list cost nothing, so its timing and churn may not transfer to signed data licenses, where each step has a cost and a counterparty.
- Early position can compound. In markets where the first credible supplier sets terms or becomes the default, a fast start can matter even before conversion data exists; the ledger records that judgment as a judgment, not as measured traction.
- One buyer can be most of the evidence. A strong first quarter driven by a single large license says more about that buyer's budget than about the market, and its renewal is a single point of failure.
- The measuring gate can change. If a company speeds up or tightens its rights review, early counts move without any change in demand, so the gate's definition must be dated along with the count.
Questions for any data licensing startup about its first weeks, SourceX included
Ask for denominators, dates and the second event, and apply the same list to every company in the space. SourceX's investors page says it shares company and commercial metrics with investors directly, and this page states none of them. Accredited investors evaluating SourceX, which is raising under Rule 506(c), can request its investor materials at sourcex.si/investors and put any of these questions to the company:
- Which gate defines a counted supplier, and what share of applicants passed it, with the rejection reasons?
- How many inquiries became a written scope, and how many distinct buyers do they represent?
- For each signed license, what conditions were open at signing, and which deliveries have been accepted?
- What share of receipts came from the largest buyer, and from buyers who have purchased more than once?
- How many weeks or months of data sit behind any growth rate quoted, and was the launch period excluded?
- Which counts were produced by a review batch, a pre-arranged partner or a one-time event?
Methods note: the MCP list replay and the illustrative figures
The replay reads git objects only. Every commit on the first-parent line of main of github.com/modelcontextprotocol/servers, up to ef9808f (fetched October 27, 2026), is compared with its own first parent, and a README version is any commit that changed README.md. An entry is a list line where a link follows a bold marker (closing ** not required) inside a section whose heading contains Official Integrations or Community Servers; its identity is the first link URL, lowercased without a trailing slash, and its listing time is the committer timestamp of the first version that contains it. Day 0 is November 25, 2024, the date of Anthropic's announcement; weeks are exact seven-day intervals, zero weeks count in medians, and percentages are rounded once, half up.
Survival means present in the last README version that still held entries (February 25, 2026, in UTC like every date here; February 24 on the committer's local clock). An owner is the first path segment of a GitHub or GitLab URL, otherwise the host, and a repeat is an owner with an entry listed in a later week. Intervals are Wilson score 95 percent intervals; zero-event bounds use 1 minus 0.05 to the power 1/n and the rule of three. The ledger, the replay and the questions were drawn up for this article on October 27, 2026; the scripts are kept for review, and the saved data holds no names, usernames or URLs.
Not measured: when each listing was requested, which the commit history does not record; renamed or moved repositories, which count as a removal and a new entry; any paid market; and any SourceX metric, none of which is public.
Frequently asked questions
Is a strong first week a good sign for a startup?
It shows reach, not a trend. A first week mixes announcement traffic, pre-arranged partners and pent-up interest, and it has no earlier period to be compared with. Treat it as the starting level and ask what the next eight to twelve weeks did.
How many weeks of data do you need before quoting a growth rate?
At least two periods for any rate, and enough weeks after the launch to see past it. In the MCP replay, week one held 2 listings while the median week from weeks 2 to 13 held 15, so a rate anchored on week one would have been wildly high and meaningless.
What should an investor ask for instead of week-one sign-ups?
Counts that have passed a defined gate, with denominators: applicants that cleared review, inquiries that became written scopes, deliveries that were accepted, and buyers who purchased again. Each should come with dates and the definition of the gate.
How do launch spikes distort lists of the fastest growing AI companies?
A ranking that compares one company's launch month with another's ordinary month rewards timing, not growth. Spikes can also come from batch processing, as with the 128 MCP listings merged on one day, so a growth list should state its window and exclude launch periods.
Does SourceX publish its traction or growth figures?
No. SourceX's investors page says it shares company and commercial metrics with investors directly, and this article states none of them. Investor materials can be requested at sourcex.si/investors, where the questions listed above can be put to the company.
Sources
- The git history of the modelcontextprotocol/servers repository, read at commit ef9808f on October 27, 2026, supplies every listing count, week, merge-day and survival figure in the replay. Source
- The README version of February 25, 2026 (UTC), the last to hold the lists, describes Official Integrations as maintained by companies building production-ready servers and notes that the server lists are no longer maintained. Source
- The README version of April 14, 2026 removes the lists and directs readers looking for MCP servers to the MCP Registry. Source
- Anthropic announced the Model Context Protocol on November 25, 2024, with an open-source repository of servers, and named early adopters such as Block and Apollo and development-tool companies including Zed, Replit, Codeium and Sourcegraph. Source
- Paul Graham's essay Startup = Growth says a good growth rate during Y Combinator is 5 to 7 percent a week. Source
- Brown, Cai and DasGupta (Statistical Science, 2001) found the standard Wald interval for a binomial proportion erratic and recommended the Wilson or Jeffreys interval for small samples. Source
- Sadeghi and colleagues (Technometrics, 2022; arXiv 2102.12893) describe novelty effects that diminish over time and primacy effects that grow with adoption in online experiments at Microsoft. Source
- Schmittlein, Morrison and Colombo's Counting Your Customers model (Management Science, 1987) uses the number and timing of a customer's previous transactions to estimate whether the customer is still active. Source
- SourceX's investors page says it shares company and commercial metrics with investors directly and offers investor materials on request. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.