What Is AEO (Answer Engine Optimization)? The Complete Guide

Answer engine optimization is the practice of making a company retrievable and quotable by AI systems that answer a question directly instead of handing back a list of links. It is a narrower discipline than the name suggests, because most of what decides whether ChatGPT, Perplexity, Claude or Google’s AI Mode names your company happens before any of them read a single word you wrote.

That second sentence is the part people skip. Plenty of teams rewrite their homepage into question-and-answer format, add some schema, wait three months, and see nothing change. The rewrite was not wrong. It was aimed at the wrong stage of a process that has five of them, and four of the other four were already broken.

This guide covers what AEO actually is, what the published evidence supports, what Google says about it in its own documentation (which contradicts a lot of what agencies sell), and how to work out which stage is failing for you. We run this work for B2B SaaS companies, so we have a commercial interest in you concluding that it matters. We have tried to write it so that it is useful even if you conclude the opposite.

Seven surfaces, seven separate crawlers and seven separate retrieval stacks. Almost none of them share an index.

The naming mess, and why it is worth two minutes

Four terms are circulating for roughly the same work. AEO, answer engine optimization. GEO, generative engine optimization. LLMO, large language model optimization. And AI SEO, which usually means whatever the person saying it wants it to mean.

GEO has the strongest claim to being a real term, because it came out of a peer-reviewed paper rather than a pitch deck. Researchers from Princeton and IIT Delhi introduced it at KDD 2024 in GEO: Generative Engine Optimization, along with a benchmark called GEO-bench, and they reported visibility gains of up to 40 percent from content changes alone. We will come back to what that number does and does not mean.

AEO won the popularity contest anyway. We use it because clients use it. The distinction that actually matters is not between the acronyms. It is between two failure modes that look identical from the outside and need completely different fixes.

What actually happens when a buyer asks about your category

Somebody types “best contract management software for mid-market legal teams” into ChatGPT. Here is the sequence, in the order it happens.

The model does not answer from memory. Well, sometimes it does, and that is a separate problem. When a live search is triggered, the system reformulates the question into one or more search queries, runs them against an index, pulls back a set of documents, splits those documents into passages, ranks the passages by how well they match the reformulated query, feeds the top passages into the model’s context window alongside your original question, and generates prose that draws on them.

That architecture has a name. Retrieval-augmented generation, first described by a Meta research team in 2020 in Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. The paper’s framing is still the clearest one available. A model has parametric memory, meaning what it absorbed during training, and non-parametric memory, meaning a searchable index it can consult at question time. AEO is almost entirely about the second one.

Two consequences fall out of this that people find counterintuitive.

First, the unit of retrieval is a passage, not a page. Your 4,000 word pillar post does not compete as one object. It competes as maybe forty separate chunks, and thirty-nine of them can be useless while one carries you. A page can rank nowhere in traditional search and still get cited constantly because one paragraph in the middle happens to answer a specific question cleanly.

Why one paragraph can carry a page that ranks nowhere.

Second, the same question does not reliably produce the same answer. Part of that is retrieval variance, because the index changes. Part of it is sampling. A clinical study published in 2025 gives an unusually concrete illustration: researchers ran identical cases through GPT-4o 250 times each across five temperature settings, 10,000 outputs in total. At temperature 0.0 the model produced an average of 4.5 unique diagnoses per case. At temperature 1.0 it produced 26.25, a 483 percent increase in divergence, with leading-diagnosis accuracy falling from 100 percent to 89.4 percent.

That was medicine, not software procurement. The mechanism transfers. If you check your brand visibility once and write down what you saw, you have recorded a sample, not a measurement. Anyone reporting AI visibility off a single run of a single prompt is reporting noise.

The retrieval chain

Here is the framework we use with clients, and the one this whole section of the site is organised around. Five gates sit between your content and a citation. It is a chain, so it breaks at the weakest link, and the symptom is identical at every gate. You do not appear. That is all you get to observe.

The five gates of the retrieval chain.

The reason this matters commercially is blunt. Gate 4 is the one every content agency sells, because rewriting pages is what content agencies do. In the audits we have run, gate 4 is rarely the binding constraint. Gate 1 fails silently and often, and gate 5 is where the economics of the whole thing live.

Gate	What has to be true	What failure looks like	How to test it in an hour
1. Fetch	The engine’s crawler can request the URL and receive rendered content	Total absence, including for queries where you are the obvious answer	Grep server logs for the AI user agents. Zero hits is your answer
2. Index	The page sits in the corpus that engine queries at answer time	Absent in one engine, present in another, with no pattern you can explain	Ask the same question in four engines and compare which name you at all
3. Retrieve	Your passage ranks for the reformulated query, not the one the buyer typed	You appear for your brand name and vanish for category questions	Search the reformulated phrasing directly rather than the natural question
4. Extract	A passage answers the question standalone, without surrounding context	Competitors get quoted from thinner pages than yours	Read one paragraph in isolation. If it needs the heading above it, it fails
5. Attribute	The engine credits you by name rather than restating your fact anonymously	Your data, numbers or framing appear with no mention of you	Search a distinctive phrase or statistic you originated and see who gets credit

Work the gates in order. There is no point tuning passage structure for a page that returns a 403 to the crawler that matters.

The part where Google says AEO is not a thing

If you are going to spend money on this, you should read what Google publishes about it. Its documentation on AI features and your website is direct in a way that should give any AEO vendor pause.

To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements.

You don’t need to create new machine readable files, AI text files, or markup to appear in these features. There’s also no special schema.org structured data that you need to add.

Google Search Central documentation

Read plainly, that kills a fair chunk of the AEO product category. No special schema. No new files. No separate technical requirements. Google’s position is that its best practices for search still apply to its AI features, and that there is no side door.

We think Google is telling the truth and that the statement is narrower than it appears. It is a statement about Google, made by Google, describing surfaces Google controls. AI Overviews and AI Mode draw on Google’s index. If you already rank, you are already eligible, and inventing a parallel optimisation track for them is mostly theatre.

ChatGPT does not use Google’s index. Neither does Perplexity, and neither does Claude. They run their own crawlers, maintain their own corpora, and apply their own retrieval. Google cannot make claims about eligibility inside systems it does not operate, and it did not try to. The generalisation from “Google needs nothing special” to “no answer engine needs anything special” is one that Google never made and that a lot of people quote it as making.

So the honest version is uncomfortable for everyone. Roughly half of the AEO checklists in circulation are describing work Google has explicitly said is unnecessary for Google. The other half describes crawler access, corpus presence and passage structure across engines that publish far less about how they choose sources, where the work is real and the feedback loop is poor.

What the evidence actually supports

Three findings are solid enough to plan around. All three are public, and none come from an AEO vendor.

Content changes do move citation visibility

The GEO paper is the closest thing to a controlled experiment we have. Same queries, same engines, content varied deliberately, visibility measured. Gains reached 40 percent. Two caveats travel with that number and usually get dropped. It is an upper bound rather than a typical result, and the paper reports that effectiveness varies by domain, which is a polite way of saying the winning tactic in one category loses in another.

Answer summaries suppress clicks, and the numbers are not marginal

The Pew Research Center tracked 900 US adults through 68,879 Google searches during March 2025, of which 12,593 produced an AI summary. Users clicked a result on 8 percent of visits where a summary appeared, against 15 percent where none did. Clicks on the sources cited inside the summary ran at 1 percent. Sessions ended outright 26 percent of the time with a summary present, against 16 percent without.

Clicks fall by roughly half when an AI summary appears. Source: Pew Research Center

Two details from that study get quoted less and matter more for how you write. The median summary ran 67 words, and 88 percent of summaries cited three or more sources. You are competing for one of a handful of slots inside a passage shorter than this paragraph.

Adoption is real, still small, and lopsided by age

The Reuters Institute’s Digital News Report 2026 puts weekly use of AI chatbots for news at 10 percent globally, up from 7 percent a year earlier, across 48 markets at roughly 2,000 respondents each. Under-35s sit at 16 percent. Trust in AI answers runs at 20 percent against 37 percent for news generally.

News consumption is not B2B software research, and the direction of travel is what to take from it rather than the absolute level. Anyone telling you AI search has already replaced Google is selling something. Anyone telling you 10 percent and climbing fast is irrelevant to a five-year plan is also selling something.

Gate 5, and why it is the one nobody budgets for

Attribution is the gate that separates this from ordinary content work, and there is now hard infrastructure data on it. Cloudflare publishes a crawl-to-refer ratio, dividing the HTML requests an AI platform’s crawlers make by the referrals that platform sends back. For the week of 19 to 26 June 2025 it recorded Anthropic at 70,900 to 1. Cloudflare notes the figures likely overstate the real ratios, because native app traffic arrives without referral headers, and says it cannot quantify by how much.

Even discounted heavily, the shape holds. Your content gets read at a volume that bears no relationship to the traffic it returns. Those ratios have improved as the AI products shipped visible citations, which tells you something useful about direction, and they started from a place where the exchange was not really an exchange.

This is why the standard measurement stack quietly stops working, and why so many marketing teams feel like they are going blind without being able to point at what broke.

What you measured before	What happens to it	What to watch instead
Organic sessions from a keyword	Falls while the buying interest behind it holds steady	Citation share on the questions that precede a purchase
Click-through rate from impressions	Becomes uninterpretable, since the answer satisfied the query	Rate at which you are named when your category is asked about
Last-touch attribution	Loses the AI step entirely, because there was no click to log	Self-reported source on inbound forms, asked as a plain question
Rankings for a target term	Stays flat while your presence in answers moves independently	Position inside answers, sampled repeatedly rather than once
Crawl budget and bot hits	Turns into a live signal rather than a technical footnote	Which AI agents fetch you, how often, and what they receive

Gate 1, in detail, because it is usually broken

OpenAI runs four separate crawlers and documents each one. They do different jobs, they respect independent robots.txt directives, and a single blanket rule can produce an outcome nobody intended.

Crawler	Token	What it does	What blocking it costs you
OAI-SearchBot	OAI-SearchBot	Surfaces sites in ChatGPT’s search features	Removal from ChatGPT search results. This is the expensive one
GPTBot	GPTBot	Collects content that may train foundation models	Exclusion from training data. No effect on being cited today
ChatGPT-User	ChatGPT-User	Fetches a page because a user’s question required it	Not an automatic crawler, so robots.txt rules may not apply
OAI-AdsBot	OAI-AdsBot	Checks the safety of pages submitted as ChatGPT ads	Relevant only if you advertise there

The distinction between the first two is the one worth internalising. Plenty of companies decided in 2023 that they did not want their content training somebody else’s model, and blocked what they could find. Some of them blocked search retrieval at the same time, and are now invisible in ChatGPT for reasons that have nothing to do with their content.

If your position is that you want to be findable without being training data, that is a supported configuration and it looks like this.

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

Anthropic and Perplexity publish their own agents and their own controls, and the same logic applies to each. Go and read your live robots.txt before you read any further into this site. We have found blanket AI blocks on companies actively paying for AI visibility consulting, added years ago by someone who has since left.

One more thing lives at gate 1. If your key content only exists after client-side JavaScript runs, assume some retrieval pipelines will see an empty shell. Google renders. Not everything does, and few publish specifics.

Where schema helps, and where it does not

Google said there is no special schema for its AI features, and that is true. Structured data is still worth maintaining, for a reason that has nothing to do with AI features.

The schema.org vocabulary lets you state facts about your company in a form that requires no interpretation. Who you are, what you sell, who founded you, which entity your brand name refers to. That last one carries more weight than most technical SEO work, because a brand name that collides with a common noun or a bigger company in another sector is an ambiguity problem before it is a visibility problem. If a retrieval system cannot tell which entity a name refers to, being quotable does not help.

Keep it accurate and keep it matching what is visible on the page. Do not expect it to be the lever. It is the part of the work with the best documentation and the smallest independent effect, which is exactly why it dominates AEO checklists.

Citations are becoming a controllable surface

One development deserves attention because it changes the ceiling on this work. Answer engines are exposing citation behaviour as a configurable feature rather than an emergent one. Anthropic’s citations documentation describes returning cited sources programmatically, with allow lists and block lists that constrain which domains can be drawn from at all.

Read that as a builder rather than a marketer. Enterprises deploying internal assistants will restrict them to approved domains. Being on that list, or absent from it, is a procurement decision made once and rarely revisited. It will not show up in any visibility tool, and for a B2B vendor it may end up mattering more than public answer share.

Working out whether you have a problem

Five checks, one per gate, in order. None require a tool and all of them can be done this week.

1. Read your robots.txt. Look specifically for AI user agents. If OAI-SearchBot or its equivalents are disallowed, stop here and fix that, because nothing downstream can work.
2. Grep your access logs for AI agents over the last 30 days. Count hits by agent. Zero means gate 1 is closed regardless of what robots.txt says. A healthy number with no citations points you at gate 5.
3. Ask four engines the same ten buying questions. Not your brand name. The questions a buyer asks before they know you exist. Record who gets named. Run it twice on different days, because of the variance problem.
4. Take the page you think should win and read one paragraph out of context. If it only makes sense with the heading above it, it is not extractable, and no amount of schema fixes that.
5. Search a statistic or phrase you originated. If it comes back attached to somebody else or to nobody, you have an attribution problem, and it is the expensive kind.

Whichever check fails first is where your budget belongs. If check 2 comes back at zero, every hour spent on passage structure is wasted.

What we would do first

Sequencing, from an agency that does this for a living, written to be usable without us.

Weeks one and two go to gates 1 and 2, entirely. Crawler access, rendering, log baselines, entity disambiguation. This is unglamorous and it is where the recoverable losses are. We have seen a single robots.txt line account for a company’s complete absence from ChatGPT.

Weeks three and four establish measurement before touching content. Pick your buying questions, run them repeatedly across engines, and record answers over time. Without a baseline you cannot tell a real improvement from sampling noise, and given the divergence figures earlier, that is not a theoretical concern.

Content work starts in month two, once you know which gate is actually binding. Then it is specific rather than general. Particular questions, particular passages, particular gaps you measured rather than guessed at.

Attribution work runs last and continuously, because it is the slowest and least controllable. Original data that other people have to cite. Named frameworks with your name on them. Presence in the third-party sources that answer engines lean on when they hedge.

Reasonable practitioners disagree with parts of this. The Google documentation above can be read as grounds for doing nothing beyond good SEO, and for a company whose buyers all start on Google, that reading is defensible. We think it understates how quickly the non-Google surfaces are compounding, and we would rather be early on something cheap than correct about something expensive.

Frequently asked questions

Is AEO different from SEO?

Partly. The technical foundation is shared, and Google states plainly that its SEO best practices still apply to its AI features. AEO adds work SEO never needed, mainly crawler access decisions across several independent AI agents, passage-level extractability, and attribution, which is the practice of being named as a source rather than quietly absorbed. If your buyers only ever use Google, the overlap is close to complete.

Do I need to add special schema markup to appear in AI answers?

Not for Google. Its documentation states there is no special schema.org structured data required for AI Overviews or AI Mode, and no new machine-readable files. Keep standard structured data accurate for entity clarity and rich results. Treat any vendor selling AEO schema as a signal about the vendor.

How long does AEO take to show results?

Gate 1 fixes can appear within days, because unblocking a crawler that was already trying to reach you is close to instant. Content and attribution work runs on months, and you cannot read it accurately without a repeated-sampling baseline. Any timeline quoted to you that does not first ask which gate is failing is a guess.

Should I block AI crawlers to protect my content?

That is a real strategic choice and it is not obviously wrong, given how lopsided the crawl-to-referral exchange has been. What you should not do is make it by accident. Training crawlers and search crawlers are separately controllable, so blocking training while staying retrievable in ChatGPT search is a supported configuration. Most blanket blocks we find were never intended to cost search visibility.

Why does the same question give different answers each time I check?

Two causes stack. The retrieved documents change as the index changes, and the model samples probabilistically when generating. Published work measuring one model at different sampling temperatures found unique-output counts rising by 483 percent between the most and least deterministic settings. One check is a sample. Report visibility from repeated runs or do not report it.

Does AEO work for small companies, or only for big brands?

Retrieval operates on passages, so a small company with one genuinely excellent answer to a narrow question can be cited ahead of a large one whose answer is buried in a brochure. That is the structural advantage available to you. It disappears in categories where engines default hard to incumbents, and category dynamics are worth checking before you commit a budget.

---