How we audit a site for LLM citations.

Every new engagement starts with a baseline read on whether the client's site is getting cited in AI products today — and if not, which competitors are taking their place. Here is the exact workflow, including the prompts. Steal it.

Step 01 · Build the query set

We start with twenty queries. They come from three places: the client's top-performing Google queries (ranked by commercial intent, not volume), the client's product category (the entity LLMs should associate with them), and their direct competitors' product names.

A typical query set for a dev-tools client looks like: "best vector database for production", "Pinecone vs Qdrant vs Weaviate", "what is the most affordable managed vector DB", and a dozen more in that shape. Mix comparison, how-to, and jobs-to-be-done framing.

Step 02 · Run each query in three products

We run the twenty queries in ChatGPT (with web search on), Perplexity, and Claude. For each run, we log three things: which sources were cited, what text was lifted from each source, and whether the client's site showed up at all.

Track the output in a spreadsheet. Columns: query, product, cited sources, quoted text, client present (Y/N), position (1..N). Sixty runs, about ninety minutes of work the first time. The agent we build in step 05 runs it in under five.

Step 03 · Score the baseline

We compute three numbers from the spreadsheet:

Citation rate. Out of sixty runs, how many cited the client? Most sites score under 10% on the first pass. The target we engineer toward is 40%+ on branded and comparison queries.

Share of voice. For each query, what fraction of the cited sources is the client vs. competitors? This is the metric the CEO actually cares about when we present findings.

Entity recognition. Does the LLM know what the client's product is? We ask each model "What does [product name] do?" and score the answer for accuracy. A product the LLM can't describe cannot get cited for the work it actually does.

Step 04 · Inspect the competitors who won

For every query where a competitor got cited instead, we open the cited URL and audit five things: page structure (headings, H1 clarity, FAQ schema), passage quotability (is there a single sentence the LLM could lift without context and attribute cleanly?), entity density (does the page use the category's primary entities explicitly?), authority signals (named author, dated updates, citations-out), and domain-level trust (external inbound links, brand mentions, Wikipedia presence).

The pattern across a hundred of these audits: the competitors winning citations aren't the ones with the best design or the most content — they're the ones whose pages read like reference material. Clear definitions. Specific numbers. Dated methodology. Nothing fluffy.

Step 05 · Build the weekly monitor

A one-time score is a snapshot. The real leverage is turning it into a weekly trend. We wire up an agent that re-runs the sixty queries every Monday, diffs the results against last week's, and drops a summary in Slack: citations gained, citations lost, new competitors emerging, queries where the entity recognition drifted.

Ninety days of this signal and you can tell exactly which content investments moved the needle. Most SEO teams don't know what worked because they never measured. This is how you stop guessing.

Step 06 · Pick three fixes and ship

The audit is worthless without prioritization. We rank the twenty queries by commercial value (dollars of pipeline each would generate if the client ranked), cross-reference with the difficulty of closing the gap, and pick the three with the best ratio. Those three become the content and structured-data work for the first month of the engagement.

Everything else goes in the backlog. You do not need to win every query. You need to win the queries that matter.