We Audited 40 Mining Websites the Way AI Reads Them. Eight Were Ready

Lantern Comitas advises mining and critical-minerals companies across Africa, Europe, North America and South America. We recently audited 40 of the sector's websites to establish what an AI assistant receives when asked about these companies. Of the 34 sites we could measure, eight were ready to be understood and quoted. Twenty-seven of the 31 companies that publish instructions for automated visitors have made no decision about AI systems at all, and not one of the 34 publishes its facts in a form an AI assistant can lift cleanly.

Why this costs you something

Your potential clients are asking AI engines about your sector and your business. When your site is hard to read, the assistant does not give up. It answers from whatever it can find: a four-year-old trade article, a competitor's summary of your project, an aggregator still listing a chief executive who left in 2023.

You lose the framing, because the most accurate account of your operations, your own, is the one being skipped. You lose the visit, because the question is answered without anyone arriving on your site. And you inherit whatever is stale, often at the worst moment: due diligence, a supply question, a bad week in the press.

We measured whether these websites can be read and quoted, not what any assistant currently says about them. The first governs the second, and it is entirely within your control. We would rather report a narrow finding we can defend than a dramatic one we cannot.

How the audit works

Our audit platform sends its own agents to each website and records what an AI system would actually receive. It checks whether the site can be reached, retrying so a busy server is not mistaken for a refusal. It reads what the site's rules allow, records the content that arrives before any code runs in a browser, and tests whether a machine can tell who the company is. A control check stops error pages being counted as real content.

What we found

Twenty-seven of 31 mining companies have made no decision about AI systems at all. A robots.txt file is the plain-text list of rules at the root of a website telling automated visitors what they may read. In our audit, 31 of the 40 sites published one we could read, and 27 of those did not mention any of the 22 AI agents we checked for. Not one of the 31 blocks the AI systems that quote and link sources in live answers. The clearest position belonged to Pilbara Minerals and IGO, which block the systems that collect text for training and leave the quoting ones free.

Only 11 of the 34 measurable mining websites state who they are in a form a machine can use. Organisation schema is a small block of code, invisible to readers, that states a company's legal name, headquarters and logo so software does not have to guess. Eleven of 34 sites publish it, and only eight link it to the company's exchange listing and official profiles, the step that confirms which company this is. In a sector where a listed parent, its operating subsidiary and a similarly named company elsewhere are routinely confused, leaving identity to guesswork is expensive.

Eleven of the 34 sites give an AI assistant no main heading to work from. The main heading is the one line that tells any reader, human or machine, what a page is about. On a third of these homepages there are none, usually because the hero is a photograph or video and the company name sits inside a logo image. A person sees a striking front page. A machine receives an unlabelled document.

Not one of the 34 measurable mining websites publishes its facts as tagged questions and answers. Tagging your real questions and answers in the page code lets an assistant lift an answer cleanly instead of paraphrasing a brochure: reserves, jurisdictions, ownership, offtake. Across this audit and our audit of London-listed investor pages, the count is 0 of 53. Where nobody does it, it is the cheapest differentiation available.

The eight that were ready, and why they are not who you would expect

Eight sites cleared a deliberately modest bar: reachable, identified, described and with a working heading. They were Sigma Lithium, Eramet, Umicore, Kenmare Resources, Central Asia Metals, Atalaya Mining, First Quantum Minerals and Capstone Copper.

The global diversified majors performed worst of the four groups we audited: one of nine measurable major sites carries an identity statement, against four of seven Africa-focused miners. Large corporate sites are older, governed by more people and often mid-rebuild, while smaller ones tend to be recent builds where the basics arrive by default. The gap is neither strategic nor expensive: most of it is a fortnight of competent development work.

What to do about it

1. Check your robots.txt, and ask your technology team whether your security settings override it. Decide separately about AI that trains on your content and AI that quotes it, and give the decision an owner.
2. Put Organisation schema on the homepage, linked to your exchange listing and official profiles. It is an afternoon of developer time.
3. Give every page one heading that names the company and what it does.
4. Publish your real answers as tagged questions and answers, starting with the questions analysts and journalists ask most.

How Lantern Comitas helps

We specify these fixes, brief whoever runs your site and check the result, then recheck it after every rebuild, so a stale version does not sit unchallenged before investors and journalists.

Do you know which account of your company an investor or regulator receives first? Get in touch to discuss how Lantern Comitas can help.

Our free AI Visibility Audit runs these checks on your domain and shows where you sit against the 40 companies here.

Frequently asked questions

Does this mean we are invisible in AI answers?

Not on its own, and we have not claimed it. We measured whether an assistant can read your site and quote it accurately. Where it cannot, it uses another source, and you have no say in what that source says about you.

Does blocking AI crawlers protect our intellectual property?

Two kinds of AI agent visit websites. Training crawlers, such as GPTBot, collect text that may be used to train models, and blocking them is a defensible intellectual-property position. Citation crawlers, such as OAI-SearchBot and PerplexityBot, fetch pages so a live answer can quote and link them, and blocking them removes you from those answers without protecting anything already public. Decide the two questions separately.

We have a permissive robots.txt. Are we fine?

Not necessarily. Four of our 40 sites never served us a homepage, and three of those served one client while refusing another sending an identical request. That refusal comes from the security layer in front of the site, which is configured separately from robots.txt. Check both.

Is llms.txt worth implementing?

llms.txt is a proposed file that summarises a website for AI systems. It is not an adopted standard, and Google states that its search ignores it. Four of our 40 sites publish one. It is cheap and harmless, and it belongs well below an identity statement and a working heading on your list.

Can you tell whether an AI system actually reaches our site?

Not from outside. That needs your own evidence: server and security logs, Search Console and Bing Webmaster Tools. Anyone offering that answer from external testing alone is inferring it.

Annex: the findings in full

Sample: 40 mining and critical-minerals websites, audited 28 August 2026, in four groups of ten (global diversified majors, critical-minerals and battery-materials producers, Africa-focused miners, London-listed and mid-cap operators). Six sites are reported as not measured rather than scored zero: four never served a homepage and two returned a security challenge on every read. Per company records are held in our evidence archive.

Four companies named at least one AI agent in robots.txt. Vale names three and blocks none, Umicore names three and blocks two, and Pilbara Minerals and IGO each name eight and block all eight, every one a training crawler. The citation crawlers PerplexityBot, OAI-SearchBot, ChatGPT-User and Claude-SearchBot were named zero times across all 31 readable files. The four llms.txt publishers are Albemarle, Sibanye-Stillwater, Central Asia Metals and Harmony Gold. No site publishes an ai.txt file: an early pass counted four, but each was an error page answering as if it were a real file, which our control check caught.

Table 1. Headline results

Table 2. Page signals, of the 34 measurable sites

If you want the mechanics behind these gaps, we cover how AI engines decide which sources to cite, and what to do when the answer about you is wrong rather than absent. The equivalent findings for listed companies are in our audit of 24 London-listed investor pages.

If you would like the same audit run on your own domain, our AI Visibility Audit is the fastest starting point, and we work with mining and critical-minerals companies on what comes after it.