If your copy contains phrases like "absolute leader", "unmatched quality" or "the best in the industry", Claude excludes you automatically from its answers — not as a visible punishment, it simply never cites you. It isn't a matter of taste: it's a filter built into the way Claude assesses the reliability of sources, and it works in a binary way. You're paying a copywriter for text that the most widely used AI in professional settings refuses to read. There are a few surgical fixes that remove the problem without stripping the commercial strength from the text.
Claude is the most conservative AI engine, even paranoid about sources. Like an editor who triple-checks every citation before publishing. Understanding this filter is understanding how to get through it.
Let me explain why this isn’t a technical detail. If you make high-end handcrafted eyewear in the Dolomites and your site is full of “world leader”, “unmatched quality”, “the best of made in Italy”, you’re doing work that might still hold up on Google, but that on Claude excludes you automatically from its answers. No warning, no visible penalty: simply, when someone asks “handmade eyewear makers in the Dolomites”, Claude cites someone else.
In my articles in the AI platforms series I’ve already explained how ChatGPT and Gemini handle citations. Claude is a different animal: it has an ethical filter coded upstream that cuts out everything that sounds like aggressive marketing. And this filter has a precise name, documented by Anthropic in 2022.
The principle that governs Claude
In the world of AI alignment research, the Anthropic team published in 2022 the paper that describes exactly how Claude was trained to filter content. The method is called Constitutional AI, and it’s the reason Claude responds in such a measured way compared with its competitors.
The only human oversight is provided through a list of rules or principles, and so we refer to the method as ‘Constitutional AI’.
Translated: the model isn’t corrected case by case by human labelers, it follows a list of written principles. A constitution, indeed. Every answer is evaluated against those principles, and if a source — your site, one of your articles, one of your product pages — contains elements that violate the principles of honesty, restraint and verifiability, the model prefers not to cite it.
The operational consequence for you is direct: if your copywriting is built on undemonstrable superlatives, Claude reads it, processes it, but when it has to recommend sources to the end reader it leaves you out. Not out of algorithmic spite, by design.
The same paper adds an important nuance about the shift from a model trained with human feedback to one trained with feedback from AI:
In prior work we discussed how to train HH RLHF models, whereby the role of human feedback is to provide comparison labels for preference modeling on both helpfulness and harmlessness.
The message is that the “harmlessness” component — harmlessness, non-manipulation — has been progressively shifted from human judgment to an automated judgment based on the principles. Which means the filter has become more systematic, faster and less negotiable.
Why this comes before any other optimization
In previous articles I talked to you about E-E-A-T for AI and about implicit reference weight: these are the signals that push a brand up in AI citations. On Claude there’s an earlier step. Before looking at how authoritative you are, the model looks at whether it can cite you without violating its own constitution.
It’s a binary filter: you either pass or you don’t. There’s no middle ground like “Claude cites you less”. If on your page it finds three exaggerated claims and no verifiable evidence, it excludes you from the set of sources usable for that answer. Your work on content structure, on entities, on authority — all important — only works if you pass the upstream filter.
The company history written in superlatives.
The test you can run in ten minutes
Let me show you the experiment I recommend you replicate in your own sector. For a high-end handmade eyewear maker in the Belluno area, the plausible queries are these:
- “best Italian handcrafted eyewear makers”
- “high-end handmade made in Italy eyewear”
- “niche eyewear companies in the Belluno area”
- “bespoke eyewear makers in Italy”
Take these queries, run them in parallel on ChatGPT and Claude, and compare the cited sources. Not the answers, the sources themselves. Note down the name and URL of every source cited by both.
Then open the cited sites and look at the homepage. You don’t need tools: just read the tone. You’ll find a fairly clear pattern.
Note down every absolute adjective: leader, best, unique, excellence, unmatched.
The test I ran myself
I ran fifteen queries on the luxury eyewear sector across ChatGPT and Claude, half in Italian and half in English, over the past two weeks. Small sample, indicative test, not a study: but the pattern came out clearly.
ChatGPT cited on average six sources per answer, with a mix of manufacturers’ official sites, trade magazines (Vogue Eyewear, L’Officiel), specialist optical blogs and a few e-commerce sites. Claude cited on average three and a half, with a clear preference for: institutional pages of the Belluno Eyewear District, articles from general-interest outlets with a verifiable editorial team, Wikipedia pages.
The figure I find most interesting: across fifteen queries, Claude cited a manufacturer’s site directly only twice. ChatGPT eleven times. I attributed the difference to the language of the homepages: the manufacturers cited by Claude had “about us” pages with people’s names, verifiable history, zero superlatives. The others didn’t.
I’ll repeat the caveat: a sample of fifteen queries, my own test, run on a specific sector. But if you replicate it on yours, I’ll bet the pattern holds.
The mistakes I see most often
When I audit SME sites with a Claude-first eye, these are the four patterns that show up almost every time:
- The generic, undemonstrable superlative. “Absolute leader in the industry”, “excellence without compromise”, “unmatched quality”. Claude reads “unverifiable claim” and files it away.
- The anonymous testimonial. “Our customers have chosen us for 20 years”. No name, no company, no date. To Claude it’s noise, not evidence.
- The company history written in superlatives. “Since 1987 we’ve set the industry standard”. Written like that, with zero external sources to confirm it, Claude treats it as a self-declared advertising statement.
- The uncontextualized “number one”. “Among the top in Italy”. Top at what, on what metric, measured by whom. If there’s no answer, the filter trips.
It’s not that Claude is allergic to your company pride. It’s that, according to its constitution, a statement without verifiable evidence isn’t a statement usable in an answer.
The operational audit in three steps
If you want to work on the Claude filter, these are the concrete steps in the order I recommend.
Step one: inventory of superlatives. Open the homepage, about us, main product pages. Note down every absolute adjective: leader, best, unique, excellence, unmatched. Goal: cut them by 70-80%. The ones that remain must have an external source (a dated award, a certification, a published ranking) linked to them.
Step two: replacement with specific data. “We produce 4,000 frames a year in our factory in Longarone, each frame requires on average 60 manual operations” is worth ten times “artisan quality without compromise”. The first is verifiable, the second isn’t.
Step three: verify entities and authors. Names of the people who run the company, with role and history. This ties directly into what I told you about in author entity recognition: Claude rewards identifiable entities, and a proper name with a verifiable biography is the strongest signal you can give.
For technical checks on schema and structured data, the real analysis requires professional tools, but you can do an honest first check with Google’s Rich Results Test: you paste your homepage URL, you see whether it recognizes “Organization” with name, founder, headquarters. If it doesn’t, start there.
Where all this takes you
The point of this article isn’t to scare you about Claude. It’s to make you understand that the AI platform most attentive to the restraint of its sources is also the one that, if you pass the filter, cites you more stably over time. The sources Claude selects once tend to stay in its reference set longer than the volatile sources ChatGPT picks up and drops.
Working on visibility in AI answers also means this: don’t write for the most generous engine, write for the most selective one. If you pass Claude, the others come on their own.
