AI Overviews and AI Mode in Search
AI Overviews and AI Mode are the two forms in which an AI answer shows up inside Google Search. A whole industry of advice about "optimizing for AI" has grown around them, and almost none of it holds up. This lesson covers what Google documents itself, how an AI answer differs from the classic results, and where in Search Console to find your own numbers for it.
Everything below is checked against the documentation on AI features in Search. That page states the main point in a single sentence: there are no additional requirements and no special optimizations needed to appear in AI Overviews or AI Mode.
The 2026 reality: AI did not kill search — it added a layer on top of it. Market shares of AI assistants and "Nx more traffic" claims circulate without a primary source: neither Google nor OpenAI publishes those figures. What you do have is your own data — Search Console reports visits from AI answers together with the rest of Search.
The big distinction: search engine vs generative engine
To stay un-confused in the new reality, hold this clean separation in your head. A classic search engine (Google, Bing) is an index: it finds and ranks pages that already exist. A generative engine (ChatGPT, Perplexity, Claude) is an interface: it reads several pages on the fly and assembles an answer. AI Overviews are a hybrid — inside Google a generative answer is shown, but it's built from indexed pages.
| Type | What it does | Where the answer comes from | What you need from your site |
|---|---|---|---|
| Search engine (Google) | A list of links | Its own index | Land in the top 10 — the user clicks |
| AI Overviews (Google AI Mode) | Generated answer + source links | From indexed pages, on the fly | Get cited as a source inside the AI block |
| ChatGPT Search | Long answer + citations | Training data + live web | Be an authority the LLM "remembers" |
| Perplexity | Research-style answer with mandatory links | Live web with a freshness focus | Checkable data and links to sources in the text |
| Claude | Long answer, enterprise-leaning | Training data + web tool | Quality expertise, structure |
Where AI answers show up
There are many generative interfaces, and their market shares change faster than the articles about them get published. It is more practical to hold the mechanics in mind than the percentages: where the answer is assembled from your indexed site, and where it comes from what the model "remembers".
| Interface | Where the answer comes from | What that means for your site |
|---|---|---|
| AI Overviews and AI Mode in Google | From indexed pages, on the fly | Ordinary SEO applies: the page must be indexed and eligible to be shown with a snippet |
| AI assistants with web access | Training data plus whatever they fetch at query time | Both indexability and how often others write about you matter |
| AI assistants without web access | Training data only | No direct influence: the data is frozen at the model's training date |
An important caveat from the documentation: AI Overviews are shown only where Google judges them additive to classic Search — and often they do not trigger at all. So the first step with this channel is to check whether an AI answer appears on your queries in the first place. You can do that by hand or with the AI overview analyzer.
How an AI answer differs from the classic results
One difference is documented explicitly and matters more than the rest: the query fans out. Google calls it query fan-out — instead of one search, the system issues several related searches across subtopics and data sources, then assembles a single answer from the results.
What follows from that in practice:
- The answer carries more and more diverse links than a classic result page — the documentation names this an opportunity for more types of sites to appear.
- You can enter the answer through a subtopic, not only through the main query. A page that closes one narrow question inside a big topic gets its own chance.
- AI Mode and AI Overviews may use different models and techniques, so their link sets differ. Check both.
The second difference is the dialogue format. In AI Mode people ask follow-up questions without repeating the previous one. This is the same point the lesson on user behaviour made: prepare answers to the nearest follow-ups alongside the main topic.
How language models work, briefly
To remove the magic: a large language model is a neural network trained on an enormous corpus of text that predicts the next word. It does not "know facts" and does not "understand" in the human sense — it is very good at guessing what comes next from patterns in its training data. The exact sizes of current models are not published by their makers, so figures like "so many trillion parameters" have no source to come from.
Three limitations follow from that design, and they are what shapes the work:
| Limitation | What it changes in practice |
|---|---|
| The model can invent confidently | So the systems lean on retrieved sources and link to them — and that is your site's way in |
| Knowledge ends at the training date | Recent events, prices and terms reach the answer only through a web search — that is, through an indexed page |
| The model reproduces what it saw most | The more others write about you, the better your odds of appearing without any "AI optimization" at all |
What it takes to be a source in an AI answer
This is where the advice industry and the documentation part ways. Google states the requirement like this: the page must be indexed and eligible to be shown in Search with a snippet — and there are no additional technical requirements. The same page says outright that you do not need to create machine-readable files or "AI text files", and that there is no special schema.org structured data to add for AI features.
Which means the work list is the familiar one:
- Crawling is allowed in robots.txt and not blocked by your CDN or host — part of the technical requirements.
- Important content exists in textual form, not only in an image or a script.
- The page is reachable through internal links on your own site.
- Structured data, where you have it, matches the visible text on the page.
- The content is written for people — the helpful content guidance, the same as for classic results.
And one caveat worth memorising: meeting every requirement guarantees neither crawling, nor indexing, nor serving. Nobody has a guarantee here, and anyone promising "placement in AI Overviews" is selling something they do not control.
The opposite job — limiting what is shown — is done with tags: nosnippet, data-nosnippet, max-snippet and noindex are described in the section on snippets. Training and grounding in Google's other systems is controlled by a separate Google-Extended user agent. Know the price: switching off the snippet switches off the classic results too.
What it gives you, and how to measure it
The main fear sounds like "AI is eating my traffic". There is no checkable answer in the form of a general percentage — but there is your own. Google says pages appearing in AI features are counted in the overall search traffic of the performance report, under the "Web" search type. There is no separate "AI Overviews" row — how those impressions and clicks roll into the totals is described in the Search Console help.
So the measurement is indirect, which is ordinary data work:
- Split out the queries where an AI answer appears and the ones where it does not.
- Compare their impressions, clicks and CTR over the same window.
- Finish the picture in analytics: Google notes that clicks from result pages with AI Overviews tend to be higher quality — people spend more time on the site. So look past the click count to what happens after the click.
The AI traffic tracker and the zero-click risk check assemble that comparison for you: they split the queries and show where impressions grow while clicks do not.
GEO and AEO: what is behind the words
Between 2024 and 2026 the industry coined two terms: GEO — generative engine optimization — and AEO — answer engine optimization. They stuck, and separate services are now sold under them.
It is worth knowing these are industry labels rather than separate disciplines with rules of their own: Google requires no separate optimization for AI features. What is useful in the words is a shift of emphasis, not a new toolkit:
| The label | What it actually means |
|---|---|
| SEO | The page ranks and earns a click. Metrics: positions, impressions, clicks, CTR |
| GEO | The same page ends up among the links in an AI answer. There are no separate requirements; what changes is what you count as success |
| AEO | The page gives a direct answer to a specific question. That is long-standing structural work: question, answer, detail |
If a "GEO" service consists of sensible structure, honest data and a named author, you are buying decent content work. If it consists of "secret markup for AI", you are being sold something that is not in the documentation.
Six conclusions you can apply today
1. First check that the page can be shown at all
There is a single condition for eligibility: the page is indexed and not blocked from showing a snippet. Before thinking about "AI optimization", check robots.txt, the noindex and nosnippet tags, and the indexing report. Sites drop out of AI answers for this reason more often than for any other.
2. Answer the question at the top of the page
This is not an AI requirement but a reader requirement, and it happens to make you easier to cite. First paragraph: one sentence of direct answer, then two or three sentences of context. The "let us start from afar" structure loses with humans and machines alike.
3. Cover subtopics, not only the head query
Since the query fans out into subqueries, the site that wins is the one with a page for each meaningful subtopic. One enormous "everything about CRM" article loses to a hub page plus linked material on specific questions. The entity coverage analysis shows the gaps.
4. Back claims with checkable data
A number with a source and a date is worth more than a paragraph of reasoning — to the reader, and to whoever decides whom to cite. The reverse holds too: the invented statistics this topic is full of damage your reputation precisely when someone checks them.
5. Do not buy "secret AI markup"
There is no special structured data for AI features and no special files to create — the documentation says so in plain words. Structured data is useful on its own merits, for rich results, but it is not a pass into an AI answer.
6. Count what happens after the click, not the impressions
On queries with an AI answer some people will not visit your site; that is a given. Which makes the remaining visits more valuable, and those are what to measure: how long someone stayed, whether they reached the goal. Impressions rising while clicks hold flat is not a failure if conversion per click went up.
Practice: look at your site the way an AI answer does
Five steps on your own data:
- In the performance report, take the twenty queries with the most impressions.
- Open each in Google and note whether an AI answer appears. You now have two groups.
- Compare CTR within the groups over one window. If CTR is visibly lower on the AI-answer queries at the same position, you have seen the effect with your own eyes rather than in someone else's statistics.
- For the AI-answer queries, look at who gets cited and how those pages differ from yours: do they answer in the first paragraph, do they have an author and data?
- Run your page through URL Inspection: is it indexed, and is the snippet allowed? That is the only formal requirement.
The glossary covers the terms: AI Overviews, E-E-A-T, SERP, featured snippet. Query types are covered by the intent analysis.
How to tell you are working this channel properly
The sign is not that you get cited — that is not fully in your hands. The sign is that you know your numbers: on what share of your queries an AI answer appears at all, how CTR behaves on those queries, and what happens to the people who do arrive.
Three questions to check yourself. Do you know whether an AI answer appears on your main queries? Have you verified that your pages are even eligible to be shown? Can you tell a CTR drop caused by an AI answer from one caused by position?
What not to expect: a lever that "switches on" citation. It exists neither in the documentation nor in reality — what exists is an indexed page that answers the question honestly.
What comes next
The next lesson is how language models work: more on where the answer comes from and why the model gets things wrong. The previous lesson, on SEO myths, is here.