In May 2026, Google did something the SEO industry had been demanding for two years: it published official guidance on how to show up in AI-generated search results. The page sits in Search Central under a new "Generative AI fundamentals" section, and it is titled Optimizing your website for generative AI features on Google Search.
Here is the awkward part. A large chunk of that document is dedicated to telling site owners that popular "answer engine optimization" tactics do not do anything. Google names them, defines the terminology, and then dismisses several specific techniques that AEO vendors have been selling.
Meanwhile, peer-reviewed academic research has shown that some of those same techniques measurably increase how often a page gets cited in AI answers.
Both things are true. Understanding why is the difference between wasting a year of content budget and actually building visibility. This article walks through what Google confirmed, what it debunked, what the independent research found, and how to reconcile the two into something you can act on.
What answer engine optimization actually means
Answer engine optimization (AEO) is the practice of structuring content so that AI answer systems quote or cite it directly, rather than simply ranking it in a list of links. Generative engine optimization (GEO) is used almost interchangeably. You will also see LLMO, AIO, and "AI SEO" describing the same work.
There is no settled academic distinction between these terms. Practitioners and vendors use them loosely, and often the choice of acronym says more about who is selling the service than about any real difference in method.
The useful distinction is not between the acronyms. It is between two genuinely different types of system:
- Google's generative features — AI Overviews and AI Mode. These run on Google's existing Search index and ranking systems.
- Independent AI assistants — ChatGPT, Claude, Perplexity, Copilot. These have their own retrieval stacks, their own source preferences, and in some cases their own crawlers.
Almost every contradiction in AEO advice dissolves once you separate these two. Google's official guidance is only about Google. It makes no claims about how ChatGPT picks sources, and it would be strange if it did.
What Google confirmed in its official guidance
AI search visibility is still SEO
Google's position is blunt. Its generative features are rooted in the same core ranking and quality systems as regular Search, so from Google's point of view, "optimizing for generative AI search is optimizing for the search experience".
This is not a rhetorical dodge. It has a hard technical consequence that Google spells out: to appear in its generative AI features, a page must already be indexed and eligible to appear in Search with a snippet. If a page is blocked, deindexed, or snippet-suppressed, no amount of AEO tactics will surface it in an AI Overview.
There is also a second, easily missed requirement. A site has to be included in Search generative AI features, which is a setting visible in Search Console. Worth checking before you spend money on anything else.
How Google's AI features actually pick sources
Google describes two mechanisms by name, and both are practically useful to understand.
Retrieval-augmented generation (RAG), which Google also calls grounding. The core ranking systems retrieve relevant, current pages from the index. The model then reviews the specific information in those pages to build a response, with clickable links pointing back to the supporting sources.
Query fan-out. The model generates a set of concurrent related queries to gather more context than the original question provides. Google's own example: a user asking how to fix a lawn full of weeds might trigger fan-out queries about the best lawn herbicides, removing weeds without chemicals, and preventing weeds in the first place.
Fan-out is the single most strategically interesting detail in the whole document, because it tells you that one well-built page answering several adjacent questions can be retrieved multiple times for a single user query. Depth beats fragmentation.
But read Google's warning carefully, because this is exactly where people overcorrect. Creating separate content for every possible fan-out variation, specifically to manipulate rankings or AI responses, falls under Google's scaled content abuse spam policy. The volume play is both against policy and, per Google, ineffective long-term. The legitimate version is one strong page that covers the obvious sub-questions in its structure.
The content standard: non-commodity or nothing
Google's guidance says plainly that creating content people find unique, compelling and useful will influence generative AI visibility more than any other suggestion in the document. It then gives an example pair that is worth memorising, because it is unusually concrete for Google:
| Commodity content | Non-commodity content |
|---|---|
| "7 Tips for First-Time Homebuyers" | "Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line" |
| Based on common knowledge, could have been written by anyone, adds little that a reader could not find elsewhere. | Carries a specific, experienced point of view that goes beyond common knowledge. |
Google explicitly says not to recycle what already exists on the internet, and not to publish what a generative model could easily produce on its own. Since Google's AI systems look across many sources, a distinct viewpoint is what makes a page worth pulling into an answer rather than skipping.
Google's stance on AI-assisted writing is unchanged and worth stating clearly: using AI tools to help create content is not itself a violation. The output still has to meet Search Essentials and the spam policies, particularly on scaled content abuse. Authorship is not the test. Quality and intent are.
What Google says you can ignore
This is the section that should change some people's roadmaps. Google's mythbusting list, stated as things that do not help visibility in Google Search:
| Tactic | Google's verdict |
|---|---|
| llms.txt and similar AI-specific files | Google Search does not use them. Keeping one neither helps nor harms your rankings. Fine to maintain for other systems that do read them. |
| "Chunking" content into small pieces | Not required. Google's systems handle multiple topics on one page and surface the relevant part. There is no ideal page length. |
| Rewriting content specifically for AI | Unnecessary. The systems understand synonyms and intent, so you do not need to capture every long-tail phrasing. |
| Chasing inauthentic mentions | Less useful than it appears. Core ranking focuses on quality while separate systems handle spam, and generative features depend on both. |
| Heavy investment in structured data for AEO | Not required for generative AI search, and there is no special schema to add. Keep using it for rich result eligibility, which is a different benefit. |
Google also added a direct caution about third-party tools that promise ranking success or claim access to internal Google metrics. No external tool has access to Google's ranking or AI systems. The advice is to use tools if they genuinely help your workflow, but to check their claims against official guidance.
So why does the academic research say the opposite?
Here is where most coverage of this topic goes wrong by picking a side.
A widely cited GEO study from researchers at Princeton, Georgia Tech and IIT Delhi, presented at KDD 2024, tested content modifications across a benchmark of roughly 10,000 queries. It reported visibility improvements in generative engine answers of up to 40% from a handful of specific strategies. The strongest performers, measured by position-adjusted word count, were adding relevant quotations, adding concrete statistics, and citing authoritative sources inline.
Those are not fringe tactics. They are close to the opposite of what Google's mythbusting section dismisses.
The reconciliation is straightforward once stated:
- Google's document is about Google. Its AI features inherit a mature ranking system that already evaluates quality, so bolt-on formatting tricks add little on top.
- The research measured generative engines more broadly. Systems with thinner retrieval layers lean harder on surface signals in the text itself.
- The tactics that tested well are good writing anyway. Specific numbers, real quotes and sourced claims make content better for humans. They are not hacks. They are the qualities Google's own non-commodity standard describes, arriving under a different name.
That last point is the one to carry forward. The genuinely effective half of AEO advice overlaps almost completely with writing well and being specific. The ineffective half is the part that involves special files, mechanical restructuring, and manufactured mentions.
Retrieval is not citation, and both are worth tracking
One distinction that materially changes how you measure AI visibility: a source that a system reads while researching an answer is not the same as a source that receives visible credit in the final response. Analysis of ChatGPT citation patterns through 2026 has found retrieval sets widening while visible citation credit concentrates on fewer domains.
Practically, that means being read is easier than being credited, and a single "are we cited?" metric hides most of what is happening. Track these as separate events:
- Retrieval — was your page accessed during the answer?
- Visible citation — did you get a linked credit?
- Brand mention — were you named without a link?
- Recommendation — were you suggested as the option to pick?
- Referral behaviour — what did the resulting visitors actually do?
For Google specifically, there is now a first-party option. Search Console includes a Generative AI performance report covering how content performs in generative features across Search and Discover. Start there before buying anything.
For the independent assistants, the honest answer is that measurement is still immature. The most reliable method remains unglamorous: take your twenty most commercially important questions, ask them in ChatGPT, Claude and Perplexity on a fixed schedule, and log whether you were named, quoted, linked, or absent while a competitor was cited. It is manual. It is also ground truth.
The traffic reality nobody enjoys discussing
Two findings are worth holding together, because they point in opposite directions.
First, AI answers reduce clicks. Pew Research Center analysis found Google AI Overviews lead to significantly fewer click-throughs to websites. HubSpot's 2026 State of Marketing data put 49% of marketers reporting search traffic decline attributable to AI-generated answers.
Second, from the same HubSpot data, 58% described AI referral traffic as high intent.
Fewer visitors, better visitors. If your business model depends on raw pageview volume against display advertising, that trade is bad. If it depends on qualified enquiries, sign-ups or sales, it may be neutral or positive. Work out which one you are before deciding how alarmed to be.
A practical checklist that follows the evidence
Ordered by expected return, based on what Google confirmed and what the research supports.
- Confirm technical eligibility. Page indexed, snippet-eligible, crawlable, site included in generative AI features in Search Console. Nothing else matters until this is true.
- Audit for commodity content. Take your top twenty pages and ask honestly whether each contains anything a reader could not get from the first three competing results. If not, that page is a rewrite candidate, not a promotion candidate.
- Add first-hand specificity. Your own numbers, your own test results, your own mistakes. This satisfies Google's non-commodity standard and the research-backed statistics tactic at the same time.
- Answer the obvious sub-questions on the page. Use fan-out as a structural prompt: what four related things would someone ask next? Cover them with real headings. Do not spin them into thin separate posts.
- Cite your sources inline and link out. Cheap, good for readers, and the one AEO tactic with independent evidence behind it.
- Make key sentences self-contained. A sentence that still makes sense when lifted out of context is a sentence that can be quoted.
- Keep structured data, deprioritise it. Useful for rich results. Not an AEO lever.
- Set up separate measurement for retrieval, citation, mention and referral, rather than one blended score.
- Skip llms.txt for Google. Harmless to keep, pointless to prioritise.
- Consider agent readability. Google's guidance notes that browser agents may read your site through visual renderings, DOM structure and the accessibility tree. Semantic HTML and genuine accessibility work now have a second payoff. Google points to the web.dev guide on agent-friendly site UX, and flags emerging protocols such as the Universal Commerce Protocol for commerce use cases. If you sell things online, this is worth a look — and it connects directly to the broader question of what AI agents can and cannot reliably do.
Frequently asked questions
Is AEO replacing SEO?
No. For Google specifically, AEO is not a separate discipline — Google states that optimizing for its generative features is optimizing for the search experience, which is SEO. For independent AI assistants, some additional tactics show measurable effect, but they sit on top of standard SEO rather than replacing it. Answer engines still need pages that can be crawled and that rank.
Do I need an llms.txt file?
Not for Google Search, which ignores them. Google is explicit that having one will neither help nor harm your visibility there. Some other systems do read such files, so maintaining one is a reasonable low-cost choice — just do not expect it to move Google rankings.
Does structured data help with AEO?
Google says structured data is not required for its generative AI search features and there is no special markup for them. It still recommends using structured data as part of general SEO because it supports eligibility for rich results. Treat it as a standard SEO investment, not an AI-specific one.
What is query fan-out?
Query fan-out is when Google's AI generates several related sub-queries alongside your original question to gather more context before answering. For a question about fixing a weedy lawn, fan-out queries might cover herbicide choices, chemical-free weed removal, and weed prevention. The practical implication is that comprehensive pages can be retrieved for multiple sub-queries at once.
How do I check whether AI search is sending me traffic?
For Google, use the Generative AI performance report in Search Console. For ChatGPT, Claude and Perplexity, ask your priority questions manually on a regular schedule and record whether you were cited, and check referrer data in your analytics. Be sceptical of any tool claiming access to internal Google ranking data, as none has it.
Will AI-assisted content hurt my rankings?
Not automatically. Google's guidance permits AI assistance in content creation, provided the result meets Search Essentials and the spam policies. The risk is not the tool. It is publishing large volumes of commodity material that adds nothing, which falls under scaled content abuse regardless of how it was written.
The short version
Google's official guidance and the academic GEO research are not really in conflict. They are describing different systems, and where they overlap, both point at the same underlying thing: specific, sourced, first-hand content that answers real questions properly.
The tactics that survive scrutiny are the ones that were already good practice. The tactics that Google debunked are mostly the ones that let you feel productive without improving anything a reader would notice. That is a useful filter to apply to any AEO advice you are sold this year, including this article.
If you are thinking about where AI is heading beyond search, two related pieces are worth reading next: our guide to running AI models privately on your own hardware, and our breakdown of why most AI agent projects fail.
Sources
- Google Search Central — Optimizing your website for generative AI features on Google Search
- Google Search Essentials — Spam policies
- Google — Creating helpful, reliable, people-first content
- web.dev — Agent-friendly website best practices
- Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (Princeton, Georgia Tech, IIT Delhi)
- HubSpot, State of Marketing 2026
- Pew Research Center, analysis of Google AI Overviews and click behaviour