What is Answer Engine Optimisation (AEO)?

Why buyers ask an AI before they ask you

The highest-intent moment in your funnel now happens somewhere you cannot see. A buyer describes their problem to ChatGPT, Gemini, Claude or Copilot. Or they type it into Google as usual and an AI Overview - or the full AI Mode - answers before the links they used to click. Whichever tool it is, it hands back a shortlist. If you are not on it, you are not compared. You get skipped.

"Answer engine" covers all of these: any AI system a buyer asks that replies with an answer instead of a list of links. They differ in brand and interface; they choose their sources in strikingly similar ways.

This is no longer an early-adopter habit. 54% of UK adults now use AI tools such as ChatGPT, Copilot or Gemini, rising to 79% of 16 to 24 year olds (Ofcom, 2026). ChatGPT logged 1.8 billion UK visits in the first eight months of 2025, roughly five times the year before (Ofcom, 2025). And it has reached buying: around 30% of UK consumers have used an AI chatbot to research a product or service before purchase, up from about 12% in early 2025 (SearchScore, 2026). In B2B it is near-universal - 89% of buyers use AI at every stage of the purchase journey (Forrester, 2026).

Meanwhile the old channel is thinning. Fewer than 1 in 3 Google searches now ends in a click to anyone's website (SparkToro, 2026), and Gartner forecasts search volume falling 25% by 2026, with brands' organic traffic halving by 2028 (Gartner, 2024).

Here is the twist that makes this an opportunity rather than a funeral: the traffic that does arrive from AI answers converts around 4x higher than classic organic search (Semrush, 2025). An assistant does the comparing before the click, so the visitor lands half-decided. Fewer visits, far warmer. The page's job has changed from persuading a cold stranger to confirming a recommendation already half-made.

How answer engines actually choose who to recommend

There is no separate AI index and no secret markup (Google, 2026). AI answers draw on two paths, and both are mundane: ordinary search indexing, and live retrieval - the assistant fetching and reading pages at the moment it answers. Both paths read the HTML your server returns. If a fact is not in that response, to an answer engine it does not exist.

It also matters which bot you let in, because the crawlers have split roles. OpenAI, Anthropic, Google and Perplexity each run separate agents for model training, for search indexing, and for user-triggered fetches (per each vendor's own crawler documentation). The practical consequence catches many sites out:

Blocking a training bot keeps your words out of a model. Blocking a search bot keeps your name out of the answer.

You can decline to train AI models and still be recommended by them. But a "block AI bots" switch flipped on at your hosting provider, or a robots.txt file copied from somewhere else, silently removes you from the very answers your buyers are reading. Check what your site actually blocks, not just what its settings say - it is one of the most common failures our diagnostic finds.

AEO, GEO, SEO: how the terms fit together

The jargon is messier than the idea. GEO (generative engine optimisation) is the academic term, coined by the researchers who first measured what lifts visibility in AI answers. AEO, AI SEO and LLM SEO are the market's names for the same practice. Use whichever you like; they describe one discipline.

SEO is not AEO's rival - it is its foundation. Because AI answers draw on the ordinary search index, everything classic SEO earns you (pages that are easy to crawl, clean structure, honest titles, a business whose details check out) is also what feeds the answer engines. AEO is the layer on top: shaping your content and your facts so an AI assistant can lift a correct claim about you without guessing, and has a reason to name you when it does.

You do not win the answer by shouting louder. You win it by being the easiest correct answer to give.

What measurably lifts your visibility in AI answers

Most "AEO ranking factor" lists are folklore. One study actually instrumented the question: researchers added different elements to pages and measured visibility across thousands of generative answers. Adding citations, statistics and quotations each lifted a page's visibility by roughly 40% (Aggarwal et al., KDD 2024). Keyword stuffing, by contrast, made things worse. Clarity and evidence win; tricks lose.

In practice, five disciplines do most of the work:

• Answer first. Open every page with the direct answer in the first 40 to 80 words, in plain language. Write that paragraph as if it will be quoted alone with your name on it - because it will be.
• Cite everything. A number beats an adjective, and a sourced number beats everything. Never invent a figure: an AI assistant that catches your site contradicting its sources has every reason to cite someone else.
• One story, everywhere. Your name, your offer, your prices and your facts should read identically on every page and every profile you control. Every inconsistency is a gap the model fills by guessing, and it rarely guesses in your favour.
• Say it in schema too. Schema (JSON-LD) is a small block of code on each page that restates its facts - who you are, what you offer, prices, FAQs - in a format AI systems read without guessing. There is no special AI version of it (Google, 2026): the same schema that wins the richer search listings is what confirms your identity for answer engines. One rule is non-negotiable: mark up only what the page visibly shows. Schema that claims what the page does not gets treated as spam.
• Keep it in the HTML. If your key claims only appear after a script runs or a "view more" click, most AI readers never see them. The test is simple: what a plain fetch of your page returns is what the answer engine gets.

Even AI judges you on speed, security and clutter

The disciplines above are about what your pages say. How your site behaves matters almost as much, because an answer engine experiences your site the way an impatient visitor does - just faster. And a recommendation is a transfer of trust: the assistant stakes its own credibility on every site it names. It will not hand its users a slow, insecure one.

Speed is the first trust signal, for both readers. Google calls a page "good" when it loads in under 2.5 seconds; past that, people leave and AI agents move on to the next source (Google, 2024). The commercial stakes are bigger than citations alone: a 0.1-second improvement lifts conversions by around 8% (Google × Deloitte, 2024). An assistant answering in real time has seconds to gather its sources. The site that responds instantly is in the running; the site that makes it wait is not.

Security reads as credibility. Assistants avoid risky recommendations, and so do the buyers checking up on them. A site with broken HTTPS, expired certificates or sloppy security basics is a risk neither will take - and in categories like law, finance and health, where the questions are most sensitive, the bar is highest.

Clutter and surveillance bury the answer. Pop-ups, cookie walls, chat widgets and a payload of trackers do not just irritate visitors; they wrap your actual content in noise an AI reader has to dig through, and a site that tracks its visitors more than it informs them gets noticed for the wrong reason. The cleanest page is the cheapest to read, and the cheapest page to read is the easiest to cite. This is Agentu's founding argument: build the site light, fast, private and honest, and both audiences - human and AI - reward you for the same reasons.

You do not need to rank on page one to be cited

Here is the most encouraging thing in the research: AI answers are not a mirror of Google's top ten. Assistants fetch and read many sources when they answer, and the same study that measured the 40% uplift found the biggest gains went to sites ranked lower in the search results (Aggarwal et al., 2024). The small firm with the clearest, best-sourced page on a specific question can be quoted ahead of the household name that outranks it.

There is a catch: assistants cross-check before they commit. What the rest of the web says about you counts - your directory listings, your Google Business Profile, your LinkedIn, your reviews, your professional register. Two things decide how that check goes:

• Consistency. The same name, the same offer, the same facts everywhere you appear. Every mismatch makes you a riskier thing to recommend, and assistants do not take risks with their answers.
• Coverage. The more independent places repeat the same story, the more confident the assistant that the story is true.

A claim that lives only on your own website is one source's word for it. A claim the web repeats back is a fact.

What about llms.txt?

You will hear a lot about llms.txt, a proposed file that gives AI systems a plain-text summary of who you are and what you offer (llmstxt.org). Our honest take: ship one - it is cheap, harmless and occasionally read - but know that it is a proposal, not a standard, and no major assistant has committed to it.

Two rules if you do. Every fact in it must match your live site; an llms.txt that oversells is a file you signed that teaches AI assistants to misquote you. And never make it the strategy. The durable investments are answer-first content and consistent, sourced facts - those survive every model release. Files like this come and go.

What does not work (and what not to pay for)

The gold rush has arrived ahead of the gold, so it is worth naming the traps:

• Anyone selling guaranteed AI citations or "submission to the AI index". No such index exists (Google, 2026). There is nothing to submit to.
• Cloaking - showing AI crawlers different content from humans. It violates search spam policy, and it torches the one asset AEO is built on: being a source a model can trust.
• FAQ schema for questions your page never visibly answers. The rule is simple: if the page does not show it, the code must not claim it.
• Optimising for one assistant. Answer engines rise and fall quarterly. The durable asset is being the clearest, best-grounded source in your niche, not gaming whoever leads this quarter.

Where the opportunity is biggest

Almost no one has done this yet. Fewer than 5% of UK SMEs have taken any action on AI search visibility (SearchScore, 2026). In most categories the "most citable source" title is sitting unclaimed, and answer engines have to recommend somebody.

The exposure is not evenly spread, either. High-stakes, considered-purchase categories lead: legal queries show 11.9x the typical rate of AI-visitor adoption, with finance and health at 2.9x (SE Ranking / Omnibound, 2025). The handover to humans is already visible in the data - of consumers who took a legal question to an AI assistant, 28% were directed to contact a lawyer (Clio, 2025). Somebody gets named in that answer. If you run an accountancy firm, a law firm or a private clinic, the queue at the door is already forming; it is just forming inside a chat window.

Why AEO is never finished

One more truth the vendors soft-pedal: AI answers are not fixed. They shift with every model release. The page that earned a citation in January can drop out of the answer in April without you changing a word. AEO is a practice, not a project, and three habits carry most of it:

• Update in place. Keep your URLs stable and refresh the facts on them. A stable, current page compounds authority; a stat that has quietly aged is a reason for an assistant to cite someone newer instead.
• Keep publishing real insight. Answer engines need sources worth naming, and they reward the site that keeps answering its category's questions with fresh, well-sourced material. One genuinely useful article a month beats a redesign a year - it is also why this Insights section exists.
• Re-verify your numbers. Every claim on your site carries a source, and sources age. A figure that was true when you published it and wrong now is worse than no figure at all: you are training AI assistants to misquote you.

How to measure whether it is working

AI answers do not show up in your analytics, so the metric has to change. The one that matters is share of answer: when the questions your buyers ask are put to the assistants they use, are you in the answer, are your facts right, and who else is named?

Measuring it is unglamorous and effective. Write down the ten questions your best clients would ask an AI assistant. Ask them, verbatim, in ChatGPT, Claude, Perplexity and Google's AI mode. Record the answers with dates. Re-run the same set quarterly. "In January no assistant mentioned us; today two of four name us and quote our pricing correctly" is a sentence your accountant can love - impressions do not close deals, recommendations do.

How to start this week

1. See where you stand. Run our free agent-readiness diagnostic. It reads your site the way an AI assistant would, in about ten seconds, with no email required.
2. Fix the foundation. A fast, clean site that AI systems can actually read is the precondition for everything above. That is what Agentu CORE builds.
3. Answer the questions your buyers actually ask. One page per question, answer first, sources attached. If you work in a sector we know well, start from an industry playbook.

Do these three things and you stop optimising for an algorithm you cannot see, and start building for the constants: clarity, speed, trust, and being the easiest correct answer to give.