SEO in the Age of AI Answers: What Changed, What Didn’t

For about twenty years, SEO had a beautifully simple mental model. Someone types a question. Google returns ten blue links. You want to be one of them, ideally near the top. Everything - keywords, backlinks, technical hygiene - served that single objective.

That model has developed a significant crack.

AI Overviews now appear on roughly 48% of tracked queries, up about 58% year on year. ChatGPT reached around 900 million weekly active users as of early 2026. A meaningful share of searches now end with a synthesised answer rather than a click, and being ranked first matters considerably less if the answer above you already satisfied the question.

The temptation is to declare SEO dead, which is both dramatic and wrong. What’s actually happened is a split: the old game continues, and a second game has started next to it with overlapping but distinct rules.

What genuinely changed

Visibility is now about citation, not just ranking. When an AI system answers a question, it typically cites two to seven sources. That’s the new front page - and it’s a much smaller front page than ten blue links.

Ranking and citation have come apart. This is the finding that should worry anyone still reporting on positions alone. Ahrefs found the overlap between top-10 organic results and AI Overview citations fell from 76% in July 2025 to 38% within eight months. Even more striking: 28.3% of ChatGPT’s most-cited pages have zero organic visibility in Google.

Read that again. More than a quarter of the pages an AI system leans on most heavily don’t rank at all in traditional search. Whatever earns citation, it isn’t purely the same thing that earns rankings.

Structure and completeness now outrank keyword placement. Traditional SEO rewarded keyword targeting and link acquisition. Generative systems reward semantic clarity, contextual completeness and extractability - content that can be understood, chunked and confidently quoted.

Zero-click became the default assumption. A growing share of informational queries resolve without a visit. Traffic can decline while brand visibility increases, which makes single-metric reporting actively misleading.

The measurement stack changed. SEO tracks rankings, impressions and clicks. GEO tracks citations, share of voice inside AI answers, and how your brand is described when it’s mentioned. Most organisations are still measuring only the first set and drawing conclusions about the second.

What emphatically did not change

Before overcorrecting, note what still works exactly as it did.

Technical fundamentals. If a crawler can’t reach, render and parse your page, no system will cite it. Site architecture, crawlability, speed, mobile rendering, clean HTML - all unchanged, and arguably more important, since AI crawlers are less forgiving than Googlebot.

Genuine expertise. Content demonstrating real experience and authority performed well before and performs well now. AI systems are trained to prefer authoritative sources; thin content that scraped by on keyword targeting has simply lost its last hiding place.

Search intent. Understanding what someone actually wants when they type a query remains the whole discipline. Nothing about generative answers changes that.

Brand. People and machines both favour known entities. Building a brand people search for by name is still the most durable SEO strategy available, and it just got more valuable.

Original data. Proprietary research, original surveys and first-hand analysis were always the strongest link magnets. They’re now also the strongest citation magnets, because an AI answering a question needs a source for the number, and there’s only one of you.

What to actually do

Structure content for extraction. Answer the question in the first 40 to 60 words, then elaborate. Use descriptive headings phrased as questions. Keep paragraphs tight. Put definitions, statistics and comparisons in clearly bounded, self-contained blocks. You’re writing for a system that will lift a chunk out of context - make every chunk survive the journey.

Implement structured data seriously. Schema markup for organisations, articles, FAQs, products, people and events gives machines unambiguous facts instead of inference. This is now table stakes.

Build entity clarity. AI systems reason about entities, not strings. Your organisation should be consistently described across your site, Wikipedia if you qualify, industry directories, LinkedIn and press coverage. Inconsistent descriptions produce a fuzzy entity, and fuzzy entities don’t get cited.

Publish original data. The single highest-leverage move available. One annual survey of your industry, published with methodology, will earn more citations than fifty listicles. It is also the one thing a competitor cannot copy.

Chase mentions, not just links. Unlinked brand mentions have been found to predict AI search visibility roughly three times better than backlinks. Digital PR that gets you described accurately on authoritative sites now works twice - once for traditional authority, once for generative visibility.

Measure both games. Add citation tracking to your reporting: which AI platforms mention you, for which queries, alongside whom, and in what terms. Then report traffic and citation separately, because a quarter where traffic dips and citations rise is not a bad quarter, and you’ll need the data to prove it.

Don’t abandon keywords. Traditional search is still enormous and still converts. GEO is additive, not a replacement. Anyone telling you to stop doing SEO is selling something.

How these systems actually pick a source

It helps to understand the mechanics, because they explain why some of the advice above works and why a lot of conventional SEO advice doesn’t transfer.

When a generative system answers a question, it typically breaks the query into sub- questions, retrieves candidate passages for each, and assembles an answer from the passages it can most confidently support. The unit of retrieval is a passage, not a page. This is the single most important structural insight available, and it has three consequences.

Your page competes chunk by chunk. A 3,000-word guide isn’t one entry in the contest - it’s forty passages, each independently retrievable. A page can be cited for one specific paragraph while the rest is ignored entirely. Which means burying your best, most quotable, most specific sentence in paragraph 34 under a heading called “Some Considerations” is a genuine cost.

Ambiguous passages get skipped. A paragraph that begins “As mentioned above, this approach…” is unusable out of context, so it doesn’t get used. Passages that restate their subject - “Core Web Vitals are measured at the 75th percentile of real users” rather than “these are measured at the 75th percentile” - travel intact and get quoted.

Corroboration beats authority. Systems favour claims they can support from multiple independent sources. A figure that appears only on your site is harder to cite confidently than one that appears on yours and three others. This is the mechanical reason brand mentions matter so much: being described consistently in many places makes you a safer thing to say.

None of this requires new tooling. It requires writing paragraphs that make sense on their own - which, conveniently, is also better writing for humans.

The uncomfortable strategic question

Here’s what most organisations haven’t sat down with yet: if a growing share of your audience gets their answer without visiting you, what is your website for?

Increasingly, it’s for two things. It’s the source of truth that AI systems read to describe you - which makes accuracy, structure and completeness commercially important in a way they never were. And it’s the destination for high-intent visitors who’ve already been pre- qualified by an AI answer and arrive further down the funnel than any visitor did five years ago.

That second group is smaller and considerably more valuable. Which suggests the sensible response isn’t to fight for volume that’s leaving anyway - it’s to make sure the machines describe you accurately, and to make sure the fewer, warmer people who do arrive find a page built for someone ready to act.