Semrush publishes a Claude Code workflow that turns Prompt Tracking exports into an AI citation outreach list, and cites its own study saying 62 percent of AI citations name no brand

Semrush published a step-by-step workflow on 21 September 2026 for turning a Prompt Tracking export into an AI citation outreach list, using Claude Code as the analysis engine. The blog post by Carlos Silva, with contributions from Faizan Ali, sits at semrush.com/blog/ai-citation-outreach and describes a project layout, an analysis command and a review process a marketing team can run without writing code.

What the workflow actually does

The setup is three files in a Claude Code project folder. CLAUDE.md describes the CSV structure that Semrush Prompt Tracking exports. brand-facts.md holds the approved positioning, pricing, competitor list and any common misinformation the team already tracks. analyze-sources.md goes into .claude/commands/ as an /analyze-sources slash command with the analysis instructions.

Once the CSV export is dropped in the folder, running /analyze-sources in Claude Code reads each cited URL, compares its content to the brand-facts file, and writes three output CSVs: opportunities.csv, manual-review.csv, and routed-reviews.csv. Every cited page gets one label from a fixed list of six: Not mentioned, Underrepresented, Outdated or inaccurate, Negative framing, No clear opportunity, or Manual review required. Semrush recommends the advanced Claude model for the run, and stresses that human validation is still required before any outreach goes out.

The numbers Semrush cites for itself

Semrush attributes the "62 percent of AI citations are ghost citations" figure to its own study with Kevin Indig, and quotes the definition in the post: "AI might cite your website in a response, but never mention you". Those are Semrush's own numbers, and Kevin Indig is the named co-author of the study rather than an independent second source.

The article shows one worked example, an H&M export of 452 unique source URLs, which after applying exclusions and duplicates falls to 330 pages the workflow analyses. That figure comes from Semrush's own tool and is not externally attributed.

The outreach step, and the one pitch example

The second half of the post covers what to do with the opportunities file. Semrush recommends a pitch structure attributed to outreach professional Sean Markey: open with a single correction, thank the writer for the current mention, offer updated information, ask permission before sending a full list of suggestions, and follow up after three or four days. Markey reports that of his last 14 pitches sent this way, 8 responses agreed to receive changes without negotiation. That is his own recent sample, not a general response rate.

Everything after that is generic outreach advice, not measured practice. Semrush is clear that the workflow ends with a human decision on whether to send anything at all.

The honest limitation

The workflow depends on a paid Prompt Tracking subscription for the input CSV. Output quality depends on the brand-facts file, which every team has to build and maintain themselves, and which is where the editorial judgement lives. Claude Code categorises against that file, so a thin brand-facts.md produces a thin opportunity list.

The workflow codifies a task most GEO practitioners were already doing by hand once a month per client: pulling every page an AI engine cites, checking whether the brand is named correctly, and flagging the pages worth writing to a human editor. A fixed vocabulary of labels and a repeatable command turn a two-day audit into a one-hour review of the CSVs.

For a team already paying for Prompt Tracking, adding a $20-a-month Claude subscription to run this against every client's export is a cheap experiment. Teams without a Prompt Tracking seat get the shape of the workflow as a template they can build on their own citation data.

Source

Carlos Silva with Faizan Ali, Semrush, AI citation outreach: a Semrush & Claude workflow, 21 September 2026. Ghost citations figure from Semrush and Kevin Indig, The Ghost Citations Study.