Google AI Overviews now generates images - and cites no one

On 14 July 2026, on Google Images' 25th anniversary, Google gave AI Overviews the ability to generate images from scratch. It is a small feature with a large implication: when the answer makes its own picture instead of borrowing yours, there is no source to send a click to. The self-contained answer just got more self-contained.

Most AI-search news is about text. This one is about pixels, and it matters more than it looks. For twenty-five years, an image result was a doorway - you searched, saw a thumbnail, and clicked through to the page it lived on. That click was a visit, a citation, a bit of credit. Google's newest change quietly removes the doorway for a growing class of visual queries.

What actually changed

Google announced that AI Overviews can now create an image directly from your prompt, powered by its latest Nano Banana model, alongside a text description and suggestions to refine the result. In Google's own words, the image is "made completely from scratch." It is rolling out over the coming weeks, in English, in the regions that already support image creation in AI Mode.

The key word is "scratch." This is generation, not retrieval. The model is not finding an existing image on a website and showing it to you - it is synthesising a new one that never existed before. That distinction is the whole story.

Why "from scratch" changes the economics

Here is the part worth sitting with, and to be clear this is our read of the mechanics rather than a statement Google made: a generated image has no source page. A retrieved image does. When Google shows you a photo it pulled from a site, there is a link back to that site - a path for the click, the credit and sometimes the traffic. When Google generates the image in the answer, there is no page it came from, so there is nothing to link back to. The visual answer is complete in itself, and it keeps the user inside Google.

Scale that across the many queries where people want a picture - "what does a mid-century sideboard look like," "show me a modern kitchen in sage green," "a diagram of how a heat pump works" - and a whole category of visual discovery stops sending anyone anywhere. It is the zero-click dynamic that already reshaped text search, arriving now for images.

"When the answer makes its own picture instead of borrowing yours, there is no source to send a click to."

Who feels this first

Not everyone is affected equally. The businesses most exposed are the ones whose value has partly lived in being the image source:

• Publishers and stock-style sites whose photos and illustrations earned image-search traffic. A from-scratch answer competes directly with "here is an image, from a source."
• Product and design brands in visual categories - furniture, fashion, interiors, food - where "show me" queries were a discovery path. A generated approximation can satisfy the browse without ever surfacing a real product.
• Explainer and how-to content that relied on diagrams and step images. A model that can draw its own diagram does not need yours.

What does not change is the value of being the named, trusted source for the underlying facts. A generated image of "a modern kitchen" does not recommend a brand of cabinetry. The answer that says which brand to buy still runs on citations - and that is where the fight for visibility stays live.

What to do about it

The instinct to panic is wrong; the instinct to re-weight is right. Practical moves:

1. Shift effort toward the surfaces that still cite you. Text answers, product recommendations and comparison-style queries still name and link sources. That is where being clear, current and corroborated pays off - so put your energy there rather than in image SEO that is losing its click.
2. Make your facts unmissable. The engines that do attribute reach for sources they can describe cleanly. Keep your category, positioning, pricing and differentiators stated plainly and consistently, on your own site and in the places models trust.
3. Treat visual identity as brand, not traffic. Distinctive product photography and diagrams still build recognition and trust with humans and can still be cited in the answers that carry links. Just stop counting on raw image-search clicks as a growth channel in affected categories.
4. Measure the shift. Watch whether you are still cited as the answer gets more self-contained. If image queries in your category quietly stop sending traffic, you want to see it in a dashboard, not guess at it months later.

What this does not change

It is worth being precise about the limits of this shift, because the panic version of the story overreaches. A generated image answers a "what does this look like" question. It does not answer a "which one should I buy" question, and it does not carry a transaction. When someone moves from browsing to deciding - "which sage-green cabinet brand is actually good," "who makes a reliable heat pump for a cold climate" - the answer turns back into named brands and cited sources. That is the moment your visibility work still decides the outcome.

Nor does it touch the parts of visual identity that build trust with a human. A distinctive product shot, a recognisable style, a diagram that makes a hard idea click - these still do their job on your own pages, in your ads, in the answers that carry links. What is disappearing is a specific, narrow thing: raw image-search clicks in categories where a good-enough generated approximation satisfies the browse. That is a real loss for some businesses, but it is a slice of the funnel, not the whole of it.

How to tell if it is hitting you

Rather than assume, look. A few signals tell you whether generated images are eating into a channel you relied on:

• Image-search referral traffic in visual categories. If you sell or publish in furniture, fashion, interiors, food, or explainer content, watch whether image-driven visits soften as the feature rolls out. A quiet decline that tracks the rollout is your answer.
• Whether you are still cited in the text answer. The image may be self-contained, but the recommendation around it often is not. If you are still named and linked in the "which brand" answer, your core visibility is intact even as image clicks fade.
• Query intent mix. Separate your "show me" queries from your "which should I" queries. The first group is exposed to generated images; the second is still a citation game. Knowing the split tells you how much of your demand is actually at risk.

The point of looking rather than assuming is that the impact is uneven. A brand whose discovery ran on image search feels this sharply; a brand whose buyers arrive through comparison and recommendation queries barely feels it at all. You want to know which one you are before you reallocate a budget.

The bigger pattern

This is one more step in a consistent direction: the answer keeps absorbing what used to be a journey. First text, now images, next likely more. Each absorption removes a doorway and raises the value of the one thing an answer engine still has to outsource - the trustworthy, nameable source behind a recommendation. The brands that stay visible will be the ones who make themselves the easiest, most reliable thing for the answer to cite, and who measure that citation directly instead of trusting a traffic number that is quietly being redefined.

Related: How to show up in Google AI Overviews →