Google just started drawing your recipe for you

You spent forty dollars on ingredients. You cooked the thing three times to get the timing right. You shot it in the good afternoon light, filmed the tricky fold, wrote the note about why the dough behaves differently in a humid kitchen. This week Google started drawing a picture of that instead, and putting the drawing above your link.

That is the plain version of what shipped. Image generation inside AI Overviews began rolling out during the week of August 10, 2026, about a month after Google first announced it, and Search Engine Roundtable documented the rollout on August 14. The images arrive in a collapsible panel. On recipe queries they show up as generated step-by-step illustrations sitting inside the answer, above the sources that fed it.

What changed in AI Overviews images

Until now the AI Overview was a text summary with links attached. It took your words and compressed them. AI Overviews images change the shape of the trade: the answer now generates the visual layer too, which was the one part of a recipe, a repair guide, or a product walkthrough that a text summary could never carry. A paragraph summarizing your brioche method still left a reason to click. A picture of the fold does not.

This is not a ranking change and it will not show in your position tracking. It is a change in what the answer surface is capable of holding, which is a different kind of problem. We have argued before that the click is no longer the unit of discovery, and the honest read here is that visual content just joined text on the list of things the answer can hold without you.

Two of those numbers deserve a caveat, because they disagree and both get quoted as if they do not. The Similarweb analysis, which covers January to April 2026 and excludes the Google mobile app, put AI Overview presence above 20% of searches. The August webmaster roundup puts it near 50%. Different windows, different measurement, four months apart. Use whichever matches your reporting period and say which one you used.

Recipes were always going to be the test case

Recipe publishing is the most punished category in search and has been for a decade. It is also, structurally, the easiest thing in the world to summarize: a list of ingredients, a list of steps, a time, a temperature. The long personal preamble that everyone mocks exists because the summary is worthless as a business and the story is the only part a summary cannot take.

Now the illustration can be taken too. That is why the sharpest objection came from a recipe site rather than a news publisher. Inspired Taste, quoted in the rollout coverage, put it directly.

“Google is directly trying to replace creators who buy groceries for testing, photography, and filming recipe videos.”

Read that as a warning label rather than a complaint. Every vertical where the proof is visual is on the same track: home repair, medical procedure explainers, assembly instructions, technique-led B2B documentation, anything where a photograph was the reason the page earned trust. Recipes just got there first because recipes always get there first.

AI Overviews images and the zero-click math

Here is the part that gets oversold in both directions. AI Overviews images will not cause a traffic cliff on their own, because the cliff already happened. US zero-click rate moved from 60.45% in 2024 to 68.01% across the first four months of 2026 in the Similarweb panel. Searches producing at least one click fell 9.51 percentage points over that span. The visual layer is an accelerant on a fire that has been burning for two years.

So the useful question is not how much traffic the images cost. It is which pages were still earning clicks in a 68% zero-click world, and whether the visual layer was the reason. In our experience the pages still pulling clicks under heavy AI Overview coverage share three traits: they contain something the summary provably cannot carry, they are the destination for a transaction rather than an explanation, or the reader has a reason to trust the specific author. Generated illustrations attack exactly one of those three, and it is the first one.

If your visual content was proof, you are exposed. If your visual content was decoration, nothing changed, because decoration was never why anyone clicked. Most content teams have never made that distinction on a page-by-page basis, and this is the month to do it. Our content marketing engagements now start with that audit rather than a keyword map.

What the objection gets right, and what it misses

The publisher objection is right about the economics. Google is generating a substitute for an asset it did not pay to produce, using a model trained in part on assets like it, and placing the substitute above the link to the original. You do not have to be sentimental about recipe blogs to see that the incentive to make the original just got worse.

What the objection misses is that outrage is not a strategy and blocking is not either. Publishers who pulled the lever on Google referral traffic have data on what happens next, and it is not a clean win: we walked through what publishers actually measured after blocking Google and the trade was worse than the pitch. The realistic position is that the answer surface is now the front page, and the fight is over what it cannot render rather than whether it should exist.

The assets a generator still cannot fake

Strip out everything that a diffusion model can approximate and look at what is left. It is a shorter list than it was in July, and it is still a real list.

One of our long-running publishing clients rebuilt around exactly this idea before the images shipped, and the BookBaby engagement is the version we can show publicly: proof-led pages, named authors, and process documentation that no summary layer can reproduce. The point was never to outrun the answer. It was to own the part of the page the answer has to send someone to.

What to do about it on Monday

Pull your top fifty visual pages and sort them into two piles: pages where the image is proof and pages where the image is decoration. The decoration pile needs nothing. The proof pile needs provenance, which means a named photographer or author, a capture date, and a visible record of the test, plus schema that makes all three machine-readable. That is a week of work and it is the highest-leverage week your content team has available right now.

Then start producing the failure photograph deliberately. Every test you run generates one. It costs nothing extra, it is the most human artifact you can publish, and it is the exact thing a model asked to illustrate a recipe will never think to draw. Pair that with the machine-access work in a proper generative engine optimization program so the proof you are producing actually reaches the engines deciding who gets named.

The uncomfortable truth is that this was coming whether or not Google shipped it in August, and it will not stop at recipes. The comfortable truth is that a system that generates the average picture of a thing has a permanent, structural weakness: it has never made the thing. Everything you build from here should exploit that. Full rollout details are in the Search Engine Roundtable coverage from August 14.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.