How we count citations — and why our numbers move slower than our competitors'.
The methodology behind Sourced. How we sample buyer-intent prompts, log verbatim model responses, weight engines, and avoid the four most common ways AI-citation metrics lie.
or every business we onboard we generate a working set of buyer-intent prompts — the questions a real customer would type into an AI assistant before choosing a service business. We fan each prompt across three large-language-model engines, log the verbatim response, detect mentions with fuzzy matching, and classify the sentiment of each one. We do this weekly. We never paraphrase what the model said.
How we pick prompts.
Prompts come from three places. First: a seed set of 20 generated at onboarding from the business’s category, services, and primary location — the questions almost every customer in that trade asks. Second: a discovery loop the agent runs each week, expanding the prompt set toward newly trending phrasings. Third: prompts you write yourself in Settings — typically the long-tail terms only an owner would know matter. On Editor and Bureau tiers the seed set grows quarter-on-quarter as the discovery loop finds new buyer behaviour.
How we score citations.
A “citation” is a mention of your business name (or a confirmed alias, including the domain stem) in the model’s response. We enforce a Jaro-Winkler similarity threshold of 0.92 so that “Quigley & Sons” matches “Quigley and Sons” but not “Quig’s Pizza.” The citation rate in your weekly report is the number of prompts where you were mentioned, divided by the total prompts times the number of engines.
How sentiment is classified.
For each cell where the business is mentioned, we run a single-word classifier (positive / neutral / negative / none) over the specific sentence the mention occurred in — not the full response. This avoids the failure mode where a hostile paragraph elsewhere in the answer drags your sentiment score down even though the mention itself was neutral.
The classifier is Claude Haiku; it errs toward “none” when uncertain, which is why the weekly report quotes sentiment counts rather than averages.
How we measure competitors.
During onboarding we discover up to ten local competitors from your category, your location, and the firms the LLMs themselves cite on your seed prompts. Each cell records every competitor mention; the competitor citation rate in the weekly report is the average across all competitors. When a competitor pulls ahead on a specific prompt we surface it in the lead story rather than burying it in a chart — that’s where the next piece of content comes from.
What we don’t do.
- We don’t measure traffic, conversions, or revenue. The value of a citation is in the buyer’s head before they ever click; attributing revenue to a specific model paragraph turns a slow real signal into a fast fake one.
- We don’t scrape the assistants’ web-search results pages. The audit calls each provider’s API directly — when a model changes its citation behaviour next quarter, the audit numbers shift with it. Customers see the shift honestly rather than smoothed.
- We don’t re-prompt with phrasings the model already cites you on. That’s the most common way competing tools inflate their citation numbers.
- We don’t count “your name appeared somewhere in a long answer” — usually as a side mention in a list of twenty — as a full citation. The score weights position and sentence-level relevance.
The third most common dodge is to ship many short thin pieces and measure publish volume rather than citation lift. We ship three long-form pieces a month on Editor. If the numbers aren’t moving, the content brief changes.
How the content gets written.
Drafts are generated by Claude Opus 4.7 with two pieces of context: a brand-voice profile calibrated against three of your own samples, and the verbatim text of the best-citation-grade response any engine returned on the target prompt. The model writes against what’s actually winning, not against what a marketing best-practice guide says ought to win.
Every draft sits in your in-tray. You approve, edit, or reject. Approved drafts get a polhia-hosted public URL within a minute and, if you’ve connected your CMS, are pushed there on the same job. Nothing publishes without an explicit click.
What changes in v1.5.
We’re testing a per-engine recency weighting that down-weights Perplexity citations from URLs older than 18 months (Perplexity favours fresh content in a way the other engines don’t, which creates a systematic over-count on older sites). And we’re adding a “cited paragraph” sample to the weekly report so you can read the verbatim sentence that named you, not just the count. Both ship before the end of the quarter.
Maintains the methodology. Revises it when an engine changes how it cites — and tells customers what changed in the same week.