Some time in the last eighteen months, a developer added a file called llms.txt to your website. Perhaps you asked for it. Perhaps a plugin did it without telling you. Either way, somebody ticked "AI-ready" off a list, and the invoice mentioned generative engine optimisation.
Here is the problem. The crawlers it was written for are, by every public measurement, not reading it.
One monitoring study covering over 500 million AI bot visits across a 90-day window found that only 408 requests targeted /llms.txt directly. GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended overwhelmingly skip the file and crawl the HTML like everybody else. As of Q1 2026, no major AI company has publicly committed to reading it in production — not OpenAI, not Google, not Anthropic, not Meta, not Mistral. Google's Gary Illyes said plainly in July 2025 that Google does not support it and has no plans to.
Meanwhile a SE Ranking study of 300,000 domains found roughly 10% adoption — of which 39.6% are plugin-generated stubs, files nobody wrote and nobody maintains.
We are not going to tell you to delete it. It costs nothing and it is not harmful. We are going to tell you something more useful: why an idea with no supporting evidence spread this fast, and what the same effort buys if you spend it correctly.
The market for effort-shaped placebos
llms.txt is popular for one reason that has nothing to do with whether it works: it is shippable in an afternoon.
It is a single markdown file. It requires no architectural decision, no content rewrite, no negotiation with anyone about what your company is willing to say in public. It produces an artefact you can point at in a status update. It looks exactly like progress.
The actual work of becoming citable looks like none of those things. It is rewriting a services page so it contains numbers instead of adjectives. It is arguing internally about whether to publish prices. It is server-rendering a catalogue that currently only exists after JavaScript runs. It is maintaining a page so it does not go stale. Every one of those is slow, contested and invisible in a sprint board.
So a proxy appears. This is not new — robots.txt meta keywords, the keyword density era, the "mobile-friendly badge" era. Each time, the industry found a cheap artefact that resembled the expensive behaviour, and sold it until the evidence caught up.
The tell is always the same: the artefact requires no decision from the business. If a tactic can be implemented entirely by a developer, without anyone changing what the company is prepared to say about itself, it is almost certainly not the thing that moves a model's answer.
What we know actually correlates with citation
Set aside the file. Here is what the 2026 evidence supports, in rough order of leverage.
1. Specific, checkable facts on the page. A model composing an answer needs something it can lift and stand behind. "Fast delivery" is unusable. "Three to five days for a landing page, one to five weeks for a store with payments" is usable, because it is falsifiable. Most B2B sites fail here before any technical consideration applies.
2. Server-rendered HTML. The crawlers are fetching your HTML, not your file. If your content only materialises after client-side JavaScript, you are gambling your presence on crawler behaviour you do not control. Everything we ship renders on the server — the Bronjoy marketplace's city, category and venue pages, the Finkonkurs coin and exchange pages, and every corporate site we build.
3. Structured data. This is the honest version of what llms.txt pretends to be: a standard, widely supported, machine-readable statement of your facts that the crawlers demonstrably do parse. Bronjoy's venue pages carry LocalBusiness entries with coordinates, hours and ratings, wrapped in ItemList, with breadcrumb and FAQ markup on top. BiliMind emits schema.org/Course on every published course. Our hackathon tool RecipeReel exports schema.org/Recipe JSON-LD, and LectureLoom exports LearningResource, because structured output is what makes a result reusable by anything other than a human reading it.
4. Freshness. Of the thousand pages ChatGPT cites most, Ahrefs found 76.4% had been updated within the previous 30 days. A Last updated date that is actually true is worth more than any file you could add.
5. Off-site corroboration. Models weight agreement across sources. First-party sites account for a large share of citations, but directories and listings account for a comparable share. If your prices on your own site disagree with your prices on a marketplace listing, you have given the model a reason to trust neither.
A four-question test for the next tactic somebody sells you
llms.txt will not be the last one. The pattern repeats every time a channel shifts, so it is worth owning a test rather than an opinion.
1. Who has to change their behaviour for this to work? If the answer is "a third party we do not control has to start reading a file", the tactic is a bet, not a plan. If the answer is "we have to say something we currently do not say", it is real work with a real payoff.
2. Can it be implemented without a single decision from the business? Tactics that need no decision are usually tactics that change nothing. The uncomfortable rule: if it does not require an argument, it probably does not require a competitor to respond either.
3. What is the evidence, and is it a measurement or a coincidence?
"We added X and citations went up" is worthless if five other things shipped that month. Ask what was held constant. In the one widely-shared llms.txt success story, JSON-LD structured data shipped in the same change — and structured data has a decade of documented vendor support behind it.
4. If it disappeared tomorrow, what would break? A tactic that leaves nothing behind when it fails is a tactic with no residual value. Server rendering, structured data and specific copy all keep paying regardless of what any assistant does next. That asymmetry is the whole argument for spending there.
Run those four questions against anything on an agency proposal. They will not always give you a clean answer, but they reliably separate work from artefacts.
The same afternoon, spent better
Here is a direct swap. Your developer has four hours. Compare:
| Option | Effort | Evidence it works |
|---|---|---|
Ship llms.txt | ~1 hour | 408 targeted fetches out of 500M bot visits; no vendor commitment |
Add FAQPage + Organization + Product/Service schema to your three highest-intent pages | ~3 hours | Parsed by every major crawler; documented, versioned, supported |
| Server-render your pricing and services pages if they are currently client-rendered | ~4 hours on most stacks | Removes the single biggest cause of invisible content |
| Replace every adjective on your services page with a number | ~4 hours, mostly arguing | The precondition for any citation at all |
Three of those four are permanent infrastructure improvements that also help human visitors, search engines and your own sales team. One of them is a file nobody fetches.
The part nobody wants to hear
The reason most businesses are not cited has nothing to do with technical readiness. It is that their website does not contain any information worth quoting.
Go to your services page right now and try to extract three sentences a stranger could repeat as fact. Not "we deliver world-class solutions tailored to your needs" — an actual claim, with a number or a boundary in it. On most sites there are none. The page is a mood, not a document.
That is why we publish entry prices for all seven services, the support period for each plan, the response time you get on each, and the list of work we refuse outright. Not because it optimises for anything, but because a company that cannot say what it charges, how long it takes, and what it will not do is a company an AI model has nothing to say about either.
The uncomfortable corollary: your competitor's willingness to be specific in public is now a distribution advantage over you. For years vagueness cost you nothing — you could keep prices off the site and hoover up "contact us" leads. Now vagueness costs you the mention that precedes the lead.
What "specific" looks like, concretely
Advice to "be specific" is itself unspecific, so here is the same paragraph twice.
Before — the version on most services pages:
We deliver bespoke, high-quality web solutions tailored to your business needs. Our experienced team leverages cutting-edge technologies to create digital experiences that drive results. We pride ourselves on transparent communication and on-time delivery.
There is nothing in there a model can lift. No number, no boundary, no claim that could be wrong. Strip the company name and it fits any of ten thousand businesses — which is precisely why it will never be quoted as evidence about yours.
After — the same three sentences, made citable:
A landing page takes three to five days and starts at $99; a store with payments takes one to five weeks and starts at $499. The price is fixed in writing before any work starts and changes only if you ask for something outside the agreed scope, which is priced before it is built. You get a working build to click through every Friday, and you own the source code and every account from day one.
Same length. Same message. Every sentence now contains something falsifiable — a figure, a cadence, an ownership claim — which means every sentence is available as an answer to a question somebody is asking an assistant right now.
The test is mechanical: read each sentence and ask whether a competitor could copy-paste it onto their own site without lying. If they could, it is not information about you. It is wallpaper.
Do this to your three highest-intent pages and you will have done more for your AI visibility than any file, plugin or retainer on the market.
A sensible position on the file itself
Keep it if you have it. Do not pay anyone to maintain it. Do not let it appear on an invoice as a deliverable. If a proposal lists llms.txt among your AI-search work, ask what else is on the list — and if the answer is thin, you have learned something useful about the proposal.
And if someone shows you a case study where adding the file coincided with a citation, apply the standard test: was anything else changed at the same time? In the one widely circulated example, the file shipped alongside JSON-LD structured data — which is the part with a decade of vendor support behind it.
What we would do with your site
If AI visibility matters to you, the work we would actually do is: server-render what is currently client-rendered, add real structured data to the pages that answer buying questions, rewrite those pages so they contain facts, and put a maintained date on them. That is a high-conversion website engagement — 3–10 days for most sites, from $99 — not a specialist retainer.
Tell us what you need. You get a fixed price and a delivery date back — free, and with no obligation to proceed.
Frequently asked questions
So should I remove llms.txt from my site?
No. It is harmless and costs nothing to leave. Just stop treating it as work done. The risk is not the file; it is the false sense that AI visibility has been handled.
Will the major AI companies adopt it later?
Possibly. If they do, adding it takes an hour. There is no first-mover advantage in a file that is fetched by a parser, so there is no reason to invest in it ahead of adoption.
What about robots.txt rules for AI crawlers — do those matter?
Yes, and they are respected far more consistently. That is a real lever: you can allow or block specific AI user agents, and Cloudflare now offers one-click controls and a paid-access marketplace. For most product and service businesses, allowing them is correct — see our piece on AI search as shelf space.
Is structured data enough on its own?
No. Schema makes your facts machine-readable; it does not make them interesting. A perfectly marked-up page that says nothing specific is still unquotable. Facts first, markup second.
How do I know whether any of this worked?
Ask the assistants the twenty questions your buyers ask, once a month, and write down whether you appear and whether what is said is accurate. It is a spreadsheet, not a platform, and it is the only measurement here that is not guesswork.
