In a meeting a few months ago, a marketing director turned their laptop towards me. On the screen sat the most valuable question in their category, and above the results sat a grey box: an answer written by Google, with three source links beside it. All three were competitors. Their own page sat just below the box, second in the organic results. They had won the ranking and lost the answer. I invented the meeting for this article; I did not invent the screen — we have watched the same scene play out across categories over the past year.
Google AI Overviews is the name of that box above the results, and it now appears regularly on Turkish queries too. This article looks at that box alone: what it is, how it works as far as Google has explained it, which queries trigger it, what it does to your clicks, and what the pages inside it have in common. The wider picture of AI search — ChatGPT, Perplexity, the Gemini app — sits in a separate guide; the general framework lives there.
What is an AI Overview?
An AI Overview is the summarised answer Google shows at the very top of its search results — assembled from several sources and written by a language model. The single box is called an AI Overview; the feature as a whole is Google AI Overviews, and links to the pages feeding the answer sit inside or beside the box. The user reads a paragraph before reaching ten blue links, and usually finds what they came for right there.
Don't confuse it with the featured snippet you already know. A featured snippet quotes one passage from one page verbatim, and you can find those exact sentences on the site. An AI Overview reads several pages, synthesises them and writes new text. The difference looks small and isn't: a snippet has one slot you either win or lose, while an AI Overview carries three or four, and being named means getting onto that short list.
How does Google AI Overviews work?
Google has said that a version of Gemini customised for Search sits behind AI Overviews. The system doesn't treat your question as one search: it breaks the question into sub-questions, runs several searches at once and merges what comes back into a single answer — Google calls this technique query fan-out.
A second point in Google's documentation matters just as much: the AI features inside Search use the same core ranking systems, and the links come from the same index. No separate "AI Overviews ranking" has been published. That comes down to two sentences in practice: if Googlebot can't read your page you have no chance in the box, and if it can, you still hold no automatic claim to a place in it.
What Google has not explained is why a given page gets picked. Everything past that point is observation: ask the same query on different days and the sources change, the same query produces a box on desktop and none on mobile, and sometimes the box disappears altogether. Treat anyone making firm claims here with caution — what we hold is a pattern, not an algorithm.
Which queries trigger an AI Overview, and which don't?
An AI Overview does not appear for every query, and that is the most useful input to your content plan. The observed pattern: questions asking for an explanation, a comparison or several steps tend to trigger the box, while brand searches, navigational queries and short questions with one definite answer usually don't.
- Definition and explanation questions: anything opening with "what is", "how do I", "why does".
- Comparisons: "X or Y", "what is the difference between", "which one is better".
- Long, constrained questions: the ones carrying budget, timeline, dimension and use case in a single sentence.
- Sensitive fields run the other way: health, finance and legal queries produce the box less often — Google says it is deliberately more cautious there.
- Brand and navigational searches: a query shaped like "brand name + login" rarely produces a box, because the user already knows the address they want.
The effect on a content plan is direct. The commercial keyword that pays your bills probably never produces a box; the questions around it do. Write down the ten questions your buyer asks before buying — that list, not your product page, is where the work of appearing in Google AI Overviews actually happens.
Where does AI Overviews stand in Turkey?
AI Overviews appear regularly on Turkish queries, though coverage is narrower than in English. Google announced the feature in May 2024, opened it in the United States first, then said it had extended it to more than a hundred countries and a long list of languages, Turkish among them. On the ground the gap shows: ask the same question in both languages and the English side produces a box more often, naming more sources inside it.
Our own Search Console data carries an interesting signal here. Over the three months to August 2026 the query "google yapay zeka optimizasyonu" brought our site 32 impressions, "yapay zeka arama optimizasyonu" 99 and "yapay zeka optimizasyonu" 136. All three are small numbers. What they share is more telling: the Turkish results page still answers those questions poorly — nobody has written a proper answer to them. AI Overviews in Turkey is not a saturated field yet; the window that started closing on the English side years ago is still open here.
What happens to your clicks when an AI Overview appears?
The short answer: impressions hold and clicks fall. When the user finds the answer inside the box they don't scroll down, so your click-through rate drops on queries that produce one — this is exactly what the term zero-click describes. How much you lose depends on query type: informational questions lose the most, purchase and brand queries the least.
You can measure this in your own data, but only indirectly. Google folds AI Overview impressions and clicks into the general Search performance report in Search Console and gives no separate breakdown; there is no screen telling you how often you appeared inside a box. What you can do: list by hand the queries you have confirmed produce one, then compare their click-through rate against the same period six months earlier. If the curve points down while your position holds, the box is the likeliest explanation.
Let's be honest: no tactic fully replaces what you lose. What is true is narrower — stay out of the box and you lose both the click and the mention; get in and you keep the mention. A brand name inside the answer is harder to measure than a click, yet it travels all the way to the purchase: buyers arriving at a meeting often can't say where they heard of you, but they remember the name.
Which conditions are known to get you named inside the answer?
Google's official position is plain: there is no separate optimisation method for AI Overviews, and the basic rules of Search apply. What that translates to in the field is five properties shared by the pages that end up inside the box — none of them a guarantee, all of them a shift in the odds.
- Indexability. Googlebot has to fetch the page and see the text in the HTML the server sends. Google states plainly that the noindex, nosnippet and max-snippet directives apply to AI Overviews as well.
- Quotable paragraphs. Each paragraph has to stand on its own when cut out of the page; a paragraph leaning on the sentence above it for context cannot be quoted.
- Question-and-answer structure. The heading should be the sentence your customer says, and the first two beneath it should answer that question plainly. The story comes later.
- A freshness signal. Publication and revision dates should be visible, and revised content should say what changed. Stale pages are rare among those named inside the box.
- Verifiable proof. Numbers, dates, measurements and sources belong inside the sentence; the model has nowhere else to check them.
Nobody can tell you that ticking this list gets you into the box. The accurate sentence is this: there is no guarantee, but the probability can be raised systematically. If five competitors compete for the same query and all five fail on the second condition, your advantage comes from a difference in form, not in technology.
When is a paragraph actually quotable?
A paragraph is quotable when it still makes sense after being cut out of the page. The test takes forty seconds: lift the paragraph out of its context and read it; if it collapses because of back-references like "this", "that" or "as we explained above", the machine can't use it either.
A concrete example. A manufacturer's product page carried this sentence: "Thanks to the advantages described above, this series is one of the most preferred in the sector." On its own it says nothing — which series, which advantages, which sector, how preferred. We rewrote the same information as: "The solvent-based series dries faster than water-based alternatives in flexible packaging print, which is why it is chosen on lines running above 300 metres per hour." The second sentence still teaches something after being cut out. The gap between the two is form, not knowledge.
The second rule is about placement. Put the answer directly under the heading; warm-up sentences and "in this article we will cover" openings stall the reader and the model alike. Don't delete the story — move it further down.
Structured data and indexability: the ground a machine reads
Structured data is not a ticket into AI Overviews; Google states plainly that no extra schema is required. Article and FAQPage markup still make the text easier for a machine to parse, and the cost is close to zero — it amounts to printing the JSON-LD equivalent of fields you already fill in when publishing a page.
Set the expectation correctly: Google restricted FAQ rich results for most sites in 2023, so you will not see a badge on the results page. You add the markup to declare the page's structure to a machine, not to win pixels. That is the call we made on our own site: the question-and-answer section of every article sits on the page as plain text, and the JSON-LD marks up that same text — there are never two versions.
On the indexability side, check three things. First, robots.txt: if a rule blocks AI crawlers wholesale, know what it actually does — blocking Google-Extended does not affect AI Overviews, because Google defined it for the Gemini app and model training, while blocking Googlebot ends both your organic presence and the box. Second, rendering: content that only exists after JavaScript runs in the browser is at risk, and the text should be visible in the HTML the server sends. Third, snippet limits: a max-snippet character cap narrows the passage a model can take, by your own hand. llms.txt, one more piece of this ground, has an article of its own; it is useful, and it will not get you into the box on its own.
Proof from the field: what changed in a manufacturer's six months
We built every condition above into a single site at once. SIM Printing Suppliers has manufactured for the press industry since 1983, and forty years of technical knowledge lived where nobody could read it: in the heads of the sales team. We moved the site from WordPress to Next.js, built it in five languages and rewrote the technical content in a question-and-answer order from scratch — every heading the customer's own sentence, the first paragraph beneath it a plain answer, every answer carrying a measurement and a number.
Six months later organic traffic had grown 15×, priority keywords sat inside the top five, and visibility across AI engines went from zero to 40,000. The structure teaches more than the number: 40,000 counted the places the brand's name appeared when the category came up, and it started at zero — because before that there wasn't a single sentence to cite. The knowledge existed; it just wasn't written down.
Let's draw the boundary too. This is a pattern, not a guarantee: not every site building the same structure moves at the same speed, and the competition in a sector plus the technical state of the site set the pace. If you want to see other measured examples with their numbers attached, our case pages are open.
Where does optimising for Google's AI answers begin?
The starting point isn't producing content, it's measuring. Until you know which questions produce a box in your category and who gets named inside it, you can't decide which page to rewrite. The good news: that measurement needs no tool and takes about forty minutes.
- Write down the ten questions your buyer asks before buying — in their words, not yours. Ask your sales team what came up most over the last month; the list writes itself.
- Put each question to Google and note two things: did a box appear, and if so, which sources were named. How many of the ten produce a box tells you directly how much your category has riding on AI search.
- Place the sentences inside the box next to their equivalent on your own page. You rarely find missing knowledge; you find a difference in form — their answer sits in the first two sentences, yours in the fifth paragraph.
At the end of that test you hold a list, and that list is more honest than a content calendar. To restate the thesis: there is no guarantee of appearing in Google AI Overviews, but the probability can be raised systematically — an indexable foundation, paragraphs that survive being cut out, headings written in the customer's sentence, and proof placed inside the text. The whole picture of AI search — ChatGPT, Perplexity and the measurement side — sits in the canonical GEO guide. This article looked at Google's box alone, because in Turkey that is still the search box your buyer opens most.