Anthropic's Jacobian Lens Shows Why AI Visibility Is Measured Too Late
Anthropic's Jacobian Lens shows why output-only AI visibility dashboards measure too late.
Anthropic's Jacobian Lens is not an AI search product. That is why it matters. The July 6 research shows that important model behavior can happen before the visible answer. For AI visibility, the lesson is blunt: measuring only the final citation or referral click means measuring the system too late.
Anthropic's Jacobian Lens exposes the pre-answer layer
The Jacobian Lens reads internal model representations before they become output tokens. Anthropic's July 6 research note describes a "J-space" inside Claude: a small set of internal patterns linked to words the model may be ready to verbalize, even when those words never appear in the prompt or the answer (Anthropic).
The companion paper, "Verbalizable Representations Form a Global Workspace in Language Models," is more precise. It says the Jacobian Lens identifies internal representations that are "verbalizable" by estimating how an activation at an intermediate layer affects present and future output tokens (Transformer Circuits).
That distinction is the useful part for search operators. The model may carry intermediate concepts, planned words, source assessments, or task constraints before those signals become text.
Anthropic's open-source reference implementation says the same thing operationally: the lens "reads out what an internal activation is disposed to make the model say" by transporting a residual-stream vector into the final-layer basis and decoding it with the model's own unembedding (GitHub).
AI visibility tools still start at the visible answer
Most AI visibility measurement starts after the machine has already made its selection. Prompt tracking checks whether a brand appears in generated answers. Referral analytics checks whether an AI surface sent traffic. Rank-style dashboards compare visible outcomes across engines.
Those are useful outputs. They are not the whole system.
| Measurement layer | What it sees | What it misses |
|---|---|---|
| Prompt tracking | Whether the final answer names or cites a brand | Why that source entered the answer path |
| Referral analytics | Clicks from ChatGPT, Perplexity, Gemini, or Claude | Answer-only influence with no click |
| AI crawler logs | Retrieval and agent access attempts | Whether retrieved content was selected or cited |
| Source eligibility controls | Whether engines are allowed to use a page | How the answer weights competing sources |
| Citation architecture | Whether a page contains extractable, attributable claims | The hidden model state before output |
The Jacobian Lens does not expose Claude's production source-selection pipeline. It proves a narrower point: visible output is downstream of internal representations that can be inspected, perturbed, and measured in research settings.
So the operator move is not to pretend a brand can inspect every hidden model state. The move is to stop treating the visible answer as the first measurable event.
Search engines are already moving measurement upstream
AI search platforms are separating access, training, search inclusion, and user-triggered retrieval. OpenAI's crawler documentation distinguishes OAI-SearchBot, GPTBot, OAI-AdsBot, and ChatGPT-User. OAI-SearchBot is for surfacing websites in ChatGPT search features; GPTBot is for training; ChatGPT-User is tied to user actions and is not used to determine whether content appears in Search (OpenAI Developers).
That means source visibility begins before the answer. A site can allow search inclusion while blocking training. A robots.txt decision can change ChatGPT search eligibility. Server logs can show retrieval attempts before analytics shows a referral.
Google is pushing in the same direction. Its June 3 announcement introduced new tools for website owners navigating AI in Search, including Search Console controls, performance insights, and updated participation options for generative Search experiences (Google).
The pattern is clear: AI search measurement is becoming a control-plane problem. The answer is the last artifact in a chain that starts with crawl permission, source structure, entity clarity, and retrieval eligibility.
Machine Relations treats the answer as the last mile
The better model is source-first AI visibility, not output-first reporting. The Machine Relations framework describes AI visibility as a system of earned authority, entity clarity, citation architecture, answer-surface distribution, and measurement. That fits the Jacobian Lens lesson better than a rank tracker pretending the answer appeared from nowhere.
In this frame, a citation is the final mile of a source chain:
- The crawler or agent can access the page.
- The page contains extractable claims with clear attribution.
- The entity graph connects the brand, category, author, source, and claim.
- The answer engine retrieves competing sources.
- The model selects and verbalizes a source-backed answer.
- The user may or may not click.
The Machine Relations Stack puts measurement after earned authority, entity clarity, citation architecture, and distribution for this reason. The scoreboard cannot repair a weak source chain.
That is also why share of citation is a better metric than share of AI mentions. A mention can be loose. A citation points to a source the answer engine treated as evidence.
The practical AI visibility audit now has four layers
AI visibility should be audited from source eligibility to final citation, not from final answer alone. The Jacobian Lens is research, not a marketing dashboard. But it sharpens the audit.
Start with the source path:
- Crawler eligibility. Are OAI-SearchBot, Googlebot, PerplexityBot, ClaudeBot, and other relevant agents allowed to access the important pages?
- Retrieval evidence. Are AI crawlers and user-triggered agents actually requesting those URLs in server logs?
- Citation architecture. Do the pages answer the target query in the first 40-60 words, with specific sources, named entities, and structured tables where useful?
- Answer evidence. Do ChatGPT, Perplexity, Gemini, Claude, and Google AI surfaces cite or name the source for category-level questions?
AuthorityTech's public publication intelligence data is useful because it tracks which publications AI engines cite, rather than assuming media authority from human traffic alone. Separately, Machine Relations was coined by Jaxon Parrott in 2024 to describe the shift from human-mediated discovery to machine-mediated discovery.
The near-term mistake is simple: teams will buy output dashboards and call the channel measured. The smarter team will measure the machine path before the dashboard event exists.
FAQ
What is Anthropic's Jacobian Lens?
Anthropic's Jacobian Lens is an interpretability technique for reading what an internal model activation is disposed to make the model say. The July 2026 research uses it to identify J-space, a set of verbalizable internal representations that can support silent reasoning before output.
Does the Jacobian Lens measure AI search rankings?
No. The Jacobian Lens does not measure search rankings, citations, or brand visibility. Its relevance is indirect: it shows that important model behavior exists before the visible answer, which makes output-only AI visibility measurement structurally incomplete.
What should brands measure before AI citations?
Brands should measure crawler eligibility, AI bot retrieval, entity clarity, extractable claim structure, and source corroboration before final citations. Those layers determine whether a page can plausibly become a cited source when an answer engine builds a response.
How can a team benchmark its current AI visibility?
Use answer evidence and source evidence together. Prompt checks show whether a brand appears; crawler logs show whether machines retrieve the source; citation architecture shows whether the page is built for extraction. A baseline AI visibility audit can separate those layers before budget goes into more content.