AI Search & Visibility
Gemini Deep Research shows you exactly which sources it read but didn't cite. For B2B marketers, that label is the most important signal you aren't tracking.
Run a topic through Gemini Deep Research and open the sources panel when it's done. You'll see two categories. Some pages are cited — named, linked, attributed in the report. Others are listed as "read but not used." Gemini visited them, processed them, drew on them as context, and then left them out of the final output entirely.
Your content may be in that second column right now. Shaping AI-generated answers that reach your prospects, with no attribution, no traffic, and no signal you can measure in any dashboard you currently own.
That is a different kind of invisibility than ranking on page two.
Three States Your Content Can Be In
When an AI research system runs, your content doesn't exist in a binary of found or not found. There are three distinct outcomes:
State 1
Not retrieved. Gemini never accessed the page. It doesn't exist in the context of this query.
No influence. No credit.
State 2
Read but not cited. Gemini accessed and processed the page, used it as background context, but did not attribute it in the output.
Influence without credit.
State 3
Cited. Gemini accessed the page, extracted a citable claim, and named it as a source in the final report.
Influence with credit.
Most content strategy work targets State 1 — getting discovered. The more significant problem for established B2B publishers is the gap between State 2 and State 3. Pages that rank, get retrieved, get read, and still don't make the citation list.
How Large Is the Gap
Larger than the industry has acknowledged. A preprint study analyzing roughly 14,000 real-world conversation logs with search-enabled AI systems found the attribution gap is structural, not incidental. For Gemini specifically, the study found that the model provided no clickable citation in 92% of answers, and that a typical Gemini query left approximately three relevant websites visited but uncited (vendor-supplied figures, unaudited). That data covers Gemini in standard search mode — but it frames the baseline tendency the Deep Research sources panel makes visible.
Separately, when Gemini Deep Research does cite, the sourcing pattern is concentrated. Research analyzing over 6.8 million citations across 1.6 million AI responses found that 52% of Gemini citations came from brand-owned websites with structured, schema-marked content (vendor-supplied figures, unaudited). The structural implication: pages without clear markup, clear authorship, and discrete attributable claims are more likely to be absorbed as background than elevated to citation.
Why This Happens: What Microsoft's Engineers Just Explained
On May 6, 2026, three engineers from Microsoft AI published an unusually candid post explaining the mechanics behind this gap. They weren't writing about Gemini — they were describing how Bing's own index is evolving to support AI-generated answers. But the underlying logic applies across every AI system that retrieves and synthesizes web content.
The engineers draw a sharp line between two jobs an index can do. Traditional search asks which pages a user should visit. Grounding — the process of anchoring an AI answer to real-world sources — asks what information an AI system can responsibly use to construct a response. Those questions sound similar. They are not.
In traditional search, the unit of value is the document. A human clicks, skims, and decides. The system tolerates imperfection because humans self-correct. In a grounded AI answer, the unit of value shifts to what the engineers call "groundable information" — discrete, supportable facts with clear provenance. When the AI synthesizes an answer, multiple sources collapse into a single statement. There is no skim-and-self-correct step. Errors compound before any human sees the output.
Jordi Ribas, a Microsoft corporate vice president, captured the stakes plainly on X: "In the era of the agentic web, the role of the web index needs to evolve to support very different needs across agents and humans."
The read-but-not-cited outcome is what happens when content passes the retrieval threshold but fails the grounding threshold. It got read. It didn't meet the bar for being used as evidence.
What Makes Content Groundable vs. Just Readable
The Microsoft engineers identify several dimensions where grounding requirements diverge from traditional search quality. For content publishers, these translate directly into reasons why pages end up read but not cited:
| Dimension | Gets you read | Gets you cited |
|---|---|---|
| Factual fidelity | Content that ranks and loads | Claims that survive chunking — atomic, self-contained, not dependent on surrounding context |
| Provenance | A byline and a publish date | Named author, clear methodology, traceable sourcing the AI can attribute with confidence |
| Freshness | Recent enough to rank | Current enough that a stale fact won't produce a wrong answer — the cost of staleness is categorically higher in grounding |
| Originality | Covers the topic adequately | Adds something the AI can't get from consensus — original data, a distinct position, a named perspective |
| Structure | Readable prose with headings | Sections where each heading answers a question and each paragraph leads with its claim |
The engineers make one point that deserves particular attention: the chunking and transformation processes that make content retrievable can distort meaning. A page that reads clearly to a human may be represented in the index as fragments that lose the original claim's precision. Dense, hedged prose — the kind that qualifies every statement before making it — is especially vulnerable to this distortion. It gets retrieved. The extractable claim isn't there.
The snippet problem
Gemini often processes search snippets rather than reading full pages. User testing confirms the model can rely on short previews rather than the full text. If your key claim is in paragraph four, Gemini may have absorbed a snippet of your page — enough to draw on as context, not enough to treat as a citable source. The claim needs to be in the first sentence of the section, not buried in the body.
The Measurement Problem
The Microsoft engineers are candid about where the field stands: decades of practice exist for measuring search quality, but measuring grounding quality rigorously is still being worked out. That gap exists on the publisher side too.
Right now, most B2B marketing teams have no instrument for detecting whether their content is landing in State 2 or State 3. Search Console shows impressions and clicks. Bing's AI Performance dashboard shows grounding citations. Neither shows you read-but-not-cited at scale. The only current way to see it is to run your own queries in Gemini Deep Research and inspect the sources panel manually — which is not a scalable measurement practice.
That's not a reason to wait. It's a reason to start building the content posture that positions you for State 3 before the measurement infrastructure catches up.
What to do Monday
Run your category in Gemini Deep Research. Check where your content lands.
Open Gemini Deep Research and run three to five queries that represent how your buyers research problems you solve. When each report completes, open the sources panel. Find your domain. Note whether your pages appear as cited sources or as read-but-not-used. That single observation tells you more about your AI visibility posture than a month of rank tracking.
For pages that appear read but not cited, apply one test: can you extract a single, self-contained, attributable claim from the first sentence of each major section? If the answer is no — if the section opens with context-setting, qualification, or scene-setting before reaching the point — that is the structural reason Gemini read it and moved on.
Rewrite those openers. Lead with the claim. Make the provenance explicit — author, date, basis for the assertion. The goal is not to game a system. The goal is to make your content usable as evidence, not just readable as content.
Sources
- Madhavan, Krishna, Knut Risvik, and Meenaz Merchant. "Evolving Role of the Index: From Ranking Pages to Supporting Answers." Bing Search Blog, Microsoft AI, 6 May 2026, blogs.bing.com/search/May-2026/Evolving-role-of-the-index-From-ranking-pages-to-supporting-answers.
- Schwartz, Barry. "Microsoft Bing on Search Indexing vs. Grounding Indexing." Search Engine Roundtable, 7 May 2026, seroundtable.com/bing-search-indexing-vs-grounding-indexing-41284.html.
- "Deep Research Listing All Sources as 'Read but Not Used.'" Gemini Apps Community, Google, 20 Nov. 2025, support.google.com/gemini/thread/389104437.
- "AI Visibility in 2025: How Gemini, ChatGPT, and Perplexity Cite Brands." Yext Blog, Yext, 29 Oct. 2025, yext.com/blog/ai-visibility-in-2025-how-gemini-chatgpt-perplexity-cite-brands.