? The definitive guide to AEO

How to get recommended by ChatGPT and cited in Google's AI Overviews

Getting recommended by ChatGPT or cited inside Google's AI Overviews is not the same job as ranking on page one of Google. Search didn't die, but the click did: AI Overviews now appear on up to roughly half of all queries, depending on whose data you use, and when one shows up, click-through on the top organic result falls by somewhere between a third and two-thirds depending on the query. This FAQ walks through the specific mechanics that get a business named in an AI answer, and how to check whether you're currently invisible.

It is written for owners, managing directors, marketing directors and practice principals of UK businesses who suspect their marketing is reporting on a game that has changed. Every figure has a source and a date at the bottom of the page. Where the research disagrees, we say so. Where something is our experience rather than a published finding, we say that too.

Last updated 25 September 2026. Written by Simon Young, Founder of Question.Marketing.

Part 1. The Basics: How AI Citation Actually Works

Before optimising anything, it helps to understand what an answer engine is actually doing when it decides who to name.

What does it mean to be "cited" by ChatGPT or an AI Overview?

Being cited means an answer engine names your business, or links to your content, as part of the answer it gives a user, rather than the user having to click through to a list of websites to find that answer themselves. This is a fundamentally different outcome from ranking: there is no "AI ranking position," there is only being named, or not. Depending on the study, somewhere between a third and around half of ChatGPT prompts trigger a live web search to build that answer, with the rest answered from what the model already learned in training. That means your content still has to be findable and retrievable at the moment the question is asked, not just well-written in general.

It helps to separate three outcomes that people tend to lump together, because they are won in different ways:

• A mention. Your business is named in the text of the answer. This is usually driven by what the model already believes about you from training and from repeated corroboration across the web.
• A citation. One of your pages is linked as a source for the answer. This is driven by retrieval: your page was found, read and judged the cleanest thing to quote.
• A recommendation. Your business is put forward as an option in response to a buying question ("who should I use for...", "what's the best... near me"). This is the one that pays, and it needs both of the above: the model has to know you and be able to find evidence for you.

A business can be cited without being recommended (your guide gets linked, a competitor gets named), and recommended without being cited (the model names you from memory and links a directory instead). The work in this guide is aimed at the recommendation, with the citation as the evidence that supports it.

What is the difference between AEO and GEO?

AEO, answer engine optimisation, is the job of building your marketing for a world where people get answers rather than lists of links. GEO, generative engine optimisation, is the mechanics underneath it: the crawler access, page structure and technical signals that replaced the plumbing of traditional SEO. You need both, in that order. A business with perfect GEO mechanics and nothing worth quoting gets nowhere, and a business with genuinely unique knowledge that the crawlers can't reach gets nowhere either.

For the record on where the terms came from: Jason Barnard of Kalicube coined "answer engine optimisation" in January 2018. The term GEO came from an academic paper published in November 2023. Anyone who tells you the two words mean exactly the same thing, or can't tell you the difference, is worth a second question.

Why has search traffic changed even though my rankings haven't dropped?

Rankings holding steady while traffic falls is the single clearest sign that AI answers are intercepting your clicks before they reach a search results page. Around six in ten searches now end without anyone visiting a website at all. The traffic hasn't shifted to a different channel and then come back to you: in sector after sector, the clicks simply stopped rather than migrated, and AI referral traffic itself is still only around 1% of total web traffic, growing by roughly 1% a month. Sector-level organic declines back this up: healthcare, banking and education each saw organic traffic fall by around a quarter to nearly a third during 2025. Waiting for "AI traffic" to arrive and replace what you've lost is not a strategy. Being inside the answer is.

You can usually see this pattern in your own data in about ten minutes. In Google Search Console, compare impressions and clicks for your most important pages over the last twelve months. If impressions are flat or rising and clicks are falling, the searches are still happening and your listing is still being shown. People are simply getting what they need before they reach you. That gap is the size of the problem, and it is a better starting number than any industry average.

Does AI decide who to recommend before it looks for sources?

Increasingly, the evidence points that way. Seer Interactive ran six behavioural tests across 362,188 AI responses, published in March 2026. Their leading hypothesis is that the model chooses which brands to recommend largely from what it already learned in training, then goes looking for sources to support choices it has already made. In Seer's own phrase, the citation is the bibliography, not the brainstorm. Seer is careful to say it can't observe the model's workings directly, so treat this as the best current explanation rather than a proven fact.

This matters because it changes where the effort goes. If the model has already decided who belongs on the shortlist before it searches, then a beautifully structured page on your own site is necessary but not sufficient. You also have to be a business the model has learned about, from sources it trusts, repeatedly and consistently. That is why third-party corroboration appears so often in this guide. You cannot schema your way into being known.

Where do LLMs pull content from on a page?

Position on the page matters more than most site owners assume. Roughly 44% of ChatGPT's citations come from content in the first 30% of a page, meaning the first 60 words often decide whether you're quotable at all. Answer engines are not reading entire pages top to bottom the way a human might skim; they favour content structured with the direct answer near the top, typically within the first third of the page. If your key answer, statistic or definition is buried under three paragraphs of preamble, you are making it harder for the model to lift it cleanly.

The same rule applies below every heading, not just at the top of the page. Retrieval systems tend to break pages into chunks and judge each chunk on its own. A section that opens with a complete, self-contained sentence answering the heading's question can be lifted out and quoted. A section that opens with "Great question" or "Before we answer that, some background" gives the machine nothing to lift.

A quick test anyone can run: take the first sentence under each heading on your most important page and read it on its own, with nothing around it. Does it answer the question in the heading? Would it make sense quoted in someone else's answer, with your name next to it? If not, that is the first rewrite.

Do all the answer engines work the same way?

No, and a strategy built around one of them will behave differently on the others. Broadly:

• Google AI Overviews and AI Mode draw on Google's own index, so your standing in ordinary Google search still feeds directly into them.
• ChatGPT answers some questions from what the model already knows and searches the live web for others, drawing on third-party search indexes and its own crawlers.
• Perplexity searches on almost every query and shows its sources prominently, which makes it the most citation-heavy of the major engines.
• Gemini, Claude and Copilot each combine a trained model with web retrieval in their own way, and each has its own crawler or search partner.

The practical consequence is simple. Test on every engine your buyers actually use, not just the one you use yourself, and expect the results to differ. A business named confidently by Perplexity can be absent from ChatGPT on the same question on the same day.

Does ranking on Google still matter?

Yes, and anyone telling you otherwise is overselling. Ranking well is part of what the machines read: Google's AI surfaces are built on Google's index, and other engines lean on search indexes when they retrieve. What has changed is that ranking is no longer the outcome. It is one input. Ranking well now means winning a bigger share of a shrinking number of clicks, and on its own it doesn't get you named or reach the buyer who never opens a results page.

Our position is that the relationship runs one way. Doing AEO and GEO work properly (answer-shaped pages, consistent facts about your business, corroboration from trusted third parties) tends to produce traditional rankings as a by-product. The old SEO playbook of keyword density, link churn and monthly ranking reports does not produce citations in return. The playbook is dead. Rankings aren't.

Part 3. Optimisation Levers: What Actually Gets You Cited

With a diagnosis in hand, the next question is which specific levers move the needle.

What is "Brand Entity Maturity" and why does it matter for AI citation?

Brand entity maturity is the degree to which an AI model can independently verify that your business is real, established and trustworthy, using signals beyond your own website. Answer engines are built to protect the quality of their answers the same way any search engine always has, because if an engine like ChatGPT started serving unreliable answers, people would stop using it. That means the model is weighing third-party corroboration: press mentions, citations from sites it already trusts, and consistent factual signals about your business across the web, not just what you say about yourself on-site. Unique content that you can be cited for, published in your own voice, is what wins here, rather than generic material that could belong to any competitor.

In practice, a mature entity looks like this:

• One consistent set of facts. Name, address, founding date, people, services and registrations are identical everywhere.
• Named, verifiable people. The founder, directors and experts behind the business have their own profiles that link back to it, and the site links out to them.
• Structured data that ties it together. Organisation and person markup on your site, with links (the "sameAs" property) to your official profiles elsewhere, so the machine can join the dots rather than guess.
• Independent references. Where a business genuinely meets the notability bar, public knowledge bases such as Wikidata can help anchor it as an entity. Where it doesn't, don't force it: a rejected or deleted entry helps nobody.
• Corroboration from outside. Press, trade bodies, directories, podcasts, partners and customers saying the same things about you in their own words.

Should I write conversational FAQ-style content instead of traditional web copy?

Structuring content around the actual questions your buyers ask, and answering them directly, is one of the clearest ways to become citable. Building content that matches a specific question and answer format is described directly as the model of longtail keyword content from the SEO era, adapted for a world where the model needs the answer, not a page to interpret. Put the direct, quotable answer in the first sentence, in the top third of the page, since that is where citation-worthy content is drawn from. Avoid vague or padded openings: an answer engine needs to lift a clean, self-contained statement, not extract meaning from three sentences of throat-clearing.

A pattern that works for each question you answer:

1. The heading is the question, phrased the way a customer would type or say it.
2. The first sentence is the answer, complete on its own, with your subject named rather than "it" or "this".
3. The next few sentences are the evidence: a number with a source, a specific example, the reason it's true.
4. Then the nuance: when the answer changes, the exceptions, what it depends on.
5. Then the next step: what the reader should do or ask next.

This is not the same as bolting a dozen generic FAQs onto the bottom of a service page. A good answer is specific to your business, your area and your experience. "How much does a boiler service cost?" answered with "It depends" helps nobody. Answered with your actual price range, what's included, and why some jobs cost more, it is quotable.

Does structured data and schema actually help?

Structured, well-organised content is part of what separates businesses that get cited from those that don't, alongside trusted third-party corroboration and clean dissemination of that content across channels the models already trust. The deeper layers of this work, building schema, structuring content properly, and getting it corroborated through public relations and third-party citation, are exactly the parts that a large share of the search industry historically skipped in favour of basic on-page tweaks and link building. If your content strategy stops at writing blog posts, you are missing the structural and citation-building work that determines whether a model trusts you enough to name you.

Be realistic about what schema does. It helps a machine read facts about you unambiguously: who you are, where you are, who works for you, what you sell, when a page was written and reviewed. It does not make an unknown business known. The types that usually earn their keep are Organisation (or LocalBusiness and its specific subtypes), Person for your named experts, Article with author and date information, Product or Service where relevant, and FAQ markup on pages that genuinely answer questions. Schema also has to be maintained. A structured data block that still lists staff who left three years ago is worse than none, because it creates exactly the contradiction the machines distrust.

Does digital PR and third-party coverage really influence AI citation?

Third-party corroboration is one of the core inputs models use to decide whether your content is trustworthy enough to cite. Content that is unique, cited elsewhere, and disseminated through channels beyond your own site is positioned as central to becoming "the cited source," as opposed to being one of many similar pages competing for the same answer. This is a slower path than simply publishing more pages, but it compounds: once a business has built up that trusted third-party footprint, competitors publishing similar content later have to work harder to displace it.

The evidence points at mentions more than links. An Ahrefs study of 75,000 brands found a correlation of 0.664 between branded web mentions and appearing in AI Overviews, against 0.218 for backlinks. Correlation isn't proof of cause, but the gap is wide enough to change where you spend. Being talked about, by name, on sites the machines trust, appears to matter more than being linked to.

Corroboration that tends to count:

• Trade and professional body directories and member profiles
• Regional and trade press, especially where you're quoted as the expert rather than advertising
• Podcast and video appearances with transcripts published on the host's site
• Review platforms, with reviews that describe the specific service
• Partner, supplier and client websites that describe what you did for them
• Industry reports and roundups that cite your data or your opinion

The test for any of it is simple: would a sceptical human reading this source believe it was written independently? If it reads like something you paid to place, assume the machine can tell too.

Should I be building a strategy around one platform, like Reddit or a single forum?

No single platform should be the backbone of your citation strategy. One clear example: the share of ChatGPT answers citing Reddit dropped from roughly 60% to 10% within about six weeks in 2025, most likely because of an upstream change to how Google served search results rather than anything Reddit did. It happened again in August 2026, when a separate tracker recorded Reddit's share of ChatGPT citations falling by around 86% in a matter of days. Building a strategy concentrated on any one property carries the same risk that mass forum posting and article spinning carried in the earlier SEO era, right before a major algorithm update wiped that tactic out overnight. Diversify the properties and formats your content lives on, so a single platform policy change doesn't erase your visibility.

The same logic applies to your own channels. LinkedIn and YouTube are useful places to publish, but both keep much of your content behind their own walls and strip or hide the structure a machine would use. Treat your own website as the primary record: publish the full version there first, with its date and author, and let the social versions point back to it.

How much of my content should be AI-generated?

Content generated entirely by AI, without unique input, is a real risk to citation-worthiness, not just a stylistic concern. A large share of content on the web today is estimated to be bot-generated, and models are increasingly built to favour unique, trustworthy content over generic material that offers nothing new to the knowledge base. The recommended approach is unique content, written in your own voice and citing your own expertise, even though it's a slower way of building a footprint than mass-producing AI content. Speed of publishing still matters, since being the first to answer a specific, emerging question is a real advantage, but speed without uniqueness produces exactly the kind of thin content that gets filtered out.

The research term for what the machines reward is information gain: does this page add something that isn't already in the pile? The Princeton paper that introduced GEO in 2023 found that adding statistics, quotations and cited sources improved visibility in generated answers. A later benchmark, C-SEO Bench, found that many of the popular "optimise for AI" rewriting tactics did little in realistic conditions, and that any advantage shrinks as competitors copy them. Read together, the lesson is that tricks decay and substance doesn't.

The most useful question to ask before writing anything: what do you know that nobody (or very few) else knows? Your prices and why they are what they are. The mistakes you see customers make every week. Your own data. The question every new client asks in the first ten minutes. That is what you should be publishing. AI tools are fine for tidying, structuring and speeding up the writing of knowledge you actually have. They cannot supply knowledge you don't.

Does it matter who the content is attributed to?

Yes. Models cite people as readily as brands, and in any subject where trust matters (health, money, law, safety, anything regulated) a named, qualified, verifiable person behind the claim carries much of the weight. Attribute your expert content to a real person, with their role and relevant qualifications stated, a short biography, and links to their professional profiles and any register entry. An anonymous "admin" byline or a generic "the team" gives the machine nothing to check.

In regulated sectors this is also the compliance position. No guarantees of outcome, no superlatives, a clear line between general information and individual advice, and a named professional who has reviewed the content. The same restraint that keeps a regulator happy is what makes a cautious model willing to quote you.

How fresh does content need to be?

Fresh enough that the machine has no reason to doubt it's current. Put a visible "last updated" date on every page that matters, name who reviewed it, and actually revisit it on a schedule. Update the substance, not just the date: changing a timestamp on an unchanged page is the kind of shortcut that erodes trust once noticed. For fast-moving subjects (prices, regulations, technology) a quarterly review is sensible. For stable reference material, once or twice a year is usually enough.

Does llms.txt help?

Not much, on current evidence. llms.txt is a proposed file that lists your most important content for AI systems. Cyrus Shepard's May 2026 review of the published research on AI citation factors found no measurable link between llms.txt and being cited. It doesn't hurt. It isn't a service. If a supplier is charging you to create one and calling that AEO, that tells you something about the rest of their offer.

Should I publish video and podcasts, or just written pages?

Both, with the written version on your own site. Video and audio are excellent for reaching people and building the kind of recognisable expertise that gets you mentioned elsewhere. But a machine answering a question reads text far more readily than it watches video. Publish a cleaned-up transcript of every useful video or podcast on your own website, as a proper page with a heading structure, date and author. It turns an hour of spoken expertise into dozens of answerable, quotable sections, and it keeps the primary record on a property you control.

Which tactics should I stop paying for?

If any of these make up most of what you're buying, it's worth asking what they're actually for:

1. Keyword-density rewrites. Stuffing a phrase into a page doesn't make it a better answer.
2. Link-buying and link churn. Mentions from trusted sources appear to matter more than links, and bought links rarely come from trusted sources.
3. Mass-produced AI pages. Hundreds of near-identical location or topic pages add nothing to the knowledge base and invite being filtered.
4. One-off schema setups. Structured data that is never maintained drifts out of date and starts contradicting you.
5. Single-screenshot reporting. One check proves nothing (see measurement, below).
6. Monthly ranking reports as the headline metric. Rankings are an input now, not the outcome.

How long does it take to see results?

Longer than a technical fix, shorter than most people fear. Access problems, once found, can be fixed in a day, and their effect shows as soon as the crawlers return. Content and entity work typically shows first movement in citation rates at 30 to 60 days, as models and indexes refresh. Meaningful shortlist presence for a core service in a defined area usually takes four to six months of consistent work. Anyone promising faster is either lucky or lying, and you can't tell which until it's your money.

It also compounds. Every answer you publish and every independent source that corroborates it makes the next citation easier to win. Traditional SEO tactics decayed the moment you stopped. This accrues.

Part 5. The Honest Counterweight

Is any of this overhyped?

Some of it, yes, and it's worth being clear about which parts.

AI referral traffic is still around 1% of total web traffic, growing by roughly 1% a month. This is a fast-growing, high-quality, still-small channel, not a replacement for everything you currently do. Anyone telling you otherwise is selling urgency.

Google's public position, set out in its Search Central guidance on AI features, is that you don't need to do anything special for them beyond making good content that follows its existing guidance. Google is half right. The fundamentals it describes (accessible, helpful, trustworthy content) are exactly the fundamentals in this guide. What that position leaves out is the measurement, the off-site corroboration and the structural discipline that separate being eligible from being named.

Nobody can guarantee a citation. The engines change how they choose sources without notice, and single platforms can swing wildly in weeks. What you can do is make your business the clearest, most corroborated and most useful answer available, measure honestly, and keep going. Anyone offering a guarantee is describing hope, not a method.

How do I tell a real AEO supplier from a rebranded SEO retainer?

There's a wave of "AEO" out there that's a traditional SEO retainer with a new sticker on it. The tells are easy to spot:

1. They conflate AEO and GEO and can't tell you the difference.
2. They promise tracking across multiple AI engines but can't produce screenshot evidence against a defined prompt set, or a stated margin of error.
3. They treat schema as a one-off setup.
4. They tell you third-party corroboration doesn't matter, usually because they can't do it.
5. Their main deliverable is still a monthly rankings report.

Ask any supplier those five questions. Then ask one more: what were you doing before this was a category? In a field this young, the honest answer tells you a lot.

Key Takeaways

Getting recommended by ChatGPT or cited in AI Overviews comes down to a handful of concrete, checkable actions:

• Confirm your content isn't blocked by a firewall or bot-management default before doing anything else.
• Put your direct, quotable answer in the first 60 words, structured near the top third of the page.
• Build unique content in your own voice rather than relying on AI-generated volume.
• Invest in third-party corroboration and PR, not just on-site content, to build the trust signals models use to verify your business.
• Diversify across platforms rather than concentrating your citation strategy on one source.
• Measure with defined prompts, sampled repeatedly, against a published margin of error, rather than a single screenshot.

And from the rest of this guide:

• Think in four gates: access, retrieval, extraction, trust. Find which one you're failing before spending on the others.
• The model often decides who to recommend before it looks for sources. Being known matters as much as being well structured.
• Make every fact about your business agree everywhere. Contradictions keep you vague.
• Attribute expertise to named, verifiable people.
• Keep ranking. It's an input now, not the outcome.
• Be honest about the size of it: around 1% of traffic today, high quality and growing.

Who wrote this

I'm Simon Young, founder of Question.Marketing. I've worked in search for thirty years, agency side and client side. I've been publishing on answer engine optimisation since 4 October 2019, 1,153 days before ChatGPT launched. I didn't coin the term (Jason Barnard did, in January 2018), but I started building for it with clients before the market did, and the dated articles are published.

We built the method in this guide on that record: questions before pitches, plain English, and measurement you can check.

Is your business the answer, or invisible?

Ask an AI about your industry tonight, in your customer's words. Does your name come up? Who comes up instead?

If you'd rather have it measured properly, the AI Visibility Audit is £495: your position across the major engines on 25 defined prompts, with screenshots, what the machines currently say about you, who is named in your place and why, and a twelve-month roadmap. Delivered in five working days, in plain English, yours to keep whether you work with us or not.