SEO for AI Search: Visibility in Google and AI Mode

AI is changing more than a single marketing channel. It is changing the way people formulate questions. Instead of a short query, longer and more specific briefs arrive, followed by further questions. For a company that means it is no longer enough to have the keyword on the page. The site has to be clear, credible and technically accessible enough for a search engine or an AI system to retrieve the relevant information from it.

Strong SEO has not disappeared, the emphasis has shifted. Google Search Central documentation answers the question directly: classic SEO applies to generative features, because they run on the same ranking and quality systems as ordinary results. SEO for AI search is therefore not a new discipline, it is the same discipline under greater pressure on source quality. Most of the work a company has already done properly can be kept. There is no shortcut, however.

What has actually changed: the route to the answer, not the rules of the game

The difference shows up in the mechanism. In a classic result a person receives a list of links and compares them personally. The page is the destination. In a generative answer the model assembles text from several sources and the user reads a finished answer. The page turns from a destination into an input.

The first consequence is a longer and more contextual query. Based on its own product data from the United States, Google reports that the average search in AI Mode is three times longer than a classic query, that AI Mode has passed one billion monthly active users globally and that its query volume has more than doubled every quarter since launch. These are figures Google published about its own product, so they indicate a direction of travel rather than independent measurement.

The second consequence: a page is not read as a whole, but as the basis for a single claim. If the answer is scattered across three paragraphs, a PDF and a salesperson's head, the system has nowhere to take it from. The third consequence, that the click stops being the only proof of visibility, is examined below.

There is also a technique Google calls query fan-out: for one query the system issues several related searches in order to surface a wider set of useful links. A single question therefore breaks apart into several partial ones. A page that tries to cover everything then has a smaller surface than a set of pages where each one answers one of them properly.

AI SEO does not start with a hack. It starts with a strong source

The market uses labels such as AEO, GEO and LLM optimisation. Some of them describe genuine changes, some are new packaging on old principles. This is where it pays to be honest even against your own business: Google Search Central documentation states explicitly that you do not need new machine-readable files, AI text files, markup or Markdown to appear in search, and that there are no additional technical requirements for generative features. Anyone selling you a package built on a new file on the server is selling an artefact the documentation has labelled unnecessary. Fixing indexing and writing five pages that answer buying questions is cheaper.

A far less exotic strategy works over the long run: a clear site structure, original expertise, evidence, current information and a technical foundation that can be crawled and indexed reliably.

Google adds a recommendation that reads like editorial policy: do not recycle what others have already said, nor what a generative model would easily produce. It looks for non-commodity content, meaning material with its own expert or experiential point of view. AI systems do not need another thousand identical articles, they need sources with their own data, process, expert judgment or calculation.

Optimise for the decision, not just for the phrase

A content strategy for a company with high-value deals covers the whole buying process. Not only “what is this service”, but also “when is it worth doing”, “what does the wrong solution cost”, “what are the risks”, “how do we compare suppliers” and “what has to be ready before implementation”. Content like that works in classic and in generative search, because it addresses the context the user genuinely needs.

A worked example, meaning a hypothetical situation rather than a real client. A manufacturing company sells a control system in the hundreds of thousands of euros and has three pages on its website: what we do, references and contact. The buying process, however, has five decision questions: return on investment, the fate of the existing machines, responsibility for integration, the length of the shutdown and support after handover. Three pages are three inputs into a fanned-out query, eight precise pages are eight inputs.

A page that answers a decision question costs an expert's time, not a copywriter's. The order is therefore set by a simple criterion: first whatever shortens the sales conversation. If a salesperson explains the same thing twenty times a year, it should be written down. If nobody asks, it should not be a page.

Publish material that cannot be cheaply commoditised

The strongest corporate content comes from where the company has real experience. Publish anonymised project outcomes, decision frameworks, benchmarks, process screenshots or original charts. A generic article titled “10 benefits of AI” is a commodity. An analysis of why you removed three tools from a specific process, reduced the number of hand-offs between people and improved data quality is a source. If the customer does not understand what you are selling, neither does the system that assembles the answer out of it, and that is the subject of our piece on a website visitors do not understand.

Google has published self-assessment questions that work as an editorial checklist: does the text provide original information, research or analysis, does it add substantial value compared with the sources it draws on, and does it demonstrate first-hand experience? As a warning signal it names content that merely summarises what others say. Within E-E-A-T, meaning experience, expertise, authoritativeness and trustworthiness, it stresses that trust matters most. Trust is not a visual property, although it is often mistaken for design, which is why we wrote separately about the fact that a brand is not a logo.

Can we write content with AI? Google does not forbid it, the method and the purpose decide. According to its documentation, generative AI is useful for researching a topic and for giving structure to original content, but generating many pages without added value may breach the spam policy on scaled content abuse, meaning mass production of pages with no benefit. The model is good for structure and speed, not for substance. The same holds as in automation, where AI will not fix a broken process, it only makes it cheaper and faster.

Technical SEO determines whether your content gets considered at all

The entire discussion about AI visibility has an entry condition that Google states unambiguously: to appear in generative features, a page has to be indexed and eligible to appear in results with a snippet, meaning a short extract. If it is not in the index, it is not in the running, and content quality will not compensate for that.

It is worth striking off the list the items that stayed there out of habit. Google states that the keywords meta tag is not used, that a sitemap is not mandatory, and it corrects a widespread myth about headings: semantic order is excellent for screen readers, but from a search perspective it does not matter whether headings are out of order. Heading hierarchy is a question of accessibility, not of ranking. Descriptive URLs and descriptive link text help more.

What appears from your site is governed by the existing controls: nosnippet, data-nosnippet, max-snippet or noindex. Search Console additionally offers a toggle for opting out of generative features. Sites that opt out receive neither traffic nor impressions from them, and the toggle is not used as a ranking signal outside those features. For a publisher this is a legitimate consideration, for a B2B company it is almost always a bad idea.

On a multilingual site the same rules apply as before, only the price of a mistake is higher. Google recommends a separate URL for each language and correct hreflang annotations, meaning the tags that connect the language versions of the same page. It warns against automatic redirects between languages, because they can prevent people and search engines from seeing all versions, and it describes IP address analysis as unreliable.

The economics: fewer clicks for the same visibility

This is the part most presentations leave out. Pew Research Center ran an independent behavioural measurement on a sample of 900 adults in the United States who installed software recording the addresses they visited. In the data for March 2025 there were 68,879 unique Google searches, of which 12,593, roughly 18 per cent, triggered an AI summary. Users clicked a classic result on 8 per cent of visits to pages with an AI summary, compared with 15 per cent of visits without one. They clicked a link inside the AI summary itself on only 1 per cent of such visits.

The numbers deserve caution, since this is an American sample and a single month, so they do not transfer to B2B or to European markets. The mechanism is clear enough, though: the same visibility generates fewer sessions. Metrics built exclusively on traffic therefore start to mislead, and the weight of branded queries, which the generative layer does not filter out, goes up.

A second perspective came from Cloudflare, which measures the crawl-to-refer ratio, meaning how many pages a platform crawls compared with how often it sends a user to the site. In the data for July 2025, Google had 5.4 crawls per referred visitor, OpenAI 1,091.4 and Anthropic 38,065.7. This is a measurement of AI crawlers across sites behind Cloudflare, not the visibility of one particular site. One thing it does say reliably: content is read far more often than it sends people back. A strategy whose only yardstick is the session is measuring a shrinking slice of the effect.

Measure visibility, not marketing promises

Google has extended Search Console with a separate report for generative AI visibility, covering AI Overviews and AI Mode. According to the trade publication Search Engine Land, it was made available to all sites worldwide on 31 August 2026. You need to know what it does not contain, though, otherwise it becomes a source of false confidence. It contains impressions only, meaning the number of times links to your site were shown in a generative feature. It does not contain clicks, click-through rate, position or query data. It can be grouped by pages, countries, dates and devices, a limit of one thousand rows applies, and data from Search Labs experiments is not included.

The practical consequence: presence can be proven, direct attribution cannot. It is therefore worth building a second layer that does not rest on one channel: branded search volume, direct traffic, a “how did you hear about us” question in the form, and query quality judged by the sales team. These are cheap and they work even when a channel stops reporting clicks.

When choosing a supplier, the most useful sentence in Google's documentation is that nobody can guarantee first place in search. The same page lists warning signals that return in a new coat in the AI context: unsolicited email outreach, a lack of transparency about the methods used, promises of top positions and claims of a special relationship with Google. It is worth setting time expectations too: according to Google some changes take effect within a few hours, others take months, and as a rule it is worth waiting several weeks before evaluating.

An advantage will therefore not come out of one prompt or one new shortcut. It will come from your website being, for people and for search systems alike, one of the best places from which the commercially important questions in your segment can be answered. And before the writing starts, it is worth knowing which part is genuinely failing today. Sometimes it is the content, often it is indexing, and surprisingly often it is simply that nobody has written down the answers a salesperson gives every week.

Key takeaways

Verify the entry condition first: a page appears in generative features only if it is indexed and eligible to appear with a snippet. Until that holds, content work has nowhere to take effect.
Do not buy artefacts that Google's documentation labels unnecessary, meaning new machine-readable files or special markup for AI. Put that money into five pages answering the customer's decision questions.
Rebuild the metrics before traffic falls. The Pew Research Center measurement shows 8 per cent of clicks on a classic result for pages with an AI summary, against 15 per cent without one. Sessions on their own are ceasing to be a good yardstick.
Use AI for structure and research, not for substance. Mass production of pages without added value falls under Google's spam policy on scaled content abuse.
Start with a diagnosis, not with buying a package: content, indexing, data and the questions that decide the deal. Without it, every further decision is made blind and usually more expensively.