Google AI Mode in 2026: How AI Search Is Changing SEO — and How Websites Stay Visible

AI Overviews generate a summary inside the conventional results page. AI Mode goes further: people can explore a topic conversationally, ask follow-up questions and phrase more complex tasks. Google splits some of those requests into several related searches, retrieves different sources and composes an answer from them. Its documentation calls the technique query fan-out; the foundation remains retrieval-augmented generation, meaning relevant indexed pages fetched from the existing Search index.

A user no longer has to search in sequence for “web agency Paderborn”, “business website cost”, “Svelte or WordPress” and “website maintenance price”. They can ask one fully phrased question instead – for an agency near Paderborn suited to a mid-sized company that cares about performance, custom development and long-term maintenance, at a budget of roughly €6,000 to €10,000.

That changes the competitive picture. A page no longer competes only to rank for one string of words. It needs enough clear, credible and specific information for a search system to treat it as a relevant source across several parts of one decision.

Asked the obvious question – whether SEO is still relevant for generative search – Google answers in three words: “In short, yes!” The reason given is that the generative features remain rooted in the core Search ranking and quality systems.

Google reports for the U.S.: an average AI Mode search there is three times the length of a traditional Search query. More than one in six searches in the U.S. now use voice or images, with image searches growing over 40% month over month in the period analysed. Planning-related AI Mode queries grew 80% faster than AI Mode queries overall in six months, according to Trends data.

On scale, Google names more than a billion monthly active users globally and AI Mode queries that have more than doubled every quarter since launch. These are a vendor’s figures about its own product, not independently audited market data – and the behavioural findings explicitly describe the United States. No comparable official numbers exist for Germany, so we do not extrapolate them here.

The direction still matters. People increasingly no longer need to translate their problem into “Google language”. “SEO agency Bielefeld price” becomes: “We get enough visitors but very few qualified enquiries. Our website is four years old and we sell across Bielefeld and OWL. What should an agency investigate first, and what would a realistic budget be?”

For companies with genuine expertise that is good news. Substance is far easier to demonstrate against a complex question than against a keyword repeated ten times.

The uncomfortable side of generative search is obvious: if the search engine answers a question well enough, the user may never open a website. There is now independent evidence for that effect – with clear limits.

Pew Research Center analysed the real browsing behaviour of 900 U.S. adults in March 2025, covering 68,879 Google searches, of which 12,593 produced an AI summary. On searches with an AI summary, participants clicked a conventional result in 8% of visits; without one, in 15%. Clicks on a link inside the summary itself accounted for 1% of visits.

A quasi-experimental study by Mehrzad Khosravi and Hema Yoganarasimhan used the staggered geographic rollout of AI Overviews and compared 161,382 matched Wikipedia article-language pairs across several language editions. The authors estimate a drop of roughly 15% in daily traffic to the exposed English-language articles, strongest for Culture and substantially weaker for STEM topics. The paper is a preprint and has not been peer reviewed.

Both findings measure AI Overviews, not AI Mode, and mostly U.S. usage. An informational Wikipedia article also behaves differently from the site of a roofing company, a software firm or a web agency where the user eventually needs a quote. “Fewer clicks on informational pages” does not automatically become “fewer enquiries for service providers” – that transfer would be an assumption, not a measurement.

The more useful metric is therefore no longer only “how many impressions did we generate?” but: for which decisions are we considered a relevant source or provider – and what qualified actions follow from that?

AI search has produced an industry of its own acronyms: AEO, GEO, LLMO, AI SEO, generative search optimisation. Many of those terms describe genuine change. The problem starts when they are sold as secret technical levers.

Google is unusually direct about this now. From Google Search’s perspective, optimising for generative search is optimising for the search experience – and thus still SEO. The guide contains a dedicated section listing practices you can ignore for Google Search.

Structured data is a good example of a nuanced statement: it is not required for generative search, and there is no special schema.org markup for it. Google still recommends keeping it as part of the overall SEO strategy because it supports eligibility for rich results. “Not required” is not the same as “useless”.

The distinction that matters most is between two kinds of statement. What a platform writes about its own system is self-reporting. What independent research measures is an observation. Both are useful, but only the first tells you what a provider recommends.

First: publish information that does not already exist everywhere. A generic explanation of web design, solar power or commercial cleaning takes a language model seconds. That is precisely why interchangeable content loses strategic value. Google names unique perspectives and non-commodity content as the lever that will influence visibility in generative search in the long run more than any other suggestion in its guide. Proprietary project data, real case studies, before-and-after comparisons, documented processes, measurements, original photography and named lessons learned are worth more than twenty interchangeable blog posts.

Second: answer buying questions completely. What does it cost, who is it for, what is included, what is explicitly excluded, how long does delivery take, which technology is used, who does the work, how does the project run, what happens after launch, which comparable projects exist? The less of that is published, the more a search system has to assemble from elsewhere. Specificity becomes a competitive asset.

Third: technical crawlability remains the prerequisite. To appear in generative features a page must be indexed and eligible to be shown with a snippet. Clean internal linking, unambiguous canonicals, correct status codes, sitemaps, indexable main content and controlled JavaScript rendering are therefore not legacy work but the entry ticket.

Fourth: think beyond Googlebot. Anyone who also wants to surface in ChatGPT Search should check that OAI-SearchBot can reach the public content they want shown. OpenAI separates this from training explicitly: OAI-SearchBot governs search visibility, while GPTBot can be disallowed separately for potential training use. ChatGPT referrals carry the parameter utm_source=chatgpt.com, so that traffic can be isolated in analytics.

The next shift goes beyond retrieving information. Google now describes agentic systems that perform tasks on behalf of people – comparing products or making a reservation, for example. Browser agents work from three representations of a website: the rendered image as a screenshot, the DOM structure and the accessibility tree.

The accessibility tree is the most interesting of those channels. It is the semantic summary of a page: roles, names and states of the interactive elements, without the visual noise – the same structure assistive technology uses. Google’s guide to agent-friendly websites accordingly recommends semantic HTML, real button and link elements instead of styled divs, form labels connected via the for attribute, stable layouts and no transparent overlays covering interactive elements.

OpenAI arrives at the same conclusion from another direction: “Making your website more accessible helps ChatGPT Agent in Atlas understand it better.” It names ARIA roles, labels and states for buttons, menus and forms.

How much this measurably helps is so far mainly an experimental result. A preprint by Said Elnaffar and Farzad Rashidi compared a conventional and an agent-ready version of the same shop prototype across 300 runs with three browser-agent models – identical catalogue, pricing and checkout flow. Strict success was 89.3% against 49.3%, and the average step count 6.49 instead of 9.31. That is a controlled laboratory experiment on a single e-commerce prototype, not a web standard and not a demonstrated Google ranking factor.

The practical value holds regardless, because the same measures improve usability for humans. Marking prices explicitly, naming buttons, describing variants consistently, building forms semantically, exposing availability clearly, returning understandable error states and not hiding important information behind purely visual interactions all pay off even if no agent ever visits the site.

In June 2026 Google announced dedicated Search Generative AI performance reports in Search Console, with separate views for Search and Discover. The data continues to flow into the overall performance report; what is new is the dedicated view of it.

According to the announcement the reports show impressions – how often URLs from your site appeared in generative features – which pages those were, countries, devices for Search results, and performance over time with hourly, daily, weekly and monthly granularity. The announcement names no separate click metric; Google only says it may add further metrics over time. Anyone promising click counts from generative features today cannot be sourcing them from here.

The rollout initially covered a subset of websites so that Google could test the reports and collect feedback. The view may therefore be missing from a given Search Console property without anything being misconfigured.

More important than the individual feature is the direction: AI search is turning from an assertion into a measurable quantity. In the same context Google warns explicitly about third-party tools that claim to use internal Google metrics – none of them has that access.

I would not publish twenty near-identical pages just because AI Mode might fan a question out into twenty sub-queries. Google assigns exactly that pattern to its scaled content abuse spam policy when the purpose is to manipulate rankings or generative responses.

I would not buy forum mentions. I would not embed invisible blocks of text for language models. I would not fill structured data with claims that are neither visible nor true on the page. I would not sell llms.txt as a guaranteed SEO advantage. And I would not rewrite readable human copy into artificial AI snippets.

Google now names several of these practices itself as things you can ignore for Google Search – among them artificial chunking, unnecessary AI text files and pursuing inauthentic mentions. That is notable, because a considerable share of current AEO and GEO offerings consists of exactly those.

For small and mid-sized businesses the consequence is rather a relief. A local company does not need 100,000 visitors but the right 100. What matters is not whether a page ranks first for the isolated keyword “web agency”, but whether the website, business profile, services and references together provide enough reliable information for a search system to consider that company for a specific situation at all. And once the user lands on the site, a few seconds should be enough to answer: why this company, for which problem, with what experience, at roughly what price, and what the next step is.

The self-assessment below condenses this article into 15 questions. It is deliberately built so honest gaps become visible – only tick an item when it is genuinely true.

The result is a VanThunder practical framework and explicitly not a Google, OpenAI or industry score. Neither Google nor OpenAI publishes a rating of this kind. The test stores nothing: reloading the page resets it.