A model names your brand at 10:00. Same prompt, same provider, same account — at 10:01 it names three competitors and not you.

Nothing changed on your website. What changed was the path the request took through the network.

Most GEO advice stops at content. The mechanics that decide whether you get cited sit lower down: which experts fired, what precision the weights were served at, and how much compute the model spent thinking. Three intersections almost nobody writes about.

Routing is the first ranking layer

Modern frontier models are sparse. A mixture-of-experts model with 400B parameters might activate 30B per token. A learned router reads the incoming token embeddings and picks which expert blocks to fire.

Your brand name is a token sequence. That sequence is a routing input.

Here is where it gets expensive. "CodeDesign", "Code Design", "codedesign.org" and "Codedesign.org" tokenize into different sequences. Different sequences produce different router logits, which activate partly different expert subsets. If your entity string is fragmented across the web, the association between your brand and your category gets spread thinly across several experts instead of concentrated in a few. No single route holds a confident representation of you.

Entity consistency was always framed as a knowledge-graph hygiene problem. In a sparse model it is closer to a signal-to-noise problem inside the router.

Quantization decides who survives the cheap tier

Almost nobody tests the model that actually answers their buyers.

Flagship checkpoints serve a minority of inference traffic. The volume runs through quantized variants — 8-bit, 4-bit, sometimes lower — powering free tiers, mobile apps, autocomplete, summarization, and every batch job where latency and cost matter more than the last two points of benchmark score.

Quantization error is not distributed evenly. Rounding hits low-magnitude weights hardest, and low-magnitude weights are disproportionately where low-frequency facts live. High-frequency knowledge — Nike, AWS, Photoshop — is encoded redundantly across many high-magnitude paths and survives compression fine. A Series A SaaS company mentioned in four hundred documents does not have that redundancy.

So your AI visibility is partly a function of numerical precision. Test on the flagship and you will consistently overstate how visible you are.

Thinking mode is a second chance, and it changes who wins

Test-time compute inverts the failure mode.

In fast mode, the model answers from parametric memory. If your brand is not in the weights at sufficient strength, you are simply absent — there is no recovery step. In extended reasoning mode, the model decomposes the question, notices its own uncertainty, and reaches for retrieval or tools. Parametric absence turns into a search query.

The consequence is a split leaderboard. Brands with weak weight-level presence but strong structured, citable, retrievable web content score materially higher in thinking mode than in fast mode. Legacy brands coasting on training-data mass sometimes go the other way — reasoning mode checks their claims and finds thinner support than the prior suggested.

Most brand monitoring samples one mode, once, and reports the number as "AI visibility."

Why one screenshot is not a measurement

Stack the three variables. Expert route, precision tier, reasoning budget. None of them are visible in the answer text. All of them move between requests.

What you get from a single manual check in ChatGPT is one draw from a distribution whose variance you have not measured. Treating it as a result is how teams end up rewriting a landing page to fix a sampling artifact.

The fix is boring and mechanical: repeated sampling of the same prompts, across models, markets, and time, with the results logged rather than screenshotted. LLM Search Console runs scheduled scans across ChatGPT, Gemini, Perplexity and others, and tracks visibility score, brand position, and sentiment per prompt per model over time.

The citation log is the part that matters for the argument above. When a model falls back to retrieval — which is exactly what happens in thinking mode and in every grounded answer — the citation log shows you the URL and the snippet it leaned on. That tells you which of your pages survive the fallback path, and which of your competitors' pages are doing the work you assumed yours were doing.

Quick wins for GEO

Pick one entity string. Use it everywhere, byte-identical.
Test the free tier, not the flagship.
Run every prompt in both fast and thinking mode.
Sample ten times before believing any answer.
Write for the retrieval fallback: structured, dated, citable.
Track citations, not just mentions.

Take the five prompts your buyers actually type. Run each one ten times, across two models, in both reasoning modes, and write down the spread. If the variance is wider than you expected — and it will be — set up scheduled tracking at llmsearchconsole.com and stop measuring your brand by screenshot. Plans start at 49€/mo.

LLM Search Console

Your Brand Lives Inside One Expert: MoE…