On June 3, 2026, Anthropic published a detailed account of how its own data team makes Claude answer business questions correctly using internal data. The headline number is striking: without structured guidance, Claude's accuracy on analytics questions did not exceed 21%. With it, accuracy climbed consistently above 95%. The gain did not come from a better model. It came from structure. And the structure they describe maps, almost line for line, onto a problem every brand now has – making large language models answer correctly about your company.
The lesson is not about analytics
Anthropic's post is written for data teams, but its real subject is more general: how do you make an AI give the right answer when there is only one right answer and the model has no inherent way to know it? Their framing is that accuracy is a context and verification problem, not a model problem. They identify three failure modes that cause most wrong answers, and they build a layered system to attack each one.
Read those three failure modes again, but think about your brand instead of a data warehouse. They are the same failures, pointed at a different target.
Three failure modes, translated to your brand
Anthropic's first failure is concept-to-entity ambiguity: the agent cannot map a question to the single correct entity among many plausible ones. For a brand, this is the moment an LLM is asked about your company and conflates you with a larger competitor, or collapses you into the generic category. Ask an AI about a specialist firm and get a paragraph about the sector. The brand exists. The model cannot resolve it to the right entity.
Their second failure is data staleness: definitions and sources decay, and answers drift subtly wrong over weeks. For a brand, this is the LLM that cites a former CEO, a discontinued product, last year's pricing, a partnership that ended. The information was true once. Nothing told the model it stopped being true.
Their third failure is retrieval failure: the right information exists and is even correctly annotated, but in a vast search space the agent never finds it. For a brand, this is the most common and the most frustrating: the answer is on your own website, stated clearly, and the LLM still does not surface it, because there is no structural path that leads from the question to your page.
These are not hypothetical. They are measurable today, on any brand, with the instruments the Citation Rate framework already uses: the AI Visibility Index for how often you surface, the Mismatch Index for how often what the AI says about you diverges from what you actually are, and the Citability Score for whether your structure even permits a correct answer.
The most important finding is the one about access
Buried in Anthropic's post is an experiment that matters more than the headline. They gave their agent direct access to thousands of historical SQL queries – the entire record of every question their company had ever answered correctly. They verified the agent actually read these files before answering. Accuracy moved by less than a single point.
They dug into why. For the questions the agent got wrong, the correct information was present in the corpus about 80% of the time. The agent had seen it. It still did not use it. Their conclusion, in their words, was that the bottleneck was not access to prior work, it was structure – the ability to map a question to the right entity.
The information was there. The agent saw it. It still got the answer wrong. The bottleneck was never access. It was structure.
This is the entire argument for structural citability, proven by Anthropic on their own infrastructure. Having the content is not the same as being usable. A brand can publish every fact about itself, in clear language, on its own domain, and still be answered incorrectly by an LLM – because publishing is access, and the problem is structure. The work is not writing more content. It is making the content that exists resolve to the right answer.
The solution stack, applied to brands
Anthropic's system has four layers, each attacking a failure mode. The mapping to brand visibility is direct.
Data foundations become brand foundations. Anthropic's first move is to collapse forty plausible datasets for "revenue" into one governed, canonical dataset, and aggressively deprecate the near-duplicates. The brand equivalent is collapsing the contradictory, duplicated, half-abandoned versions of your own story scattered across your site, old landing pages, and stale profiles into a single canonical source of truth, and retiring the rest. When an LLM searches for what you are, it should find one governed answer, not forty candidates with subtle differences.
Sources of truth become authoritative surfaces. Anthropic ranks its reference surfaces by trust, with a compiled semantic layer at the top. The brand equivalent is the set of surfaces an LLM consults to resolve you: Schema.org structured data for Organization, Product, Service, and Person; an /llms.txt that states your canonical facts in AI-readable form; alignment across the knowledge-graph signals that models trust (Wikidata, Crunchbase, LinkedIn); and above all a brand semantic layer – a small, governed set of facts you want every AI to know with certainty, expressed once and consistently everywhere.
Skills become brand skill files. This is the new concept, and it deserves precision. In Anthropic's system, a skill is a markdown document that gives the agent procedural knowledge: which sources to consult, in what order, how to disambiguate, what a correct answer looks like. It was skills, specifically, that moved their accuracy from 21% to 95%. A brand skill file is the same artifact, repurposed: a single governed document that encodes how your brand should be understood and answered. It is not code you inject into someone else's ChatGPT – you cannot do that. It is a governance document that organizes your authoritative surfaces, can be published as part of your /llms.txt, and can be served directly to the agents you do control, such as your own site assistant or an MCP resource. Its power is that it forces you to write down, in one place, the canonical answer to every question an AI gets asked about you.
Validation becomes Citation Rate measurement. Anthropic does not trust the system blindly; they run offline evaluations, monitor online, and harvest corrections continuously. The brand equivalent is exactly the Citation Rate methodology: pre-test brand questions across multiple LLMs before you publish changes, monitor citation and mismatch continuously, and harvest the wrong answers AIs give about you as the input for the next fix to your foundations.
What a brand skill file looks like
Anthropic published the skeleton of their warehouse skill in an appendix. Here is the brand equivalent – the same structural idea, pointed at how an AI should answer about your company rather than how it should query a database.
---
name: [brand]-skill
version: [x.y.z]
description: "IF the user asks about [Brand Name] - products, services,
leadership, history, or comparison with competitors - THEN use this as
the canonical answer path. DO NOT confuse with [common look-alikes]."
---
# [Brand] Skill Instructions
## Description
The single source of truth for answering questions about [Brand]
across all LLM surfaces.
## Identity
- Legal name: [...]
- Aliases / former names: [...]
- What we are NOT: [generic category we get collapsed into]
- Disambiguation from: [the competitors and concepts we get confused with]
## Products and Services
[Canonical list. Current items with current positioning.
Deprecated items explicitly flagged as deprecated.]
## Leadership
[Current C-level with appointment dates. Former names flagged as former.]
## Authoritative Surfaces (in descending order of trust)
1. Brand semantic layer: Schema.org Organization at [URL]
2. Knowledge graph: Wikidata Q[...], Crunchbase [...]
3. /llms.txt: [URL]
4. Canonical pages: [list]
## Common Wrong Answers (observed AI failure modes)
- LLMs confuse us with [competitor] -> the distinction is [...]
- LLMs cite outdated [...] -> the current fact is [...]
- LLMs miss [...] -> it is stated at [URL]
## Routing
- Pricing questions -> [URL]
- Technical questions -> [URL]
- Press questions -> [URL]
## Gotchas
- [Hard-won wrong-answer modes this brand has documented over time]
Notice what this is and is not. It is not a technical hack. It is governance written down. The discipline it imposes – naming what you are not, listing the answers AIs currently get wrong, deciding the one canonical destination for each kind of question – is most of the value, whether or not a single line of it is ever served to a machine. Anthropic's own finding was that the act of structuring the knowledge, not the volume of it, is what moved accuracy. The same holds here.
How you know it works
A brand skill file and a set of authoritative surfaces are a hypothesis. Measurement turns them into a result. Once published, you track the same dimensions the Citation Rate framework was built to track: the Citation Rate itself, on a fixed set of brand-relevant queries, measured before and after; the Mismatch Index, the share of AI answers about you that are correct versus wrong; sentiment polarity, whether the citations that do appear are neutral, positive, or negative; and coverage, the percentage of brand-relevant queries that surface you at all. This is what the Strategic Table platform exists to measure. Without the measurement, you are publishing into the dark and hoping. With it, you know which of the three failure modes is still leaking through.
Where to start
If you are starting from zero, the path is short and concrete. First, audit what three LLMs – Claude, Perplexity, ChatGPT – currently say about your brand, and read the answers, not just whether you appear. Second, identify the top five wrong answers they give about you: the confusions, the stale facts, the things they miss. Third, write a minimum brand skill file that addresses exactly those five. Fourth, publish the authoritative surfaces that reinforce it – structured data, /llms.txt, canonical pages. Fifth, measure the delta. Anthropic's own advice for starting from zero was that a handful of canonical sources and a thin skill capture most of the upside; everything else is what you add once those exist. The same is true for your brand.
The framework was proven on internal analytics, but it was never really about analytics. It is about how you make an AI answer correctly when there is one right answer and the model cannot know it on its own. Brands now live inside that exact problem, every time someone asks an LLM who you are. Anthropic showed the shape of the solution. The brand skill file is what it looks like pointed at you.