What is answer engine optimisation (AEO)?

An answer engine is any system that replies to a question with written prose instead of a list of links. Google's AI Overviews and AI Mode are answer engines. So are ChatGPT, Gemini, Perplexity and Copilot on the occasions they search the web before replying. Answer engine optimisation, usually shortened to AEO, is the work of making a site something those systems can retrieve, trust and quote without distorting it.

The term is new enough that it is being sold as a whole new discipline. It is not. Most of what makes a page quotable is what made it rankable, done more carefully. The parts that are genuinely different are worth knowing precisely, because they are the parts a supplier can either do or wave at.

You will also see it written answer engine optimization. That is the American spelling of the same thing.

What happens before the answer is written

Almost every grounded answer is produced in two steps. First the system retrieves: it turns your question into one or more of its own queries, runs them against an index, and pulls back a small set of documents. Then it generates: it writes an answer from the material it has in front of it, usually citing some of the sources it used.

Those are two separate gates, and a page can fail either one. A page nobody can retrieve is never considered, however well written it is. A page that is retrieved but says nothing liftable gets read, discarded, and the answer is assembled from somebody clearer. Almost every argument about AEO makes more sense once you ask which of the two gates is the problem.

The first gate is ordinary technical search work

Retrieval runs on indexes. For Google's AI features the index is Google's, which is why the company's own documentation on AI features in Search, at developers.google.com/search/docs/appearance/ai-features, points back at Google Search Essentials for eligibility. For assistants that search the open web, retrieval usually runs through a search API belonging to a conventional engine. Either way, a page that is not indexable is not in the running.

That makes the first half of AEO indistinguishable from a technical audit. The list is familiar and it is not optional.

• The page returns a real status code and is not blocked by a stray disallow rule or a leftover staging header
• Content is in the HTML rather than assembled by a script that only runs for some clients
• Canonical tags point where you think they point, so the version that gets retrieved is the version you wrote
• One page answers one question, instead of four pages competing over the same one
• The site is fast enough on a mid-range phone that nothing times out on a slow connection

In our experience this is where most of the lost ground actually is. A site that cannot be crawled cleanly does not have an AI problem. It has the problem it already had, with a new symptom.

The second gate is whether a passage survives being lifted

The generation step does something a search result never did. It takes a passage out of your page, strips it of its surroundings, and puts it in somebody else's paragraph next to two competitors. Writing that depends on the heading above it, the image beside it and the three paragraphs before it does not survive that. Writing that states the thing plainly does.

So the practical test for any page is simple. Take any two consecutive sentences out of the middle of it. Do they still mean something on their own, and are they still true on their own? If the answer needs the rest of the page, the page will be read and not quoted.

One set of facts, in every place you appear

The dullest part of this work carries most of the weight. A model assembling an answer about your business sees your site, your profiles, your directory listings, an old press release and whatever a data aggregator decided about you three years ago. Where those disagree, the safest thing it can do is quote a business whose facts do not disagree.

So the fixes are unglamorous and they are the job.

• One legal name, spelled and punctuated the same way everywhere
• One address and one phone number, including on the contact page, which is the page that most often contradicts the rest of the site
• One description of what the business does, short enough to be repeated verbatim
• One list of services, with the same names on the site as in the markup and the profiles
• Retired services and old office addresses taken down rather than left to be found

Structured data states what prose only implies

Schema.org markup, documented at schema.org, is how a page says a fact rather than implying it. Organisation, service, article, breadcrumb and FAQ types cover most of what a business site needs. The value is not a ranking effect. It is that a machine reading your markup does not have to infer your address from a paragraph, or guess whether the name in the footer is the company or the parent group.

Two rules keep it honest. Mark up only what is visible on the page, because markup that describes content a reader cannot see is what Google's structured data guidelines exist to rule out. And validate it after every template change, because structured data breaks silently. Nobody reports a schema error. It just stops being read.

Write the answer first, then earn the rest of the page

Answer led writing is not a content format, it is an ordering decision. The question goes in the heading, the answer goes in the first two sentences under it, and the explanation, the caveats and the trade-offs follow for the reader who wants them. You lose nothing by putting the answer at the top. The people who were going to read the whole thing still do.

1. Use the question as the heading, worded the way a person would ask it out loud
2. Answer it completely in the first two sentences, with no throat clearing and no promise to explain later
3. Put the qualifying detail next, because the caveat is what makes the answer usable rather than repeatable
4. Keep one idea per paragraph, so a quoted fragment lands on a whole thought rather than half of two

There is a line here worth respecting deliberately. Mass producing a page for every phrasing of the same question is what Google's spam policies describe as scaled content abuse. The demotion, when it comes, costs more than the citations were worth. One good answer beats forty variants of it.

Decide what your robots.txt says to AI crawlers

Robots.txt is now a published specification, RFC 9309, at rfc-editor.org/rfc/rfc9309, rather than the informal convention it was for years. Alongside the crawlers that build search indexes, the large providers document separate user agent tokens for the crawlers that gather training and grounding material. Google's Google-Extended token is the documented example.

This is a real decision with a real trade-off, and it is worth taking on purpose rather than inheriting from whatever your platform shipped. Blocking those crawlers is a legitimate position. It also removes you from the answers. The other thing to be clear about is that robots.txt is a crawling instruction and nothing else. It is not a privacy control, it does not remove a page from an index, and a URL blocked in robots.txt can still be listed if other sites link to it.

llms.txt, and what it is not

The llms.txt proposal, at llmstxt.org, suggests a plain markdown file at a known path that lists a site's important content for language models. We implement it where it costs almost nothing, because it does no harm and the file is easy to generate. We also say plainly what it is: a proposal, not a standard, with no obligation on anyone to read it. Anybody selling llms.txt as the reason your citations will improve is selling a text file.

Measuring something that is not a ranked list

There is no rank tracker for answer engines, because there is no rank. What there is, and what we think is the only honest measurement available, is a repeated observation. Write down the questions your buyers actually ask. Ask them of the same assistants, on a schedule, and record what comes back.

• Whether you are named at all, and in what company
• How you are described when you are named, which is frequently the finding that matters most
• Which page of yours was cited, if any, and whether it is the page you would have chosen
• Which third party sources keep appearing for your category, since those are where your facts need to be correct

Answers vary between runs for the same prompt, so a single check tells you very little and a trend over months tells you something. Treat any supplier who reports a single AI answer as proof with the scepticism you would apply to a single screenshot of a ranking.

What this work cannot do

We would rather say this before the proposal than after it.

• Nobody can guarantee a citation in any assistant, because these systems are not ranked lists and they change without notice or documentation
• Nothing here substitutes for having something worth citing, and a business with nothing specific to say is not held back by its markup
• Answers will still sometimes be wrong about you, and the realistic goal is that the correct version is the easiest one to find
• There is no paid placement to buy inside these answers, and anything sold as one is advertising with a different label

Where we would start

If you are doing this yourself, the order matters more than the effort. Most of the value sits in the first three steps and most budgets get spent on the fourth.

1. Write down the questions, then ask them and record the answers, so you have a baseline before you change anything
2. Reconcile the facts into one agreed version and fix every place that contradicts it, starting with the contact page
3. Add and validate organisation, service and FAQ markup on the templates rather than on individual pages
4. Rewrite the openings of the pages you already have, before writing any new ones
5. Ask again in a quarter, with the same prompts, and compare

The uncomfortable finding, most of the time, is not that a competitor is being cited. It is that the answer describes your business using facts you stopped being true about two years ago. Tech Conversion

Questions people ask about this

What is answer engine optimisation?

Answer engine optimisation is the practice of structuring a website so that AI answer systems can retrieve it, read it and quote it accurately. Where search engine optimisation competes for a position in a list of links, answer engine optimisation competes to be the material an answer is written from. The foundations overlap heavily, and the emphasis on factual consistency is what differs.

Is AEO different from SEO, or just a new name?

It shares most of its foundations. Crawling, indexing, clear writing and authority serve both. What differs is emphasis: answer engines reward facts that agree with each other everywhere you appear, structured data that states rather than implies, and passages that hold their meaning when lifted out of the page. Anyone selling it as wholly new work is usually selling you your own technical SEO twice. Treat AEO, SEO and the content behind them as one programme with two kinds of result: rankings and quotations.

Does blocking AI crawlers protect my content?

It stops compliant crawlers fetching your pages, which is a legitimate choice. It also removes you from the answers those systems generate, so it is a trade rather than a free protection. Robots.txt, specified in RFC 9309, is a crawling instruction only. It does not remove anything from an index and it is not a privacy or security control.

How do I know whether any of this is working?

Write down the questions your buyers ask, ask them of the same assistants on a fixed schedule, and record whether you are named and how you are described. Answers vary between runs, so a single check proves nothing and a trend across months means something. Treat a supplier who shows you one favourable screenshot the way you would treat one ranking screenshot.