AI Answer Engine Optimization: What Actually Changed

The most repeated piece of answer engine optimization advice is to mark up your pages with FAQ structured data so that AI assistants can lift the answers. Google stopped showing the FAQ rich result on 7 May 2026, removed its documentation in June, and the old documentation URL now redirects to a changelog entry that explains why.

Every page currently ranking on the first US page for ai answer engine optimization still recommends some version of that tactic. None of them mentions the deprecation. None of them quotes a single line of Google documentation at all.

That is the state of the category, and it is the reason this article exists. AEO is a real question with a mostly boring answer, wrapped in a vocabulary that changes every six months. This piece separates the three parts: what the term means to the only party publishing guidance, which tactics are aimed at features that no longer exist, and the short list of things that genuinely work differently when a machine reads your page instead of a person.

Answer engine optimization, defined by the search engine

Google addresses the term directly, in a section of its generative AI guide titled "What about 'AEO' and 'GEO'?". The definition is short and it is not flattering to the category.

"AEO" stands for "answer engine optimization" and "GEO" for "generative engine optimization". These are both terms you may see used to describe work specifically focused on improving visibility in AI search experiences. From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

That last clause is the whole argument in seven words. Google is not saying AI search does not matter. It is saying the thing you would do to be visible in it is the thing you were already supposed to be doing, and that a separate discipline with a separate budget line is a marketing construct rather than a technical one.

The same guide defines two mechanics worth knowing by name, because vendors use them loosely. Grounding links are the "prominent, clickable links to relevant web pages that support the information in the response". Query fan-out is "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query". Google's example: a query about fixing a weed-filled lawn fans out into herbicide, chemical-free removal and prevention queries.

Fan-out is where a lot of AEO advice goes wrong. The obvious inference is to build a page for every fan-out variant. Google's spam policies call that scaled content abuse when the pages exist primarily to manipulate rankings, which is the same trap covered in our piece on what an AI SEO agency actually sells.

There is a second first-party document, added on 5 June 2026, on evaluating third-party SEO tools. It names the category explicitly in its list of services to think critically about: "Promising improvements for AI experiences and search formats (also known as 'AEO' or 'GEO' tools)". Google has now written about AEO twice, and both times the framing is about how to evaluate the people selling it.

The tactic that died in May 2026, and is still in every AEO guide

FAQ structured data is the load-bearing recommendation of the AEO vertical. The logic sounds right: answer engines want question-and-answer pairs, FAQPage markup produces question-and-answer pairs, therefore mark up everything. It has not been true on Google for over a year, and the timeline is documented in Google's own changelog.

You do not have to take this from a blog post. The removal is visible in one command, because the old documentation URL is now a redirect:

curl -sIL -o /dev/null -w '%{url_effective}\n' \
  https://developers.google.com/search/docs/appearance/structured-data/faqpage
# https://developers.google.com/search/updates#removing-faq-rich-result

The changelog entry it lands on states the reason: "Removed documentation for the FAQ rich result feature. Why: The FAQ rich result feature is no longer shown in Google Search results."

Two things follow, and they point in opposite directions. Adding FAQPage markup to a page today buys you nothing in Google Search, because there is no feature left for it to make you eligible for. It also costs you nothing, and other consumers of structured data still parse it, so nobody needs to run a removal project.

What is worth changing is the belief underneath. If a proposal treats FAQ markup as the mechanism by which assistants find your answers, the person writing it has not checked a Google source since 2023. That is a cheap and specific way to read an AEO pitch, and it works better than arguing about philosophy. The same test applies to how-to markup, which went the same way three years earlier.

What structured data is still worth doing

Structured data has not stopped mattering. It has stopped being the AEO lever it gets sold as. Google's position on the generative side is one sentence:

Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add. However, it's a good idea to continue using it as part of your overall SEO strategy, as it helps with being eligible for rich results on Google Search.

Read that as a division of labour. Structured data earns you search appearance features. It does not earn you a citation in an AI answer. The types that still map to live features include Article, Organization, LocalBusiness and its subtypes, Product, Review, Event, JobPosting, Recipe, QAPage, Video and ProfilePage. The QAPage documentation is still live, and it is not a replacement for the retired FAQ type: it describes a page where users can submit answers to a single question, not a page where you publish your own answers.

For most sites the honest minimum is small. One Organization block sitewide, one Article block per post, and the specific LocalBusiness subtype if you have a physical location, which is the pattern we work through in the vertical guides such as SEO for restaurants and SEO for dentists. Here is the article block, rendered server side:

// app/blog/[slug]/page.tsx
export default async function Page({ params }) {
  const { slug } = await params
  const post = await getPost(slug)

  const jsonLd = {
    '@context': 'https://schema.org',
    '@type': 'Article',
    headline: post.title,
    datePublished: post.date,
    dateModified: post.updated ?? post.date,
    author: { '@type': 'Organization', name: 'Your company' },
    mainEntityOfPage: `https://example.com/blog/${slug}`,
  }

  return (
    <>
      <script
        type="application/ld+json"
        dangerouslySetInnerHTML={{
          __html: JSON.stringify(jsonLd).replace(/</g, '\\u003c'),
        }}
      />
      <article>{/* ... */}</article>
    </>
  )
}

The replace is not decoration. JSON.stringify does not escape angle brackets, so a title containing markup would inject it into the page. The rest of the Next.js specifics, including why this is a plain script tag and not the next/script component, are in our Next.js SEO guide.

The five things Google says you can ignore

Google's guide carries a section headed "Mythbusting generative AI search: what you don't need to do". It lists five things, and between them they cover most of what is sold as AEO methodology:

• Special files and markup. "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them." A clarification added on 15 June 2026 spells out the consequence for llms.txt files: keeping one "will neither harm nor help your site's visibility or rankings in Google Search".
• Chunking. "There's no requirement to break your content into tiny pieces for AI to better understand it." Google adds that there is no ideal page length.
• Rewriting content for AI. "You don't need to write in a specific way just for generative AI search." The stated reason is that the systems understand synonyms, so missing long-tail variants is not the failure mode it is sold as.
• Chasing mentions. Seeking inauthentic mentions across the web "isn't as helpful as it might seem", because the same ranking and spam systems apply.
• Overfocusing on structured data. Covered above.

We went through this list in detail in our article on LLM SEO, so the point here is narrower: compare it against what the AEO pages ranking today recommend. The tool vendor at position 2 builds its playbook around content structuring and schema. The enterprise platform at position 6 sells extractability. The CRM guide at position 7 frames AEO as a different discipline with different metrics. All three are arguing for things Google's own document tells you to deprioritise, and none of them cites it to disagree with it, which would be a perfectly respectable thing to do.

The content half, in Google's own words

If the technical answer to AEO is "the same as SEO", the content answer is more specific than it first looks, and Google's guide is unusually concrete about it. The instruction is to create "valuable, non-commodity content", and the document does not leave the term abstract. It gives an example of each.

Commodity content, in Google's example, is a post titled "7 Tips for First-Time Homebuyers". The guide's objection is precise: content like that "is often based on common knowledge, which could originate from anyone, and typically adds little unique insight for readers". Non-commodity content, in the same passage, is titled "Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line". One could have been written by anybody. The other could only have been written by someone who did the thing.

That distinction has teeth in an AI answer context. A generative system assembling a response from several sources has no reason to quote the fourth restatement of common knowledge, because the first three already said it. It has a reason to quote a first-hand account, because nothing else on the web contains it. Google states the mechanism directly: "a first-hand review provides a unique perspective based on personal experience, whereas a summary of existing content simply restates information already available elsewhere".

The practical version for a small team is uncomfortable but cheap. The things you know that nobody else does are your own numbers, your own failures, and the specific reason you made a decision that looked wrong. A pricing post built from your actual invoices beats a pricing post built from four competitors' pricing posts, which is the approach we took in what SEO actually costs and in the vertical guides. This is also the one part of AEO that no tool can do for you, and it is the reason "AI writes 100 articles" products produce pages that never get cited.

One more instruction from the same guide is worth quoting because it cuts against the usual advice: "Write content for your human audience". Not for the retriever, not for the fan-out query set, and not in the answer-first template that AEO tooling generates by default.

What genuinely differs: your page gets read by three different machines

Here is the part that is real, and it is not about words at all. Google's guide now links to a web.dev article on agent-friendly sites, and that article describes something no AEO guide in the top 10 mentions: an agent does not see your page the way a browser user does, and it uses up to three representations at once.

Screenshots. The agent renders the page and runs a vision model over it. Useful for grouping and hierarchy, but web.dev notes it is "slow and expensive (in terms of used tokens), making it better as a backup when the structure is confusing".

Raw HTML. The agent reads the DOM: nesting, hierarchy, attributes, and the raw strings. This is how it decides that a button inside a product container belongs to that product.

The accessibility tree. This is the interesting one. It is the browser's own semantic summary of roles, names and states, and web.dev describes it for an agent as "a high-fidelity map that ignores the visual 'noise' of CSS to focus on pure utility". You can look at exactly what an agent would get in Chrome DevTools.

The practical rules that follow are unglamorous and testable:

This is accessibility work with a new commercial justification, and web.dev says so plainly: everything that makes a site agent-ready "also makes sites better for humans". If you want the forward-looking end of it, the same article points at WebMCP, a proposed standard for sites to expose actions to agents, and Google's guide points at the Universal Commerce Protocol for transactional flows. Neither is something to build a 2026 content strategy on. Both are worth knowing exist.

The one thing that beats all of it is still server rendering. If your answer only exists after JavaScript runs, a retriever fetching raw HTML sees an empty shell. Check it the crude way, from a machine that is not your browser:

curl -s https://example.com/your-page | grep -c "a distinctive sentence from your answer"
# 0 means the text is not in the HTML a retriever receives

Our ChatGPT citation article covers this failure in more depth, including why it is the most common reason a page that ranks well is never quoted.

Access, which is the only thing that is truly per-engine

The second genuinely different layer is permission, and it is the one place where "answer engine" is a useful category, because each vendor publishes its own contract. Two examples make the point better than a table of every crawler.

OpenAI documents four agents. OAI-SearchBot is the one that decides whether you can appear in ChatGPT's search results, and OpenAI's guidance is explicit: "Please use OAI-SearchBot in robots.txt for managing Search opt outs and automatic crawl." About the user-initiated agent it says the opposite of what most people assume: "Because these actions are initiated by a user, robots.txt rules may not apply."

Anthropic documents three agents and takes the opposite position on the equivalent case. Its page states that Anthropic's bots "respect 'do not crawl' signals by honoring industry standard directives in robots.txt", Claude-User included. One rule cannot express the same intent to both vendors, which is why a single Disallow line copied from a blog post is a bad idea:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: ClaudeBot
Disallow: /

That file says "index me for answers, do not train on me", which is a defensible position and not the same as blocking everything. The per-crawler detail, including how to verify a crawler is genuinely who it claims to be and how a CDN setting can override your robots.txt, is in our LLM SEO article.

Measuring it without buying an AEO tool

Two days before this article was published, the free option got considerably better. Google's help documentation for the Generative AI performance report in Search Console now opens with a note: "As of August 31, 2026, we've rolled out these insights to all websites worldwide." The staged rollout that made this report unusable for most small sites is over.

Know exactly what it gives you before you build a dashboard on it. It reports impressions, defined as "how many times links to your site were shown to a user in a generative AI feature on Google Search", across AI Overviews and AI Mode, broken down by page, country, device and date. It does not report clicks, click-through rate or queries. Search Labs experiments are excluded. If your property shows nothing, the documented reasons are too few impressions or that the site has been excluded from generative AI features, which is worth checking if anyone applied a nosnippet directive during the 2024 panic about AI training.

For the assistants Google does not report on, the only honest method is sampling, and it is manual work: a frozen set of prompts, run on a schedule, with the citations recorded each time. We wrote the procedure up, including the sample size that makes the number stop bouncing, in how to measure LLM brand visibility yourself.

The third source costs nothing and nobody uses it. Your server logs already record every AI crawler that fetched a page, with a timestamp:

grep -Ei 'OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-SearchBot|PerplexityBot' \
  access.log | awk '{print $1, $7}' | sort | uniq -c | sort -rn | head -20

A page that is never fetched cannot be cited, and that distinction is the difference between a content problem and an access problem. Most sites that believe they have the first have the second.

What nobody selling AEO can promise you

Google's guidance on third-party tools is unusually blunt for a first-party document, and it is the right thing to read before signing anything:

Third-party tools don't have access to our internal ranking data. They can't guarantee performance. Any predictions are their own and like predictions generally, may not happen.

It adds that "Google doesn't evaluate third-party services, so be wary of such claims and those making them", that "using a service or tool doesn't guarantee ranking success", and that whether or not you use a tool, "we strongly encourage using our first-party tool, Google Search Console".

Google's page on hiring an SEO carries the sentence that settles the question of outcomes: "No one can guarantee a #1 ranking on Google." The same page now includes an AEO-specific vetting question, asking whether a provider's advice on "AI experiences (also known as 'AEO' 'GEO' services)" is aligned with Google's official guidance on generative AI features. That is a question you can ask on a sales call and score honestly, and it is worth pairing with the budget arithmetic in what SEO actually costs.

The pattern to watch for is a proposal that is unfalsifiable. Citation counts from a vendor's own panel, share-of-voice percentages with no published method, and visibility scores that only that vendor computes cannot be checked by you or by anyone else. Impressions in Search Console, crawler hits in your logs, and a prompt panel you run yourself can be. Whether any of it moves is not something we or anyone else can promise, and results vary by site, market and starting authority.

What to do this week

Ranked by ratio of effect to effort, for a site that already publishes:

1. Open the Generative AI performance report in Search Console. It is now available to every property, and it takes two minutes to see whether you have any impressions at all.
2. Run the curl | grep check on your three most important pages. If the answer text is not in the raw HTML, fix that before anything else.
3. Grep a week of access logs for the AI user agents. Access problem or content problem, answered in one command.
4. Audit your robots.txt per crawler rather than in one block, and decide search and training separately.
5. Stop treating FAQ markup as an AEO tactic. Leave it in place if it is there, and do not add it to new templates expecting an answer-engine effect.
6. Look at your own pages in the Chrome DevTools accessibility tree. If the interactive elements are unnamed div soup, that is a real defect for both agents and screen readers.

None of this is a new discipline. It is the same technical hygiene that decides whether you get indexed at all, which is the point Google keeps making and the AEO vertical keeps declining to hear. If you are doing your own search work rather than hiring it out, the ordered version of that hygiene is in our DIY SEO guide.

FAQ

What does answer engine optimization mean?

Answer engine optimization is the practice of trying to get your content used and cited by systems that answer questions directly, such as AI Overviews, AI Mode, ChatGPT and Perplexity, rather than only ranking in a list of blue links. Google's own definition adds a qualifier worth keeping: from its perspective, "optimizing for generative AI search is optimizing for the search experience, and thus still SEO".

How is AEO different from SEO?

For Google Search, it is not a different practice. Two things do genuinely differ across the wider set of answer engines. Access is per vendor, since OpenAI, Anthropic, Perplexity and Google each publish their own crawler rules and they do not agree with each other. Measurement is different, because a citation in an assistant answer produces no click and no Search Console query row, so it has to be sampled rather than reported.

How do I optimize my content for AI answer engines?

Make sure the answer exists in the HTML a crawler receives without executing JavaScript, allow the search-oriented crawler of each vendor you want to appear in, answer the question in the first sentence of the section rather than after four paragraphs of preamble, and publish something that is not already available on twenty other sites. Google's guide is direct about that last point: it contrasts commodity content based on common knowledge with content that carries first-hand experience.

Do I still need FAQ schema for answer engines?

No. Google stopped showing the FAQ rich result on 7 May 2026 and removed the documentation in June 2026, so there is no Google Search feature left for the markup to make you eligible for. Existing markup is harmless and other parsers still read it, so there is no reason to run a removal project, but adding it to new templates for AEO reasons is aiming at a feature that no longer exists.

What is the best answer engine optimization tool?

There is no way for any tool, ours included, to see inside a ranking or retrieval system, and Google states plainly that third-party tools "don't have access to our internal ranking data" and "can't guarantee performance". The two sources that are not guesses are free: the Generative AI performance report in Search Console, and your own server logs. A paid tool is worth it when it saves you the manual work of running a prompt panel, not when it sells you a proprietary score.

How long does answer engine optimization take to work?

Nobody can tell you honestly, and results vary. The mechanical parts have their own clocks that you can at least observe: OpenAI notes it can take around 24 hours after a robots.txt change for its systems to adjust, while indexing and re-crawling by any engine happens on a schedule you do not control. Anyone quoting a fixed timeline to a citation is quoting a number they cannot know.