Google put AI recipe illustrations in AI Overviews — and pulled them in four days

On 13 August 2026, the recipe site Inspired Taste searched Google for its own spaghetti squash recipe and found AI-generated illustrations that had appeared on their own inside the AI Overview. The images summarised the steps of the dish, and no user had requested them (Search Engine Roundtable (2026)).

The site's objection had three parts: their article contained a dedicated section on safely cutting the squash, and the illustration oversimplified it; it displaced creators who buy groceries for testing and shoot photos and video; and a searcher might assume the illustrations came from one of the sites cited in the AI Overview.

This AIO is directly trying to replace creators who buy groceries for testing, photography, and filming recipe videos. We are doing all of the work and Google is directly trying to over simplify everything and replace us with AIOs like this.

— Inspired Taste, quoted by Barry Schwartz / Search Engine Roundtable (2026)

On 17 August, Google's VP of Product for Search, Robby Stein, replied on X: "This was a small experiment we're no longer running." Adam from Inspired Taste wrote that they were "extremely happy" to hear the experiment had stopped (Search Engine Land (2026)). Four days passed between the sighting and the shutdown.

It isn't — and Google had to draw that distinction explicitly. On 14 July 2026, Google announced it was bringing image generation into AI Overviews. Built on the Nano Banana model, the feature was to roll out over the coming weeks in English, across all regions that already supported image creation in AI Mode (Search Engine Land (2026)).

When shutting the experiment down, Stein added that this was not the Nano Banana feature announced in July — that one only triggers when users explicitly ask to generate an image (Search Engine Land (2026)).

The distinction sits in two words: user-triggered, or unprompted. An image that appears because someone asked for it is a tool. An image placed on top of a search result nobody asked for is a substitution. Google kept the first and shut down the second.

Google's choice of recipe queries for the experiment wasn't arbitrary. The structure of a recipe query is distinctive: sequential steps, a visual need at each step, and a concrete outcome at the end. Beauty has a query family with exactly that skeleton.

• "How to use retinol" — sequential, carries warnings, step by step
• "Skincare routine order" — numbered steps, product sequence
• "How to contour" — shows hand motion and facial zones
• "How to apply foundation" — explains application technique
• "How long to leave a hair mask on" — involves timing and sequence

These are a beauty brand's most expensive assets. Model, lighting, shoot, edit — plus multiple variations to cover a range of skin tones. If an AI illustration summarises that investment "well enough", the click never happens and purchase intent stays on the search page.

Beauty also carries a risk the recipe experiment didn't: skin tone representation. An AI illustration of foundation application assumes a single skin tone. If a brand shoots each shade separately while the summary collapses it into one generalisation, the loss lands on both sales and representation. We've discussed when visual production is essential and when it's waste in our piece on studio photoshoots.

Three reasons. First: shutting down an experiment doesn't mean the interest behind it closed. What surfaced is which content type Google treats as summarisable by an AI image — step-by-step, procedural content that requires visuals.

Second: the gap between being cited and receiving traffic shows up here too. The most striking item in Inspired Taste's objection was that a searcher might believe the illustrations came from one of the sites cited in the AI Overview. The site is listed as a source while the image shown isn't theirs. That's the visual face of the pattern we covered in citation versus recommendation.

Third, and perhaps most important: this is a documented case of a small publisher's objection producing a result within four days. What ended the experiment wasn't regulation or litigation — it was one recipe blog's public objection and the industry making it visible.

• Make application content resistant to illustration: showing the step isn't enough — explain why it happens in that order and in that way. An illustration can summarise the "how"; it can't summarise the "why".
• Put skin tone diversity at the centre of the content. The one thing a single generalisation can't substitute is how the real result looks on different skin.
• Write safety and warning information into the text, not into the image. The strongest point in Inspired Taste's objection was that the illustration simplified a safety step.
• Track impressions and clicks separately for your application pages in Search Console. If impressions hold steady while clicks fall, a summary layer may have entered above that query.
• Don't tie your visual production plan to one channel — content that can be summarised on a search page still holds full value in ads and social, as we cover in our creative production service.

In short: Google shut this experiment down, but it's now clear that step-by-step content is seen as summarisable. Producing what an illustration can't imitate — real skin, real texture, real outcome — defends both sales and visibility.