Everyone in GEO is optimizing for the 1T-parameter model in the demo. Almost nobody is optimizing for the model that actually decides whether you get mentioned. In 2026 that is rarely the same model. Between the user's question and the final answer sit a synthetic-data pipeline that decided what the model "knows," a format fine-tune that decided what shape the answer takes, and a 7B router that decided which sources were even worth reading. Three filters. Your brand has to survive all three, and most brands only ever measure the last one.
1. Synthetic data is the new crawl, and it only keeps what the teacher explains
Frontier labs hit the data wall. The fix is synthetic data: a strong teacher model generates reasoning paths, comparisons, and worked answers, and the student is trained on those. The under-discussed consequence is that your brand no longer enters the weights by being on the web. It enters by being in the teacher's explanations. If the teacher answers "best rank-tracking alternatives for AI search" and reasons through three vendors, those three get millions of synthetic examples. The fourth vendor, the one it paraphrased as "other tools," gets nothing.
That means today's citation share is tomorrow's training distribution. A brand that is consistently named in current frontier answers compounds; a brand that is consistently summarized without a name gets distilled out. This is why raw "mention count" is a lagging metric. The leading metric is named vs. paraphrased mentions across models, which is exactly what LLM Search Console separates in its competitor tracking. A paraphrase is not a win. It is a signal that you are about to be deleted from the next generation.
2. Fine-tuning is now about format, and format is a schema you either fit or don't
Fine-tuning has stopped being how models learn facts (RAG handles that) and started being how they learn shape: comparison tables, three-item shortlists, pros/cons blocks, JSON for downstream agents. Those formats have implicit schemas. A shortlist row needs a category, a price band, a primary use case, and a differentiator. If a model retrieves your page and cannot fill those four slots from it, the format fine-tune does not produce a half-empty row. It drops the row and picks a competitor whose page filled the slots.
The hidden link: format adherence is a retrieval filter. Your content is not competing on quality. It is competing on whether its attributes can be extracted into the answer template with zero inference. Pages that state "Pricing: from $X/month. Category: AI visibility monitoring. Best for: mid-market SEO teams" survive the template. Pages that say "flexible plans for teams of any size" do not.
3. The 7B model with great CoT beats the 70B one, and it has no room for you
The parameter-count lesson of 2026 is that a small model with strong chain-of-thought often outperforms a large "dumb" one. So agent stacks moved the cheap, high-frequency steps to small models: query classification, source reranking, tool selection. A 7B reranker running quantized on a single GPU now decides which ten of fifty retrieved chunks the big model gets to read.
Small models have small parametric memory. A 7B model has essentially no long-tail brand knowledge. It reranks on lexical and structural signals in the chunk itself: does this passage directly answer the sub-question, does it name entities, is it self-contained. If your brand only appears in a header and a footer while the body chunk talks about "our platform," the reranker scores the chunk as an anonymous passage and cuts it. The frontier model never sees you, not because it forgot you, but because a model that never knew you did the pruning.
4. Measuring the gap between the demo model and the deployed stack
Put these together and you get the GEO failure mode that no screenshot catches: a brand that appears in a direct ChatGPT test, is paraphrased in Perplexity, and is absent from an agentic product built on a small-model router, all for the same query on the same day. Testing one model tells you nothing about the pipeline.
The fix is to run the same prompt set across models on a schedule, track named vs. paraphrased mentions per model, and watch the divergence. When a small or mid-tier model drops you that a frontier model keeps, you have a chunk-level structure problem, not a content problem. When a frontier model paraphrases you, you have a training-distribution problem coming in six to twelve months. LLM Search Console runs exactly this loop: prompts across ChatGPT, Perplexity, Gemini, and Claude, competitor share of voice, and citation source tracking, so the divergence is visible before it becomes permanent.
Quick wins for GEO
Make every body chunk self-identifying. Name the brand and the category inside the paragraph, not just the H1. Rerankers score chunks, not pages.
Publish the schema, not the vibe. Category, pricing band, best-for, and one differentiator, stated as plain attribute sentences on your product page.
Track named vs. paraphrased. A paraphrase is a leading indicator of removal from synthetic training data.
Test small models, not just flagships. If a 7B or 8B open model cannot place you, the agentic stacks built on it won't either.
Ship comparison content that fills the template. Third-party comparisons with the same four attributes for you and your competitors feed the teacher model exactly what it needs to name you.
Schedule it. One-off checks measure a moment; the distributions you are fighting shift monthly.