Claude Opus 5 visibility: what marketing teams should measure

To evaluate brand visibility in Claude Opus 5, build a controlled prompt set around real customer decisions, record complete answers with their conditions, and separate mentions from qualified recommendations, citations and factual accuracy. Anthropic announced Claude Opus 5 on July 24, 2026. That date matters: early observations describe a newly launched model, not a durable market baseline. The practical goal is not to "rank in Opus" once. It is to learn where your public evidence helps Claude answer buyer questions and where ambiguity, stale facts or missing proof creates risk.

Start with the customer journey, not the model name

A model launch can tempt teams into running a handful of vanity prompts such as "What is the best software?" Those prompts rarely mirror a serious purchase. Build clusters for category discovery, problem diagnosis, shortlist creation, product comparison, risk review, implementation and branded verification. Add the constraints that influence an actual choice: company size, geography, budget, integration, security needs and intended outcome. Preserve a fixed benchmark panel, but reserve a smaller exploratory panel for new language found in sales calls and support tickets. This design remains useful if Anthropic changes the model lineup later.

Define four separate visibility outcomes

• Mention: the answer names the brand, regardless of whether the context is favorable or useful.
• Qualified recommendation: the brand is suggested for a defined need and the stated reason is supportable.
• Citation: a source attributable to the brand or an independent authority is linked or clearly referenced.
• Accuracy: material claims about price, availability, capabilities and policies match current approved facts.
• Competitive position: alternatives, comparison criteria and omitted brands are logged without turning one answer into a market-share claim.

Create evidence Claude can interpret and buyers can verify

Publish one clear source for each decision-critical fact. Category pages should explain who the product is for and when it is not a fit. Comparison pages should name criteria and evidence rather than declaring victory. Documentation, security pages, pricing, policies and release notes need visible dates and accountable owners. Keep brand identity consistent across the website, profiles and reputable third-party listings. Structured data may clarify visible entities, but it cannot replace readable proof. Use the schema generator to draft markup and the schema validator to check that it agrees with the page.

Run a reproducible Opus 5 test

1. Freeze the prompt text, audience, market, language, account state and collection date.
2. Run each prompt enough times to observe variation without claiming statistical certainty from a small sample.
3. Save the complete answer, links, model label, mode and any visible retrieval indicators.
4. Code mentions, recommendations, citations, accuracy problems and competitor context separately.
5. Route factual errors to the owner of the relevant source page, then retest after a meaningful update.
6. Repeat the same panel on a documented cadence and annotate model or product changes.

Account for model lifecycle changes

A report should identify the exact model exposed by the product at collection time, not merely say "Claude." Anthropic maintains an active-model and deprecation reference, which teams can use as a change log input. If an alias moves, a model is retired or a product mode changes, open a new comparison period instead of blending observations. Keep historical results, because they explain why a chart moved; do not treat them as a promise about the current experience. Where the precise underlying model is not disclosed, label that uncertainty rather than guessing.

Turn observations into an evidence backlog

Prioritize fixes by buyer importance, frequency and business risk. A wrong legal limitation deserves attention before a missing mention on a broad brainstorming prompt. For every gap, ask whether the source is absent, unclear, contradictory, inaccessible or simply not selected in the sampled answer. Start with an AI visibility scan, keep the full prompt-and-answer record, and use repeated matched tests to distinguish a persistent evidence problem from ordinary response variation.