Case study
We audited ourselves first.
The first CiteGap case is our own site, demandfieldnotes.com, run through the same 30 buyer-style questions across ChatGPT, Perplexity, and Gemini, measured twice over two days. No client signed anything. No vendor self-report is involved. Day 1 is measured and published below; the Day 2 retest and the two blocked engines land when they are measured.
What Day 1 actually measured.
The run started November 12, 2026. Of the 90 planned cells (30 buyer-style questions × 3 engines), 15 produced measurable answers, all on Perplexity. ChatGPT was blocked at the entry point: our proxy exit IP is rejected by OpenAI's edge, with the rejection records archived in the run log. Gemini accepted no anonymous submission, and its retest path needs a clean Google account we did not have. Both remediations are listed in the log instead of being papered over.
Named in 2 of 15 measured answers: one name-only, one substantive. 4 of the 15 answered about a different product that shares our name, a weekly tracking SaaS called CiteGaps.
Discovery shelf: absent. In the 9 measurable discovery and use-case questions, competitors named included Profound, Semrush AI Visibility, Peec AI, Otterly, AthenaHQ, Scrunch, and Ahrefs Brand Radar. Competitors outweighed us roughly 3 to 1 in the answers we measured.
Where content exists, citability works. The one pros-and-cons question produced an accurate, well-sourced answer that cited our own pages, alongside a wrong-domain attachment and the same-name tracker. Entity disambiguation, not content quality, is our gap.
Trust signals: near zero. The only third-party proof the engine surfaced for us was one Product Hunt submission without substantive reviews. No uptime page, no client quotes, no independent coverage. Logged, and now on our own fix list.
Measurement basis: 30 buyer-style questions per engine, single Day 1 pass, 15 of 90 cells measured. Every figure above is traceable to the per-question logs. The Day 2 retest and the two blocked engines ship to this page when they are measured, not before.
The mechanisms we expect the self-audit to expose.
Pages engines cannot parse do not get quoted
Structured data and a plain answer-first opening make a page liftable. Pages that bury the answer under marketing copy get skimmed or skipped. Our own pages are not exempt, and the audit checks them the same way it checks a client site. What a 1,885-page experiment found about JSON-LD.
Absence at the comparison moment hands the deal to whoever showed up
When a buyer asks an engine to compare options, the engine names vendors it can back with sources. If you are not in that answer, the sale started without you. Our own category is no softer: buyers comparing AI visibility audits will ask engines exactly that question. Why being named beats ranking.
Engines repeat what they can cite, and hedge over what they cannot
If the only material about you is your own marketing copy, the answer stays vague, and vague praise does not survive a purchase decision. Decision-stage questions about trust, results, and risks only get confident answers when citable proof exists. What happens when AI recommends an unknown vendor.
Stale facts make engines hedge, and hedging loses deals
Engines hold snapshots from different weeks. If pricing, category, or company details disagree between snapshots, the safe output is a qualified answer. The audit compares what engines state against what the site states today, and a mismatch becomes a fix ticket. The full method is documented on the methodology page.
What sits behind the published numbers.
Each published figure from the self-audit traces to the evidence pack the pipeline produces on every run:
Per-question answer logs. Every answer from every engine, for both days: the question asked, the full answer text, who was named, and in what order.
Citation capture. Which sources each answer cited, recorded per round, not reconstructed later.
Answer snapshots. Where an engine offers no stable public link, the answer text is captured as-is and archived.
Fix tickets. The ticket the audit produced for each gap, so a published claim maps to a change someone actually made.
The pack is archived and checkable. Internal file paths stay internal; what gets published are the numbers and the measurement basis, not the storage layout. On paid runs, the raw answers are part of the deliverable, so a client can re-derive every figure in the report.
You would work with the person who ran this audit.
Demand Field Notes is a lab of one. The questions, the runs, the report, and the reply to your email all come from the same person. Ask us anything about the method, or start with the free check on your own homepage.
Method details first? Read the methodology. Full pricing lives on the pricing page.