Source Provenance Mining Framework

Overview

A method for extracting the source pages behind AI answers and converting them into a ranked citation target list.

Business problem

Teams measure whether AI mentions them but never capture which pages the model actually read, so content and outreach effort is aimed by guesswork instead of at the sources feeding the answers.

Decision supported

Where to spend citation-earning effort: which owned pages to strengthen, which existing listings to deepen, and which third-party sources to pursue.

Inputs & outputs

Step-by-step process

1. 1
   Define unbranded decision-stage questions

   Ask what a buyer asks at the point of choosing. Branded questions inflate the apparent citation rate and reveal nothing, because they hand the model the answer.

2. 2
   Sample engines with grounding on

   Run each question across the target engines with web search active, so the response returns real source URLs rather than training memory.

3. 3
   Capture every source URL

   Record all cited URLs per question and engine, not only the ones naming the brand. The pages that describe the market without you are the valuable ones.

4. 4
   Aggregate and rank

   Count citation frequency and how many distinct engines cite each page. Multi-engine pages represent shared consensus about the market; single-engine hits are noise.

5. 5
   Triage into three buckets

   Owned pages already cited (strengthen and reuse their structure), third-party pages already mentioning you (deepen the entry), and pages citing competitors only (the outreach target list).

6. 6
   Route to owners and re-sample

   Assign each bucket to content or digital PR, then re-run on a cadence so source drift shows up as an event rather than a surprise.

Maturity model

1. L1
   Blind

   Nobody knows which sources AI reads about the category.

2. L2
   Mention-aware

   The team tracks whether the brand is named but discards the source URLs.

3. L3
   Source-aware

   Source URLs are captured and reviewed manually for a small question set.

4. L4
   Triaged

   Sources are aggregated, ranked and split into owned, mentioning and competitor-only buckets with owners assigned.

5. L5
   Continuous

   Source sets are sampled on a cadence, drift is alerted on, and outreach is prioritised by citation frequency.

KPIs

• Share of high-frequency sources that mention the brand
• Number of multi-engine sources earned per quarter
• Competitor-only source count trend

Common mistakes

• Recording only the sources that already mention you, which throws away the target list
• Using branded questions, which flatter the result and hide the gap
• Treating a single sample as the source set when AI answers vary run to run
• Chasing low-frequency sources cited by one engine ahead of pages cited by four
• Reading an uncited answer as a failure rather than as evidence the model replied from memory

SearchScore insight