
As of November 2026, Google is again showing AI-generated images inside AI Overviews for some searchers, after pausing the same test in August. If your Singapore team is asking whether this changes where clicks come from, the honest answer is: not yet in any way you can measure from your own dashboard, but it does change how you should describe, illustrate and track the pages you want AI answers to use. Anyone weighing a first push into local seo singapore work should read the return of this feature as a signal about image assets, not as a new ranking lever.
Three things are true at once, and most comment on this story collapses them into one.
The feature exists again. Search Engine Roundtable reported in mid-November 2026 that its author could reproduce AI Overview image generation in some browsers but not others, and that in other browsers the AI Overview replied that it could not generate images directly. That uneven behaviour is the point: you should plan for a feature that appears for a share of searchers, not for everyone.
The feature has a history. Google announced it in July, and the test was paused the next month. When the pause was announced, Robby Stein of Google described it as a small experiment the company was no longer running, and drew a line between it and the Nano Banana feature announced in July, which only triggers when a user explicitly asks for an image.
The measurement has not caught up. Nothing in Google's own reporting breaks out images drawn inside AI Overviews. So the practical question for a team is not "did we get an AI image" but "would we know if we did".

Real, so far, for part of the audience. The mid-November report quotes Google's original announcement, that the company was bringing image generation directly into AI Overviews in Search, and then records what the reporter saw: images generated in one browser, a refusal in another.
Two details decide how much this should change a plan. First, the announcement text is about getting a custom visual from a simple text prompt, made from scratch. That is a generative answer, not a stock image pulled from an index. Nothing in the feature, as reported, gives a publisher credit inside the picture.
Second, the earlier rollout was scoped. When the feature first went out, it was rolling out in English, for regions that already supported image creation in AI Mode, and it used Google's latest Nano Banana model. Singapore readers searching in English sit inside that language and region logic by default, which is why the pause and the return both land here rather than somewhere else.
A feature that appears in one browser and not another, that was paused before and has resumed without a dated announcement of its own, is a test by another name. Treat statements about its permanence as provisional.
The first pause followed public objection from recipe creators, and the specifics are worth keeping because they point at where the next objection will come from.
Search Engine Roundtable reported that a recipe site, Inspired Taste, posted about being shown an AI Overview with AI step cartoons while checking how its own spaghetti squash recipe appeared in Google. The site argued that the AI Overview was trying to replace the creators who buy groceries for testing, photography and filming recipe videos. Its post also flagged a subtler concern: a searcher might assume one of the websites cited in the AI Overview had produced the illustrations.
That assumption is the real risk for any brand, not just a food blog. When a generated visual sits next to a citation, credit blurs in the reader's mind. The publisher supplied the knowledge; the answer supplied the picture. Readers rarely separate the two, and nothing on screen invites them to.
Google's response was to call the experiment small and stop it. The company then continued with the separate Nano Banana path that responds when a user explicitly asks for an image. Most summaries lose that distinction between an image generated unprompted inside an answer and one generated on request, and it is the line that decides whether a brand's visual identity is at stake.
Less than the headline implies, more than nothing.
The reporter's own reading of the earlier announcement was that it would lead to fewer clicks on links and fewer people looking at Google Images from publishers. That is a reasonable prediction, stated as one. It is not a measurement, and anyone building a budget on it this quarter is building on a hunch.
What you can say with confidence is narrower and more useful:
For a Singapore business that sells through search, the useful shift is to stop counting visits as the only score and to start asking whether the pages you want cited are the ones an engine can actually quote. That is the same discipline a competent seo consultant applies to any query cluster, applied to a surface where the answer, not the list, is the result. It is also why a team reviewing its local seo singapore pages this quarter should start from what the pages say, not from what they show.

This is the part that decides whether the story is actionable.
Google's generative AI performance reporting covers AI Overviews and AI Mode impressions. It does not break those impressions down by whether an image was generated, and it cannot be extended into full prompts, internal query fan-out or standalone conversions. Its availability also depends on the client account. Where the report is missing, no workaround turns it into data.
So a team asking "how do our AI images perform" is asking a question the tooling cannot answer yet. Three things can be done instead, and they are worth doing regardless of whether the image feature stays.
First, track the answers themselves. Sample the questions your buyers type, record whether your brand is mentioned, whether the description is accurate, and whether your site is among the cited sources. Those are text-level observations and they work today.
Second, keep the denominator honest. A mention rate is only meaningful when it is mentions divided by valid answers collected under the same conditions. Changing the questions between rounds, or slipping the brand name into them, manufactures an improvement that does not exist.
Third, separate retrieval from citation. A page that an engine retrieved is not the same as a page it visibly cited in the answer. If your reports merge the two, you will overstate your presence in exactly the way a stray AI cartoon overstates a recipe site's authorship.
You can read more about how AI search reshapes a measurement plan in this guide to AI search strategy for marketing teams.
If you accept that the feature is real but unmeasurable, the checklist writes itself. Four moves, no new tools required.
None of this protects an image asset. It protects the page, the citation and the credit, which is where the value sits. A seo consultant brought in to review the cluster will start from the same four moves.

Topkee's position is deliberately narrower than the news, and that is the useful part. Topkee's GEO service starts from real buyer questions and monitors brand mentions, description accuracy and source presence in generative AI answers. Its verifiable connection to Google AI Overviews is the AI Overviews and AI Mode exposure data supplied by Google Search Console's generative AI performance report. That data cannot be extended into full prompts, internal query fan-out or standalone conversions, and its availability depends on the client account.
The service runs across three monthly tiers, with the same four analyses and four monthly deliverables at each: query fan-out analysis, sentiment and brand description, competitor comparison and AI browsing analysis. The tiers differ in capacity: thirty, sixty or ninety active tracked questions, and twenty, forty or sixty knowledge articles a month. The scope assumes one brand, one primary site, one market and one content language. The tracked-question figure is a cap on questions running at the same time, not a monthly query count, and the same question on a different engine counts as sample size rather than a new question.
Topkee Media also produces images and edits them for brand handbooks, ad creatives and social assets. That belongs to advertising creative and content production, and it is a different matter from Google generating images inside search results. The two should not be conflated, and a Topkee GEO report showing improved mention rates says nothing about AI image exposure. Topkee Media was founded in Hong Kong and has served multinational and local businesses with advertising and marketing growth services since it established its Singapore company.
Pick one query cluster you already rank for, sample the AI Overview for each question in it, and write down what you see, whether the brand is mentioned, whether it is cited, and whether an image appeared. Compare that baseline with your Search Console generative AI impressions for the same weeks. If the impressions move and the answers do not, you are looking at volume; if the answers change and impressions stay flat, you are looking at the surface itself.
For teams that want a second pair of eyes while they build that baseline, the local seo singapore page sets out the keyword clusters, page mapping and tracking that underpin it. Recent reporting on how Google's organic referrals have shifted, such as Ziff Davis's fall in Google referrals, is a useful companion read before you set a target for the quarter.
As of November 2026, some searchers see AI-generated images in AI Overviews and others do not. Search Engine Roundtable reported in mid-November 2026 that its author could reproduce the feature in some browsers, while other browsers returned a message that images could not be generated directly. Google has not published a dated announcement of the return.
The test was stopped after recipe creators objected publicly. Inspired Taste reported that an AI Overview showed AI step cartoons for its recipe and argued the feature was replacing the people who buy groceries, test, photograph and film recipe videos. Google described the test as small and said it was no longer running it.
No. The answer can cite your page in its text and still carry a generated illustration. The two are separate: your page supplies the information, the model supplies the picture. The risk is that a reader assumes the cited site drew the image, so marking your own visuals clearly and publishing them in context matters.
Not at present. Google's generative AI performance report covers AI Overviews and AI Mode impressions, with no breakdown for generated images, and its availability depends on the client account. The workable substitute is sampling the answers themselves and recording brand mentions, description accuracy and source presence.
Not on the strength of this news alone. The feature affects how answers look, not which pages earn citations. Spending decisions should follow your own sampled answers and impressions over several weeks, not a single report, and any push into local seo singapore work should still be judged by qualified enquiries rather than by visibility alone.