When a potential client asks Claude to analyze your industry or compare the leading operators, the model builds the answer starting from what it has learned — and if your data isn't in those sources, you never show up in that analysis. Having an optimized website isn't enough: Claude trusts industry bodies, technical publications, and structured data on institutional sources, not your website. Building your presence in those datasets is the concrete way to appear in the analyses your potential clients are already reading today.
Anthropic is stricter than OpenAI on source quality. You should know, in fact, that Claude prefers 1 institutional source over 100 blogs.
I’m telling you this right away because it’s the starting point for everything else in this article. If you sell mountain guiding services in the Dolomites or run a refuge above Cortina d’Ampezzo, and you wonder why Claude never mentions you while ChatGPT pulls you up now and then, the answer isn’t “write more content.” It’s understanding that Claude has a different trust threshold, and that within that threshold there’s enormous room for those who publish structured data and technical documentation about their industry.
In this article I’ll explain how Claude’s Artifacts mechanism works (the mode in which the model builds documents, code, and analyses inside the conversation), why it rewards those who have data published cleanly, and what you can do in your industry to end up inside those analyses.
What Claude does when a user asks it for an analysis
Open Claude and ask: “give me an analysis of the market for UIAGM-certified mountain guides in Italy, focused on the Dolomites, output in a table.” Claude doesn’t respond with a paragraph. It opens an Artifact — a separate panel where it generates the table — and fills it with names of mountaineering schools, guide cooperatives, average daily prices, required certifications. Those names come from the model: from training, from the documents Claude saw during learning and refinement.
In the research on how Anthropic trains its models, there’s an important detail about the composition of the data Claude uses to generate answers like those.
These conversations are similar in distribution to, but distinct from, those appearing in the PM and RL training data.
Translated: the conversations the model sees during evaluation have a distribution similar to those in training, but they aren’t identical. It’s a technical detail, but for your business it means one specific thing: Claude generalizes from patterns learned on high-quality data, it doesn’t retrieve your web page in real time. If your industry data didn’t enter the pattern, it doesn’t come out in the Artifact.
The practical consequence is that the game isn’t played only on classic SEO or backlinks. It’s played on how readable, structured, and citable your data is from the sources Anthropic considers reliable: institutional bodies, industry federations, technical publications, official registries.
Why Claude filters you more than ChatGPT
In the previous articles in this series I talked about how E-E-A-T is reinterpreted by AI engines and how author recognition as an entity works. Those mechanisms apply to all models, but Claude applies them with a higher threshold.
The reason is structural: Anthropic built its training around a principle of caution about sources. It means that when the model has to generate an analysis of the “Dolomites mountain guides” industry, it discards amateur blogs, forums, and sites without an identifiable author more aggressively, and leans more on sources like the National College of Mountain Guides, the CAI, the international UIAGM, ISTAT for tourism data, and the publications of the Autonomous Provinces of Trento and Bolzano.
For a refuge manager above Misurina or a mountaineering school in San Vito di Cadore, this is an opportunity: if your data reaches institutional sources, you have a privileged channel into the analyses Claude generates.
Data scattered only across your own website.
The test you can run in 15 minutes
Here’s a simple audit you can do today without paid tools.
- Open Claude and run 5 queries like: “list of managed refuges in the Sesto Dolomites,” “UIAGM mountain guides in the province of Belluno,” “mountaineering schools in Cortina d’Ampezzo,” “average daily rates for a Dolomites mountain guide,” “comparison of historic refuges Cortina vs Val Gardena.”
- For each answer, note whether your brand appears. If so, in what position (first named, in the middle, last).
- Ask Claude for the source of what it just said. It won’t always give precise URLs (it’s not Perplexity), but it will say things like “institutional sources of the Italian mountaineering industry” or cite reference categories.
- Open Google’s Rich Results Test, paste your homepage URL, and check whether you have valid “Organization” and “LocalBusiness” schema. If it’s not there, you’ve found a first gap.
- Go to Wikidata and search for the name of your refuge or your guiding company. If there’s no entry, you’ve found a second gap.
Binary threshold: if you appear in 0 out of 5 queries, you’re invisible to Claude. If you appear in 1-2 out of 5, you exist but you’re marginal. If you appear in 3+ out of 5, you’re inside the pattern. An indicative test, not a study: but the signal is clear.
Publish an “operational data” page with altitude, number of beds, opening period, staff certifications, average rates, contacts — in a readable format (a clean HTML table, not images).
The pattern I saw across 20 mountain queries
I ran a longitudinal observation over the past few months, comparing how Claude, ChatGPT, and Perplexity respond to the same type of query about the Italian alpine world. 20 themed queries (refuges, guides, schools, via ferratas, comparisons between areas), repeated across the three engines.
The pattern that emerged is clear-cut: ChatGPT cites 8-12 different names per query, mixing institutional sources and enthusiast blogs. Perplexity cites 5-8 sources with URLs, often drawing from TripAdvisor, trekking blogs, outdoor magazines. Claude cites 3-5 names per query, and in 16 out of 20 cases those names traced back to entities with a strong presence on institutional sources: guide colleges, federations, provincial registries, or historic brands with a Wikipedia or Wikidata entry.
Stated limitations: not a large sample, queries in Italian, observation period of a few weeks. But the pattern is too consistent to be a fluke. Claude rewards those who have verifiable institutional authority, not those who have more content.
For the refuge manager or the mountain guide reading this, the takeaway is simple: your presence on AI engines depends less on how much you publish on your website and more on how much your data circulates in the sources Anthropic trained to consider reliable.
The mistakes I see most often
When I run a GEO audit for a business in the mountain sector, I almost always find the same patterns.
- Data scattered only across your own website. The refuge has the page with altitude, number of beds, opening period, rates — but that data doesn’t exist anywhere else. For Claude, unreconciled data = questionable data.
- No Wikidata entry or Wikipedia article. For historic refuges (we’re talking about structures with 80-100 years of history) it’s a self-defeating gap. A Wikidata entry can be filled in for free and becomes a node that many AI models use as an anchor.
- Membership in the guide college present but no structured author page. The mountain guide holds the UIAGM qualification but on their website there’s no “About” page with structured Person data, certification year, specializations. Without those signals, Claude doesn’t build the entity “mountain guide X.”
- Press releases and industry data given to magazines without a return link. The refuge has a record season, tells an outdoor magazine, an article comes out. But the article doesn’t link the refuge’s website and doesn’t cite structured data. That handoff is lost.
What to do in the next 30 days
If you run a refuge, a mountaineering school, or you’re a UIAGM guide in the Dolomites, these are the three actions with the highest effort/result ratio.
- Publish an “operational data” page with altitude, number of beds, opening period, staff certifications, average rates, contacts — in a readable format (a clean HTML table, not images). Mark it up with LocalBusiness or Lodging schema via the Rich Results Test.
- Create or update the Wikidata entry for the structure/company. Link coordinates, founding date, current manager, historic events. It’s the point through which reconciled entities pass.
- Send your structured data to the guide college of your province, to your reference CAI section, and to the publications of the Autonomous Provinces. When a piece of data appears on an institutional source, Claude absorbs it in subsequent training cycles.
This is only a first step; real analysis requires professional tools and a competitive mapping across all the relevant queries in your industry. But if you start here, within a few months you enter the pattern.
The thread that ties it all together
Coming back to the opening point: Claude filters more strictly than OpenAI on source quality, and the way to exist inside that filter isn’t to fight it, it’s to populate it with your data. The analyses Claude generates in its Artifacts are a mirror of who made it all the way to the institutional sources. Your visibility in AI answers goes through there.
In the following articles in this series on Claude and Anthropic we’ll look at how Claude’s citation system works, why Anthropic handles web search differently from ChatGPT, and which metrics to use to measure your positioning on the model. If you want to brush up on the fundamentals, start from how Google’s knowledge graph connects to your brand’s entity: it’s the first brick that Claude uses too.
