Open-Weight Models Are Half the AI Internet. Most AEO Strategies See None of It.

Twelve months ago, the combined global usage share of DeepSeek and Qwen was about 1%. By January 2026, it was roughly 15%. That is the fastest adoption curve in AI history. And almost none of it shows up in the dashboards marketing teams use to monitor AI visibility.

When most brands talk about getting recommended by AI, they mean ChatGPT, Gemini, Claude, and Perplexity. Those four are the visible ocean. Below the waterline, a second AI internet has been quietly forming, built on Llama 4, Qwen 3.6, DeepSeek V4, Gemma 4, GLM-5.1, and Mistral Small 4. It powers thousands of consumer apps, internal enterprise copilots, RAG systems behind support portals, and on-device assistants that never call a public API. Whatever you optimized for the closed-model leaderboards is invisible there. And that blind spot is getting bigger, not smaller.

The Two AI Internets

There is a useful way to think about the current AI landscape. The closed ecosystem (ChatGPT, Gemini, Claude, Perplexity) is the surface where AI visibility gets discussed because it is the surface where it can be measured. The open ecosystem is the harder half: six open-weight model families, hundreds of forks and fine-tunes, thousands of internal enterprise deployments, and consumer apps in markets most Western brands do not natively monitor. As of mid-2026, that open half accounts for at least a third of all AI inference and is growing faster than the closed half.

Hugging Face now hosts more than 2 million models, 500,000 datasets, and 1 million demo apps. 92.5% of model downloads are for sub-1B parameter models, which means they are being run locally and embedded into products, not called over an API. The mean downloaded model size jumped from 827M parameters in 2023 to 20.8B in 2025, driven by quantization and mixture-of-experts architectures. Open weights are not a research curiosity anymore. They are the substrate of an entire production layer of AI.

Why Open Models See Your Brand Differently

Both closed and open ecosystems start from Common Crawl. 64% of the 47 LLMs Mozilla analyzed used at least one filtered version of it. What diverges is what each model does with that raw data. Closed labs run private quality classifiers and layer in curated proprietary data. Open models like Llama 4, Qwen, and DeepSeek lean on public filtered datasets such as RedPajama-V2, FineWeb, and Dolma. These public pipelines are aggressive about deduplication and quality scoring, which is good for model performance but punishing for brands whose web presence is thin or lives mostly on their own domain.

Common Crawl itself is not a representative sample of the web. Its crawler prioritizes domains that are heavily linked to, so Facebook, Google, YouTube, and Wikipedia dominate the graph and long-tail industry sources get sparse coverage. When public filtering pipelines layer on top of that bias, brands that are well known to ChatGPT can vanish from open-weight model recall entirely.

Then there is quantization. The 4-bit and 8-bit quantized versions of Llama 4 and Qwen that ship into production apps trade fidelity for speed. The first capability to degrade is long-tail entity recall, which is exactly where most B2B brands live. Your brand might be in the full-precision weights and silently absent from the version a developer actually deployed.

The Self-Hosted Enterprise Blind Spot

Gartner reported a 340% jump in enterprise private LLM development through 2025. 65% of Fortune 500 companies have deployed LLM-based engagement tools. 44% of organizations cite data privacy as the top barrier to using public AI, the exact wedge that pushes them toward self-hosted Llama, Qwen, and Mistral deployments inside the corporate firewall.

From an AEO perspective, this is a category change, not an incremental one. When a Fortune 500 buyer asks their internal procurement copilot which vendors to shortlist, that copilot is often a fine-tuned Llama or Qwen running inside the corporate firewall. There is no API to monitor. No prompt logs to scrape. No public ranking to track. If your brand was not in the base model and is not in the RAG corpus the company assembled, you are not on the shortlist. And the buyer has no idea you exist.

The Dario Problem

There is a political dimension to this story that the AEO industry has largely ignored. Anthropic CEO Dario Amodei has been among the most vocal advocates for restricting open-weight models. He told the US Senate that open source AI is on a dangerous path. In June 2026, he published a lengthy essay arguing that governments should have the power to block AI models from deployment if they present unacceptable risks. OpenAI and Anthropic have since aligned in Washington on warning policymakers about the risks of powerful Chinese open-weight models specifically.

On a recent episode of Unsolicited Biz Advice, I called this out directly. Dario has been doom and apocalypse blogging about how AI is going to be the Terminator. The result is not just bad PR, it is a policy agenda that, if successful, would entrench the exact duopoly that is already extracting monopoly-level pricing from the market. Anthropic has reportedly throttled or banned certain use cases entirely. Companies working in genomic sciences have found they cannot do the work they need to do because Anthropic is afraid of bioweapons. That is not safety. That is a company using safety rhetoric to limit competition.

Nathan Lambert at Interconnects AI put it plainly in July 2026: the action Anthropic is effectively asking for is the wholesale banning of pretty much all Chinese open-weight models in the US. That would demolish the open model economy that is emerging with inference companies, fine-tuning companies, and new products built on top of them. It would also be futile. If the models are not banned in China as well, it is very easy for a bad actor to still use a banned open-weight model, which negates the safety argument entirely.

The critics of this agenda include David Sacks, the Trump administration AI adviser, who has said the scrutiny could amount to regulatory capture: rules intended to improve AI safety could instead entrench the largest companies by making it harder for competitors to release models. That framing is correct. The closed labs stand to benefit enormously from greater scrutiny of open-weight models. That conflict of interest should be front of mind whenever Anthropic or OpenAI publish safety arguments about open-source AI.

My view is simple: open always wins. It won in operating systems. It won in databases. It won in cloud infrastructure. The idea that AI will be the exception, that a handful of closed labs will permanently control access to intelligence, is not a safety argument. It is a business model dressed up as one.

RAG Changes the Visibility Math

Most production deployments of open-weight models are not bare. They sit behind a retrieval layer. Documents get embedded, queries hit a vector database, and the LLM synthesizes an answer from retrieved chunks plus its parametric knowledge. The closed-model conversation about training data presence matters less here. What matters is whether your content was ingested into the RAG index.

That sounds like an opportunity, and it is, partially. If a developer wires your docs into their internal copilot, you get cited every time. But the practical pattern is that companies index their own internal docs, the public docs of incumbents they already trust, and a curated knowledge base. New entrants are absent from both the model weights and the RAG corpus. The asymmetry compounds.

The brands that win this layer ship their content in formats that get pulled into RAG indexes by default. That means clean Markdown documentation, public API references with structured schemas, llms.txt files, and content that is easy to chunk and embed. If your top-of-funnel page is a JavaScript-rendered marketing site with no direct factual claims, you are invisible to retrieval too.

What to Actually Do About It

The practical response is not complicated, but it requires treating open-weight model visibility as a separate workstream from closed-model AEO, because the optimization levers are different.

Start by testing your visibility on at least one open model directly. Run Qwen 3.6 or Llama 4 against your category prompts via Hugging Face Inference, Together AI, Groq, or Fireworks. The cost is trivial and the signal is real. If your brand surfaces on ChatGPT but not on Llama 4, you have a training-data gap, not a hallucination problem.

Then audit your presence in the public sources that public filtering pipelines preserve. Wikipedia is not optional. Crunchbase, GitHub READMEs that describe what you do, structured Schema.org markup, and at least one canonical entry on each of the major review aggregators in your category are the sources RedPajama, FineWeb, and Dolma are biased toward keeping.

Publish for retrieval. Maintain a comprehensive plain-Markdown documentation site, expose llms.txt and llms-full.txt, and structure your highest-value pages so they survive chunking. If you only have one set of marketing pages and they are JavaScript-rendered, you are double-blind on the open layer.

If you sell internationally, earn citations in non-Western sources. A single Zhihu post or Baidu Baike entry does for Qwen what a TechCrunch piece does for ChatGPT. The same logic applies to NAVER for Korean models.

And accept that the self-hosted enterprise layer will stay opaque. You cannot monitor what runs inside a corporate firewall. What you can do is track inbound traffic for AI user agents you do see (PerplexityBot, GPTBot, ClaudeBot), monitor support tickets and sales calls for mentions of internal AI tools, and instrument your demo flow for unusual referral patterns. Self-hosted copilots cannot be queried, but their downstream behavior leaves traces.

The closed-model leaderboard is the easy half of AEO. The brands that recognize the open layer early and treat training-data presence as a portfolio play across both ecosystems will have a structural advantage over those that keep optimizing for the dashboards and wonder why their pipeline numbers do not move.

AEO Updates is published by The Prompt Group. Editorial decisions sit with the AEO Updates team, and any commercial relationship that touches a story is labelled on the page.