Citation Audit Across ChatGPT, Perplexity and Claude

Across 200 Japanese buying-intent prompts run through ChatGPT search, Perplexity and Claude in the same week, only 22 percent of the cited URLs appeared in more than one of the three answers. These are not one channel with one fix. They are three retrieval stacks that happen to share an interface, and this audit produces three separate remediation lists, because merging them would hide which platform each fix is actually for. Why the lists differ: ChatGPT search resolves through Bing's index alongside OpenAI's own fetchers. If your Japanese pages are thin or missing in Bing Webmaster Tools, nothing you do on the page moves ChatGPT, and we have watched teams spend a quarter rewriting copy for a problem that lived in an index nobody had checked. Perplexity runs its own crawler and its own index and rewards pages that answer in structure. Claude retrieves through Brave, a smaller index with different Japanese coverage, where a domain absent from Brave is not a candidate at all, whatever it does on Google. The part most audits skip: We run every prompt twice, once where the assistant retrieves and once where it answers from what it already holds. For consumer questions in Japanese, retrieval often does not fire, and the ungrounded answer is what your buyer actually reads. That answer is built from training data, so it describes the company as it was rather than as it is. We have seen a food client's discontinued line recommended by name, a beverage brand attributed to the wrong parent after a merger, and a skincare range credited with an efficacy claim it has never made anywhere. The third is not an inconvenience. Under 薬機法 the exposure sits with the brand, not the model vendor, and it is the finding we lead with when it appears. We verify fetch rather than assume it. We request your pages with the published user agents from the published address ranges and read what comes back. On Japanese sites that hydrate client-side, the fetcher routinely receives a shell with a loading state, and the assistant cites a competitor whose HTML arrived complete. What you get: A citation ledger per assistant, showing what was cited for what, and where you appeared. Index-presence checks in Bing and Brave with the specific gaps named. A fetch verification matrix: status code and rendered text length per user agent. The grounded and ungrounded answers side by side for every prompt in both languages, with factual errors about you marked and traced back to the source that most likely produced them. Everything ships as JSONL with model identifiers and timestamps. What is not in this: No ongoing measurement. This is one week read properly. If you want the trend, take the tracking engagement instead, and do not take both at once. No content production. No outreach to OpenAI, Perplexity or Anthropic on your behalf, because there is no queue to join, and anyone offering to submit your brand to a model is describing something that does not exist. Who this is not for: Brands that want one number for a board deck. The output is three lists, and if the ChatGPT fix and the Perplexity fix get averaged into a score, the audit was wasted. Brands whose product line changes monthly, where a snapshot dates before you can act on it. And early-stage companies with no Japanese third-party footprint, where the answer is known before we start and we would be charging you to write it down. How it runs: Three weeks, most of it capture and verification. One session to agree the prompt set before anything runs, because a prompt set written without your sales team is a set of questions nobody asks. Written deliverable and a working readout, in Japanese or English.