AI Overviews and AI Mode surface images pulled directly from indexed pages, and for once the platform documents exactly how it chooses. We mapped the documented signals, the disqualifiers, and the popular claims the evidence does not support.
Home › Research › Image selection in AI search
When an AI Overview, search result or Discover card shows an image for a page, what decides which image gets picked? Image SEO advice is drowning in speculation about "AI vision models", so we wanted to separate what the platform documents from what the industry sells.
We anchored on primary platform documentation: Google Search Central's "AI features and your website" and its March 2026 update to "Image SEO best practices", which for the first time spells out the preferred-image signals. We corroborated the documentation change through independent reporting by Search Engine Journal, surveyed practitioner commentary on original-versus-stock imagery, and graded every claim by evidence tier. We then ran the resulting check-set against a live site, our own, and found real failures, which we publish below.
The mechanics are documented and unglamorous: eligibility prerequisites most sites fail on, three explicit preferred-image signals, and a short list of disqualifying image traits. The claims with the biggest price tags attached (AI vision optimisation, stock-image penalties) are the ones with the least evidence behind them.
Before any selection logic runs, an image must be eligible for indexing at all. Google's documentation is unambiguous: images are indexed from the src of an img element (including inside picture elements). CSS background images are never indexed. A visually stunning hero applied as a CSS background does not exist as far as Google Images, Discover previews or AI Overview image selection are concerned.
Descriptive filenames and information-rich alt text are the second layer: Google uses them, along with computer vision and page context, to understand what an image shows. A photographer whose portfolio images are named DSC_4031.jpg with empty alt attributes has met neither condition. In our experience auditing small business sites, eligibility failures are far more common than selection failures, and they cost nothing to fix.
In a March 2026 documentation update, Google spelled out the mechanism behind image selection for search results, Discover and, by extension, the AI surfaces drawing on the same index. Selection is automated, but three metadata signals let you indicate a preference:
| Signal | Where it lives | Notes |
|---|---|---|
| primaryImageOfPage | schema.org structured data (WebPage) | The most explicit signal; sits naturally in existing JSON-LD |
| image property | On the page's main entity (Article, Product, LocalBusiness) | Doubles as general entity enrichment |
| og:image | Open Graph meta tag | Also controls social sharing previews; the most commonly present, and most commonly broken |
Google does not rank the three signals against each other, and we will not invent a ranking. The guidance on what the chosen image should be is equally concrete: relevant, representative of the page content, high resolution (at least 1,200px wide for Discover-quality treatment, 16:9 preferred there), and not a generic logo, not an image containing text, not an extreme aspect ratio.
Most small business pages send either no preferred-image signal or a broken one, a dead og:image URL, a logo, or a text-heavy banner. The mechanism to influence image selection is documented and free; it is simply unused.
Google's AI features documentation is deliberately deflationary: AI Overviews and AI Mode surface images from indexed pages under the same image SEO fundamentals, with "no additional requirements" and "no special schema.org structured data that you need to add". Anyone selling "AI image optimisation" as a distinct technical service beyond the fundamentals above is selling something the platform explicitly says does not exist.
We say this as a business that sells schema implementation: ImageObject and primaryImageOfPage markup are worth doing as part of a coherent entity strategy, and we price them accordingly. They are not a secret AI-citation switch, and the documented fundamentals must come first.
The most marketable claim in this space, that AI surfaces penalise stock imagery and reward original photography, is practitioner consensus without a dataset. Multiple independent commentators assert it, it aligns with Google's general preference for representative and non-duplicated content, and it is directionally sensible for E-E-A-T. But we found no controlled study on image origin driving AI Overview selection, and Google's guidance does not say it. We hold it at 60% confidence and treat it as a sensible default, especially for visual businesses, rather than a proven lever.
The practical resolution: an original, representative hero image is worth having for reasons that are documented (representativeness, uniqueness, resolution). Whether it additionally buys AI selection preference over stock is unproven, and we will not charge anyone on the basis that it does.
Our free AI Visibility Snapshot includes the image eligibility and preferred-image checks from this research. One clear page, within 24 hours.
Get My Free Snapshot →Part of the AI Visible Research: Our Findings series. We grade every claim by evidence tier and publish the grades, including the popular claims we decline to sell. Questions or challenges to the method are welcome via our enquiry form.