The short answer
Google's own documentation settles most of this argument. Its page on AI features and your website says: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." It goes further. "You don't need to create new machine readable files, AI text files, or markup to appear in these features." And: "There's also no special schema.org structured data that you need to add."
So most of what gets sold as AEO and GEO is SEO with a new invoice attached. Not all of it. Three things really did change, and none of them is a writing technique.
Three labels, one discipline
| Term | Stands for | What it claims to be | What it is in practice |
|---|---|---|---|
| SEO | Search engine optimization | Getting pages found and ranked | The whole discipline, including AI features |
| AEO | Answer engine optimization | Getting picked as the direct answer | A goal, not a method. No standards body, no paper, no documented origin |
| GEO | Generative engine optimization | Getting cited inside a generated answer | One 2024 academic paper, plus a lot of vendor content that has not read it |
Wikipedia's entry on generative engine optimization notes that no consensus definition separating these terms existed as of early 2026, and that practitioners use them interchangeably. That is an accurate description of the market. Two vendors selling the same product will disagree about which of the two acronyms they sell.
One term has a paper trail and one does not. GEO traces to a named paper with authors, a venue and a public dataset. AEO traces to nothing we could find: no first use anyone cites, no author, no standards document. Every page that claims an origin story for AEO cites another page that claims an origin story for AEO. Absence of a founding document is not proof the idea is empty, but it does tell you which of the two you can check.
What is genuinely new
Citation is a different outcome from ranking
Ranking puts a link in a list. Citation puts your name inside a paragraph the reader may never click through from. These overlap, but they are not the same set, and nothing in your rank tracker tells you which questions name you.
The Pew Research Center measured the click side of this directly. In a study of 900 US adults whose browsing was tracked through March 2025, 68,879 Google searches produced 12,593 with an AI summary. Users clicked a standard result on 8% of those pages, against 15% of pages without a summary. They clicked a link inside the summary itself on 1%.
That is the whole business case for measuring citation separately. The data is US only and seventeen months old, so treat it as a floor rather than a forecast. It also comes from watching real browsers instead of a keyword tracker, with the method published, which is more than can be said for most numbers in this niche.
Crawler control forked into three questions
"Can bots read my site" used to be one question. It is now three, and the answers are set in different places.
| Question | What controls it | What it does not control |
|---|---|---|
| Can my content train a model? | GPTBot, Google-Extended, ClaudeBot, CCBot |
Whether you appear in that company's search product |
| Can my content be retrieved for a live answer? | OAI-SearchBot, Claude-SearchBot, PerplexityBot |
Model training |
| What can Google show from my page? | nosnippet, data-nosnippet, max-snippet, noindex |
Gemini training, which is Google-Extended's job |
Google states the boundary plainly. Its crawler documentation says Google-Extended "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search". It governs training for Gemini Apps and Vertex AI, and grounding in those products. Blocking it does not remove you from AI Overviews.
OpenAI draws the same line in its bots documentation. GPTBot "is used to crawl content that may be used in training our generative AI foundation models". OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features". They are separate robots.txt tokens, and blocking one does nothing to the other. We wrote up the consequences of getting that backwards in does blocking GPTBot remove you from ChatGPT.
One question is now several searches
The same Google documentation describes a "query fan-out" technique used by both AI Overviews and AI Mode. The engine issues several related searches across subtopics and data sources before it writes a response. Google says this lets it surface a wider and more diverse set of links than a single query would.
This breaks a habit rather than a tactic. Rank tracking assumes one query maps to one result set that you can hold still and watch. If one user question quietly becomes several searches behind the scenes, the page that ends up cited may rank for none of the terms you track. The fix is not a new discipline. It is checking the answer itself instead of inferring it from positions.
An answer engine can name you without sending anyone
Referral analytics only sees a visit. A brand mention inside a generated answer with no click is invisible to Google Analytics, to your server logs, and to your rank tracker. This is the one real measurement gap, and it is why the AI visibility tooling category exists at all.
What is not new
Everything else on the standard AEO and GEO checklist is ordinary SEO, and some of it is worse than ordinary.
- Write clear answers near the top of the page. Good advice since featured snippets launched. Not a new discipline.
- Use structured data. Still worth doing for the rich results Google actually still shows, but Google says explicitly that no special schema is needed for AI features. We sorted through which types still earn something in structured data that still matters.
- Add FAQ schema. Google's FAQ rich result stopped appearing in Search on 7 May 2026, and the HowTo rich result was removed in September 2023. Selling either as an AI visibility tactic in 2026 is selling a deprecated feature.
- Publish llms.txt. The proposal is real and reasonable, and several AI labs publish one for their own documentation. Publishing a file and reading a file are different things, and no major AI search product has documented reading it. Our take is in llms.txt, what it is and who reads it.
- Increase entity mentions across the web. This is digital PR with a new noun.
There is one academic paper behind GEO, and it is more interesting than the marketing around it. It tested nine content edits on a benchmark of 10,000 queries, against a generative engine the authors assembled from GPT-3.5-turbo and the top five Google results. Keyword stuffing scored below doing nothing. The paper is worth reading properly rather than quoting the 40% headline, and we will take it apart in a separate post.
What we measured
We ask Google real buyer-intent questions live and read the AI Overview back, including the citation URLs. Bright Data lists that call at $0.0015, which is $1.50 per thousand. That is list price, not a figure we diffed against our own balance.
On a set of commercial buyer-intent questions, every single one returned an AI Overview. A 100% presence rate, against roughly 15 to 20 percent suggested by provider documentation and consistent with Pew's 18%. Query type is almost certainly the reason. Commercial questions phrased as questions are close to the worst case for a publisher and nothing like an average keyword mix.
A nine question scan of stripe.com cost us $0.0135 and found stripe.com cited in 44% of the questions Google answered. Nine questions is nine questions. It is enough to tell you the shape of a result and not enough to report as a rate.
One disclosure: SEOBuilder, our product, asks seven answer engines the same buyer questions, Google AI Overviews and Google AI Mode, ChatGPT, Perplexity, Gemini, Microsoft Copilot and Claude. Anyone claiming to measure these should still be asked how, and at what cost per question, ourselves included: we read the consumer products directly, and Claude through Anthropic's API rather than claude.ai.
What to do this week
- Read your robots.txt against the three questions in the table above. Most sites have accidentally answered one of them and think they answered a different one. The AI crawler checker will show you which tokens you are actually setting.
- Pick ten questions a buyer would type before choosing your category. Not keywords. Questions.
- Check whether Google answers them with an AI Overview and who it names. That is the number to put in front of a client, and the method is in how to measure AI visibility.
- Stop paying for tactics that Google has documented as unnecessary. If a vendor's deck leads with FAQ schema or llms.txt, ask for the primary source.
If you are choosing tooling rather than doing this by hand, we compared what is on the market in the best AI visibility tools of 2026, including what each one actually queries.