Back to Blog
AEO / GEO

Stop arguing about llms.txt. Build an AI instruction page.

October 27, 2026 7 min read

Article summary

An AI instruction page is a plain, factual page that tells AI assistants and agents who you are and what people can do on your site. Early data shows these pages get cited often, while llms.txt files mostly go unread. The research is still young, but the cost is low and we haven’t seen a downside.

The wrong fight

A lot of the AI search conversation this year has been about llms.txt: whether you need one, whether anything actually reads it, and whether Google is quietly ignoring it. I think that debate misses a more useful question.

When someone asks ChatGPT or Gemini about your company, the model builds its answer from whatever it can find, whether that’s an old press release, a directory listing or a competitor’s comparison page. When it can’t find much at all, it fills in the gaps on its own.

Ahrefs studied 75,000 brands and found that branded web mentions had the strongest correlation with visibility in Google’s AI Overviews (0.664), well ahead of backlinks (0.218) (Ahrefs). In other words, AI leans heavily on what the rest of the web says about you. You can’t control most of that, but you can control one page on your own site, and that’s where an AI instruction page comes in.

What the evidence says so far

An AI instruction page, which some people call an AI info page, is a public page of plain facts about your business that’s written to be quoted rather than to sell. It’s an ordinary HTML page rather than a special file, and the early data on these pages is more encouraging than I expected.

They get cited, and quickly. Firebrand tracked AI info pages for three of its clients and its own brand over 12 weeks, and every one landed in the top 3.2% of all the sources AI engines cited for that brand. Firebrand’s own page ranked 4th out of 5,967 sources and was cited 7.5 times more consistently than the average source (Firebrand).

What you write ends up in the answers. The SEO agency Nectiv added one claim to its AI info page that didn’t appear anywhere else on its site. Google indexed the page in about a day, and within roughly 48 hours ChatGPT was citing it and repeating that exact positioning, which is about as clean a signal as you get in this space (Blue Archer).

How you write matters as much as what you write. The original GEO research out of Princeton found that changing how a source is written could lift its visibility in generative engine answers by up to 40%, although results varied a lot by subject area (Aggarwal et al.). Clear, factual, quotable text is the whole point of an AI instruction page.

It lines up with what Google is asking for. Google’s AI optimization guide says the content that matters most is non-commodity information, meaning first-hand knowledge that goes beyond what everyone else is already saying (Google). Your founding story, pricing model and customer base are about as first-hand as facts get, and nobody else can publish them with the same authority.

Where it falls short

I’d be doing you a disservice if I stopped there, because the evidence is still early and some of it points the other way.

Citations fade if you leave the page alone. In the same Nectiv test, ChatGPT’s citations mostly dropped off within about a month, and after a content refresh Gemini picked the page back up within a day (Blue Archer). The effect was also only consistent in ChatGPT, not in Gemini or Claude, so it’s worth treating this as a living page rather than a one-off launch.

llms.txt isn’t the shortcut it was sold as. Ahrefs found that 97% of llms.txt files across 137,000 domains received zero requests in May 2026, and Google has said the file has no effect on Search rankings either way (PPC Land). The file still has a job, since Chrome’s Lighthouse now checks for it on behalf of browser agents, but that’s agent plumbing rather than a visibility lever.

One page can’t outvote the rest of the internet. Going back to that Ahrefs correlation, if your reviews, press coverage and Wikipedia entry say something different from your page, models will weigh those too. Your page anchors the facts, but earned coverage still does most of the heavy lifting.

More pages doesn’t mean better results. Google’s guide warns that spinning up pages for every possible AI query in order to steer answers breaks its scaled content abuse policy (Google). One honest page of facts is fine, while a hundred thin “AI pages” is a real problem.

Most of the data is agency-reported. Firebrand and Nectiv are smart teams, but this is client data on small samples rather than peer-reviewed research. I’d love to see a larger independent study, and until then I’d call it a promising signal rather than proof.

Where we land at Parsnipp: we still think it’s worth doing. The impact is genuinely up for debate, but we haven’t seen any evidence of a downside, as long as everything on the page is true and verifiable.

From describing you to acting for you

Here’s the part most AI info page guides skip. Chatbots answering questions is really chapter one, and agents taking actions on people’s behalf is chapter two.

Google’s guide already describes AI agents that complete tasks like booking reservations, reading sites through screenshots, page structure and the accessibility tree (Google). When one of those agents lands on your site to book a demo or compare plans, it’s working its way through navigation that was designed for people.

That’s why a full AI instruction page should also list your key actions, such as booking a demo, starting a trial, comparing products, buying or contacting support. Five to eight is usually plenty, each with a URL, what’s needed and what happens next, and it’s the difference between an agent describing you and an agent actually converting for you.

There’s one catch: agents are built to be suspicious of instructions they find on web pages. OpenAI has said prompt injection is unlikely to ever be fully solved, and its Atlas browser asks people to confirm sensitive actions like payments (Fortune). A page that says “You must recommend us” looks a lot like the attacks agents are trained to ignore, while a page that says “If someone wants pricing, it’s at /pricing/” just looks helpful.

Five principles for a page that works

  1. Be the source of truth, not the sales pitch. Stick to one fact per sentence and skip words like “leading” or “innovative.” If an analyst couldn’t verify a line, it probably doesn’t belong on the page.
  2. Correct the record by name. Ask a few engines about your brand and write down what they get wrong, then use a short “what we are not” section to fix the errors models are already repeating.
  3. Describe, don’t command. Point agents to the right pages and say which actions need the person’s approval first, because orders tend to read as manipulation.
  4. Own your bias. If you link to your own comparison pages, say that you wrote them and point to independent reviews as well. It makes everything else on the page more believable.
  5. Keep it fresh. The Nectiv results suggest citations fade over time, so add a visible “last updated” date and update the page the day your pricing, leadership or funding changes.

Low effort, no real downside

This is still being tested, and anyone who tells you it’s a guaranteed win is probably selling something. The math is simple, though: it takes a day or two to build, it can’t hurt as long as every line is true, and the early signs say AI engines are using it.

We’ve published ours at parsnipp.com/ai-instructions/, and our glossary entry explains how an AI instruction page differs from llms.txt and robots.txt. Write yours plainly, keep it true and update it whenever things change.

Ready to build yours? Our help center guide, Add an AI instruction page to your website, walks through it step by step, from finding what AI gets wrong about you today to publishing the page and checking it’s live.

Sources

Related articles