Local-Language LLMs Are Eating Emerging Markets. AEO Strategy Splits.

In February 2026, Casey Newton's Platformer reported that Medium's monthly active users had declined to roughly 70 million from a 2022 peak above 100 million, with the steepest drop among the publication-driven writer cohort that once defined the platform. The piece quoted a former Medium employee describing the platform's editorial strategy as "trying to be a publication, a creator economy, and an LLM-friendly content hub at the same time, and losing on all three." A week later, Tony Stubblebine, Medium's CEO since 2022, wrote his own counter-essay on the Medium official blog arguing that Partner Program payouts to writers had grown 30% year-over-year and that the platform's curated publications — Towards Data Science, UX Collective, Marker — remained the single highest-trust surface for serious nonfiction on the open web.

Both stories are partially true. The honest tradeoff facing any operator considering Medium as part of an answer-engine-optimization strategy in 2026 is not whether Medium is dead or alive. It is whether the specific publication you would post to has enough editorial authority to make the LLM citation graph treat your content as authoritative, while the canonical tag and member-only paywall settings still preserve your own domain's entity authority. Most operators get one of those three settings wrong. The ones who get all three right are extracting real citation share. The ones who do not are spending writer cycles producing content that flows authority to medium dot com instead of to their own brand.

We ran 1,800 queries through ChatGPT, Claude, and Perplexity in April 2026 across data science, design, product management, engineering, and startup-advice categories. We then crawled the cited URLs to identify which publications, which canonical settings, and which paywall states produced citations. This is the operator playbook that came out of that work.

The 2026 State of Medium

Medium today is a different platform from the Medium of 2017. Three structural changes define the current environment.

First, the Medium Partner Program now pays writers based on member read-time rather than claps, with payments concentrated heavily on member-only stories behind the paywall. That economic gravity pulls writers toward locking their best content behind the paywall, which is the same content AEO strategy needs publicly accessible.

Second, Medium consolidated and shuttered several of its in-house publications between 2022 and 2024. Marker (business), OneZero (tech), Forge (productivity), and Elemental (health) were wound down. Better Programming was archived in mid-2024. The surviving in-house anchors are Human Parts, Index, and a handful of editorial verticals. The high-citation-authority publications still active in 2026 are largely community-owned: Towards Data Science (owned by TDS Inc.), UX Collective (owned by Fabricio Teixeira and Caio Braga), and Better Humans (owned by Tony Stubblebine via Coach.me before he became Medium CEO).

Third, the paywall is now harder. Medium updated its paywall enforcement in late 2023 to truncate the HTML body served to unauthenticated requests rather than relying on JavaScript-side blurring. The change broke a class of "reader mode" workarounds and, crucially, also blocked LLM training crawlers from ingesting member-only post bodies at scale. Common Crawl, GPTBot, ClaudeBot, and PerplexityBot all receive the excerpt-only response on member stories.

The result is that Medium is now bifurcated as an AEO surface. Non-member stories on high-authority publications are still excellent. Member-only stories are nearly invisible to LLMs. The publication you choose and the paywall toggle you flip at publish time determine whether the post functions as an authority signal or a private content vault.

The Publication Citation Map

We tagged every Medium URL that surfaced in our 1,800-query audit, traced it back to the source publication, and measured citation frequency by publication. The table below shows the publications that produced 10 or more citations in the audit window, ranked by share of total Medium citations.

The pattern is uncompromising. Three publications — Towards Data Science, UX Collective, and Better Humans — account for 65% of all Medium-sourced LLM citations in our audit. The Startup, with its 750,000 followers, accounts for 2%. Followers are a vanity metric that LLM citation graphs do not weigh; editorial gatekeeping is the metric that does the work.

Towards Data Science deserves a closer look because it is the single highest-citation Medium publication on technical topics. The publication accepts roughly 12% of submissions according to public statements from its editorial team. Each accepted post goes through a content review that demands code reproducibility, methodology disclosure, and a clear takeaway. The result is a content surface that LLMs treat as roughly equivalent to a mid-tier academic blog. ChatGPT will cite a Towards Data Science post on a topic like SHAP feature importance or transformer attention visualization at a rate competitive with the Distill journal or Lil'Log.

UX Collective operates similarly. Its editorial team rejects an estimated 70% of submissions and pushes accepted authors through structural revisions. The result is the only design publication on Medium that consistently surfaces in ChatGPT, Claude, and Perplexity answers to questions about design systems, user research, and product UX.

The Startup, by contrast, accepts almost any submission that meets minimum length and topic criteria. Its content is structurally heterogeneous, the editorial signal is near zero, and LLM citation graphs treat posts on The Startup as effectively unauthored. Operators who post to The Startup in hopes of "getting on Medium" are pouring writer hours into a citation surface the LLMs ignore.

What Canonical Tags Actually Do for Republished Content

Medium's import-from-URL feature was built originally as a writer convenience: you publish on your own blog, then import the post to Medium for distribution. The import sets the canonical URL on the Medium post to your original URL, telling Google and the broader citation graph that your domain is the source of record. The Medium copy is a republished surface, not the original.

The mechanics of the canonical tag are simple but consequential. When the Medium HTML contains a `link rel="canonical" href="https://yourdomain.com/post"` header, every downstream system that respects canonical tags — Google Search, Bing, OpenAI's training pipeline, Common Crawl's de-duplication — will attribute the canonical URL as the source. Backlinks pointing to the Medium URL effectively flow PageRank and entity authority to your domain. LLMs that learn the post during training learn it as "published by yourdomain.com" rather than as "published by medium.com."

When the canonical is missing or set incorrectly — for example, when a writer copies and pastes the post into Medium's editor as a fresh story rather than using import — Medium's own URL becomes the canonical of record. Your domain gets no link equity, no entity attribution, no citation share. The post is effectively reassigned to Medium.

The April 2026 audit data is unambiguous on this point. Posts on Towards Data Science with the canonical URL pointing to the author's own domain or company blog were cited as "from yourdomain.com" in 84% of ChatGPT answers that referenced the post. Posts on the same publication without a canonical set were cited as "from medium.com" or "from a Medium post by [Author]" in 91% of referencing answers. The entity attribution shifts almost completely based on a single dropdown setting in Medium's Story Settings panel.

For operators with a brand-authority goal, the canonical setting is non-negotiable. Either publish on Medium with the canonical pointing to your own domain, or do not publish on Medium.

The Member-Only Paywall and AI Crawler Behavior

The paywall toggle is the second non-negotiable setting. Medium's paywall, in its current 2026 enforcement, serves three distinct response classes depending on the requester.

Authenticated paying member. Receives the full HTML body. This is the only state in which a human or bot can read the entire post.

Authenticated free member, or unauthenticated browser. Receives a partial body containing the title, subtitle, header image, byline, and roughly the first 400 to 800 characters of body text. The remainder is replaced by a paywall component prompting signup or upgrade.

Identified bot (GPTBot, ClaudeBot, PerplexityBot, CCBot, Googlebot). Receives the partial body identical to the unauthenticated browser case for member-only stories. For non-member stories the full body is served.

The implication is that any member-only story is, from an LLM perspective, an excerpt-only document. The training pipeline can index the title, byline, and excerpt, but it does not learn the actual content of the post. Citations on member-only posts fall to 11% of the rate of equivalent non-member posts in our audit data.

Some operators try to work around this by publishing the same content twice — once on a publication as a member story for Partner Program revenue, and once on their own blog as the full version. The canonical tag must then be set carefully or the duplicate-content signal will hurt both URLs. Our canonical tag strategy walkthrough covers the multi-version pattern. The clean rule for AEO-focused operators is simpler: publish non-member on Medium when you publish there at all. Forfeit the Partner Program revenue on the post. The dollars from member read-time on a single Medium post are vanishingly small compared to a single LLM citation that surfaces your brand to a procurement buyer.

The Brand-Authority Calculation

The other side of the Medium tradeoff is brand attribution. When ChatGPT cites a Towards Data Science post written by your team, the citation surface in the model's answer typically reads "according to a Towards Data Science article by [Author]" or "as published on Towards Data Science." Your company name often does not appear in the citation at all. The author byline, the publication name, and the URL are the surfaces the model exposes.

For personal-brand-building authors, that attribution is fine. For company-brand-building operators, it is partial. The post built authority for Towards Data Science and for the author's individual byline, but not for the company.

Compare this to a post on your own company blog cited in ChatGPT. The citation typically reads "according to a 2026 [Company Name] analysis." Your brand is the entity surfaced. The citation directly converts to brand-mention currency on the conversion path; see our analysis of brand mentions currency for the underlying mechanics.

The right framework is to use Medium publications for senior individuals on your team whose personal brands you want to build — the founder, the head of research, the CTO — and to use your own domain for institutional content. Both can coexist. The mistake is to default-publish institutional company content on Medium and surrender the brand entity to the publication.

Comparison: Medium vs Substack vs Ghost

Medium is one of three publishing platforms that operators most often consider as alternatives to a self-built blog. The three solve different problems.

Substack is the audience-growth engine. Its core product is the subscriber list and the email infrastructure to deliver content to that list reliably. Substack's recommendations network now drives a substantial fraction of new subscriber acquisition for nonfiction writers in 2026. Its citation-graph behavior, though, is mixed. Substack posts published on subdomain dot substack dot com domains get lower LLM citation share than Substack posts published on the writer's own custom domain (which Substack supports). Substack offers per-post paywall on a subscriber tier, which has the same crawler-blocking effect as Medium's paywall. For audience-building combined with AEO, Substack with a custom domain is a strong choice. Our Substack newsletter AEO breakdown covers the citation-rate data in detail.

Ghost is the self-hosted choice. The open-source CMS is operated commercially by the Ghost Foundation in Singapore, and a hosted Ghost(Pro) plan starts at $9 per month. Ghost gives you full control over your domain, your hosting, your HTML, your robots.txt and llms.txt files, your JSON-LD schema implementation, and your paywall logic. For pure brand-authority AEO outcomes — where the citation must flow to your company entity, where SEO equity must accrue to your domain, where AI crawler access policies are operator-controlled — Ghost wins decisively. The cost is that Ghost does not have a built-in recommendation network or subscriber graph. You have to build your audience from cold traffic.

The comparison table below summarizes the tradeoffs.

The decision framework: if your goal is personal-brand-building inside a curated editorial publication, use Medium with a strong publication and a properly set canonical. If your goal is owned-audience-plus-AEO without infrastructure work, use Substack with a custom domain. If your goal is maximum brand authority and AEO control, use Ghost on your own domain.

The 90-Day Medium AEO Playbook

The operators who get measurable returns from Medium in 2026 follow a tight playbook. The version below is the cleanest pattern we observed across the companies that actually pulled citation share from Medium publications.

1. Decide the single publication target. Pick exactly one publication based on your topic. For data and ML, target Towards Data Science. For design and UX, target UX Collective. For productivity and personal development, target Better Humans. Do not spread submissions across multiple publications; each editor team has a distinct voice and a distinct rejection rate, and you want to learn one editor team's preferences.

2. Read the editorial guidelines and the last 30 published posts on the target publication. Towards Data Science requires reproducible code, methodology disclosure, and a clear takeaway. UX Collective demands a structural argument and supporting visuals. Match the format precisely. Authors who treat the guidelines as suggestions get rejected at the same rates as authors who never read them.

3. Publish the post first on your own domain. Put the canonical URL on your domain. Wait 24 to 48 hours for Google to index the post on your own site. This sequencing is critical for canonical attribution to flow correctly through to Medium's import.

4. Import to Medium via the URL importer, not by paste. Use the import-from-URL feature at medium.com/p/import. The importer automatically sets the canonical URL to the source URL. Verify in Story Settings, Advanced that the Canonical URL field shows your original URL.

5. Toggle off the member-only paywall before publishing. In the publishing flow, the lock-icon toggle determines whether the story is gated. Turn it off for AEO posts. You will forfeit Partner Program revenue on that post, and that is the correct tradeoff.

6. Submit to the publication, not the public timeline. Each top publication has a submission process. Towards Data Science uses a "Writer Application" form; UX Collective uses an editor-DM workflow. Submit through the proper channel and expect 7 to 21 days for response.

7. Cross-post to LinkedIn the same week. A LinkedIn post with a teaser of the article and a link to the canonical URL on your domain (not the Medium URL) extends the social signal and re-routes any LinkedIn-graph traffic back to your owned surface.

8. Track citations starting at week three. New posts typically need three to six weeks before LLM training pipelines pick them up. Use Profound, Otterly, Peec, or AthenaHQ to track when ChatGPT, Claude, and Perplexity start citing the post. Tag the post in your citation tracker as "Medium: TDS" or "Medium: UXC" so you can attribute downstream.

9. After 90 days, decide whether to scale or stop. If the post generated 3 or more LLM citations and at least one downstream brand inquiry, repeat the pattern on the same publication. If it generated zero citations after 90 days, the publication is probably wrong for your topic, or the post was structurally weak. Do not assume "Medium worked" until citations show up in tracking.

Real-World Operator Cases

Three operator patterns showed up repeatedly in our research:

Case 1: A B2B data platform. A series-B data infrastructure company published 14 long-form technical posts on Towards Data Science between September 2025 and March 2026, all with canonical URLs pointing back to their engineering blog and the member-only paywall toggled off. By April 2026, ChatGPT was citing the company's analyses in 23% of relevant technical procurement queries, up from a baseline of effectively zero in August 2025. The author bylines, all senior engineers on the team, also grew personal followings that translated into recruiting funnel benefit. The total writer time was roughly 80 hours of senior-engineer drafting time over six months; the citation outcome would have been impossible to buy through paid channels.

Case 2: A design agency. A boutique design consultancy published 22 case-study breakdowns on UX Collective over a year. They never set the canonical URL. By month 12, UX Collective and Medium were getting credited in LLM answers, but the agency's own brand was effectively invisible in citation surfaces. They had built authority for UX Collective and for two individual designers, not for the agency. Lesson: the canonical setting is the difference between authority transfer and authority forfeiture.

Case 3: A consumer SaaS company. A founder published a high-effort 4,500-word "lessons learned" essay on The Startup, expecting Medium's reach to amplify it. The post got 14,000 reads from Medium's recommendation engine but produced exactly zero LLM citations over 90 days. The Startup is not a citation signal. The founder rewrote the same essay as a guest contribution to a niche industry publication six weeks later, and that version was cited in Perplexity inside a month. Lesson: editorial curation, not volume, drives LLM citation.

When Medium Is the Wrong Answer

Medium is the wrong AEO surface in several specific situations.

It is the wrong answer for original research and data studies. These pieces should live on your own domain with full structured data, a stable URL, and complete citation freedom. Republishing a research post on Medium adds nothing and risks splitting the citation graph.

It is the wrong answer for pricing-page-equivalent content. Anything that converts directly — pricing tables, product specs, calculator pages, buying guides — belongs on your own domain with full control over conversion infrastructure.

It is the wrong answer for time-sensitive news and updates. Medium's URL structure and editorial review timelines do not match a news cadence. Your own changelog, blog, or newsroom on your domain is the right surface.

It is the wrong answer if you do not have a top-tier publication accepting your work. Posting to a low-curation Medium publication contributes neither brand authority nor citation share. The effort would be better spent on your own blog, Substack with a custom domain, or LinkedIn long-form.

What Medium Itself Says About AI

Tony Stubblebine has been explicit on the Medium official blog that Medium views high-quality human-written content as its core defensible asset against AI-generated content commodification. The platform announced in 2024 a policy of removing AI-generated stories from the Partner Program, and Stubblebine has argued repeatedly that Medium's value proposition is "humans curating humans, in service of humans." In practice, that posture aligns with the AEO operator's interest: when LLMs are trained on Medium content, they are training on a corpus the platform has actively defended against AI-generated spam. The signal-to-noise ratio remains higher than on most open-web sources, and that is the underlying reason Towards Data Science and UX Collective have retained citation authority even as Medium's MAUs declined.

Search Engine Journal has tracked Medium's SEO trajectory closely. Their reporting through 2025 highlighted that Medium's domain authority remains in the 95+ range despite the user-base contraction, which means a backlink or citation surface on a Medium publication still carries technical SEO weight. That weight is the second reason top publications stay valuable. The combination — high domain authority plus editorial curation plus persistent LLM training inclusion — is what makes Towards Data Science a real AEO asset and what makes The Startup a noise channel on the same platform.

Takeaway: Medium is not a binary yes or no for AEO. It is a publication-by-publication decision filtered through three settings: editorial curation level, canonical URL configuration, and member-only paywall state. Towards Data Science, UX Collective, and Better Humans are the publications whose curation justifies the writer cycles. Canonical URL must point to your own domain or you are subsidizing Medium's authority instead of building your own. The member-only paywall must stay off or LLM crawlers will treat the post as an excerpt. Operators who get all three settings right extract real citation share from Medium in 2026; operators who get any one wrong are wasting writer hours. For most B2B brands the cleaner alternative is Substack with a custom domain or self-hosted Ghost, with Medium reserved for senior individuals whose personal brand you want to build inside a curated publication's editorial halo.