Writing for machines without stopping writing for people
Two stories from the same week. The *Seattle Times* and *Newsday* sued OpenAI and Microsoft for infringement. And authors are arguing over the split of the settlement reached with Anthropic, as publishers and agents claim a share.
You can read those as lawsuits. It is more useful to read them as what they are: the ongoing negotiation over the terms on which published text feeds models. And that negotiation reaches anyone who publishes, not only large media.
Meanwhile a fact does not wait for the outcome: for many people the first answer to a question is no longer a list of links, it is a generated paragraph. If your company is not in that paragraph, it does not exist for that query.
Crawlable is not the same as citable
Classic SEO optimises for being found and clicked. When the reader is a model, the objective changes: you want it to understand you well enough to cite you correctly. Those are different things and they are optimised differently.
A model that cites you wrongly is worse than one that does not cite you. If it answers that your company is somewhere it is not, or sells a product it does not, that error propagates without you being able to correct it. And it happens for a banal reason: on your site that information was implicit, and the implicit does not get quoted.
It happened to us. Our own site claimed a portfolio of products without linking to any of them, with brands spread across three domains and nothing connecting them. A human reader inferred it from context. To a model they were four unrelated companies.
What we changed, concretely
Three things, none of them a trick:
- An explicit entity graph. Every property of the group has a stable identifier and edges stating what belongs to whom: who publishes what, which company is a subsidiary, who founded the business. No two descriptions of the same object can contradict each other, because there is only one and everything else references it.
- Files meant to be read by machines. `llms.txt` and its long form: the site map in quotable prose, with the company's facts written so they can be copied without being interpreted. Legal entity, address, founding year, what each brand is and how it relates to the parent.
- Answers visible, not merely structured. The FAQ is on the page as text a person reads, and additionally marked up so a machine understands it. Marking up an answer that is not visible is asking a model to trust something the visitor cannot check.
What unites the three: saying explicitly what used to be inferred. Nothing in the content is new; there is simply less to guess.
The uncomfortable part of the deal
Becoming citable means accepting that some of your content will be consumed without a visit. It is a trade with real losers, which is why media companies are suing: for a business that lives on page advertising, a summary that avoids the click is a direct loss.
Our calculation is different and worth stating plainly: we do not live on visits, we live on someone deciding to work with us. A model correctly explaining what Valfiguer does to someone who will never open our site is still valuable. If we lived on advertising this decision would be another one, and the companies suing have reasons that are not ours.
Being citable is a trade: fewer visits in exchange for appearing where there are no longer links.
There is a counterweight worth attention too. Law schools are telling students to put AI away, and the reason is exactly this: when the summary replaces the source, the ability to check is lost. Publishing for machines should not make that worse; it should do the opposite, putting facts where they can be verified.
How you measure this, which is the awkward part
Here is the honest problem: there is no Search Console for generative models. No panel tells you how often you were cited, on which queries, or whether it was accurate. Anyone selling you an "AI visibility" dashboard is selling an estimate with a nice interface.
What you can do is more manual and considerably more reliable:
- Ask directly. Put the ten questions a prospective customer would ask about your company to several models and read the answers. Not a metric — an audit, and it catches the errors that matter.
- Watch the error, not the volume. Being cited less is ambiguous data. Being cited saying something false is a concrete failure you can trace back to the page that allowed it.
- Look at referral traffic. When a model links, that link shows up in your analytics like any other. It is a fraction of the citations, but it is the fraction you can count.
Our provisional conclusion, stated with the uncertainty it deserves: the most useful signal is not how often you appear, it is whether what is said about you is correct. The first you do not control. The second you do, and it is what gets fixed when the graph gets fixed.
What we do not intend to change
None of this justifies writing for the algorithm. The obvious temptation is volume: if models read, give them plenty to read. It is a mistake, and a penalised one — Google treats mass-produced content made primarily to rank as abuse, and that penalty lands on the whole domain, not the page.
Our rule is boring: an article gets published if someone who already knows the subject learns something. Otherwise it does not, however much the calendar says it was due. Text a model can quote but which says nothing only gets you quoted saying nothing.
The part that stays the same as ever: write well, say checkable things, and link where they come from. It turns out that is also what makes a model understand you. Not a coincidence — clear text is clear to every reader.
