Answer Engine Optimization: writing sections that work as answers

Answer Engine Optimization, or AEO, is the work of getting a section of a website used as the direct answer to a specific question. The systems in question are answer engines such as ChatGPT, Perplexity, Google AI Overviews, Google's AI Mode and Copilot.

The difference from classic search engine optimization lies in the unit of work. Search engine optimization works on a page that should appear in a list of results. Answer Engine Optimization works on a section that answers one question completely.

Where the term Answer Engine Optimization comes from

The term predates generative AI search. It can be traced to early 2018, when Google began placing answers above the list of results. A trade article in Search Engine Watch from February 2018 uses it, and a conference talk at BrightonSEO in April 2018 carries it in the title.

One difference from the related term Generative Engine Optimization stands out. GEO comes from an academic paper with a documented method. AEO has nothing comparable. The term grew out of consulting practice and is still used mainly by providers who sell the corresponding services.

Featured snippets as the origin

A featured snippet shows an extract from a website as the answer to a question, above the first organic result. Sitting there earned visibility without holding the top position. A working method grew out of that: identify questions, write short self-contained answers, build the section so it can be lifted out.

The format itself has shrunk sharply. Ahrefs analysed one million randomly drawn search results and published the numbers in July 2025. The share of results carrying a featured snippet fell from 15.41 percent in January 2025 to 5.53 percent in June 2025. Over the same period the share with an AI Overview rose from 3.93 to 27.43 percent. The two curves run almost as mirror images.

The working method survived the decline. What changed is the destination the section travels into.

What became of the voice search story

The second historical driver was voice assistants. The assumption was that spoken queries would soon make up most of search and return exactly one answer each.

That assumption rests on a figure that never existed. AEO guides still claim that ComScore predicted half of all searches would be voice by 2020. ComScore never published this and has confirmed it does not recognise the number. The claim traces back to Andrew Ng, then at Baidu, in 2014. He spoke about images and speech together, referred to the Chinese market, and named a five-year horizon.

A prediction now circulates that is attributed to a company that never made it. The voice search reasoning still appears in AEO guides regardless. It no longer carries the topic.

What an answer engine is

An answer engine answers a question instead of returning a list of possible sources. It reads several sources, assembles a text from them and names some of the sources.

Which systems belong in this group

• ChatGPT with web search enabled
• Perplexity
• Google AI Overviews above the result list
• Google's AI Mode as a separate surface
• Copilot
• Claude with web search
• the classic featured snippet, which needs no language model at all

The list shows why the term is blurry. A featured snippet and a generated ChatGPT answer come about in completely different ways. What connects them is the shape of the result.

What Google says about AEO

One statement belongs here that is missing from most guides on the subject.

On 10 July 2026 Google published a guide to optimizing for generative AI features. It contains a section on the terms AEO and GEO. The answer given is that from Google Search's perspective, optimizing for generative AI search means optimizing for the search experience, and therefore remains SEO.

The same guide explicitly advises against chasing so-called AEO and GEO hacks. It names chunking content into small pieces, creating extra text files such as llms.txt, and pursuing inauthentic mentions.

That is awkward for an article with this title, and it still belongs up front. In our experience teams gain more from knowing the dispute over the term than from managing another discipline.

The useful distinction therefore runs between levels of work. The term is weak. The working method behind it is not.

AEO, GEO and SEO compared

The three levels build on each other. Without indexing, a section is never retrieved. Without a usable section, there is nothing to quote. Without authority beyond your own domain, the brand stays one source among many. The article on Generative Engine Optimization covers the brand level.

The section that can answer

This is the practical part. A section that can answer covers one question completely, without anyone having read the paragraph before it.

The six rules below match what Microsoft recommended in February 2026 for AI analysis in Bing Webmaster Tools: clear headings, tables and question-and-answer sections, evidence for claims, current information, and consistent statements across text and images.

Put the question in the heading

The heading of a section should contain the question the section answers. This sounds obvious and is rarely done.

Not this:

„Cost structure"

Better:

„What does a backlink cost in German-speaking markets?"

The second version names the topic, the question and the scope. It stays understandable when it appears alone in a table of contents or as the single line quoted in an answer.

Put the answer in the first sentence

The first sentence after the heading carries the answer. Reasoning, limits and examples follow.

Not this:

„There is no blanket answer to the cost question, as a range of factors come together that carry different weight depending on the project."

Better:

„An editorial backlink on a German-language trade site usually costs between 200 and 500 euros. The price depends on visibility, topical relevance and placement within the text. The upper end is open, and large consumer publications sit well above it."

The first version announces an answer. The second gives it and explains afterwards.

Make the section readable without its predecessor

A section that opens with „this" or „that method" loses its reference the moment it is lifted out. Write every section so it stands alone.

Not this:

„This only works if the page has already been indexed."

Better:

„An answer engine can only cite a page that is indexed. Google names indexing as the technical minimum requirement for appearing in AI Overviews and AI Mode."

One fact per sentence

Two facts in one sentence are harder to lift out than two sentences carrying one fact each.

Not this:

„We have worked with over a thousand publishers across various languages for several years now, covering smaller niches as well."

Better:

„GetMentioned has more than a thousand publishers connected. They span eight languages. Smaller trade niches are among them."

Lists and tables instead of prose

As soon as several points of equal rank follow one another, they belong in a list. A comparison of two or more quantities belongs in a table. Both forms are easier to lift out than a paragraph with the enumeration built into the sentence.

Not this:

„Points to check are topical relevance, organic visibility, price performance and the pattern of outbound links."

Better:

Four points belong on the checklist before booking:

• topical relevance to your own target page
• organic visibility of the publisher domain
• the relation between price and expected effect
• the pattern of outbound links

Date every figure

A number without a date is hard for an answer engine to place and worthless to the reader. The date belongs in the same sentence.

Not this:

„Most searches now end without a click."

Better:

„Between January and April 2026, around 68 percent of US Google searches ended without a click on a result. SparkToro measured this through a user panel."

Not every question deserves a section. Answering every conceivable question produces a page without a focus.

Where the questions come from

• Search Console, the queries report, filtered for question words
• the questions sales and support actually receive
• the follow-up questions that keep coming in first calls
• the questions competitors answer and your own page does not
• the basic question words, meaning who, what, when, where, why, how and how much

The second and third sources are underrated. They deliver the questions asked by people with buying intent, and those rarely show up in a keyword tool.

One finding from the data helps with the selection. In November 2025 Ahrefs analysed which queries trigger an AI Overview at all. Questions and longer phrasings trigger them far more often than single-word queries. Almost all affected queries carry informational intent. Writing for answer engines therefore means working on fully phrased questions.

Which questions are not worth it

• questions another page on your domain already answers, because the two pages then crowd each other out
• questions whose answer is identical for every industry, because they carry no experience of your own
• questions unrelated to your offering, because a citation there has no effect

Structured data and what is actually established

Barely an AEO guide gets by without recommending schema markup. The evidence on this is better than its reputation, and it says something else.

What the correlation studies show

Pages cited in AI answers carry JSON-LD more often than pages that are not cited. Across six million addresses, Ahrefs finds that 53 percent of cited pages use structured data. An analysis by AirOps covering 16,851 queries measures a 38.5 percent citation rate for pages with JSON-LD against 32.0 percent without.

These numbers show a relationship. They do not show a cause.

What the intervention test found

Ahrefs is so far the only party to have measured the intervention itself, publishing the analysis on 11 May 2026. It compared 1,885 pages that added JSON-LD afterwards with 4,000 matched control pages, over 30 days before and after the change.

The relationship from the correlation studies disappears once the markup is added deliberately. The obvious explanation is that structured data turns up where a website is carefully maintained anyway.

Ahrefs names one limitation itself. Only pages that already had more than a hundred citations were examined. Whether markup helps a page that is invisible so far is a question the study does not answer.

Google and Microsoft say different things

This disagreement belongs in the article because it is almost never mentioned.

In the guide of 10 July 2026 Google writes that structured data is not required for generative AI search and that there is no special schema.org markup you need to add. For rich results in classic search it remains useful.

Microsoft sees it differently. Fabrice Canel, principal product manager at Bing, said at a conference in March 2025 that schema markup helps Microsoft's language models understand content. That statement comes from a conference rather than from official documentation.

In its documentation Microsoft has recommended visible structure since February 2026, meaning headings, tables and question-and-answer sections. Schema.org does not appear there.

OpenAI does not mention schema.org once across its entire developer documentation. Its only recommendation to site owners is to allow OAI-SearchBot in robots.txt and not to block requests from the published IP ranges. The only markup OpenAI explicitly recommends is ARIA, and that serves an agent operating a page rather than citation.

So when a guide claims schema markup is required for AI search, the disagreement is worth knowing. The two largest providers contradict each other publicly on this question. No independent study measuring the effect cleanly exists so far.

llms.txt

The llms.txt file is a 2024 proposal to provide a machine-readable content index for language models. The numbers on it are clear.

In June 2026 Ahrefs analysed 137,210 domains. 28 percent had an llms.txt. 97 percent of those files were never requested once during May 2026. Of the requests that did occur, the larger share came from SEO tools rather than AI systems.

Google Search does not evaluate the file, by its own account. The guide states that creating one neither helps nor harms visibility in Google Search, because Google Search ignores them. The corresponding changelog entry is dated 15 June 2026.

The countermovement inside the same company is worth noting. Lighthouse, the audit tool in Chrome, has carried an llms.txt check under agentic browsing since May 2026. It reports a server error on retrieval and marks a missing file as not applicable, because providing one is optional. Two products from the same company, two positions, both officially documented.

Our assessment is a sober one. An llms.txt costs little and does no harm. As a measure with a predictable effect it does not hold up today.

What AEO does not deliver

No guaranteed citation

OpenAI writes in its own help centre that there is no way to guarantee top placement in ChatGPT Search. No format and no markup obliges a system to use a particular source.

No substitute for authority

A well-built section beats a badly built section of the same authority. Against a source a system trusts more, it does not reliably win.

Honesty is in order here, because we sell placements ourselves. No study with a documented method has yet measured whether backlinks or mentions causally lead to more citations in AI answers. What exists are correlation observations without disclosed samples. Anyone claiming otherwise does not have the evidence.

One related finding is established, and it too is only a correlation. Seer Interactive tracked 42 organisations across 3,119 informational queries. Brands cited in AI Overviews recorded 35 percent more organic clicks than competitors that were not cited. Cited brands tend to be the better known ones, so no proof of cause follows from this.

What can be said is this. A system assembling an answer from several sources needs indications of reliability. Your own website is only one voice among them. With GetMentioned you can compare suitable publishers by SEO data, price and topical fit and book placements directly. That is the level GEO operates on, and it sits beyond what a single section can achieve.

Visibility is not a visit

A mention in an answer produces a visit less often than a position in the result list. SISTRIX measured the effect for the German market in February 2026 across more than 100 million keywords. The click rate of the first position drops from 27 to 11 percent on queries with an AI Overview. Extrapolated, that amounts to roughly 265 million organic clicks per month lost in Germany.

At the same time search volume has not collapsed. Alphabet reported search revenue up 19 percent for the first quarter of 2026 and described the number of queries as an all-time high.

Anyone doing AEO should set expectations accordingly. The value sits in the mention itself, in being perceived as a provider, and in preparing later searches. It sits in the direct click far less often.

Figures from AEO guides that do not hold

Several numbers circulate on this topic that dissolve under checking. We traced the most common ones.

One pattern runs through all six cases. The numbers are not invented. They measure something other than what the sentence quoting them claims, or they are out of date. A simple rule follows for your own work. Every figure in an AEO guide needs a source, a date and a method before it goes into a presentation.

How to measure the effect

Measurement follows the same logic as any work on AI visibility. It consists of a fixed list of questions asked regularly across several systems, and an analysis of whether your brand appears in them and with what statement. The article on AI visibility describes the procedure in detail.

A narrower analysis is enough for AEO. Three questions are useful:

• Does your section appear as the answer, and in which system?
• Is the answer reproduced correctly?
• Which questions does a competitor answer instead?

Sensible checking intervals run to several weeks. Daily checks produce movement without insight, because the same question can draw different answers between two retrievals.

Conclusion: AEO is a writing craft

The term Answer Engine Optimization does not describe a discipline of its own. Google Search now says so itself. What it describes is a level of work, namely the question-and-answer unit in the visible text.

At that level the work is unglamorous and well documented. The question in the heading, the answer in the first sentence, one fact per sentence, lists instead of enumerations in prose, every figure with a date. No markup replaces this, and the only intervention test so far shows that markup does not amplify it either.