AI citation

It is an attribution: this sentence came from that URL. It is not a ranking, not an endorsement, and not stable — the same prompt asked twice can return a different citation set, because generation is non-deterministic and retrieval sets vary run to run.

It is also not the same as a click. A large share of AI answers resolve the buyer's question in place, and the citation is seen rather than followed. Which means AI referral traffic systematically undercounts citation volume, and treating referral sessions as your only AI metric will make a growing surface look like a dead one.

Nobody outside these companies knows the selection function, and anyone quoting a precise multiplier for a given tactic is quoting a small unaudited sample. What is defensible comes from mechanics rather than measurement:

• The page has to be retrievable. Reachable by the engine's crawler, present in the initial HTML, indexed.
• The passage has to contain the specific claim. Models cite where the fact is, not where the topic is.
• The passage has to be self-contained enough to attribute. A spec value with no product named beside it cannot be tied back to your SKU.
• The claim has to survive comparison. When three sources disagree, specificity and verifiability tend to win — a certification with a standard number beats "industry-leading quality."

None of that is a growth hack. It is the same list that makes a page useful to a human who is checking your claim.

Consumer citation strategy revolves around review roundups and best-of lists. The B2B equivalent is narrower and considerably more winnable: the queries are specification queries, and almost nobody answers them well.

"What breaker fits a Square D QO load center," "what replaces a discontinued Bell & Gossett seal kit," "which A2L-compatible line set do I need for this condenser." Those answers live in cross-reference tables, fitment data, and compatibility notes that most distributors keep in a rep's spreadsheet rather than on a page. Publishing them as structured, self-contained content is the highest-yield citation work available in distribution, precisely because the competitive set is thin.

Three imperfect instruments, used together. Server logs show which retrieval crawlers fetched which URLs and how often. Prompt testing panels re-run a fixed query set on a schedule and count how often you appear. Analytics shows the referral traffic that does arrive.

Run the panel against queries a buyer would actually type, not brand-name vanity prompts, and keep the query set frozen so movement means something. Any single run is a sample of a probabilistic system; the trend across a quarter is the only part worth reporting.