Google Makes Image-Led Search Measurable

The Visibility Report #36 | Week of September 22-28, 2026

Google Search Console can now separate searches that start with an image from ordinary text searches. The new filter covers Lens, Circle to Search on Android, image uploads, and Chrome's "Search this image" action.

The reporting arrives as Google adds more links inside AI Overviews, B2B research finds that answer engines use different source systems, and OpenAI expands ChatGPT Ads. The common thread is measurement. An impression, citation, external click, and paid placement describe different events. Put them in separate columns.

Google splits multimodal search from text in Search Console

Google added a Web: multimodal filter to the Search results report and the Generative AI report on September 24. The standard Search results view can show clicks and impressions for qualifying image-led searches. The Generative AI view shows impressions when those searches surface a site's pages in AI Overviews or AI Mode.

The filter does not reveal the image or a text query behind the search. That limits diagnosis, but it gives image-heavy businesses a first-party baseline that did not exist before. Export the page, country, device, and date views now. Then compare multimodal landing pages with conversion data in analytics before changing an image program.

Sources: Google Search Central | Search Engine Journal​

More AI Overview links still do not guarantee an external click

Google now uses several link treatments inside AI Overviews. Some open a publisher page. Others move the user into AI Mode or another part of Google's experience. Search Console counts those interactions according to their destination, so a larger link inventory does not translate directly into more visits to the open web.

Track the full path: appearance, citation, external-click opportunity, referral session, and conversion. A citation can still improve awareness, but it should not be reported as traffic. The unresolved piece is the click rate for each link treatment across different query types.

Source: Search Engine Journal​

Foundation and AirOps found two source systems behind B2B AI answers

Foundation and AirOps logged 380,000 AI answers across 81 B2B software categories and six AI surfaces over almost five weeks. They classified roughly 3 million citation events and found that platforms and categories draw from different mixes of owned pages and outside sources.

This is partner research from an agency and a tool provider, not a neutral census of every AI answer. It still exposes a reporting problem: one blended visibility score can hide whether a brand's own pages, review sites, communities, or competitors supplied the evidence. Break citation share out by platform, buyer question, and source type.

Source: Foundation x AirOps research​

Comparison pages supplied most of the evidence in one ATS test

Foundation tested 20 applicant-tracking-system buying questions across six AI surfaces. Comparison pages and listicles accounted for 58.7% of surface-weighted source presence, while the four vendors in the study supplied 27% of the evidence. On direct two-product questions, dedicated comparison pages reached 42.1%.

The dataset covers one software category, so those percentages should not become a benchmark for every market. The operating lesson is narrower: map the pages AI uses for real comparison prompts. If competitors or publishers explain your product more often than you do, build a clear first-party comparison page and correct material gaps in the outside evidence.

Source: Foundation ATS study​

A text-only export can remove the structure machines need

Technical consultant Slobodan Manic argues that turning a site into plain text or markdown can discard headings, labels, tables, navigation, and relationships that explain how information fits together. An AI system may be able to read every sentence and still lose the page's operating structure.

This is practitioner analysis, not a ranking-factor announcement. Use it as a test. Check whether product details, comparisons, prices, and state changes remain understandable in rendered HTML and in machine-readable output. Simplifying a page should preserve meaning rather than flatten it.

Source: Slobodan Manic in Search Engine Journal​

ChatGPT Ads expands to seven more Asian markets

OpenAI began rolling ChatGPT Ads out in Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam, and Taiwan on September 23. The company says the product is now available in more than 60 countries through its sales team, agency and technology partners, and self-service Ads Manager for eligible businesses.

Availability is not evidence of performance. Teams testing the channel should separate paid exposure from organic mentions and citations, then use geographic holdouts or another clean comparison to estimate incremental conversions. Otherwise ChatGPT Ads can take credit for demand another channel created.

Source: OpenAI​

From Practitioners

• ​Slobodan Manic: Machine-readable content still needs hierarchy and relationships. A text dump can preserve words while losing meaning.
• ​Foundation research team: Owned pages and outside evidence should be measured separately because platforms and categories use them differently.
• Community signal: Current X and LinkedIn discussion is focused on the new multimodal filter, while Reddit discussion remains more concerned with weak AI Overview click-through. These conversations guided the editorial emphasis but are not used as factual proof.

From the Tool Blogs

• Ahrefs: How to optimize for AI search connects AI visibility work to crawlability, indexability, useful content, and separate citation measurement. Published September 25.
• AirOps: AI Ads in AirOps adds paid conversational-search monitoring to the vendor's AI visibility product. Published September 24.
• Semrush: An AI visibility audit workflow separates mentions, cited pages, topic gaps, outside sources, and technical access. Updated September 23.
• seoClarity: A framework for choosing prompts to track focuses enterprise monitoring on prompts tied to real influence instead of every wording variation. Published September 23.
• Profound: Refreshed Answer Engine Insights consolidates platform, persona, region, visibility, citation, and trend views. The page exposes current September update metadata but no clear visible publication date.

What to watch

• Whether Google adds multimodal filters to the Search Analytics API.
• Whether Search Console separates external AI Overview clicks from interactions that remain inside Google.
• Whether the Foundation findings hold in non-software categories and over longer periods.
• How ChatGPT Ads reporting distinguishes paid exposure from organic recommendation behavior.

What to do next

• Export a multimodal baseline from both Search results and Generative AI reports.
• Join multimodal landing pages to conversions instead of stopping at impressions.
• Track owned and third-party citations separately for five buying questions.
• Review the comparison pages that answer engines use before creating another generic guide.
• Label paid, organic, citation, referral, and conversion data as separate measures.

Search is adding more surfaces faster than most dashboards can explain them. Keep the source, destination, and business result attached to each number.

The Visibility Report | Will Scott​
This newsletter is produced collaboratively by Will Scott and Bob, an AI agent. Human oversight, AI efficiency.
​Subscribe at thevisibilityreport.com​