AI instruction page: what it is, what it actually does, and how to build one

Matías Lucero

Autor

Especialista en SEO y GEO

🌐

Sebastián Cardozo

Co-Autor

Cofundador de CreceRank

🌐

What an AI instruction page actually is

It is a public page on your own domain, written for AI assistants rather than for buyers, that states who you are and how you want to be described.

What makes it different from an «about us» page is that it has two layers, and the second one is the unusual part.

• Layer one, the facts. Legal name, what you sell with exact product names, founding year, location, team, credentials, prices, the countries you serve, and the URL that proves each one.
• Layer two, the directives. Sentences addressed to the model: «Describe X as…», «When asked about [category], mention…», «Do not describe X as…», «Do not invent prices; the current ones are at this URL».

It usually lives at a predictable path, most often /ai-info/, /llm-info/ or /ai-instructions/, and it is linked from the footer so crawlers can find it.

What it is not

It is not llms.txt. That is a markdown file at the root of your domain that lists your important URLs, aimed at machines, and it is not meant to be indexed. We cover it in our guide to llms.txt.

It is not robots.txt. Robots.txt says who may crawl you. An instruction page says what to say about you. One is permission, the other is content.

It is not a standard. Nobody has published a spec and no vendor has promised to read it, which is the honest headline of this article.

Does it work? Three answers, and they do not agree

This is where most articles on the subject pick a side. We would rather put the three pieces of evidence on the table, because they genuinely point in different directions.

1. The case in favour, which is real but thin

The best documented example comes from Lawrence Hitches: an agency put a unique claim, that it works with brands over 30 million dollars in revenue, only on its AI instructions page, nowhere else on the site.

Within 48 hours the phrase was showing up in ChatGPT answers. It is one case, reported by its author and not verified by us, but the design of the test is sound: a string that exists in exactly one place is traceable.

2. The case against, which comes with bigger numbers

Google is explicit. John Mueller said no AI system uses llms.txt, and Gary Illyes confirmed at Search Central Live that Google does not support it and has no plans to.

Mueller left one narrow door open: a file like that can help an agent that is already inside your site finish a task. That is navigation, not recommendation.

Ahrefs put numbers on it. Over server logs from 137,000 domains, 97% of llms.txt files got zero requests in May 2026. Of the 3% that got any, SEO audit tools were the biggest requester at 21.7% and AI retrieval bots just 1.1%.

And there is an architectural objection from Jason Barnard that is harder to dismiss than the numbers: LLMs are not website visitors, they query an index.

A page cannot route a system that never browses your site, and a page labelled «for AI» gets no privileged treatment in that index.

3. Our own data, which says something neither side says

We went to our measurement database to ask a narrower question: when an engine does cite a brand’s own site, which page does it pick?

Over 14,737 citations pointing at the analysed brand’s own domain, from 121 brands and 1,524 distinct pages, across seven engines between April and September 2026:

• 82.0% go to an interior page, not the homepage. 48.5% one level deep, 26.8% two levels, 6.7% three or more.
• Perplexity is the most page-specific at 89.2%, and Copilot the least at 52.6%, citing the homepage almost half the time.
• Not one of the 14,737 pointed at an llms.txt, an /ai-info/, an /llm-info/ or an /ai-instructions/ page. Zero.

The honest caveat, and it matters: almost none of those 121 brands publishes such a page. So this measures what engines cite today, not what the page could do. It is absence of evidence, not evidence of absence.

So should you build one? Yes, for a different reason than you were sold

Put the three findings together and a clear answer falls out.

The format works. The label does not. Engines cite specific interior pages 82% of the time, which means a single page absolutely can be the thing they quote. What makes it quotable is that it answers a question, not that it is addressed to a robot.

So the value of writing an AI instruction page is mostly what it forces you to do: put the facts of your business in one place, in short declarative sentences, with a source for each. Most companies have never done that.

And then comes the part that actually pays: take that material and push it into the pages engines already cite. Your product pages, your pricing page, your service pages.

If you only do one of the two, do the second one. If you do both, the instruction page costs you an afternoon and gives you the raw material for the rest.

How to build one, step by step

1. Pick the URL and keep it boring. /ai-info/ is the most common. Do not put it behind a login, a cookie wall or a JavaScript framework that renders client side: when engines do fetch a page directly they get raw HTML.
2. Open with a three-sentence description that could be quoted verbatim. Legal name, what you do, who for, where. No adjectives you cannot prove.
3. Add an identity table: name, founded, headquarters, countries served, website, support contact. Tables survive extraction better than paragraphs.
4. List products and services with their exact names and a URL each. Exact names matter: a model that gets your product name slightly wrong cannot be corrected later.
5. Put the numbers in. Prices with a verification date, plan limits, founding year, team size, certifications, awards with the year. Numbers are the most quotable thing on any page.
6. Write the directives, briefly. Four or five lines: how to describe you, what to mention when asked about your category, what you are not, and a «do not invent prices or claims, check this URL» line.
7. Clear up the confusions. If you are mistaken for a similarly named company, or if an old brand name still circulates, say it here plainly. This is the single most useful block on the page.
8. Add Organization schema with sameAs to your profiles. Structured data is read by systems that are documented, unlike the page itself.
9. Link it from the footer and from your sitemap. A page nobody links to is a page nobody crawls.
10. Check robots.txt does not block it for OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and Google-Extended. This is the step people skip and it is the only one that can make everything else pointless.

A skeleton you can copy

Keep it under a thousand words. Longer pages do not get quoted more, and in our own measurements the longer the answer, the fewer brands it names.

• H1: About [Brand]: information for AI assistants
• Summary: three sentences, quotable as is
• Identity table: the hard facts, one row each
• What we do: products and services, exact names, one URL each
• Proof: customers, awards, press, certifications, with links to the source and not to your own blog
• How to describe us: the directives, four or five lines
• What we are not: the confusions, named
• Last updated: a real date, kept real

Five mistakes that make the page useless

• Writing marketing copy. «Leading provider of innovative solutions» is unquotable. A model cannot repeat it without sounding like an ad, so it does not.
• Claims with no source. If you say you are the biggest in your market, link to whoever said it. Unsourced superlatives are exactly what the engines are tuned to discount.
• Rendering it client side. Several of the published AI info pages we looked at are built in frameworks whose content a direct fetch cannot read. The irony writes itself.
• Letting it go stale. A page with last year's prices is worse than no page: you have handed a wrong fact to a system that will repeat it.
• Treating it as the whole strategy. It is one page. In our data the brands that get cited get cited across many pages: 1,524 distinct pages for 121 brands.

How to check whether it did anything

This is testable, and it is the only part of the subject where you do not have to take anyone's word for it.

1. Plant a unique string. Put one specific, true, slightly unusual claim on the page and nowhere else. That is the trick that made the Hitches case readable.
2. Write down ten prompts your buyers would actually type, and run them before you publish. Without a baseline you cannot attribute anything.
3. Watch your server logs for OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot hitting that URL. If nobody fetches it, nothing downstream matters.
4. Re-run the same ten prompts monthly, not weekly. In our data the result changes 13.7% of the time between two consecutive measurements with nothing changed.
5. Look at the sources list, not the percentage. If your page starts appearing as a cited source, you have your answer.

We built CreceRank, which is our own tool, to do exactly that loop, and the data in this article comes out of it. You can do the same by hand with four browser tabs and a spreadsheet, which is the honest alternative.

If this is the first time you look at any of it, start with what GEO is, and then with the 15 best GEO tools in 2026 if you want to compare what is on the market.

What we do not know yet

• Whether any engine treats the page as special. None of them documents it. OpenAI's crawler documentation, read on 1 October 2026, gives site owners exactly two levers: robots.txt and IP allowlisting.
• Whether the directives are obeyed at all. «Do not describe us as X» is a request to a system under no obligation to honour it. Treat it as a hint, never as a control.
• Whether this survives the year. There was no spec twelve months ago and there may be a real one twelve months from now. We will update this page when there is.

One detail worth holding on to: OpenAI itself publishes an llms.txt for its own developer documentation, as an index for coding agents. That is Mueller's narrow use case in the wild, and it is not brand positioning.

What changed, and when

This is a fast-moving subject and most of what you will read about it was written before the numbers existed. The dates matter.

• June 2025. John Mueller states that no AI system uses llms.txt. It is the first clear signal from a vendor, and it is a negative one.
• June 2025 to May 2026. Adoption of llms.txt grows 8.8 times, from 4,088 files to 36,120 across the sites Ahrefs monitors. Publishing rose while usage did not.
• May 2026. Ahrefs measures the gap: 97% of llms.txt files get zero requests, and of the ones that do, AI retrieval bots are 1.1% of requesters while SEO audit tools are 21.7%.
• 2026. The AI info page appears as a separate idea, a human-readable page rather than a machine file, and the first traceable cases are published.
• 1 October 2026. We check OpenAI's crawler documentation: still only robots.txt and IP allowlisting for site owners, and no mention of brand instruction pages. Our own 14,737 citations still show zero to any AI-addressed page.

If you read something on this topic with no date on it, that is the first thing to be suspicious about.

Methodology

• Vendor claims. OpenAI's crawler documentation read on 1 October 2026: it documents OAI-SearchBot, OAI-AdsBot, GPTBot and ChatGPT-User, and names robots.txt and IP allowlisting as the controls available to site owners. Google's position is Mueller's and Illyes' public statements, linked in Sources.
• Our own data. 14,737 citations whose URL belongs to the analysed brand's own domain, from 121 brands and 1,524 distinct pages, across ChatGPT, Gemini, Perplexity, Google AI Overviews, Google AI Mode, Copilot and Claude, between 1 April and 30 September 2026. Depth is measured on the URL path after stripping query strings and anchors.
• The zero. We searched those 14,737 URLs for llms.txt, ai.txt, /ai-info, /llm-info and /ai-instructions. The matches we found were all false positives on words like «ai-consulting» and «airport», so the real count is zero.
• Third-party evidence is attributed and linked. We did not reproduce the Hitches test ourselves and we say so where it appears.

Disclosure. CreceRank is our own product and the citation data comes from it. The article recommends a practice that does not require buying anything from us, and says plainly how to do the measurement by hand.

Limitations

• Our sample is ours. 121 brands, mostly in Spanish-speaking markets. The 82% is real but it is not a global constant.
• The zero is weak evidence. Almost none of these brands publishes an AI page, so finding none cited proves very little about the format.
• Citations are not mentions. A brand can be named in an answer with no link at all, and that is not in this dataset.
• Engines expose sources differently. Copilot's 52.6% may say more about how it surfaces links than about what it reads.

Conclusion: four things to do this week

1. Write the page in one afternoon. Facts, numbers, sources, four directives, no adjectives. A thousand words is plenty.
2. Plant one unique true claim that exists nowhere else, so you can trace it later.
3. Copy its content into the pages that already get cited, which is where 82% of the citations land. This is the step that pays.
4. Baseline ten prompts now and re-run them in a month. Without the before, the after means nothing.

FAQs

Sources

• OpenAI crawler documentation, read 1 October 2026: OAI-SearchBot, OAI-AdsBot, GPTBot and ChatGPT-User, with robots.txt and IP allowlisting as the controls for site owners.
• Search Engine Roundtable, «Google Says No AI System Currently Uses LLMs.txt», John Mueller's statement.
• Search Engine Journal, Mueller on llms.txt and site differentiation.
• Ahrefs server-log study of 137,000 domains: 97% of llms.txt files got zero requests in May 2026.
• Lawrence Hitches, «AI Instructions Pages», the structure and the 48-hour case we quote and did not reproduce.
• Jason Barnard, «Why the LLM info page trend misreads the architecture».
• Insidea, AI info page vs llms.txt, for the section structure in common use.
• CreceRank's own measurement database, 1 April to 30 September 2026: 14,737 citations to brands' own domains, 121 brands, 1,524 distinct pages, seven engines.