The Citation Graph
Doesn't Read
Your Drafts.

The argument of the last Shift was that AI citation positions reinforce themselves. Brands cited as authoritative today become more likely to be cited tomorrow. Brands absent from the citation pool now become harder to surface in future cycles, and the gap compounds.

That logic does not stop at the brand level. It applies, with uncomfortable precision, to the editorial those brands publish. Every piece you ship is either entering the citation graph or being forgotten by it. There is no neutral state. The piece you wrote last quarter is, right now, either accumulating credibility or accumulating absence, and almost no founder treats it as a live position rather than a finished artefact.

This is why most founder editorial fails. Not because the writing is bad. Because the writing is treated as the work, when it is roughly 30% of the work. The other 70% (the part that decides whether the piece is cited, surfaced, retrieved, remembered) is engineering, and almost no founder does it.

What follows is three failure modes a founder will recognise, and the diagnostic discipline underneath all of them.

The most common failure mode in founder editorial is the easiest one to describe.

A founder spends three weekends on a piece. It is good. It is original. It says something they actually believe. They publish it on a Tuesday morning, post the link to LinkedIn with a three-sentence framing, and watch the engagement: forty likes, six comments, two reshares, a private DM from someone who liked it. By Friday, the post is buried. By the following Tuesday, the piece is invisible. Three months later, the founder concludes that editorial doesn't work for their business and stops writing.

What actually happened is not that editorial doesn't work. It is that the founder treated publication as the end of the project, when publication is the start of it. The piece existed for 72 hours. After that, it depended entirely on systems the founder never engaged with: the LinkedIn algorithm's decay curve, the absence of inbound links, the lack of any signal to search engines or LLMs that this piece was worth indexing prominently.

The mechanics of editorial decay are unforgiving. A LinkedIn post peaks in reach within four hours and has a median half-life of less than 24 hours. A piece without inbound links has no way to communicate authority to Google's ranking systems, regardless of how good the content is. A piece without structured data is parseable to humans but illegible to the systems that decide which humans see it. None of these are fixable retroactively without significant effort. All of them are trivial to address before publication.

The failure is not in the writing. It is in the assumption that good writing earns its own distribution. It does not, and has not for at least a decade.

What the discipline looks like, in operating terms: the piece needs at least five direct outreaches to named individuals on the day it ships, three trade-press pitches in the first week, and a fragmentation cadence on social; the same piece broken into four or five separately-shareable pulls over thirty days. None of this is glamorous work. All of it materially changes whether the piece survives its first month.

The deeper point is what not doing this work costs. A piece that doesn't get cited in its first 30 days does not just fail to travel. It actively trains the systems around it that this author and this topic are not authoritative. Future pieces by the same founder, on the same topic, start from a slightly worse position. The absence compounds.

The second failure mode is harder to spot because the founder did everything they were supposed to.

They wrote carefully. They published properly. They shared with thought. They did the outreach. The piece even got modest pickup; a couple of newsletter mentions, a handful of LinkedIn reshares from people whose opinion they respect. By all visible metrics, the piece worked. Six months later, they Google their own argument, and find nothing. They ask ChatGPT about the topic they wrote 4,000 words on, and the model has never heard of them. They check Perplexity. The pieces it cites are weaker than theirs, by writers with smaller audiences. Their own work is nowhere.

This is not a writing problem. It is a structural illegibility problem, and it is the silent killer of mid-market founder editorial in 2026.

A piece that does not present its key claims as extractable units (schema-marked headers, FAQ blocks, structured methodology, well-formed citations) is a piece that requires inferential work to summarise. Models can do that work, but at the margin, they prefer not to. Given a choice between two pieces making similar claims, the citation graph picks the one whose structure makes the citation cheap. Beautiful prose is expensive to extract. Bullet-pointed methodology is free.

This is the part of the new search reality that mid-market founders, raised on the idea that quality content wins, find hardest to accept. The citation graph is not adjudicating quality. It is adjudicating extractability. The two correlate weakly at best.

The technical inputs are unromantic and well-documented: Article schema with proper author and publisher metadata; FAQ schema for any question-and-answer block; BreadcrumbList schema for navigation; Open Graph metadata for social previews. A robots.txt that explicitly permits the AI crawlers you want indexing your work (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended) is no longer optional. None of these are technically difficult. Most CMSes generate them automatically, but most automatically-generated implementations are incomplete in ways the founder cannot see without checking.

There is also a content-shape question that sits underneath the technical one. A piece argued in beautiful flowing prose, with no extractable units, is a piece the citation graph cannot easily quote. A piece that includes a clearly-labelled methodology section, a CFO objection box, a worked example with a table, and a sidebar with three numbered moves, is a piece that gives the graph four extractable surfaces. The first piece may be better writing. The second piece is more citable writing. They are not the same skill, and conflating them is the most expensive mistake in modern founder communication.

The remedy is not to write worse. It is to write the same argument with deliberate extraction surfaces built in; pull-quotes, structured asides, methodology blocks, FAQs at the foot. The argument lives in the prose. The citations live in the structure. Both jobs need doing.

The third failure mode is the most uncomfortable, because it is the one a founder typically discovers too late to do anything cheap about.

A founder asks ChatGPT about themselves, or their company, or their category. The model responds; fluently, confidently, with citations. The description is wrong. Not factually wrong, exactly. Tonally wrong. Positionally wrong. The model has decided the company is a marketing agency when they have repositioned as a strategy consultancy.

This is not a hallucination. It is path-dependency.

Citation graphs are sticky. In 2023, Stanford researchers (Liu et al.) published "Lost in the Middle," demonstrating that large language models suffer from profound primacy bias during retrieval. They exhibit a U-shaped performance curve, heavily weighting information placed at the very beginning of a sequence and effectively ignoring what comes after.

When applied to a founder's digital footprint over time, this mechanism anchors the AI's prior. The first cluster of references the model encounters about a person or brand gets weighted disproportionately in subsequent answers. If the earliest material positioned a founder as a digital marketer, the model will keep describing them that way long after they have rebranded as a web intelligence advisor.

The dormant Medium account from 2021. The LinkedIn posts from when they were figuring out their voice. The guest podcast appearance they barely remember. These are not warm-up. These are foundational. By the time the founder is writing their flagship piece, the model has often already decided who they are.

This is documented for retrieval behaviour and emerging in training-data analysis; the behavioural extension to founder positioning is inferred, not directly measured. But the inference is expensive to ignore.

This produces the most painful version of the COI logic, because the cost is hidden. The founder writes their best work, properly distributes it, gets it cited correctly, and the citations are read against an existing model-position that is years out of date. The piece does not move the needle as much as it should. The founder cannot see why, because the model never tells them what its prior was.

The remedy is not subtle, but it is uncomfortable. It involves auditing what the major models currently say about you and your company, identifying where the position is wrong, and deliberately publishing into the gap; pieces specifically designed to give the citation graph fresh, well-structured, extractable signal that contradicts the outdated prior.

The three failure modes look different. A founder experiencing publish-and-pray has a discoverability problem. A founder experiencing illegibility has an extractability problem. A founder experiencing path-dependency has a positioning problem. Three different headaches.

But the underlying physics is the same in all three. Each is a symptom of treating editorial as a finished artefact rather than a position in a graph that compounds. The graph rewards founders who treat each piece as a deliberate move in a long game; it punishes founders who treat each piece as a one-off broadcast.

Until 2024, the assumption that publishing was the work was not catastrophic. Editorial sat on a website. Search engines indexed it linearly. A founder who wrote well and shared modestly could expect a fair shake at the discoverability lottery, because the lottery was largely run on the surface signal of the piece itself.

That world is gone. The citation graph has memory. It rewards consistency, structure, extractability, and authority signal; not effort, eloquence, or sincerity. Founders who do not internalise this difference will keep producing editorial that they think is excellent and the world will keep treating as approximate.

The discipline that addresses all three failure modes is the same discipline. Treat each piece as a position in a compounding system. Engineer its discoverability before publication. Build extraction surfaces into the writing itself. Audit what the systems currently say about you, and publish into the gap deliberately. None of this is glamorous. All of it is the work that the writing was always supposed to be the start of.

The piece you publish next week will either compound or evaporate. The citation graph is already counting. It does not read your drafts.