The six elements of an AI fingerprint, and how we audit each one
Written by
Odin Lund
Category
Playbook
Date
Share this article
Owned surfaces: website, content, schema
Website: the pages the models retrieve and cite. Measure: pages with at least one citation, and retrievals per month. For the client in our log, 63 pages are cited and the site draws 4,319 retrievals in 30 days. Content: pages written to answer one customer question each, in the shape the models retrieve (a direct answer first, then the facts that support it). Measure: live pieces and how many are cited; 31 pieces, most of them cited within their first month. Schema: structured data the engines can parse, from Organization and LocalBusiness to FAQ and Article. Measure: coverage across the pages that get retrieved; 355 pages carry it now, none did in June.
These three are the fastest to fix because the business controls them outright, and they compound: a page that states the facts plainly, carries schema and answers one question is the page the models come back to.
Brand signal: data and facts, brand statements
Data and facts: the verified numbers the models should repeat, such as results, prices, dates and credentials. Measure: facts verified and signed off, and whether every source states them identically. The client agreed 16 facts; each now appears in the same form on the site, in its listings and in its structured data. Brand statements: the claims the models currently make about the business, set against the claims the business makes about itself. Measure: alignment between the two, read claim by claim; 90.2% for this client, with the gap turned into a correction list.
This is the element most businesses have never looked at. The models have already formed a description of you from whatever they found. Until you read it, you are not managing your business in AI search; the sources are.
Earned media: the sources the models trust
The third-party sources that carry the business: directories, review platforms, press, forums and specialist publications. Measure: citing domains and retrievals from them; 46 domains and 18,742 retrievals in 30 days for the client, about four times the retrievals from its own site. In most categories a small number of domains account for most citations, so the audit question is specific: which sources do the models rely on here, and is the business present on each with the same facts?
The controlled evidence agrees. In the GEO study (Aggarwal et al., KDD 2024), adding citations to authoritative sources and quotable statistics raised visibility in generative answers by up to 40%; the models reward content that other trusted sources corroborate. Earned media is how a business gets corroborated.
How to run the audit, and sources
Start from the customer’s questions, by stage, and run them on each model. Record whether you are named, your position, and the source behind every claim about you. Map each claim to one of the six elements. Score each element on presence (is it there), consistency (does it match the others) and structure (can a machine extract it). The lowest-scoring element with the most claims attached is where to start; for most businesses that is data and facts, because a fact conflict poisons every element that repeats it. Then re-run the read on a fixed cadence and treat the audit as a living document, not a report.
Sources
SourceNode tracking, aisearchmonitor.com, client values at 17 September 2026.
Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024 (arXiv:2311.09735).
Schema.org vocabulary for Organization, LocalBusiness, FAQPage and Article.
Share Blog
