How Google AI Overviews Are Changing SEO in 2026 — and How to Get Cited

For two decades the goal of SEO was a position on a list. Somebody typed a query, Google returned ten links, and the work was to be as high on that list as possible. That model still exists, but it is no longer the whole picture. Since 2024 Google has rolled out AI Overviews — a generated answer at the top of the results, built from several sources and shown before any link — and expanded them into AI Mode, a conversational results page with no traditional list at all. At the same time a meaningful share of buyers now begin supplier research in ChatGPT, Perplexity or Gemini rather than a search box.

This article is about what that has done to the numbers, what the pages being cited have in common, and the specific work that gets a business into the answer rather than beneath it.

What changed in the numbers

Three effects show up consistently in the data through 2025 and into this year.

• Zero-click searches grew. When the Overview answers the question, many people stop there. Studies across large keyword sets found organic click-through on informational queries falling by roughly 20–35% when an Overview was present, with the steepest drops on definitional and "how to" searches.
• The effect is uneven. Commercial-investigation queries ("best CRM for dental practices"), local queries and anything with a brand name were affected far less. People still click when they intend to buy, compare or contact.
• Cited sources changed who wins. The Overview typically links to three to eight sources. They are usually drawn from pages that already rank on page one, but not always in rank order — a well-structured page at position seven can be cited above position one. Being cited puts you in front of the visitor before any list does.

So the honest summary is: informational traffic is down and will stay down; commercial traffic is intact; and there is a new, separate competition to be one of the sources the answer is built from.

How the citations are chosen

Google has said little in public, but the mechanics are visible from testing and consistent with how retrieval-augmented systems work generally.

The system does not read your page top to bottom. It fans the query out into several sub-questions, retrieves passages — sections, not whole documents — that appear to answer each one, ranks those passages by relevance and by how much it trusts the source, and then writes an answer that quotes or paraphrases the best of them. A page gets cited when one of its sections is the clearest self-contained answer to one of those sub-questions, on a site the system considers a credible entity for the topic.

That has a direct implication: a page can be excellent overall and never be cited, because no single section stands alone. The long, meandering "ultimate guide" that ranked well in 2019 is often the worst format for retrieval.

What cited pages have in common

Looking at pages that are repeatedly cited across our clients' markets, the pattern is consistent enough to act on.

1. The heading is the question, and the first sentence is the answer. Not a preamble, not "in this section we will explore" — the answer, in 40 to 60 words, followed by the detail.
2. Definitions are explicit. "Interaction to Next Paint is…" beats three paragraphs that circle the idea. Answer engines love a sentence they can lift verbatim.
3. Structure carries meaning. Steps are numbered lists. Comparisons are tables. Specifications are bullet points. The structure tells the retriever what kind of answer the passage is.
4. There is original substance. A number the page measured, a decision framework it invented, an example from real work. Restated consensus is rarely the best passage available.
5. A real person wrote it. A named author with a role, a page describing them, and consistent authorship across the topic. Entity trust is assessed at the site and author level, not just the page.
6. The technical layer is clean. Indexable, fast, canonicalised, no duplicate versions of the page competing with each other. A page that is not reliably in the index is not in the retrieval set.
7. The markup agrees with the page. FAQPage, Article and Organization schema that matches the visible content, so a machine can confirm what the passage is and who is speaking.

A practical playbook

Here is the sequence we use. It is unglamorous, and it works.

1. Fix retrievability first

Run the technical audit before touching a word of content: index coverage in Search Console, crawl errors, canonical conflicts, duplicate URLs (trailing slashes, parameters, www and non-www), soft 404s, render-blocking resources, and Core Web Vitals in the field. A site with two indexable versions of every page splits its own authority in half. We see this on a third of the sites we audit — including sites owned by agencies.

2. Restructure the pages that matter

Pick the ten pages that map to the questions your buyers actually ask. Rewrite each into question-led sections with an answer-first paragraph, add a definition block where a term is central, convert prose comparisons into tables and prose processes into numbered steps. Keep the narrative — humans still read — but make each section survive on its own.

3. Add the machine-readable layer

FAQPage schema for the questions you answer, Article schema with a real author, Service schema on service pages, Organization schema with consistent name, logo, contact and sameAs links. Validate it, and make sure it matches what is on the page. Our guide to schema for service businesses covers the six types worth the effort.

4. Become a consistent entity

Answer engines weigh whether the web agrees about who you are. Same business name, same description of what you do and same contact details across your site, your Google Business Profile, LinkedIn, directories and any press. Author pages that list credentials. Mentions on other sites that are not paid for.

5. Measure the right things

Search Console does not yet separate Overview impressions from ordinary ones, so track proxies: impressions and average position for the target queries; referrals from chatgpt.com, perplexity.ai and gemini.google.com in analytics; branded search volume; and — the only number that matters — enquiries by source. Spot-check the target queries monthly and record who is cited.

What to stop doing

• Publishing thin "ultimate guides". Five thousand words of restated consensus is neither read by people nor lifted by machines.
• Blocking crawlers you actually want. Cloudflare's one-click "block AI bots" setting blocks GPTBot, ClaudeBot, PerplexityBot and others. That removes you from those assistants' answers entirely. Note the distinction: Google-Extended controls Gemini training, but AI Overviews are built by the ordinary Googlebot, so blocking Google-Extended does not opt you out of Overviews — and blocking Googlebot removes you from search altogether.
• Optimising for a single ranking. Position one on a query with an Overview may now get fewer clicks than being cited inside the Overview at position five.
• Duplicating your own content. Two domains with the same articles, or one domain with several URLs per page, hands the retriever an excuse to pick a version that is not yours — or none.

Where this is going

AI Mode is a conversational results page: a query, an answer, follow-up questions, and links woven into the text. Agents that research and shortlist suppliers on a user's behalf are already in limited use. In both, the currency is the same — clear, trusted, structured passages from a well-defined entity. The work described above is not a tactic for one Google feature. It is what being findable now means.

Ranking is necessary. Being cited is the prize. The pages that win both are the ones written so that any single section could be quoted on its own, on a site whose technical foundation never gives the retriever a reason to look elsewhere.

This is the core of our SEO and AI search visibility service: fix the foundation, restructure the pages that matter, add the markup, and report on citations and enquiries every month — including the months that disappoint.