Google Tests AI Snippets in AI Overviews: Page Descriptions Written by a Neural Network

⚡ The gist in 5 seconds

• Google is testing source cards in AI Overviews where the page description is written by a neural network rather than the site owner.
• The summary appears to the right and below the AI answer. No need to enable anything: it's either there for your query or it isn't.
• This is a test. There's no official launch, and no countries, languages, or devices have been named.

🔍 What happened

Google is trying out a new type of page description in AI Overviews. In the source cards — to the right and below the AI answer — the usual snippet is replaced by a short summary generated by a neural network. SEO specialist Brodie Clark spotted this, followed by Barry Schwartz of Search Engine Roundtable. Clark states it plainly: he had never seen such summaries in AI Overviews citations before.

There's one difference from a regular snippet, but it's an important one: the text in the card doesn't match the page verbatim. Google paraphrases what it considers relevant to the specific query. That means the same page will get different descriptions for different questions. Where the model takes the text from — the description, markup, or straight from the page body — there's no definitive answer.

How Google builds AI answers can be seen from the AI Mode documentation. The answers rely on retrieved web pages — that's grounding. The model splits the question into subtopics, searches for each in parallel, and stitches the results into a single answer — query fan-out. In its May guide on optimizing for generative AI, Google confirmed it can pull information from pages in its index. So reports Search Engine Land.

The scale is this: according to Comscore, the AI Overviews block appeared in 39.4% of desktop Google searches in the US in June and 25.8% in July 2025. The coverage of the overviews themselves grew from 7 countries in 2024 to 229 in 2025 — according to a study by Aral, Li, and Zuo. France, Turkey, China, and Cuba remained without it. Russia wasn't mentioned in the annotation, and there's no official data on the block being shown in Russia.

🧠 Our take

Our assessment: Google isn't testing the summary out of the goodness of its heart. AI Overviews is growing, and with it grows the discontent of sites losing clicks. A query-tailored summary looks like an attempt to give people a reason to click. That's a hypothesis, not a fact from the source — but the logic is simple: the summary promises an answer, and the click promises the details.

What it changes for sites: the description stops being the main lever. Now what matters is the first screen and direct answers under subheadings. If it's unclear from the summary why one should click through, there will be no click — and that's no longer about meta tags, it's about text structure. From experience, rewriting the entire site for this is pointless, but the first 2–3 paragraphs of key pages are well worth it.

Who wins: sites with a clear benefit right at the top of the page. Who loses: those whose essence is hidden on the third screen behind an intro. We believe the summary will pull out exactly what the model deems relevant — and a page without a clear thesis in the first screen risks getting an incoherent fragment in the card.

⚠️ What we don't know

The source doesn't say where the model takes the text from: the description, markup, or only the page body. No countries, languages, or devices for the test have been named. Everything rests on observations from SEO specialists, not on a statement from Google. The Comscore figures look odd: 39.4% in June, 25.8% in July — the source doesn't explain the drop within a single summer. And most importantly: whether AI Overviews is shown in Russia is officially unknown. You'll have to check yourself — on your own queries, with your own eyes.