Ranking #1 in traditional search is no longer the final goal. In the age of AI search, your content must survive a "grounding pipeline"—a complex extractive filter that decides which exact sentences will feed the generative model and which will be thrown away. Building on the pioneering SRO research from Dan Petrovic and DEJAN AI, we've analyzed how Gemini's grounding pipeline creates the winners and losers of AI visibility.
Total word budget per query distributed across all sources
Overall content survival rate through the AI filter
Median words a #1 ranking source can expect to contribute
Selection Rate: The New CTR
Selection Rate Optimization (SRO) addresses a fundamental shift in user behavior. Where traditional SEO focused on Click-Through Rate (CTR), SRO focuses on Selection Rate (SR). This is the measure of how often an AI system selects a specific source from the results it has retrieved.
In our Bay Area industrial audits, we consistenly see that ranking at the top of Google is only step one. If Gemini retrieves your data on "precision CNC machining" but chooses the competitor's price table for its final answer, your selection rate is zero—and your brand remains invisible in AI Mode.
How the Gemini Grounding Pipeline Operates
Google's Gemini grounding pipeline operates in a rigorous, multi-step sequence that decomposes a human prompt into machine-operable sub-queries.
- Query Fanout: The model breaks a complex prompt (e.g., "Find a semiconductor supplier with cleanroom ISO-5") into single-intent sub-queries.
- Retrieval: Google’s index returns 5–20 high-relevance sources for each sub-query.
- Extractive Summarization: This is critical. Google pulls exact sentences from your page—it does not paraphrase. Each sentence is scored for semantic proximity to the fanout query.
- Context Assembly: High-scoring "grounding chunks" are assembled into snippets and supplied to the model.
Key Insight: Because snippets are query-dependent, the same page will yield entirely different extractions for different user prompts. You must optimize for the "fanout intent," not just the primary keyword.
The ~2,000 Word Grounding Budget
Research indicates that each AI query operates under a fixed grounding budget of approximately 2,000 words. This budget is remarkably consistent, and your share of it is determined by your rank.
| Rank | Median Word Share | Percentage of Budget |
|---|---|---|
| #1 | 531 words | 28% |
| #2 | 433 words | 23% |
| #3 | 378 words | 20% |
| #4 | 330 words | 17% |
| #5 | 266 words | 13% |
The #1 result gets more than twice the real estate of the #5 result. You're competing for share of a fixed pie.
Density Beats Length
One of the most pivotal findings is the content "survival rate." On average, only 32%of a page’s content survives the grounding filter. However, this varies dramatically based on total page length:
- Short Pages (<1K words): 61% content survival rate.
- Medium Pages (1-2K words): 35% content survival rate.
- Long Pages (3K+ words): 13% content survival rate.
Grounding typically plateaus at about 540 words. If you are an industrial supplier seeking to be findable for complex specs, a dense 800-word product page is significantly more effective than a verbose 4,000-word whitepaper.
What Gets Selection vs. What Gets Filtered
The AI grounding algorithm is highly tuned to find factual, actionable evidence. In our work with Bay Area manufacturers, we've identified the specific elements that trigger selection:
Selected (High SR)
- ✓Factual Precision: AS9100D, ISO-5, ITAR, NADCAP certifications
- ✓Feature Specificity: exact tolerances (e.g., +/- 0.0001")
- ✓Lead Information: Declarative statements in the first 2 paragraphs
- ✓Service Data: Pricing, MOQs, and concrete lead times
Excluded (Low SR)
- ✕Structural Noise: ToC entries, headers, and navigation link text
- ✕Marketing Fluff: Generic claims like "world-class quality"
- ✕Boilerplate: Legal copy, copyright footers, and generic menus
- ✕Promotions: Time-sensitive sales or temporary banners
Primary Bias: The Worldview Factor
The biggest lever on SR isn't on-page—it's Primary Bias. This is the model’s internal perception of your brand’s relevance for a specific topic, formed during its initial training and fine-tuning.
If an LLM "believes" a brand is the authority in "Silicon Valley hardware prototyping," that brand receives an implicit boost in Selection Rate when its pages are retrieved. Detecting this bias requires specialized tools like the Tree Walker Algorithm, which identifies "high-uncertainty" tokens where a model is least confident about associating your brand with a concept.
Practical SRO Strategies for 2026
- Front-Load Authority: The grounding pipeline has a heavy positional bias. Put your most important specs and value propositions at the absolute top of the page.
- Write Grounding-Friendly Copy: Each sentence should be self-contained and factual. Avoid "it" and "this"; use specific nouns that anchor the sentence in context.
- Eliminate Noise: Clean, well-structured HTML isn't just for developers—it's for the AI extraction layer. Remove structural artifacts that compete for grounding share.
- Address Fanout Angles: Structure your product pages to answer multiple "intent facets"—quality, compliance, location, and capacity.
SRO is the new battleground for industrial visibility.
As AI search continues to abstract the web, your content's "survival rate" is the only metric that determines whether a buyer sees your brand or your competitor's. If you aren't optimizing for the grounding pipeline, you are effectively invisible.
