How to ‘optimise for AI’ when nobody is sure what that means

Publishing your own content remains a highly effective marketing tool. Companies that want to raise their profile and lead conversations in their industry need a strong website and a steady flow of thought leadership articles. But, on its own, that is no longer enough to guarantee visibility with the audiences that matter. Increasingly, people searching for products or services never reach websites at all: instead of using a search engine and clicking through links, they put a question directly to an artificial intelligence (AI) chatbot, trusting the answer it provides.

The scale of this search shift is significant. Google’s AI Overviews, the summaries above its search results, has surpassed two billion users a month, and ChatGPT has passed 900 million weekly users. The challenge has moved from the ability to be found by a search engine to being visible, and quotable, inside an AI-written answer.

That work has a name, Answer Engine Optimisation (AEO), but few agree on what it involves. Traditional Search Engine Optimisation (SEO) involves optimising online content for rules that search engines such as Google and Bing publish. Chatbot makers don’t publish an equivalent, and may never be able to, because nobody knows exactly how these tools choose their answers. Even the labs building them call the models a black box. That means the first task with AEO is figuring out what it means so you can apply it in practice.

Three FirstWord clients have asked us for AEO help in the past year, each from a different starting point: a pharmaceutical group undergoing a rebrand, a private-equity firm responding to a content audit and a global manufacturer worried that AI-driven answers were misleading their customers. Here’s what each asked for, and what we did. First, though, what we know.

What good AEO looks like, as far as anyone can tell
At FirstWord, we’re inquisitive editors and journalists by background, and our internal AI experts spent months experimenting with the models and reading the research. There are no published rules; you infer what works by asking the same question in enough variations to see patterns. Academics have done this at scale, benchmarking thousands of queries and measuring what moved. They have discovered that adding statistics, clear citations and quotations lifts the visibility of a page by up to 40 per cent, for example.

When you ask a chatbot a question, it does roughly what a human would: runs a search, usually against Google’s index, then writes an answer from the results. A page that can’t be found in search can’t be quoted, so good SEO is the foundation of AEO, not something to be replaced by it.

AI models favour explicit, extractable and attributable content: short paragraphs, headings that clearly describe the section underneath, acronyms that are defined on first use and facts that are grouped so they can be lifted cleanly for use in an answer. Where possible, they also prefer a named author or spokesperson with a real title, because models weigh an attributable claim more heavily than an anonymous one.

The most subtle of these is making content than can be easily ‘excerpted’; smaller passages that can be pulled out and still make sense in isolation. A paragraph with an undefined “it”, “we” or “this” is ambiguous so the model might ignore it, limiting AI visibility for that company. It also pays to start a page with a plain summary of its contents, because that is the first thing a model reaches for when it needs to understand what any page is about. A clear summary gives your company, rather than a rival, a good chance of framing the AI answer. With all that in mind, what does optimising for AI look like in practice?

The rebrand: a website built for two types of readers at once
A pharmaceutical group came to FirstWord mid-rebrand, with a new logo, new colours and a website that would be the first public expression of the change. The brief was to edit every page against two objectives: bringing the copy into a new tone of voice and then optimising it for both traditional search and chatbots.

Our first move was to build an editorial checklist for the new site, drawing on established SEO rules, emerging AEO techniques and plain editorial common sense. It gave everyone one standard to write and check against.

Sometimes objectives disagreed. The new tone of voice was conversational, including a preference for saying “we” rather than naming the company. Optimisation wants the opposite. If we write “FirstWord provides this”, a chatbot can quote it and the attribution remains. “We provide this” leaves the model unsure who “we” is.

The checklist can’t resolve that, so an editor must – by weighing readability against extractability. The compromise was to name the company where a summary or a standalone claim needed it, and use “we” where the surrounding copy already made the subject obvious. That balance is an editorial judgement made on the final read-through, and it’s the part a tool can’t do for you.

The private-equity firm: an AI audit, but no one to act on it
The private-equity firm was further along. It had commissioned an audit of its appearance in AI answers and been handed a prescriptive content guide: abstracts at the top of each page, at-a-glance fact boxes, pull quotes, a bank of questions and answers at the bottom. What it needed was help turning this into strong, compelling content that was friendly to machine visitors while retaining interest to human ones.

That’s where FirstWord came in. We carefully rewrote a page in the new format, keeping it genuinely readable rather than merely compliant, and the firm is now running this against a comparable page written the traditional way. The aim is to obtain real data on which changes influence how and where the company shows up. This client is measuring, not guessing, and we’ll learn what works alongside them, applying those changes to the next pages we write.

The manufacturer: invisible to AI and misquoted
A global manufacturer came to us worried on two counts. Its content was barely surfacing in AI answers, and when the models did describe the company, they leaned on third-party sources that were often wrong: an incorrect product spec here, an outdated feature set there. Uncorrected, that becomes the version of the company millions of users read first.

A consultant advised the firm to restructure the whole site for AI bots, with fact boxes and bullets throughout. That advice is fine up to a point. Applied mechanically to every page it produces a site that reads more like a spreadsheet and has the potential to drive visitors away. It echoes the early days of SEO, when writers crammed in keywords, making web content unreadable.

Our role here was to help the manufacturer publish clear, well-structured pages that models could find and quote correctly. The goal was for the company’s own content to outrank the incorrect pages. We did this by using structure where it genuinely helped, but never at the expense of readability. That judgement is the value we add on top of the checklist.

What our checklist does, and what it can’t
The checklist FirstWord uses synthesises the best current thinking from SEO and AEO into something an editor can work through. Does a webpage page open with a summary a model can lift? Is the company named? Do sentences stand alone as much as possible? It’s a floor, not the finished job, and our caveat is always that a checklist can’t guarantee a ranking or a citation, because those depend on the workings of the model’s black box.

AEO remains elusive. There is no surefire checklist and given how AI systems work there may never be. What we can tell you is that we’ve done the reading, run the experiments, worked on live sites and arrived at a set of principles that constitute what best practice looks like while the ground is still moving. Tangible, replicable and evolving, they beat being baffled by the black box.