Nostr’s Retention Problem: A Structural Diagnosis

Note: The author is value for value only, has no ads, no sponsors, and no financial or personal ties to any Nostr or Bitcoin developers.

Abstract

Nostr, a decentralised social protocol designed to resist censorship and eliminate centralised gatekeeping, retains between 1% and 2% of new users by week one, with retention rounding to less than 1% by week three. This analysis presents evidence that the retention failure is structural rather than a consequence of poor user experience. Independent analysis of zap receipt events from five public relays (capturing approximately 3% of total network zap activity) across two periods (February to April 2023 and March to May 2026) shows a Gini coefficient of 0.92 to 0.96 in the distribution of zap value across recipients within the sample, a concentration that persists across a tenfold difference in network activity. Analysis of kind-3 (contact list) events shows that in October 2025, 55.4% of users carried identical follow lists, with 4,175 accounts sharing the exact same contacts, a pattern consistent with client-pushed default follow lists during onboarding. By mid-2026, the identical-list rate dropped to 15.7% (largest group: 588), but concentration persists: 21 accounts are now followed by more than 30% of all users in the sample. All 20 of the most-followed accounts are Bitcoin influencers or Nostr developers. An initial engagement comparison of 20 default-listed accounts against 20 manually verified organic accounts shows an approximate 15:1 disparity in average engagement per note; this comparison does not control for content type, posting frequency, or topic, and should be treated as indicative rather than definitive. Taken together, these findings describe a Cantillon-like dynamic: New user attention flows first and disproportionately to a small number of accounts that were pre-selected by client developers and placed into new users’ feeds. The protocol was designed to eliminate centralised data distribution, and it succeeds at that. But default follow lists reintroduce it at the client level, concentrating attention through the same structural position the protocol removed from infrastructure. This analysis concludes with recommendations for client developers and the protocol community.

Note: This work is in progress; a companion user experience survey is open until June 22, 2026 at https://daniella.io/polls

Author’s Note

This article has been in progress for several weeks. Over the weekend of May 29, 2026, a conversation broke out on Nostr about exactly the problems described in the data below. I am publishing the current findings now rather than waiting, because the community is asking these questions and the data can contribute to the answers.

As per the request of my viewers,[9] tomorrow (Tuesday, June 2) I am publishing a Nostr onboarding tutorial for beginners as well as sharing the anonymous Nostr User Experience Survey I wanted to use to complete this work at https://daniella.io/polls. The goal is to complement the relay data with direct user testimony not only from current users, but also from people who signed up and decided to leave. If you use or have used Nostr, your two minutes are valuable. The survey closes June 22, 2026. The final version of this article, incorporating all remaining data and the survey results, will be published by the end of June 2026.

This analysis follows the spirit expressed by Derek Ross on May 31, 2026:

“Decentralization is fucking hard. It’s hard for designers. It’s hard for developers. It’s hard for infrastructure runners. It’s hard. All of it. But fighting and arguing doesn’t make it any less hard. All it does is divide us and put us against one another… Seek solutions. Push the conversation forward. Discuss things in the open.”[1]

This work seeks solutions by discussing things in the open backed by data and feedback.

1. Introduction

Nostr retains almost no one. Week 1 retention sits at 1% to 2% across all measured cohorts from February through May 2026, rounding to less than 1% by week three [2]. New pubkey counts averaged approximately 34,600 per day in April 2026 (these figures reflect new pubkeys, not confirmed human users; bots, Fediverse bridge accounts, and automated clients inflate the count). Almost none return.

The community already identifies the problem. “People quit because they feel like they’re yelling into a black hole. There are only a small handful of influencers who get any engagement at all,” wrote one user. The prevailing developer diagnosis is different: That retention fails because “apps are broken.”[3]

Yet, retention remains critically low despite continuous UX improvements across clients over three years. The evidence points to structural attention concentration at the client level, specifically default follow lists that pre-select a small number of accounts for new users. These lists inflate the visibility, engagement, and zap income of the listed accounts while rendering smaller accounts, or those that are not listed, invisible. The data shows that 21 accounts appear in the follow lists of more than 30% of all users in the sample, and in an initial comparison these accounts receive approximately 15 times more engagement per note than accounts outside the list. New users arrive, encounter the same accounts dominating their feeds, find themselves publishing into silence, and leave. The data is consistent with this sequence, though causation is currently inferred.

The protocol itself works. Censorship resistance is empirically verified: When the 30 largest relays go offline, over 90% of posts remain accessible [4]. Infrastructure has grown 66% through a bear market. [2][4] The protocol distributes data effectively. However, the social layer built on top of it does not distribute attention.

2. Methodology

2.1 Zap Distribution Analysis

Zap receipt events (kind-9735) were queried via websocket from five public relays (relay.damus.io, nos.lol, relay.nostr.band, relay.snort.social, nostr.wine) across two 90-day periods: February to April 2023 (peak adoption) and March to May 2026 (current). Events were deduplicated by event ID. Recipient pubkey and sats amount were extracted from the embedded kind-9734 zap request. For the 2023 dataset, 10 events dated March 16, 2023 were excluded; each claimed exactly 100 BTC but carried bolt11 invoices encoding only 1,001 sats, confirming test or spam events during early NIP-57 implementation. The Gini coefficient was calculated to measure concentration. [7]

Note on sample size: This analysis queries five relays. The Pensieve indexer (stats.andotherstuff.org), which connects to dozens of relays with automatic NIP-65 relay discovery, reports approximately 30 times more zap activity over a comparable 90-day window in 2026 [2]. Our five-relay sample therefore captures roughly 3% of total zap activity. The Gini coefficient measures distribution shape within a sample and remains valid as a concentration measure at this sample size. However, if the remaining 97% of zap activity flows to a substantially different set of recipients, the network-wide Gini could differ. This is stated as a limitation.

2.2 Default Follow List Analysis

Kind-3 (contact list) events were queried from the same relays for two periods: October 2025 and the last 30 days ending June 1, 2026. October 2025 was selected because client onboarding flows were documented at the time as pushing pre-selected follow lists to new users. [8] The 2026 period tests whether the pattern persists after clients adjusted their onboarding flows.

The analysis identified: Accounts appearing in a high percentage of all follow lists in the sample (thresholds: >50%, >30%, >20%); groups of users with 100% identical follow lists (indicating a client wrote the same list for multiple users); the size of the largest identical group; and follow list size statistics. If large groups of users carry the exact same follow list, the most straightforward explanation is that a client wrote that list during onboarding rather than each user independently choosing the same accounts. [7]

2.3 Engagement Comparison

The top 20 most-followed accounts from the 2026 default follow analysis were compared against 20 organic accounts with follower counts ranging from 49 to 1,806. Twenty organic accounts were compiled and manually verified by the author as real, active humans who post regularly and do not appear on any default follow list. For each account, up to 20 recent notes (kind-1) were queried from all five relays in parallel, and reactions (kind-7), replies (kind-1 referencing the note), and reposts (kind-6) were counted with deduplication by event ID. Of the 20 default-listed accounts, 3 had no posts in the 30-day window and were excluded from the engagement calculation. Of the 20 organic accounts, 5 had no posts in the 30-day window and were excluded. The final comparison is 17 default-listed accounts versus 15 organic accounts.

The organic group was manually verified rather than randomly sampled. This is a deliberate methodological choice. Random sampling from Nostr’s user base produces unusable data: An earlier iteration of this analysis selected accounts at random and returned bots, spam accounts, and automated clients posting timestamps, rendering the comparison meaningless. Nostr’s public key model means that a large proportion of pubkeys in the network are inactive or belong to non-human actors. In this data environment, random sampling does not produce a representative sample of human users; it produces a sample contaminated by accounts that are not comparable to the default-listed group in any meaningful way. Manual verification, confirming that each account belongs to a real, active human who posts regularly, was the methodological response that ensured every account in the comparison group is a legitimate data point. The limitation is sample size, not selection method: A larger manually verified sample would strengthen the findings, and the final version of this analysis will attempt to expand the organic group. The direction of the disparity is unlikely to change with a larger sample, but the magnitude may.

2.4 Qualitative Evidence

Public posts from a community discussion on May 31 to June 1, 2026, and responses to a Nostr User Experience Survey (daniella.io/polls, open until June 22, 2026) provide qualitative context. The survey includes anonymous responses from people who chose to leave Nostr; results will be published after the survey closes on June 22, 2026.

3. Results

3.1 Zap Concentration

The Gini coefficient measures how equally or unequally something is distributed across a group. It ranges from 0 (perfectly equal: Every recipient receives the same amount) to 1 (perfectly unequal: One recipient receives everything). It is most commonly applied to income, but can measure inequality in any distribution.

Metric	Feb-Apr 2023	Mar-May 2026
Zap events	94,901	9,934
Total sats	35,063,952	2,941,395
Unique recipients	3,903	1,458
Gini coefficient	0.96	0.92
Top 1% share of value	72.7%	60.7%
Top 10% share of value	95.2%	89.0%
Bottom 50% share of value	0.2%	0.9%
Median received (sats)	117	100

Source: Independent analysis of kind-9735 events from five public relays. [7]

Zap activity contracted approximately tenfold from 2023 to 2026, and total value contracted approximately twelvefold. But within this sample, the concentration persisted. The Gini was 0.96 at peak adoption and 0.92 in the current period. When the network had the most users and the most value flowing through it, concentration was at its most extreme: The top 1% of recipients (39 accounts) captured 72.7% of all zap value while the bottom 50% (1,952 accounts) received 0.2%.

The median recipient received roughly the same amount in both periods: 117 sats in 2023, 100 sats in 2026. Even when the system carried twelve times more total zap value (2023), the median recipient received roughly the same amount.

For context, South Africa has the highest national income Gini on Earth at approximately 0.63 [5]. No published Gini baseline exists for voluntary tipping systems (YouTube Super Chats, Twitch donations, Patreon, etc.), so a direct comparison to a peer system is not possible at this time. Some concentration in voluntary tipping is expected; most users never tip at all, and a small number of creators attract disproportionate support on every platform. The national income comparison is not apples-to-apples but is offered to illustrate the scale: Within the population of Nostr users who do receive zaps, the distribution in this sample is more concentrated than the most unequal national economy ever measured.

The question is, why?

3.2 Default Follow Lists: The Mechanism

When a new user joins Nostr through a client, they need something in their feed. Clients solve this by pre-selecting a list of accounts for every new user to follow automatically. This is a reasonable UX decision with structural consequences.

Metric	October 2025	Last 30 days (2026)
Users with follow lists	9,272	9,125
Median follow list size	3	75
Average follow list size	34	223
Users with identical follow lists	5,135 (55.4%)	1,434 (15.7%)
Largest identical group	4,175	588

Source: Independent analysis of kind-3 events from five public relays. [7]

In October 2025, the median follow list contained three accounts, and 55.4% of all users with follow lists shared an identical list with at least one other user. This pattern is extremely unlikely to arise organically. It is difficult to construct a plausible scenario in which 4,175 users independently choose to follow the same three accounts in the same order. The most likely explanation is that a client onboarding flow wrote a pre-selected follow list to the user’s contact list event before the user made any choices of their own. A documented tutorial from October 2025 confirms that at least one major client’s onboarding flow included a default opt-in to a curated follow list at that time [8]. Other explanations, such as a viral tutorial or copy-paste guide, cannot be excluded but would still point to a structural rather than organic pattern.

By 2026, the identical-list rate has decreased to 15.7%, and the median list has grown to 75 accounts, suggesting that onboarding flows have changed. But 21 accounts are now followed by more than 30% of all users, and the top five most-followed accounts reach 41 to 46% of users. The uniformity of those percentages suggests they appear together as a set in one or more clients’ onboarding flows.

All 20 of the most-followed accounts in the 2026 sample are Bitcoin influencers or Nostr developers. At least seven are developers or technical founders (35%). Not one is a non-Bitcoin content creator. Some concentration of Bitcoin-focused accounts is expected on a protocol built and adopted by Bitcoiners; early adoption on any platform reflects its founding community. The structural concern is not that Bitcoin accounts dominate early, but that default follow lists lock that dominance in. On a platform without default lists, the most-followed accounts would shift over time as the user base diversifies. Default lists prevent that turnover by ensuring every new user, regardless of their interests, begins with similar accounts. This is the composition of the list that shapes every new user’s first experience of the protocol.

The downstream effect on feeds is direct. If the same set of accounts appears in every new user’s follow list, those accounts dominate every new user’s feed. They will appear in “most liked,” “most zapped,” “most reposted,” and trending lists, because they are the ones with a structurally guaranteed audience. This creates the experience users describe: The same accounts, always, everywhere. As one small-account user put it: “A way to not hide small accounts like mine would help.”[10]

A feed composition analysis (sampling what actually appears in a typical user’s timeline) would quantify this more precisely.

3.3 Engagement Inequality

Twenty default-listed accounts (the top 20 most-followed in the 2026 sample) were compared against 20 manually verified organic accounts, all confirmed as real, active humans who are not on any default follow list. Three default-listed accounts and five organic accounts had no posts in the 30-day analysis window and were excluded from the engagement calculation.

Group	Accounts with data	Avg engagement per note
Default-listed	17	29.85
Organic	15	1.99
Ratio		15:1

Source: Independent analysis of kind-1, kind-7, and kind-6 events across five public relays, deduplicated by event ID. All organic accounts manually verified. [7]

Accounts on the default follow lists receive approximately 15 times more engagement per note than the organic accounts in this comparison. This ratio is indicative, not definitive: The comparison does not control for content type, posting frequency, topic, or timing, and the organic group was manually selected rather than randomly sampled for reasons explained above. A larger sample with explicit controls could produce a different magnitude. The 15:1 ratio also includes accounts that are well-known outside of Nostr and would receive more engagement on any platform. The structural signal is clearest in the cases where the pattern breaks.

Within the default-listed group, the range is enormous. The highest-engagement account received 242 engagements on a single post in 30 days. At the other end, the creator of one of Nostr’s most widely used clients, followed by 33% of all users in the sample, averaged 4.4 engagements per note across 20 posts. This account posted constantly, had thousands of structural followers, and generated barely more engagement per post than the highest-performing organic account (3.1 per note, with a fraction of the followers). The structural followers are not engaging.

This finding points to a metric more revealing than raw follower count: Engagement rate relative to followers. An account followed by 33% of the network but averaging 4.4 engagements per note has a worse engagement rate than organic accounts with a few hundred followers averaging 2 to 3 per note. Default follow lists inflate follower counts, but engagement rate reveals the hollowness. Follower count on Nostr is not a signal of audience; it is largely a signal of structural position.

At the bottom of the organic group, accounts that have been active for nearly three years, building real projects and communities, averaged 0.8 to 0.83 engagements per note. These are real people doing real work, publishing to near-silence.

The data describes a structural engagement gap. Most content creators who depend on audience engagement for their livelihood would not choose to build on a platform where structural invisibility is the default experience.

3.4 Retention

Cohort	Size	W1	W2	W3	W4
Feb 16	587.7K	2%	0%	0%	0%
Mar 2	346.5K	2%	1%	1%	1%
Mar 23	484.9K	1%	0%	0%	0%
Apr 20	665.3K	1%	1%	1%	–
May 4	592.9K	2%	–	–	–

Source: JeffG (2026), stats.andotherstuff.org. Select cohorts from 13 weekly cohorts, all showing the same pattern. [2]

The 0% entries in the table do not mean literally zero returning users; they mean the percentage rounds to less than 0.5% of the cohort. The cohort sizes reflect new pubkeys as reported by the source and include bots, Fediverse bridge accounts, and automated clients alongside human users. If the denominator is inflated by non-human pubkeys, the actual human retention rate could be higher, though still very low given the absolute numbers.

For context, retention is low across all mobile apps. The industry average day-seven retention for social media apps is approximately 9% to 13%, with strong performers reaching 15% to 20% [11]. Most new apps lose the vast majority of users within a week. Nostr’s 1% to 2% is not unusual in direction, but it is an order of magnitude below the category average. The relevant comparison is not perfection; it is the baseline that other social apps, many of them also small, underfunded, and competing for attention, manage to achieve.

Retention remains critically low despite continuous UX improvements across clients over three years. The structural explanation accounts for this: Regardless of how polished the client becomes, if the onboarding flow funnels every new user’s attention to the same small set of accounts while leaving the new user invisible, nothing changes for them.

4. Discussion: The Cantillon Effect on Nostr

In classical economics, the Cantillon effect describes how newly created money benefits those closest to its point of entry into the economy. When a central bank expands the money supply, the institutions with first access to the new money spend it before prices adjust, gaining a real advantage over those further from the source. Proximity to the point of issuance, not productive contribution, determines who benefits.

The data in this analysis describes an analogous dynamic. When a new user joins Nostr, their attention is the “new money.” The default follow list determines who receives it first. The accounts on that list accumulate followers, reactions, and zaps not necessarily because their content is better, but likely because they are positioned at the point of entry. By the time a new user might discover smaller creators organically (if they ever do), the attention economy has already allocated its value, and the new user has already begun to leave.

Attention is not zero-sum in the way that money is: Following one account does not unfollow another. But feed space and scrolling time are finite in practice. A new user’s first session is limited. If their feed is pre-populated with 20 accounts they did not choose, those accounts occupy the real estate in which organic discovery would otherwise occur. The scarcity is not in the follow action itself but in the time and attention a new user allocates before deciding whether to stay or leave.

To be clear: The analogy describes the structural outcome, not the intent. Default follow lists exist because the alternative (an empty feed) is considered worse. Client developers attempted to solve a real problem. The question is if a better onboarding flow could help.

The parallel holds across three dimensions:

• Concentration persists regardless of supply. Just as the Cantillon effect operates whether the money supply expands by 1% or 100%, Nostr’s attention concentration within the sampled relays persists across a tenfold difference in network activity. The Gini was 0.96 at peak adoption and 0.92 in the bear market. More users entering the system did not distribute attention more broadly within this sample. It concentrated it further.

• The median participant is unaffected. The median zap recipient received 117 sats during peak adoption and 100 sats in the bear market. Even when twelve times more total value flowed through the system (2023), the median recipient received roughly the same amount.

• The mechanism is structural, not conspiratorial. Default follow lists are a reasonable onboarding decision, just as monetary policy serves legitimate economic functions. Nobody set out to create a Cantillon effect on Nostr. The concentration is an emergent consequence of a structural choice. This makes it harder to see, but also amenable to structural solutions.

Vitor Pamplona, the developer of Amethyst (one of Nostr’s most popular Android clients), identified a broader dimension of the problem: “The hard truth is that people are tired of social media in general. And if we keep onboarding them into Twitter-like interfaces we will keep losing them. Somebody will come up with something better than this and I hope they do it soon. We need some brand new ideas not just recycled UIs.”[10] Pamplona’s diagnosis points to a cause that is partly independent of this analysis: If social media fatigue is driving people away regardless of platform, then fixing follow lists alone will not fix retention. This analysis does not claim otherwise. But the two explanations are not mutually exclusive. Structural concentration may compound the fatigue Pamplona identifies: Users arrive already tired of social media, encounter a system that reproduces the same dynamics they were trying to escape, and confirm their suspicion that the game is the same regardless of the platform. Fixing the structural layer does not guarantee retention, but leaving it in place guarantees that users who are willing to try something different will not find it.

The engagement data provides the strongest evidence for the structural interpretation. Causation is inferred, not proven, and this analysis cannot fully separate the effect of default follow lists from pre-existing fame. But the cases where the pattern breaks are revealing: A developer whose client is used across the network, followed by a third of all users, averaging 4.4 engagements per note. The structural followers were assigned, not attracted, and they are not reading. That finding cannot be explained by fame. It can be explained by default follow lists populating contact lists with accounts that users never chose and do not engage with.

5. Recommendations

These recommendations address the structural mechanism identified in this analysis. They are not directed at any individual or company. The author is value for value only, has no ads, no sponsors, and no financial or personal ties to any of the developers. These recommendations are untested in the Nostr context; measuring their impact (for example, tracking the Gini over time after onboarding changes) would be the way to validate them.

5.1 For Client Developers

The core problem is that current onboarding flows prioritise speed over agency. A new user is given a pre-populated feed or pre-populated groups, and is expected to engage with it. The alternative is to prioritise user agency from the first interaction. This does not mean removing default follow lists entirely; it means onboarding differently.

Nostr’s defining feature is the absence of an algorithm. This is a major advantage compared to centralized social media and the onboarding experience should reflect it. Rather than presenting a pre-built feed, clients could walk new users through the process of building their own feed, explicitly. This means explaining that they are in control of what they see, and guiding them through a selection process (individual accounts, hashtags, topics, communities) that results in a feed they explicitly chose.

This is harder to build and may initially feel slower for the user. But the current approach optimises high time preference for the first five minutes, at the expense of low time preference for the first five weeks. A user who understands that they built their own feed has a reason to return to it. A user whose feed was built for them by a client has no such attachment, and when the pre-built feed does not match their interests, they leave.

Some clients have introduced follow packs: Curated groups of accounts organised by topic. This is a step in the right direction, but the current implementation reproduces the same concentration in a different shape. Because the most visible accounts on the platform are already the most visible, they tend to appear across multiple follow packs. There is no mechanism for a smaller creator to easily request inclusion in a relevant pack; adding oneself requires contacting the curator directly, and Nostr’s DM infrastructure does not always deliver reliably. A structural improvement would be to build a request mechanism into follow packs: Any account could request inclusion in a relevant pack, and the curator would review the profile and approve or deny. This keeps a human in the loop to prevent spam while opening access to creators who are currently invisible simply because no curator knows where to place them. The curator’s role becomes more valuable, not less, because they are actively building a quality-filtered directory rather than relying on who they already know.

Where topic-based onboarding already exists, the structural concentration data suggests it is not yet effective. If the top 5 most-followed accounts still reach 41% to 46% of all users, the defaults are still driving the outcome regardless of what options the onboarding flow presents. The test is whether the Gini decreases over time.

Transparency matters independently of the onboarding design. If a client writes a default follow list, the composition of that list and the criteria for inclusion should be public. This allows the community to audit the structural choices that shape attention distribution.

5.2 For the Protocol Community

The protocol could support structural solutions through NIPs that distinguish user-chosen follows from client-suggested ones in the contact list event, making default follow list dynamics visible and measurable at the protocol level. A follow-list provenance tag would allow researchers and users to see how much of their social graph was chosen versus pre-populated.

Distribution metrics should become standard. The Gini coefficient of zap distribution, measured periodically and published alongside user counts and event volumes, would create visibility into whether structural changes are working since what gets measured gets managed.

6. Limitations

The zap distribution analysis samples five relays, capturing roughly 3% of total zap activity as measured against the Pensieve indexer [2]. The Gini coefficient is valid as a measure of distribution shape within the sample. However, if the remaining 97% of zaps flow to a substantially different set of recipients (for example, through client-specific or paid relays not in our sample), the network-wide Gini could differ from our measurement.

The default follow list analysis detects identical lists, but lists differing by one or two accounts would not register as identical. The analysis queries all kind-3 events in each period, which includes both new users creating their first follow list and existing users updating theirs; the term “users” rather than “new users” is used throughout to reflect this. The “why” behind the identical lists is inferred: 4,175 users with the same follow list is not organic behaviour, but proving the client pushed the list requires either the client’s source code or a developer statement. A documented tutorial from October 2025 confirms the onboarding flow had a default opt-in following curated lists at that time [8].

The analysis does not apply explicit spam filtering; kind-7 reactions and kind-6 reposts are standard Nostr event types and not inherently artificial, but the possibility that some engagement events are automated cannot be excluded.

The daily new pubkey counts cited in this article (approximately 34,600 average) include bots, Fediverse bridge accounts, and automated clients alongside human users. The retention percentages are based on raw pubkey cohorts as reported by the source [2]. If the denominator is inflated by non-human pubkeys, the actual human retention rate could be higher, though still very low given the absolute numbers.

All community quotes are from public Nostr posts and reflect individual experiences, not representative survey data. The companion survey at daniella.io/polls is designed to address this gap.

7. Conclusion

The protocol distributes data across 1,184 relays in more than 40 countries. When the 30 largest relays go offline, over 90% of posts remain accessible [4]. This works.

The clients themselves have improved substantially. Multiple community members independently report that apps work fine for them. The UX has come a long way, and several Nostr clients are well-designed, functional applications. The problem this analysis identifies is not app quality.

Within the five-relay sample, the social layer distributes attention to a Gini of 0.92. The top 1% of zap recipients in that sample capture over 60% of all value. Organic accounts average 1.99 engagements per note. This does not work.

The most likely explanation, supported by the engagement data, is that client onboarding decisions pre-select a small number of accounts for every new user to follow, creating a Cantillon-like dynamic in which new user attention benefits those positioned at the point of entry, regardless of content quality or community contribution. Causation is inferred, not proven; other factors including pre-existing fame and social media fatigue contribute. But the structural pattern is consistent across every dataset examined. The protocol eliminated centralised data distribution. The clients reintroduced gatekeeping of discovery.

Many of Nostr’s users came here because they understood this dynamic in fiat systems. They understood that proximity to the money printer, not merit, determines who benefits from monetary expansion. They understood that structural position in a financial system matters more than individual effort. They built a protocol to escape that logic. But the clients built on the protocol reproduced it. The currency changed from fiat to attention. The concentration did not. And people who spent years learning to recognise a rigged game are not going to spend long on a platform before they sense it again, even if they cannot name exactly what they are sensing. The 1% to 2% retention rate is consistent with users deciding quickly that the platform does not offer them something different.

This is not a protocol failure but rather a client implementation choice that can be changed. The structural interventions are within reach: Onboarding that prioritises user agency over speed, transparency in default follow list composition, discovery mechanisms that let any creator request visibility in relevant follow packs, and protocol-level support for measuring and distinguishing organic from pre-populated social graphs.

The question for Nostr is whether the clients built on top of it will distribute attention as effectively as the protocol distributes data.

Take the Survey

The relay data tells one side of the story. Your experience tells the other: I am running an anonymous Nostr User Experience Survey at daniella.io/polls. Which clients you use, what your first experience was like, why people leave, what matters most. The survey closes June 22, 2026.

References

[1] Ross, D. [@derekross]. (2026, May 31). Decentralization is fucking hard. It’s hard for designers. It’s hard for developers. It’s hard for infrastructure runners. It’s hard. All of it [Post]. Nostr. https://primal.net/e/nevent1qqs8h8dga4evmrcusfe5ryfzj89l6e4zdrfmx3dyz95r66956vxkk0g2ajnus

[2] JeffG. (2026). Nostr stats. And Other Stuff. https://stats.andotherstuff.org/

[3] sdbtc. (2026, May 31). [Reply to ODELL post on Nostr growth]. Nostr (kind-1 event via Primal). https://primal.net/e/nevent1qqs03xenudt4c606w935pfmj2lerm9f9eajmy5nq8xqpjc3kk2xq2egh3znyf

[4] Wei, Y., & Tyson, G. (2025). An empirical analysis of the Nostr social network: Decentralization, availability, and replication overhead. arXiv. https://arxiv.org/abs/2402.05709v2

[5] World Bank. (2024). Gini index. World Bank Open Data. https://data.worldbank.org/indicator/SI.POV.GINI

[7] Liberati, D. (2026). Independent analysis of Nostr kind-9735 (zap receipt), kind-3 (contact list), kind-1 (note), kind-7 (reaction), and kind-6 (repost) events. Queried via websocket from five public relays (relay.damus.io, nos.lol, relay.nostr.band, relay.snort.social, nostr.wine). Events deduplicated by event ID. Raw data and scripts available on request.

[8] Liberati, D. (2025, September 30). NOSTR explained for beginners [Video]. YouTube. https://www.youtube.com/watch?v=-EhXdsJr8Hw&t=1282s

[9] [Viewer request prompting Nostr onboarding tutorial]. Nostr (kind-1 event). https://primal.net/e/nevent1qqsgcy2gs7pelskyswsvn7tn94vuamyg438vvd8muk4lt0q7763y2aq2k3hss

[10] Pamplona, V. (2026, May). [Post on social media fatigue and Nostr onboarding]. Nostr (kind-1 event via Amethyst). https://primal.net/e/nevent1qqsgmwgf3sjl525hgcty3jplrhq9r4f7jhtajcxdjfzz38cat9spw7cj055a4 Reply from Bitcoin Awareness cited.

[11] UXCam. (2026). Mobile app retention benchmarks by industry. https://uxcam.com/blog/mobile-app-retention-benchmarks/ ; GetStream. (2026). 2026 guide to app retention: Benchmarks, stats, and more. https://getstream.io/blog/app-retention-guide/