In the Net, Episode 02
LinkedIn was founded in 2002 as a professional network. The promise was simple: your résumé, your connections, your visibility, all in one place, all yours. Twenty-three years later, your connections export in a format that imports into almost nothing, your reach is allocated by a 150-billion-parameter language model, and your professional profile is searchable by Google only as far as LinkedIn allows.
This is an architectural review of the platform on which it is published, written with full awareness of the irony, and with the expectation that the architecture under examination will respond to this post the way it responds to every post containing references to off-platform writing: by reducing its distribution.
The Promise
LinkedIn's original architecture, conceived by Reid Hoffman and a handful of co-founders in 2002, took the professional rolodex and gave it network effects. The promise was straightforward and, for a meaningful number of users, has been kept. People find jobs through LinkedIn. Suppliers meet clients. Products get distributed. Recruiters reach candidates they would not otherwise have reached. The original mechanism, the recommendation graph, is genuinely useful when it works as advertised.
What changed is not the original promise. What changed is the architecture that now sits between the user and the promise: an algorithmic layer that decides who sees what, an export layer that decides what leaves, and a regulatory position that decides what the platform owes its users in the first place.
The Hooks
The reach algorithm punishes external links
The most rigorous public dataset on the matter is Richard van der Blom's 2026 Algorithm Insights Report, drawn from approximately 1.3 million posts. Its headline finding: a single external link in the post body reduces median reach by 18.8 per cent. Marketing analyses (LinkBoost, MelanieGoodman/Substack, the broader "growth-hacker" literature) report higher numbers, up to 60 per cent, but their methodology is opaque and their datasets smaller; van der Blom's figure is the one that survives serious scrutiny.
The traditional workaround, recommended for years by every social-media-marketing practitioner on the planet, was to put the link in the first comment instead of the post body. This worked, until early 2026. LinkedIn's algorithm now identifies what researchers have begun calling "bridge behaviour": a post whose body is clearly designed to funnel readers into a comment containing a link. The pattern is detected, and the parent post receives a similar reach reduction to one with the link in-body. The escape that was the standard advice is no longer an escape.
The mechanism is simple in intent and opaque in execution. Posts with external links signal "user about to leave the platform." The platform's recommendation system, optimised for time-on-platform, demotes them. The official LinkedIn position, repeated by senior product staff, is that links do not reduce reach; the measured outcome across multiple independent datasets is that they do.
The compounding effect on personal-promotional content is observable but harder to publish at scale. In my own posting data, repeated references to my own writing (newsletter, book, blog) correlate with a reach reduction of roughly 80 per cent over a quarter, n=1, with the obvious caveats. The pattern is consistent with the platform's algorithmic preference for posts that look like content rather than posts that look like distribution. (For balance: a Q1 2026 analysis by Saywhat, on roughly 400 thousand posts, found that posts with multiple external links performed considerably better than posts with none. The methodology has not been publicly released. The 2026 research has not yet converged.)
The algorithm punishes thematic breadth
The second mechanism is less discussed and more structural. LinkedIn's recommendation system maintains a topic fingerprint for every creator, derived from what the creator posts, what they engage with, and what other users save from their content. The fingerprint informs the algorithm's decision about who else might be interested in a given post: the more narrowly focused the fingerprint, the more confidently the algorithm distributes new content from that creator into adjacent audiences.
The corollary is the punishment of breadth. A creator who writes on three different domains, a system architect who occasionally posts on management, a developer who shares photography, a consultant whose interests do not respect the algorithm's clustering, is treated by the system as a noisy signal. The fingerprint becomes ambiguous; the distribution becomes timid. The platform's preferred creator is the one who can be filed cleanly into one or two adjacent topic clusters and stays there.
This produces a structural conformity pressure. If you want reach on LinkedIn, you write on the topics LinkedIn has decided you write on. Diversification of voice is, on this architecture, a tax.
The reach itself is computed by a foundation model
In January 2025, LinkedIn's FAIT (Foundation AI Technologies) team published a paper on arXiv (Firooz et al., 2501.16450) detailing 360Brew, a 150-billion-parameter decoder-only foundation model built on top of Mixtral 8x22. The model is trained on LinkedIn's Economic Graph: member profiles, job descriptions, posts, interaction histories across 5+ surfaces. It performs more than 30 predictive tasks (feed ranking, job matching, people-you-may-know, advertising) using a single shared architecture rather than the previous patchwork of task-specific models.
The architectural elegance is real. The accountability properties that come with it are not. A 150-billion-parameter model is opaque by construction: its optimisation function is private, its training data unrevealed, its individual decisions unreproducible. There is no mechanism by which a user can ask "why was this post seen by 2,000 people instead of 20,000?" and receive an answer that could be checked against the model. There is no appeal. There is no audit. One model decides, in unobservable ways, what 1.1 billion users see when they open the application.
This is not a complaint about machine learning. It is a complaint about the absence of governance around an architecture that has become the world's largest professional gatekeeper.
The data export gives you names without the graph
LinkedIn's data export, accessible under Settings → Data Privacy → Get a Copy of Your Data, is a GDPR-compliant download of "your information." The "your" is doing significant work in that sentence. The native CSV format includes first-degree connections only. The second-degree graph (the connections of your connections, which is the actual network) is absent. Email addresses appear only for connections who opted in to share them, which in practice is a small minority. Conversation history is a separate, larger request that is well-known to be incomplete and to omit the full thread structure. For networks above 2,500 connections, exports are routinely incomplete; LinkedIn's own help documentation acknowledges this for networks above 5,000.
The export is, formally, complete. It is, operationally, a list of names with most of the relationships removed. A list of names is not a network. The relationships are what make the network work, and the relationships are what the export does not give you.
The API exists for partners, not for users
A user who wished to build their own export tool, or who wanted to use a third-party service to keep a portable, queryable record of their professional network, is structurally prevented from doing so. LinkedIn's official API is restrictive by design. Most useful endpoints are gated behind the LinkedIn Partner Program, which is itself selective: marketing-platform integrations, recruiting-tooling for HR enterprise software, and compliance-archival tooling for financial services are routinely approved; competitive intelligence, lead generation, data enrichment, and market research are routinely denied (Microsoft Learn, restricted-uses documentation).
The Terms of Service prohibit accessing, storing, displaying, or facilitating the transfer of any LinkedIn content obtained through scraping, crawling, or any technology outside the official APIs. The Marketing API additionally forbids combining member data with any other data, including the user's own data, to create or supplement profiles, leads, or reference tables. In practice, a user who wants to enrich their own copy of their own LinkedIn data with their own CRM is contractually prohibited from doing so.
The architecture of the API is not an accident. It is the second-order lock-in: even if a third-party tool wanted to mitigate LinkedIn's first-order lock-in (the export gap), the API terms forbid the kind of integration that would make mitigation possible. The Walled Garden defends its walls not only against the user leaving, but against tools that might help the user leave.
Your content trains the model that ranks you
On 3 November 2025, LinkedIn updated its Terms of Service. The substantive change: by default, member content (profile data, public posts, interactions) is used to train AI models, including content-generating models offered by LinkedIn and shared with affiliated Microsoft entities for Microsoft's own model-training activities. For users in the European Union, the European Economic Area, Switzerland, Canada and Hong Kong, this means data flows to Microsoft Copilot and related services for training. The setting is on by default; an opt-out exists in account settings; data shared on the platform before 3 November 2025 has already been used and cannot be retroactively removed.
The architectural observation is precise: the same recommendation system that decides what reach a user's content receives is built on a foundation that was trained, in part, on that content. The user contributes the data that trains the model that ranks the user's data. The opt-out, where it exists, is opt-out from future contribution; it is not opt-out from being ranked by a model whose training has already been completed using earlier contributions.
Whether one welcomes this architecture, finds it neutral, or finds it troubling is a question of taste and politics. The architectural fact is straightforward: the user's labour produces the platform's product, and the platform's product is the layer that decides who hears the user.
The Standing
LinkedIn has approximately 1.1 billion registered users and is the dominant professional network in nearly every market in which it operates. The closest national competitors (Xing in DACH, Viadeo in francophone markets) are an order of magnitude smaller. There is, for practical purposes, no second professional network of comparable scale that a user could move to and retain even partial network effects.
Microsoft, LinkedIn's parent company since the 2016 acquisition, was designated a Digital Markets Act gatekeeper by the European Commission in September 2023, with full DMA obligations from March 2024. LinkedIn itself was not designated. The platform is regulated as a Very Large Online Platform under the Digital Services Act, which imposes content-moderation transparency obligations but not data-portability or interoperability requirements. The Commission's reasoning has not been publicly detailed; the practical effect is that LinkedIn faces no mandatory algorithmic transparency, no enforced data portability beyond GDPR's already-weak provisions, and no required interoperability with competitors. The architecture is, by regulatory choice, off the gatekeeper list.
This matters because the obligations the DMA places on the six designated gatekeepers (Apple, Alphabet, Amazon, Meta, ByteDance, Microsoft for its core platforms) include exactly the structural requirements that would constrain the lock-in mechanisms described above: algorithmic transparency, real data portability, interoperability for messaging and contacts. LinkedIn carries the lock-in mechanisms without the obligations.
A Note on Alternatives
The natural follow-up question, when one critiques a dominant platform, is: what about the competitors? In the DACH region the question has a specific shape: what about Xing, which preceded LinkedIn in the German-speaking market and still has roughly 22.5 million members across Germany, Austria, and Switzerland?
The architectural answer is unkind. Xing is the same Walled-Garden model in a smaller, declining regional market. The export, the algorithm, the API restrictions carry the same structural patterns as LinkedIn, with the additional caveat that Xing's official API is meaningfully thinner: third-party scrapers exist precisely because the platform offers no usable bulk-export interface. By March 2025, LinkedIn had reached 22 million users in Germany alone, drawing level with Xing in active relevance while continuing to grow as Xing contracts. Xing's revenue declined 12 per cent year-on-year in the first half of 2024, with the Talent Access segment (its core recruiting product) down 8 per cent.
There is a second, sharper observation. In August 2024, New Work SE, Xing's parent, delisted from the Frankfurt Stock Exchange. By June 2025, Burda Digital SE completed a full squeeze-out at EUR 105.65 per share, taking the company entirely private. The user base, the connections, the professional histories that 22.5 million people had built on the platform, changed ownership without any mechanism by which those users could be consulted, vote, or refuse. The Walled Garden, in the period after the IPO and before whatever next chapter Burda has in mind, is the sole property of a single private holding.
This is the deeper version of the lock-in question. The everyday lock-in is algorithmic and architectural. The structural lock-in is that the platform you have spent ten years building inhabits an ownership structure in which you, the user whose work makes it valuable, have no voice. Switching from one Walled Garden to another exchanges one such ownership structure for another, often smaller and more concentrated. The architectural lesson is the same: the exit is not "another vendor"; the exit is "another architecture".
The Exit That Isn't
You can leave LinkedIn. You cannot take what you built.
You can request your data export. You will receive your first-degree connections, without their email addresses (mostly), without the second-degree graph that turned the rolodex into a network, without the conversation history that documented your professional relationships, and in a CSV format that no other professional network is positioned to import. You can hand-rebuild your network elsewhere, but the relationships your LinkedIn ten years built are not portable; they are observable only inside LinkedIn.
You can delete your profile. The export you took with you is the list of names. The audience you accumulated, the second-degree introductions you might have leveraged, the inbound recruiters who looked at your profile last week without contacting you, the people who saved your post for later, the colleagues who follow you for one specific reason: all of that stays on LinkedIn. The exit door exists. It opens onto a car park.
The platform is, in formal terms, GDPR-compliant. It produces an export when asked. The export contains the data the regulation requires. The fact that the export does not constitute a usable network is not, formally, the platform's responsibility. It is, operationally, the entire point.
The Price
The price of staying is algorithmic submission. Every external link, every reference to your own writing, every link in your profile bio compounds against your reach in ways that are measurable across multiple independent datasets and consistent with the platform's stated preference for on-platform engagement. The user who wants visibility on LinkedIn writes for LinkedIn, in the format LinkedIn rewards, with the cadence LinkedIn prefers, knowing that any deviation costs reach that the platform decides whether to grant.
The price of leaving is the audience you cannot take with you. The names yes; the network no. The professional gravitational mass that took years to accumulate, gone the moment you walk out, because the gravitational mass was never yours.
In the Würde-Verhältnis between user and platform, the structural balance has settled. The user contributes the content (their professional history, their writing, their network connections) and the platform owns the discovery layer that decides whether the content reaches anyone. This is not a complaint about advertising or about the existence of an algorithm. It is an observation about who, in this architecture, owns the result of whose work.
The Escape Route
The escape route is not "leave LinkedIn." The escape route is "stop letting LinkedIn be canon."
Whatever one thinks of the platform, X reaches a meaningfully larger general-topic audience than LinkedIn. The architectural reason is structural rather than political. X's distribution is engagement-driven: roughly half of the For-You timeline is drawn from accounts the user does not follow, and reposts (Retweets) are weighted by an order of magnitude greater than any other engagement signal in the open-source ranking algorithm. A post that receives strong engagement signals can move from its author's follower base into the broader graph in a way that LinkedIn's fingerprint-driven distribution does not permit. The link penalty exists on X (around 30 to 50 per cent for non-Premium accounts, with a partial reform announced in October 2025) but, against the larger general-topic audience, registers far less than LinkedIn's compounding fingerprint-plus-link-plus-self-promotion penalty. X is, architecturally, a different walled garden, with different distribution mechanics that produce, for general-interest content, materially different reach. The platform serves a different market than LinkedIn does and is rather more polarising in its public reception; whatever one's view, the reach asymmetry is the empirical observation.
Your own domain with Open-Graph tags is your real profile. A static or lightly-dynamic site under a domain you control, with Open-Graph metadata that lets your content render properly when shared on any platform, is the only profile that does not depend on a third party's algorithm to exist. The site is directly searchable by Google, you control the metadata, you decide what redirects and what does not, and the content survives any platform's policy change about you.
The honest qualification: in 2026, a personal domain has almost no organic growth on its own. RSS and Atom subscriptions, the historical mechanism by which readers built durable habits around independent sites, are vestigial. Almost no one outside niche developer communities subscribes to feeds in 2026. A personal site is the durable archive; it is not, on its own, the audience.
A self-hosted newsletter is the audience nobody can throttle, but the audience must be built. Mailcow on FreeBSD or Linux gives you SMTP infrastructure under your control. Listmonk or Buttondown handle subscriber-management UX. The audience-list belongs to you, deliverability is your responsibility (which is also the right place for it), and no recommendation algorithm sits between you and your readers. A newsletter at 1,000 subscribers reaches roughly 1,000 inboxes; a LinkedIn post at 1,000 followers reaches whatever 360Brew decides to grant on a given Friday.
The honest qualification: building those 1,000 subscribers in 2026 is genuinely difficult. The acquisition channels are limited. Most newsletters of this kind grow through direct mention from existing platforms, which loops back to the question of where the audience starts. A newsletter is a long-term position, not a short-term substitute for distribution.
Treat LinkedIn as one channel, not the channel. Post here for distribution; archive on your domain for permanence; treat X as the platform whose distribution rules do not punish you for being broadly interested. Reference your domain in posts the way large publications reference their print edition: not as a competitor to the platform, but as the canonical, durable form of the work. The platform may throttle the link; the work outlasts the throttle.
The structural assessment is harder than the architectural one. For its users, LinkedIn is, on this analysis, structurally hostile: a Walled Garden that rewards thematic narrowing, penalises external links, drosselt comments that reroute the link penalty, and uses members' content to train the model that ranks them. The other walled gardens with comparable reach (Facebook, Instagram, TikTok) serve different content markets and audiences and are not realistic distribution channels for engineering and architecture writing of the kind discussed here. X, with all its public controversy, is the platform whose distribution architecture remains structurally compatible with general-topic, link-carrying, occasionally self-promotional professional writing.
The honest acknowledgement: none of these alternatives reproduce the specific gravitational mass that LinkedIn carries among B2B decision-makers in 2026. That mass is the lock-in itself.
Coda
The network you built belongs to them. The reach you did not earn is the reach they can take back. The exit door opens onto a car park, and the car park is precisely as well-lit as the platform decides.
This post will, I expect, demonstrate the thesis. By the time you read it, the algorithm will have decided how much of the audience LinkedIn lets it have. The thesis does not require a particular outcome to be confirmed. It requires only that the outcome be observed.
