GEO after the hype: what actually makes content visible in AI search?

Content becomes visible in AI search through the same properties that made it discoverable before: pages that crawlers can reach, that search systems can index and understand, and that answer a real question with specific, credible information. Google says so directly. Its May 2026 guide to generative AI features states that, from Google Search’s perspective, optimising for generative AI search “is optimizing for the search experience, and thus still SEO”.

What has changed is the shape of the question and the evidence available to you. AI features break one query into many related searches, which favours sites that answer the sub-questions coherently, and they report less than classic search does. Other assistants run their own retrieval and publish little about how they choose sources. Treat GEO as the outcome of a well-structured body of knowledge. This article separates what is established from what is sensible practice, and both from what is still speculation.

Three labels for one problem

Generative engine optimisation (GEO), answer engine optimisation (AEO) and AI search optimisation are overlapping market terms. They are used interchangeably, and none has a settled definition. The mechanism matters more than the label. In this article, AI search means a system that generates an answer and links to its sources, such as Google’s AI Overviews and AI Mode, ChatGPT search or Perplexity. Visibility means your page is retrieved and shown as one of those sources.

Three outcomes are easy to conflate. Being cited as a source, being mentioned by name in an answer, and having your content used to train a model are separate outcomes, and each has its own crawler or setting.

The term GEO comes from research. GEO: Generative Engine Optimization, by Pranjal Aggarwal and colleagues at IIT Delhi and Princeton, was posted to arXiv in November 2023 and accepted to KDD 2024. It framed visibility in generated answers as something content creators could optimise, and in its experiments the strongest methods added citations, quotations and statistics to the source text. The caveats deserve equal weight. The paper measured visibility with its own metrics, on its own benchmark, mainly against a simulated engine that used a language model to answer from top search results, with some validation on Perplexity. Its findings describe that set-up. They offer hypotheses to test against production engines, and they do not measure traffic.

What Google has put in writing

Google is one ecosystem, and it is the one that has documented its position most fully. In May 2026 Google Search Central published Optimizing your website for generative AI features on Google Search, announced in a post by John Mueller. Its central claim is that SEO best practice stays relevant because Google’s generative AI features “are rooted in our core Search ranking and quality systems”. Three points from the guide and its companion documentation shape everything else.

• Eligibility is ordinary eligibility, plus one setting. A page must be indexed and eligible to show in Search with a snippet. The site must also be included in the Search generative AI control in Search Console, under Settings > Search generative AI. The options are Include, Exclude or inherit from a parent property. Include is the default, and Google says the control reached all websites worldwide by 31 August 2026. It covers AI Overviews, AI Mode and generative AI features in Discover. Google says it is not used as a ranking or inclusion signal for the rest of Search, and training is governed separately through Google-Extended.
• Several popular tactics are unnecessary for Google. The guide’s list of things you don’t need to do includes llms.txt files and other “special” markup, “chunking” content, rewriting content for AI systems, seeking inauthentic mentions, and overfocusing on structured data. Google Search ignores llms.txt. There is no special schema for AI features, though structured data remains worth using as part of SEO.
• No outside tool sees inside. The guide states: “No third-party tool has access to our internal ranking or AI systems.”

The fundamentals Google lists on its AI features and your website page are unglamorous. Allow crawling in robots.txt and at the CDN or hosting layer. Make content findable through internal links. Provide a good page experience. Keep important content in text, supported by good images and video where relevant. Make structured data match the visible page.

Google Search has no separate AI layer to optimise for, and that matters for budgets. Spend on AI-specific files or markup does nothing for Google Search, while spend on the crawlability and clarity of important pages serves classic results and AI features at once.

Query fan-out rewards coherent topics

The mechanism most specific to AI search is query fan-out. Google’s guide defines it as “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”. Its documentation describes both AI Overviews and AI Mode searching across subtopics and data sources to build a response.

Take a hypothetical example. Someone asks which commercial heat pump suits an older office building. The response may draw on related searches about sizing, insulation, planning constraints, running costs and maintenance. A manufacturer whose site answers each of those sub-questions on clearly structured, well-linked pages has several chances to be retrieved. A manufacturer with one brochure page and a PDF datasheet has one, and the PDF may hold the most specific facts in the least accessible form.

This is why AI search rewards content architecture as much as individual pages. Topical depth, descriptive headings, internal links between related answers and consistent vocabulary make a site easier to draw from across a set of related questions.

Google draws a line. Making pages for every fan-out query “primarily to manipulate rankings or generative AI responses in Google Search” violates its scaled content abuse spam policy. Before adding a page, check that it answers a question a real customer has, with information the organisation is well placed to give.

Other assistants run their own retrieval

Google’s documentation describes Google. ChatGPT, Perplexity and other assistants retrieve and select sources through their own systems, and publish far less about how that selection works. We will not guess at their ranking, and we would be wary of anyone who claims to know it.

What they do document is access. OpenAI’s crawler overview says OAI-SearchBot is used “to surface websites in search results in ChatGPT’s search features”, and that sites opted out of it will not be shown in ChatGPT search answers. GPTBot is a separate crawler for content that may be used in training, and each setting is independent. Perplexity’s crawler documentation recommends allowing PerplexityBot in robots.txt for a site to appear in its search results, and says that bot is not used to crawl content for foundation models.

So decide deliberately for each crawler. A robots.txt rule or CDN bot setting written to limit training can also remove a site from AI search answers. Checking which bots are allowed, and at which layer, is among the cheapest AI discovery work available.

What you can and cannot measure

Google launched Search generative AI performance reports in Search Console on 3 June 2026 and says they reached all websites worldwide by 31 August 2026. The Generative AI performance report (Search) covers AI Overviews and AI Mode. It shows impressions, meaning how many times links to your site were shown in a generative AI feature, broken down by page, country, device and date. Several links from one site in the same feature count as one impression at property level.

The report shows no clicks and no queries. Clicks from AI features sit inside the Web search type of the main Performance report, mixed in with other search clicks. Google says it may add metrics over time.

Third-party AI visibility tools estimate visibility from outside, including across assistants Google’s reports do not cover. They are useful as directional evidence when the method is understood: which prompts, how often, and from where. Google’s statement that no third-party tool can see its internal systems applies to all of them. We looked at what it takes to turn that kind of visibility data into content decisions in closing the loop from AI visibility to content action.

What cannot yet be reduced to a formula

Several things remain beyond any checklist.

• Source selection is opaque. No assistant publishes how it chooses which retrieved sources to cite, and the systems behind them change on the provider’s schedule. A tactic that appears to work this quarter carries no guarantee for the next. Google’s own documentation notes that indexing and serving aren’t guaranteed.
• Generated answers vary. The same question asked twice can return different answers and different sources, so a single observation says little.
• Attribution is weak. Google reports AI feature impressions without queries or clicks, so connecting an appearance to a visit or an enquiry requires inference.
• Correlation is easy to find. A page that is cited often may also be authoritative, detailed, recent and well linked. Observational studies can show what cited pages have in common, and they struggle to show which property caused the citation.
• The weight of off-site reputation is unknown. What others publish about an organisation plausibly shapes how assistants describe it, and no provider has documented how much. Google advises against seeking inauthentic mentions.

A fall in AI feature impressions can mean a content problem, a change in the engine, a change in demand or a competitor improving.

Where to start

1. Check access for every crawler you care about. Review robots.txt and CDN or firewall bot rules for Googlebot, OAI-SearchBot and PerplexityBot, and decide separately about training crawlers such as GPTBot and Google-Extended. Confirm the Search generative AI setting in Search Console is where you intend it to be.
2. Put important information in text. Specifications, eligibility rules, prices and policies held only in PDFs, images or client-side widgets are harder for search systems to use. Start with the pages that matter commercially.
3. Map questions to topics. List the questions buyers ask before they contact you, group them into topics and sub-questions, and find where the site has no clear answer. Fill the genuine gaps, and resist generating a page for every variant.
4. Make who and what unambiguous. Use consistent names for the organisation, products and people. Show authorship, dates and sources for claims, and keep structured data aligned with the visible page.
5. Set a baseline before changing anything. Record Generative AI report impressions by page alongside Web search clicks, annotate content releases, and treat third-party visibility data as a second, directional view.
6. Compare with competitors from the outside. How rivals structure and evidence the same topics shows where your own content architecture is thin, using only public evidence.

Crawler rules usually sit with engineering or infrastructure, topic architecture with content strategy, structured data with developers, and measurement with SEO or analytics. AI discovery depends on all four, so it needs a named owner who can coordinate across them. Our AI discovery and GEO page sets out how we assess this from publicly observable evidence, comparing a site with its competitors, and where that evidence stops.

GEO after the hype is information architecture done with discipline: reachable pages, clear answers, coherent topics, visible evidence and honest measurement. Each of those pays off in classic search and AI search alike.