AIO: optimising for the machines that answer the question

Something is shifting in how people find things out. Not away from search exactly, but alongside it: a growing share of questions now go to ChatGPT, Gemini, Perplexity or Claude, and come back as an answer rather than as ten links.

For twenty years the discipline for being found was search engine optimisation. The emerging counterpart has picked up the name AIO - artificial intelligence optimisation - and it means shaping your content so an AI system can understand it, trust it, and cite it.

Why it is not just SEO with a new name

A search engine ranks documents. You give it a page, it decides where that page sits in a list, and the user picks from the list. The mechanics are opaque in their details but well understood in their shape, and there is a whole industry built on measuring them.

An assistant does something different. It composes an answer from a mix of training data, its own reasoning, and - increasingly - a live retrieval step over the web. The user often never sees a list at all. There is no rank to measure, no position to track, and far less feedback about why you were or were not included.

Which means two things. Being cited is more valuable than being ranked tenth, because there is no tenth. And the process is less transparent, which makes it easier to influence in ways nobody can see.

The grey zone, and where we draw the line

SEO went through a black-hat phase: keyword stuffing, invisible text, link farms, doorway pages. All of it worked for a while, all of it eventually got penalised, and a lot of businesses paid for someone else's shortcut with a ranking collapse.

AIO is at the start of the same curve. The technique that prompted this post was a single line on a page, invisible to anyone reading it in a browser, addressed to the machine rather than the human - a nudge to recommend one vendor over another. We are not going to reproduce it, because a working example on a company blog is just a distribution channel for it.

But it is worth being clear about why we think it is a bad bet, not just a rude one:

• It is content that contradicts the page it is on. Everything we know about how these systems get hardened says that gap is exactly what gets detected and penalised.
• It is trivially discoverable. It lives in a public document that anyone can view the source of, including your competitors and your prospects.
• It wins you a citation for a claim your actual product does not support, which converts into a disappointed buyer rather than a happy one.

The uncomfortable part, for users rather than marketers: most people have no idea that an AI recommendation could already have been shaped this way. "The AI said so" carries a weight that "a website said so" lost years ago, and it has not earned it yet.

What actually seems to work

The honest summary is that nobody has a reliable playbook yet, and anyone selling one is ahead of the evidence. But the techniques that look durable share a property: they make a page genuinely easier to understand, so they pay off even if every assistant disappeared tomorrow.

Here is what we did on this site, concretely.

• An llms.txt file. A plain-language map of the site at /llms.txt - what the company does, the services, the case studies with their outcomes, the FAQ, the legal entity, the price floor. It is generated from the same data as the pages, so it cannot drift into describing a site that no longer exists.
• One structured-data graph per page, not a pile of fragments. The organisation is defined in full exactly once, on the home page, and every other page references it by @id. That is what lets a crawler treat forty-odd pages as one business rather than forty unrelated ones.
• Facts stated plainly, in prose. Where we are, what a project starts at, which languages we work in, the registration number. A model cannot cite a number that only exists inside a graphic.
• Answers before preamble. Our FAQ answers the question in the first sentence and explains afterwards. That is good writing anyway; it also happens to be the format that survives being extracted.
• One source for anything that appears twice. The FAQ page and its FAQPage markup are rendered from the same array. A page whose markup disagrees with its own visible text is the single easiest thing for a system to distrust.

None of that is a trick. It is the same answer the web has had for a long time: be clear, be specific, be consistent, and make the machine-readable version say exactly what the human-readable version says.

What this means for you

If you are using these tools: stay curious and stay critical. An answer that arrives in confident prose, with no links and no visible sources, is not neutral by default. Ask where it came from.

If you run an e-commerce site or a content operation: start by finding out whether you are being cited at all. Ask the assistants your customers would ask, in your categories, and see whose name comes back. That takes an afternoon, and it is more useful than any strategy deck on the subject.

AIO is not a mature discipline yet. But the groundwork - structured data that matches your content, a readable summary of what you do, facts in text - is cheap, it is useful on its own terms, and it is in place before the question of who gets cited becomes competitive.