Before we run the audit, we need to make sure we're asking the right questions about the right competitors to the right buyers. This document presents what we've learned about Anthropic's market — your job is to tell us what we got right, what we got wrong, and what we missed.
Before we measure how often AI answer engines cite Claude in the frontier-AI and enterprise-LLM space, these three signals tell us whether those crawlers can reach, refresh, and cleanly parse your content in the first place.
AI search is reshaping how buyers in the frontier-AI and enterprise-LLM space discover and evaluate solutions — increasingly, the first shortlist a CTO or CISO sees is assembled by an assistant, not a search page. Companies that establish citation visibility now gain a first-mover advantage that compounds: early citations become self-reinforcing as AI platforms learn which domains to trust. Anthropic enters this shift from an unusually strong position — a widely recognized frontier lab that is, notably, also the vendor behind one of the very crawlers deciding who gets cited.
This Foundation Review presents three things we need you to validate before the audit runs: the competitive landscape that shapes how we construct head-to-head queries, the buyer personas that determine which search intents we test, and the Layer 1 technical baseline that determines whether AI platforms can access and parse your content at all. Think of it as a shared read of your market — here's what we've concluded, and where we need your judgment to confirm or correct it.
The validation call is a working session with real stakes: your answers decide which inputs drive the buyer query set across the AI platforms we select. Two kinds of decisions come out of it — input validation (are the right entities in the right tiers, and is every persona a real buyer?) and engineering triage (which technical fixes can start before results come back?). The Pre-Call Checklist below aggregates every one of those questions into a single printable page.
Purpose This is the foundation for your GEO visibility audit in the frontier-AI / enterprise-LLM category. It captures our outside-in read of Anthropic's market — competitors, buyers, capabilities, and frustrations — plus the technical state of your site. Everything here becomes an input to the buyer queries we run against AI platforms. It is deliberately not a content-gap analysis; that requires the audit's response data and comes later.
Your Job Confirm what's right, correct what's wrong, and tell us what's missing. The purple boxes throughout the document are the specific questions where your answer changes what the audit does. You know your deals better than any outside analysis can — treat those as the agenda.
Confidence Badges Every entity carries a confidence badge. High means directly sourced (from your site, reviews, or public landscape data). Medium means inferred and worth a closer look. Scan for medium badges — those are where we most need your validation.
→ Anthropic's identity splits two ways: the company (Anthropic) versus the product (Claude), and — more consequentially — the commercial surface (pricing, product pages, platform docs) has migrated from anthropic.com to claude.com / platform.claude.com, while this audit crawled anthropic.com. Do buyers and AI engines resolve "Anthropic" and "Claude" to the same entity, and which domain should the audit measure — anthropic.com, claude.com, or both? If citations now accrue to claude.com, auditing anthropic.com alone measures a research-and-news surface, not the commercial one where deals are won.
5 personas: 3 decision-makers, 1 evaluator, 1 influencer — the buyers whose distinct search intent drives every query the audit generates.
Critical Review Area Personas drive the entire query set — each buyer searches differently, so a wrong role or influence level sends the audit chasing the wrong intent. Scrutinize the influence badges and the one inferred persona below before the call.
Data Sourcing Name, role, department, seniority, influence, veto power, and technical level are pulled directly from the KG. Role descriptions, buying jobs, and query focus areas are synthesized from those fields to show how each persona would actually search — validate the synthesis, not just the labels. Four of five personas are review-mined at high confidence; one (Director of AI Enablement) is LLM-inferred at medium confidence.
→ Is the CTO and the Chief AI Officer one budget-holder in your deals, or two seats at the table? If AI strategy sits with a separate Chief AI Officer, we split Marcus into two personas with different intents — platform reliability vs. a transformation mandate.
→ Does the CISO hold a hard veto in your AI deals, or sign off after IT? If security can independently kill a deal, we weight governance and residency queries as gating criteria, not evaluation factors — which reorders the whole security query cluster.
→ Does the VP of Engineering control the platform budget, or only recommend? If Ravi actually signs the contract, we reclassify him as a decision-maker and add validation-stage queries on API reliability and coding-agent performance.
→ Is there a distinct AI Enablement / Innovation buyer in your deals, or is this really a CTO-plus-CISO purchase with no separate champion? Sofia is the only inferred persona — if she isn't real, we drop her adoption/ROI query cluster entirely rather than reporting visibility for a buyer who never shows up.
→ Does Legal/Compliance gate the deal on IP indemnification and data-use terms, or only review the paper at the end? If the GC can veto, we promote copyright-indemnification and data-privacy queries into the must-answer set rather than treating them as secondary.
Missing Personas? These roles sometimes appear in enterprise frontier-AI deals — do they show up in yours? A Head of Data / ML Platform (if model-infrastructure decisions sit apart from app-layer AI and warrant their own query cluster); a Procurement / Vendor-Risk lead (if a formal security-and-legal questionnaire gates regulated deals); a line-of-business economic buyer — Head of Support, Sales Ops, or Research (if a specific Claude use case is funded outside central IT). Who else shows up in your enterprise evaluations?
5 primary + 5 secondary competitors — tier assignments determine which vendors get head-to-head query matchups versus category-awareness coverage.
Why Tiers Matter Tier assignments decide which vendors get direct-comparison queries like "Claude vs OpenAI for enterprise" or "best LLM vs alternatives," versus lighter category-awareness coverage — roughly 6–8 head-to-head queries per primary competitor. Two primaries carry medium confidence: Microsoft Copilot (which may compete on M365 bundling more than model quality) and Cohere (whose presence in your regulated-enterprise deals we can't confirm from public data). Moving either to secondary would shift those queries out of the direct-comparison set.
→ Three tier questions decide the head-to-head set. (1) Does Microsoft Copilot show up as a model-quality rival in your deals, or as M365 bundling pressure at the seat level? (2) Does Cohere actually appear in your regulated-enterprise evaluations, or is it a category listing rather than a real rival? Moving either to secondary shifts ~6–8 head-to-head queries per vendor out of the direct-comparison set. (3) Since Bedrock distributes Claude, should Amazon Nova be tested model-vs-model rather than platform-vs-platform? And are any names here missing entirely — the set skews to public landscape listings rather than your own win/loss data.
11 buyer-level capabilities mapped — these determine which capability queries the audit tests and where competitive differentiation is pressed.
Consistently accurate, nuanced answers on complex, high-stakes work without hand-holding — a model I can trust for analysis, not just chat.
An AI that can navigate a large codebase, run multi-step engineering tasks end to end, and actually ship working code with an agent like Claude Code.
A model that says "I don't know" instead of confidently making things up, and won't go off the rails on sensitive prompts.
Feed it entire contracts, codebases, or research corpora at once and have it reason across all of it without losing the thread.
SSO/SCIM, audit logs, role-based access, custom data retention, and a compliance API so IT can actually govern how employees use it.
Guarantee my business data isn't used to train the model, and stand behind copyright indemnification so legal will sign off.
A clean, reliable API with tool use, MCP, and SDKs my team can build agents on without fighting the platform.
Run it where our data and compliance rules require — our cloud, our region, ideally on-prem or air-gapped.
It should already plug into the tools we live in — Slack, Microsoft 365, our IDE, our data — not sit in a separate tab.
Generate and understand images, voice, and video — not just read them — so one vendor covers all our modalities.
Predictable, competitive token costs for high-volume workloads — the top-tier model shouldn't blow up our budget versus cheaper alternatives.
Prioritization Seven capabilities are rated strong: Reasoning & Response Quality, Agentic Coding & Software Development, AI Safety, Alignment & Low Hallucination, Large Context Window & Long-Document Handling, Enterprise Security, Admin & Governance, Data Privacy & IP Protection, and Developer Platform, API & Tool Use. The audit tests all 11 capabilities, but competitive-differentiation queries will emphasize 3. Which of these seven best represents where Anthropic actually wins deals?
→ The two weak ratings carry the most consequence: if premium Opus pricing is genuinely losing deals to DeepSeek, Llama, and Nova, we build value-vs-quality queries; if buyers pay the premium for frontier quality, we press differentiation instead. Likewise, is the lack of native image/audio/video generation a real deal-loss factor or an accepted trade-off? Are Deployment Flexibility and Ecosystem Integrations truly only moderate versus Microsoft and Google — and should any strong capabilities merge (e.g., Enterprise Governance + Data Privacy & IP, which buyers often evaluate as one security conversation)?
10 pain points: 6 high, 4 medium severity — the buyer frustrations whose first-person language becomes the phrasing of the queries we run.
→ Is "runaway token cost" really high-severity for your enterprise segment, or does frontier quality neutralize price objections at the deals you actually win? Severity ordering decides which frustrations we test first. And are we missing category-specific pains — model-deprecation / version-migration churn (models retired on a schedule, breaking pipelines), evaluation difficulty (buyers can't objectively benchmark models before purchase), or vendor lock-in / portability fear? Which of these actually stalls your deals?
What our crawl found on anthropic.com — the technical fixes engineering can act on now, independent of the audit results.
For Engineering Good news first: robots.txt is confirmed open to every major AI crawler — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and Googlebot are all allowed, so nothing is blocked at the door. No critical blockers were found. Engineering should focus on one high-severity item — the stale evergreen explainers with no refresh date — and three medium structural fixes: the sitemap URLs that 308-redirect off-domain to claude.com, the JS-only Economic Index dashboard, and the non-semantic hub-page headings. All are actionable now.
What we found: Across the 25 inventoried content-marketing pages (news, engineering, research), the category freshness average is 0.25 — 17 of 25 carry a visible date older than 180 days and 12 are older than 365 days. This includes Anthropic's most-linked evergreen technical assets — "Building effective agents" (Dec 19, 2024), "Introducing the Model Context Protocol" (Nov 25, 2024), "Raising the bar on SWE-bench Verified" (Jan 6, 2025) and "How we built our multi-agent research system" (Jun 13, 2025) — none of which show any "last updated" refresh signal even though they remain live, frequently-cited reference pages.
Why it matters: AI answer engines concentrate citations on recently-updated pages. ConvertMate found 76.4% of ChatGPT's most-cited pages were updated within the prior 30 days, and Ahrefs measured AI-cited content as 25.7% fresher on average than typical results. Anthropic's evergreen agent/MCP/coding explainers are exactly the pages that surface for "how to build AI agents," "what is MCP," and coding-evaluation queries; without a visible refresh date they lose freshness credit to newer third-party write-ups even when Anthropic is the primary source.
Recommended fix: Add a visible "Last updated" date and a lightweight quarterly refresh cadence to the ~10 evergreen engineering/research explainers that still attract traffic (agents, MCP, context engineering, SWE-bench, multi-agent). Update model/version references inline and re-timestamp. Leave dated launch announcements as-is, but keep the evergreen guides current.
What we found: www.anthropic.com/sitemap.xml still lists product URLs — /product/claude-code, /product/enterprise, /product/security, /product/claude-cowork — and at least one engineering URL (/engineering/claude-code-best-practices) that now return HTTP 308 Permanent Redirect to a different registrable domain (claude.com, code.claude.com). The bulk of Anthropic's commercial surface (pricing, solutions, API/platform docs, customers) has migrated to claude.com / platform.claude.com and is no longer served on the audited anthropic.com domain, yet the sitemap continues to advertise the moved paths.
Why it matters: Sitemap entries that redirect cross-domain waste crawl budget and split link/citation authority: inbound links and prior citations pointing at anthropic.com/product/* bounce to a second domain, so neither domain accumulates the full signal. AI crawlers that resolve the sitemap encounter redirect chains instead of canonical 200 responses, reducing the share of Anthropic's own commercial content that is cleanly indexable under the brand's primary domain.
Recommended fix: Regenerate the anthropic.com sitemap to drop paths that permanently redirect off-domain, and either (a) keep canonical commercial content reachable under anthropic.com, or (b) if the claude.com split is intentional, publish and cross-submit a claude.com sitemap and add rel=canonical / consistent internal linking so authority consolidates on the intended domain.
What we found: Several commercially relevant hub/index pages have heading structures that don't function as passage labels: /learn/claude-for-work repeats a generic "Resources" heading roughly eight times with no descriptive nesting; /system-cards exposes no article H2s at all (content is a bare table, and nav labels surface as the only H3s); /economic-index emits a single run-on H1 and no section headings; and the homepage uses full-sentence strings ("Anthropic is built on hard questions.") as H2s. A minor rendering artifact also duplicates the model name in the H1 on /claude/fable and /claude/mythos ("Claude Claude Fable 5").
Why it matters: LLMs use heading text as passage boundaries and standalone labels when selecting what to quote. Repeated "Resources" headings, missing section headings, and sentence-as-heading patterns give extraction engines no clean, self-describing anchor to cite, lowering the odds these pages are quoted even when their underlying content is relevant.
Recommended fix: Give hub/index pages a single descriptive H1 and unique, noun-phrase H2/H3 section headings; replace repeated "Resources" labels with specific ones; ensure /system-cards and /economic-index expose real section headings around their tables/dashboards; and fix the duplicated-word H1 on the /claude/fable and /claude/mythos model pages.
The following items could not be assessed through our analysis method (rendered markdown). We recommend your engineering team verify these manually before the validation call.
What to check: The /economic-index page returned only its H1 plus a "Loading data" placeholder in rendered output — the actual economic dashboard content did not appear in the fetched HTML/markdown, consistent with client-side rendering of the main data payload. Because our method reads rendered markdown rather than executing a full browser, we flag this for confirmation rather than asserting it definitively. If the substantive content is injected only after JS executes, non-JS crawlers and some AI ingestion pipelines will index an empty shell — making an otherwise citable, data-rich asset invisible for economy / AI-labor-impact queries where the Economic Index is a differentiated primary source.
Recommended action: Verify /economic-index (and any other dashboard/interactive pages) with JavaScript disabled or via a rendering crawler such as Screaming Frog in JS-rendering mode. If the data is client-side only, add server-side rendering or a static, crawlable summary of the key figures within the initial HTML response.
What to check: Our analysis reads rendered markdown, so JSON-LD / microdata blocks in the raw HTML head are not visible to us. We could not confirm whether model pages carry Product schema, whether news/engineering posts carry Article schema, or whether FAQ-bearing model pages carry FAQPage schema. Schema presence is recorded as null for every inventoried page rather than asserted absent. Appropriate schema helps search and AI systems parse entities, dates, and Q&A pairs reliably — a low-cost structural win, but it must be verified against raw HTML before any claim is made.
Recommended action: Run the model pages, a sample of /news and /engineering posts, and the FAQ-bearing pages through Google's Rich Results Test or a structured-data validator (or view-source). Add Article, Product, and FAQPage schema where absent, prioritizing model pages and evergreen guides.
What to check: Meta description and Open Graph / social-card tags live in the raw HTML head, which is not present in the rendered markdown our method reads. We cannot confirm their presence or quality on any inventoried page, so meta_description is recorded as null and has_og_tags as false-by-default (unverified) across the inventory. These tags shape how pages are summarized and previewed in search and when shared into the tools where buyers discuss vendors.
Recommended action: Spot-check meta descriptions and OG/Twitter card tags across the model pages, top /news posts, and /learn hub pages using a social-preview debugger or view-source. Add or de-duplicate where needed.
Partial Sample This analysis covered 41 pages on anthropic.com. Two data limits to note: schema coverage is null across all 41 pages (our method can't see raw-HTML JSON-LD — see the verification checklist) and 9 freshness scores are unscored (7 structural/reference pages plus 2 undated product pages). Just as importantly, the bulk of Anthropic's commercial surface now lives on claude.com / platform.claude.com and was outside this crawl — which is why the audit-domain question in the Company Profile matters before we finalize scope.
Why Now The window to establish GEO visibility is open but closing:
• 94% of B2B buyers now use LLMs during the buying process (6sense, November 2025) — for a frontier-AI vendor, discovery has already moved into the assistant.
• Early citations compound: domains AI platforms learn to trust now get cited more often as retrieval and training data accumulate.
• Gartner predicts 90% of B2B buying will be AI-agent-intermediated by 2028 (Gartner, October 2025) — increasingly, the "buyer" reading your pages is an agent, not a human.
• Frontier-AI vendor comparison is still early-innings in GEO — acting now means competing against inaction, not against entrenched strategies.
The full audit will measure how often Claude surfaces — and how often it doesn't — across the buyer queries that define this category: "best enterprise LLM for regulated industries," "most accurate model for high-stakes analysis," "Claude vs OpenAI vs Gemini for coding," "AI with copyright indemnification." You'll see exactly which of these return your competitors but not Anthropic across the AI platforms your buyers use. Because Layer 1 already shows your crawlers are open and your product pages are current, you're starting from a stronger technical baseline than most — so the fixes in this document compound your advantage rather than just closing a gap.
Start Now — Engineering Three technical fixes your engineering team can start before we even meet — none depend on the audit results: (1) regenerate the anthropic.com sitemap to drop the /product/* and /engineering/* URLs that 308-redirect off-domain to claude.com; (2) fix the non-semantic hub headings (the repeated "Resources" labels on /learn/claude-for-work, missing section headings on /system-cards and /economic-index) and the duplicated-word H1 on /claude/fable and /claude/mythos; and (3) verify /economic-index renders server-side (or add a static summary) and confirm JSON-LD schema + meta/OG tags via Google's Rich Results Test. Your robots.txt is already confirmed open to GPTBot, ClaudeBot, PerplexityBot, and Google-Extended — no action needed there. These improve your baseline visibility before we measure it.
Two jobs before we meet. The questions on the left require your judgment — no one knows your business better than you. The engineering tasks on the right don't require the call at all.