AI Overviews Drew the Answer. Google Has Paused It

AI Overviews generated images without being asked, and for four days in August 2026 nobody could say on whose authority. On August 13 a recipe publisher documented Google’s AI Overviews rendering AI-generated “step cartoon” illustrations for a spaghetti squash query — no image request, no label, no attribution. On August 17, Google’s Search VP confirmed the behavior was real, was not the image feature announced in July, and is no longer running.

That correction matters more than the sighting. It establishes, on Google’s own record, that AI Overview image generation has at least two distinct code paths: an announced, opt-in path that triggers only when a user explicitly asks for an image, and an unannounced, auto-triggered path that briefly decided on its own that an answer deserved an illustration. The second one is paused. The capability it demonstrated is not going anywhere.

This post reconstructs the timeline from primary statements, then makes the argument that outlasts the sighting: a generated image is not just an attribution problem, it is a measurement problem. A text citation is a link that can be present or absent, and therefore counted. A synthesized image has no source URL at all — so the entire citation-tracking paradigm has nothing to count, and no publisher control has anything to act on.

On Thursday, August 13, 2026, the recipe publisher Inspired Taste posted on X an example of a Google AI Overview showing AI-generated step-by-step cartoon illustrations for a spaghetti squash recipe query. The searcher had not asked for an image. The illustrations carried no label identifying them as AI-generated and no attribution to any source. Search Engine Roundtable picked up the post the next morning, Friday, August 14, embedding the publisher’s screenshot and text.

The sourcing here deserves precision: this is trade press reporting a publisher’s own X post, and there is no Google primary source describing the behavior before Google’s August 17 response. We are relying on a screenshot and the account of the publisher who saw it, corroborated by the fact that Google subsequently confirmed the experiment existed rather than disputing the sighting.

Inspired Taste’s stated objection went beyond lost clicks. In the publisher’s own words: “There is also risk that a searcher might think that one of the websites cited in the AIO did the illustrations. It is very confusing and extremely oversimplified.” That is a stated concern from the affected publisher, not a measured outcome — no user study sits behind it — but it names the confusion mechanism precisely: the AI Overview cited real websites for its text while displaying images none of them made.

On Monday, August 17, 2026, Google’s VP of Search, Robby Stein, responded on X to the circulating sighting. His statement did two things at once: it confirmed the auto-generated recipe images were real and Google’s, and it severed them from the feature Google had announced in July.

Read carefully, that is a significant admission delivered in a casual register. It means the recipe images were not the July-announced capability reaching production. The announced feature is opt-in and user-requested. The sighted behavior was auto-triggered, unannounced, unlabeled, and is now paused. Search Engine Roundtable updated its original Friday story to Monday afternoon to carry the confirmation; its headline now carries “(Test Paused)”.

The distinction cuts both ways for publishers. The reassuring half: the alarming version — Google deciding unprompted that your recipe deserves a synthetic illustration — is not a shipped feature, and Google stopped it when surfaced. The unsettling half: it ran at all, in production, with no announcement, no label, and no attribution, and the only reason anyone knows is that one publisher happened to catch it and post the screenshot. Google tests features in live search constantly; that is normal practice. What this particular test demonstrated is that the auto-triggered variant exists and works — the decision to run it is a dial Google controls, not a technical boundary.

The feature Stein referenced was announced on Tuesday, July 14, 2026, in a Google blog post marking 25 years of visual search, bylined to Brad Kellett, Senior Engineering Director for Search. Google described bringing image generation “directly into AI Overviews in Search” using its “latest Nano Banana model” — without naming which tier of the Nano Banana family, which per Google’s June 30 developer announcement spans four models. The rollout was promised “over the coming weeks in English, for all regions that currently support image creation in AI Mode”; Google did not publish an exact region list on that page.

We analyzed what that July announcement means for brand visual control and for extending GEO to visual assets in our July piece on image generation in AI Overviews and brand visuals — none of which this post re-argues. What this post adds is what happened after: the two events of August, and the measurement problem the July announcement did not address.

Here is the argument that should outlive the sighting. AI-visibility measurement — including our own 1,000-query AI Overviews citation study — rests on one operation: for a given query, check whether a given site appears among the cited sources. The citation is a link. A link is either present or absent. Presence and absence can be counted, trended, and reported. The entire discipline of measuring AI search visibility is built on that binary.

A generated image breaks the binary at its root. The pixels are synthesized fresh by a model; they are not “from” any crawled page. There is no URL attached to them, so the question every citation tracker asks — was I cited or not? — cannot even be posed. This is not a gap a tool vendor closes with a feature update to link detection. The object being measured does not exist. A tool could, in principle, detect that an AI Overview contains a generated image at all — but it could never attribute that image to a source, because there is no source to attribute it to.

The same absence disables publisher-side controls. Google’s documented directives for AI features — nosnippet, data-nosnippet, max-snippet, noindex, per the Google Search Central AI-features page as checked at the time of writing — all govern whether Google may show content sourced from your page. A synthesized image is sourced from no page. Nothing on that documentation page addresses it, and there is no crawl-time signal a publisher can set that reaches a generation decision. That absence claim is narrowed to the one page we checked — but that page is Google’s own canonical list of publisher controls for AI features.

This is why we call it a measurement gap before a traffic problem. The broader zero-click numbers are real and live in our zero-click AI Overviews data update, and this post deliberately repeats none of them — because for generated images there are no numbers to repeat. When a text citation replaces a click, analytics show the loss. When a generated image replaces your photograph, nothing registers anywhere: not in Search Console, not in a rank tracker, not in an AI-visibility dashboard. The displacement is invisible by construction.

“AI Overviews generate images now” is not one undifferentiated fact. It is four distinct surfaces with four different answers to the questions that matter to a publisher: is there a source URL, can existing tools count it, is there a documented control, and what has Google actually said. Laid side by side, the spectrum is stark.

Sources for the table: Google’s July 14, 2026 announcement; Robby Stein via X on August 17, 2026 as reported by Search Engine Roundtable; and the Google Search Central AI-features documentation page as checked at the time of writing. The two bottom rows share every structural property except one — who decides an image gets generated. That single difference is why the paused experiment was the more alarming of the two: opt-in generation answers a user’s request; auto-triggered generation substitutes for a publisher’s asset without anyone asking.

To test whether the measurement gap is real rather than rhetorical, we surveyed the marketing pages of AI-visibility and citation-tracking tools — Profound, Peec AI, Otterly, Semrush AI Visibility, Am I Cited, and several aggregator roundups of the category — at the time of writing. The descriptions are uniform: tracking brand mentions, citations, and sources inside AI-generated text answers across ChatGPT, Perplexity, and Google AI Overviews. None of the pages surveyed describes tracking whether an AI Overview generated a synthetic image in response to a target query.

That is a narrowed absence claim, and we hold it to its limits: it covers the pages we surveyed, not every tool in existence, and marketing copy is not a feature audit. But the uniformity is telling. The category describes itself entirely in the vocabulary of citations and mentions — the countable objects — because that is what the underlying methodology can see.

Project this forward and the stakes compound. Generated visuals in answer engines are, on Google’s own July framing, an expanding surface — announced for AI Overviews after existing “in AI Mode for some time,” as Search Engine Roundtable noted in its July 15 corroboration. Barry Schwartz’s own read at the time: “I do think this will lead to fewer clicks on links and fewer people looking at Google Images from publishers.” That is his stated opinion, not a measurement — but if visual answers grow while measurement stays citation-shaped, the industry’s dashboards will keep reporting a shrinking slice of what AI search actually shows users. The gap does not stay constant; it widens with every visual surface that ships.

The obvious rejoinder is that the industry already has an answer in flight: content provenance. Google states that Nano Banana model outputs carry SynthID watermarking, checkable “through the Gemini app, Gemini in Chrome or Search.” On May 19, 2026, Google announced SynthID verification expanding from the Gemini app to Search, with vendor-stated cumulative figures as of that date of over 100 billion images and videos watermarked and 50 million Gemini-app verification uses. The same May 19 post said C2PA Content Credentials verification “will come to Search and Chrome in the coming months” — and at the time of writing, nearly three months later, we could not confirm that C2PA verification has actually shipped inside Search. Treat it as announced, not confirmed shipped.

Four days before Google’s pause statement — on August 13, the same Thursday as the sighting — Google open-sourced Credentio, an on-device C2PA validator we covered in our Credentio guide. The provenance infrastructure is real and advancing. But look at what it actually delivers for this problem, and the fit is partial at best.

Three limits stack. First, provenance marking is searcher-facing: it can tell a user an image is synthetic, but it gives no publisher any signal that their query surfaced a generated substitute for their content — the measurement gap is untouched. Second, credentials are fragile in exactly the way this surface gets consumed: as our EU provenance-marking guide put it, screenshots and re-encodes strip C2PA metadata — and an AI Overview image is precisely the kind of asset that gets screenshotted and re-shared. Third, the regulatory hook runs to the provider: under EU AI Act Article 50(2), in force since August 2, 2026, the machine-readable marking obligation for EU users sits with Google as the model provider — it mandates marking, not attribution, and gives publishers no new lever. For completeness on the economics: Google’s stated developer-API price for Nano Banana 2 Lite is $0.034 per 1K-resolution image on the Gemini API surface — a developer list price, not a consumer Search cost, and Google has never said which Nano Banana tier serves AI Overviews.

Honesty first: the available responses are modest, because the gap is structural and the levers sit with Google. That is not a reason to do nothing — it is a reason to do the cheap, durable things that build your evidence base and force the question onto your vendors’ roadmaps.

The consistent thread: build the record now. Google paused this experiment because it was caught and confirmed within four days — publisher documentation demonstrably moved the needle. If auto-triggered generation returns, in Search or in any other answer engine, the publishers with dated screenshots, credentialed originals, and vendors already on notice will be the ones who can respond with evidence rather than anecdote. This measurement-first approach to AI search is the core of our agentic SEO services — including honest scoping of what current tooling can and cannot see.