Bing’s AI-Guided Image Search: A Publisher Testing Protocol
Bing announced an opt-in AI-guided image experience on US desktop. Use this protocol to test grouping, labels, sources, and publisher representation.
Confirmed May 2026: Microsoft announced an AI-guided image-search experience in Bing for US desktop users who opt into a “New Version.” The announcement describes grouped images, labels, summaries, and sources intended to help people explore a visual topic.
Microsoft did not publish a dedicated publisher markup, selection formula, or guarantee of inclusion. The appropriate next step is a bounded observation study, not a new image-SEO rule. SearchEngineAnswer has not completed the sample described below, so this article reports a protocol rather than results.
Define the test environment
Record country, language, device, browser, signed-in state, SafeSearch setting, date, and whether the new experience is enabled. Capture the same queries in the conventional and AI-guided views when both are available.
Build a fixed sample that includes brand entities, products, people, places, events, diagrams, how-to tasks, current topics, and ambiguous visual queries. Add publisher-owned queries where you know the canonical media and destination.
| Layer | Observation | Boundary |
|---|---|---|
| Grouping | Cluster names and image membership | Visible sample only |
| Summary | Claim text and cited sources | Do not infer image ownership |
| Image | Thumbnail, crop, caption, duplicate | Rendered result can differ by environment |
| Destination | Landing URL and canonical owner | Redirects may intervene |
Audit each group
- Save the group label, explanatory summary, visible sources, and full screenshot.
- Record each image’s position, crop, apparent source, target URL, and canonical landing page.
- Check whether duplicate images from different hosts are treated as one work or several results.
- Compare the label and summary with the visual evidence inside the group.
- Record images that contradict the group or have the wrong subject.
- Preserve queries with no AI grouping or no publisher-owned result.
A correct-looking cluster can still misattribute an image, send the user to a syndicator, or summarize a visual using unrelated surrounding text. Audit the media, the source page, and the generated explanation as separate objects.
Prepare publisher pages without inventing rules
Use stable image URLs, crawlable landing pages, accurate alternative text, descriptive visible captions, clear ownership, appropriate licensing information, and media dimensions suited to the page. Keep the primary image meaning consistent with the title and surrounding copy.
Use image and licensing structured data only where it accurately represents visible content and supported fields. The announcement does not say that a particular schema triggers AI-guided groups. Ordinary image discoverability practices remain defensible; a claimed AI-grouping optimization does not.
The technical SEO launch checklist covers indexability, rendering, media, and canonicals. The small SEO experiment method provides a release log and decision rule.
Report results with the right boundary
Publish the query set, environment, sampling dates, inclusion rules, screenshots where permitted, and coding rubric. Report prevalence only inside that sample. Separate publisher representation, source accuracy, label accuracy, click destination, and normal search performance.
Do not claim that appearing in a group caused traffic without a controlled release and appropriate measurement. A visual can be present yet receive no click, and a searcher can obtain enough information from the group to avoid visiting the source.
Repeat the sample on a schedule because the experience, index, and generative layer can change. A rerun is a new observation, not proof that the previous state was stable.
Ask a question or join the discussion