SEO for AI Search in 2026: how to get found in AI Overviews and answer engines

A few years ago 'doing SEO' meant climbing Google's SERP and hoping to enter the top three results. Today, when someone searches online, they no longer get ten blue links: they get an answer. Written by an AI that has read the web on their behalf and decided which sources are worth citing. The others are left watching.

This article is a map for understanding what's happening and, more importantly, what to do to avoid becoming invisible in the era of answer engines.

The leap no one predicted

Numbers tell a story that until a couple of years ago seemed science fiction. In March 2026 nearly half of Google searches, about 48%, trigger an AI Overview, the box at the top of the page with the AI-generated answer. It was 31% in February 2025. Growth has been vertical.

The effect on traffic is exactly what you'd expect. When an AI Overview appears, organic click-through rate plunges: a Pew Research study on 68,000 real queries measured a 46.7% relative decline in clicks. Ahrefs, on an even larger sample, talks of 58% reduction in early 2026. In practice, your organic position is worth much less than before, even if you're first.

64% of Google searches now end without any click. Zero-click is not new, but AI has structurally accelerated it. Gartner estimates that by end of 2026 a quarter of organic search traffic will shift toward chatbots and voice assistants.

But there's a second truth, less told. Brands that get cited inside AI Overviews get on average 35% more organic clicks and 91% more ad clicks compared to non-cited brands. Total traffic drops, but the share that remains concentrates on sources the AI chose. The game is no longer 'being first': it's being the cited source.

SEO, GEO, AEO: not just an acronym change

Two new acronyms have appeared lately: GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization). The line between them is fine, but the core idea is the same: optimise a site so an AI cites it when answering a user.

The difference from traditional SEO isn't cosmetic. A classic search engine orders ten results, a language model synthesises one. And there's an extra complication: LLMs are non-deterministic. If you ask ChatGPT the same question five times you get five slightly different answers, with different combinations of sources. There's no 'first position' like on Google. There's a citation frequency: how often your brand appears when AI answers questions in your sector.

This changes the whole way of measuring success. The KPI is no longer keyword ranking, but share of voice in AI answers: in how many of the possible outputs does your brand appear?

The good news is classic SEO doesn't die: it remains the foundation. Studies show that between 40% and 76% of sources cited in AI Overviews come from Google's organic top 10. Translation: if you don't rank, no one cites you. SEO is the necessary condition; GEO is the sufficient one.

Technical basics: make sure AI can read you

Before worrying about what you write, you must verify AI crawlers can access your site. Sounds trivial, but it's the most common problem. Many sites block GPTBot, ClaudeBot, PerplexityBot or GoogleOther in robots.txt, often without knowing it, because some Cloudflare default settings do it automatically. Check the file. If you want to be cited, you must allow reading.

The second point is schema markup. AIs don't read your HTML like a person: they look for structured data that remove ambiguity. Types like Organization, Article, FAQPage, HowTo and BreadcrumbList in JSON-LD format are the grammar with which a model understands that 'Acme Ltd' is a company, not just a word, and that that block is a frequently asked question, not a random paragraph. A BrightEdge study found that pages with structured data get 30% more clicks in traditional results; for AI search the effect is even more marked, because structured data is raw material ready to extract.

Then there's llms.txt, an emerging standard on the model of robots.txt designed to declare site structure in LLM-readable form. By early 2026 about 844,000 sites had adopted it. Honesty here: the concrete impact on citations isn't yet proven, and several analysts recommend adopting it more for its low cost than for proof of effectiveness. The pragmatic advice is simple: do it, it takes half an hour. But don't consider it your GEO strategy.

Finally server-side rendering. If your site depends entirely on client-side JavaScript, many AI crawlers won't see the content. SSR or SSG is the minimum requirement to be indexable by generative models.

How to write content AI chooses to cite

Here we enter the editorial part, and it's where most companies go wrong. AIs don't reward elegant prose or keyword-stuffed SEO: they reward content that can be extracted.

In practice this means three things. The first is the direct answer in the first 40-60 words of each section. The inverted pyramid format, with the answer first and explanation later, isn't a journalistic fad, it's how models look for the chunk to cite. Not coincidentally, a recent analysis showed that 44% of LLM citations come from the first 30% of text: those who put key information at the end simply aren't read.

The second is factual density. Data, statistics, citable numbers with verifiable sources every 150-200 words. A vague article, even if well-written, gets ignored. An article full of numbers with attribution becomes raw material for AI.

The third is demonstrated authority: the famous Google E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness). Author bio with real credentials, concrete case studies, first-hand data. Models are getting very good at distinguishing expert content from content written 'based on general knowledge'. The first gets cited, the second doesn't.

At site architecture level, topical authority also matters: a single isolated article isn't enough. Models reward sites that cover a topic in depth, with pillar pages and clusters of linked articles. If you talk about 'business automation' you must have ten coherent contents on that world, not just one. It's the signal AI uses to understand if you're an authoritative source or a generalist who wrote one article by chance.

And then freshness: content from 2024 not updated progressively loses ground against a 2026 article on the same topic. A visible 'last updated' date and a quarterly refresh cycle on key contents make a concrete difference, especially on Perplexity which weights recency more than others.

Each engine has its tastes

Something many ignore: ChatGPT, Perplexity and Google AI Overviews don't work the same way.

Google AI Overviews tends to pick from those who already rank well organically, so classic SEO counts a lot. ChatGPT Search rewards encyclopaedic, well-structured content with high overall domain authority. Perplexity is the most 'journalistic': it gives enormous weight to freshness and source transparency, so much that it's often used as a second-reading engine precisely to verify others' citations.

In practice this means content designed to be cited everywhere must have three qualities together: source authority (for Google), clear structure and completeness (for ChatGPT), constant updating with verifiable citations (for Perplexity). It's not a technical trick: it's a different way of thinking about what you publish.

The new measurement dashboard

If you're still measuring only keyword positions, you're looking at the wrong phenomenon. In 2026 the measurement dashboard must include three things. The first is how often your brand gets cited in AI answers, monitorable with tools like AthenaHQ, Kalicube and SGE Checker. The second is referral traffic from AI domains (ChatGPT, Perplexity, Copilot and Gemini now regularly appear in Google Analytics referrals). The third is conversion rate of that traffic.

This last point is crucial and often overturns the apocalyptic narrative about CTR decline. Traffic arriving from an AI Overview or ChatGPT converts much better than classic organic: those who click after reading an AI summary have already done part of the search and want to deepen, they're not exploring. The 'lost' traffic was largely informational traffic that didn't convert. What remains is more qualified.

Where to start, concretely

If you manage a business site and are wondering what to do tomorrow morning, the priority order is fairly clear.

Start with a technical audit: check robots.txt, verify AI crawlers aren't blocked, add (or review) schema markup on key pages, install llms.txt. These are interventions taking days, not months, producing effects within a few weeks.

Then look at existing content with new eyes. Do your pillar pages start with a direct answer in the first paragraphs? Do they have citable data? Do they show who's behind it, with verifiable name and skills? If the answer is no, you've found where to start rewriting.

Finally, rebuild your measurement system. Add to classic tools the monitoring of AI citations and referral traffic from generative engines, and start reasoning in terms of share of answer instead of ranking. It's the KPI of the next three years.

The real question

SEO for AI search isn't a passing trend to chase with some CMS tweaks. It's a restructuring of how the web is discovered, read and used. Those moving now, even without huge budgets, simply putting order in technical basics and the way they write, are building an advantage that in two years will be hard to bridge.

The web isn't disappearing behind AIs: it's reorganising around them. The difference between who gets cited and who doesn't is being decided now.