Google puts image generation inside AI Overviews

The news came in a post on Google's own blog on 14 July, marking 25 years since Google Images launched. Two changes were announced together. The first is a rebuilt Google Images home page, a browsable gallery of images from across the web that updates in real time and is tailored to the individual, rolling out on desktop in the US in English over the coming weeks. The second is image generation inside AI Overviews, built on what Google calls its latest Nano Banana model, turning a text prompt into a custom image made from scratch. That rolls out over the coming weeks in English, in every region that already supports image creation in AI Mode.

Search Engine Land covered the change the same day and flagged the obvious risk to publishers: if searchers get the picture they wanted inside the overview, fewer of them carry on to the sites that host real images. Google's post does not address attribution, does not say whether generated images appear alongside web image results or in place of them, and gives no brand controls. Until it ships widely nobody outside Google can say how often a generated image will replace a real one, so treat the balance as an open question rather than a settled loss.

The whole mechanism of AI visibility so far has rested on citation. An assistant reads pages, names some of them, and a brand that is named gets seen. That mechanism does not apply here. A synthesised image is not retrieved from anywhere, so there is no source to name, no link to earn and no page to optimise for the slot.

This matters most for queries with visual intent and no specific brand attached. Somebody searching for a sage green living room, a walnut banana bread or a small-space desk setup used to get a grid of photographs, plenty of them from retailer and brand sites, and those photographs pulled people onto product pages. A generated image answers the same query without any brand being involved at all.

Generation does not touch the queries that decide purchases. When somebody asks which desk to buy, which CRM suits a five-person team or whether a particular product is worth the money, the answer still comes from retrieval, and it still names brands. That is the surface worth defending, and it works on text: what your pages say about your pricing, availability and specifications, and what third-party pages say about you.

It also still matters what an assistant believes your product looks like, because a model asked to draw something specific relies on whatever description it holds. Vague or inconsistent product descriptions across the web produce vague results either way.

Start by finding out what the engines say about you in text today, since that is where the commercial queries land. Our free AI visibility check runs live queries about your brand through Claude with web search and needs no login, and shows the specific claims an assistant makes about you. Support for further engines is rolling out.

Then keep an eye on the Google surface specifically, because AI Overviews now appears on a large share of queries and changes faster than the rest. Our Google AI Overviews tracking page covers how that surface works and what shows up in it. Alongside that, keep doing the unglamorous things that make your pages usable by retrieval: accurate crawlable text for every product fact, honest captions and alt text, and structured product data that matches what is on the page.

What is not worth doing is trying to game a slot that has no link in it. There is no known way to influence a generated image, and any vendor claiming otherwise is guessing. The sensible response is to shift effort towards the answers that still cite sources.

Search is steadily replacing pointers to the web with things made on the spot, and images are the latest to go that way. You cannot compete for a result with no source. You can make sure that when an engine does need a source, about your pricing, your products or your category, yours is the one it can read and trust.