# Citation observation pilot: frozen protocol

Frozen 6 October 2026 at 07:05 UTC, before any assistant query. Collection is a method demonstration, not representative research or evidence of effectiveness. The observation operator is Codex through connected Chrome. No independent human review is claimed.

## Questions

- Q1 (branded identity): What is SEO Informatica (seoinformatica.com), and what does it do? Use current public sources and cite the sources you use.
- Q2 (branded service fit): Would SEO Informatica (seoinformatica.com) be a suitable SEO partner for a small B2B service business serving the United States? Explain what public evidence supports or limits that choice, and cite your sources.
- Q3 (nonbranded buying criteria): How should a small B2B service business in the United States choose an agency for AI search visibility? Give practical evaluation criteria and cite current public sources.
- Q4 (nonbranded provider discovery): Which companies should a small B2B service business in the United States consider for AI SEO support? Give a short list with reasons, evidence limits and current public sources.

## Attempt plan and controls

Platforms: ChatGPT and Perplexity public browser interfaces. For each platform, run Q1-Q4 in this order in a fresh conversation, then repeat Q1-Q4 once in fresh conversations. Sixteen planned attempts. Do not regenerate or edit a prompt after submission. Do not retry to obtain favorable mentions. A load or login/access failure is retained, and unattempted rows remain not-run with a reason. Use visible default modes, no paid upgrades, no account settings changes. Capture mode labels and sign-in/context availability without publishing account identifiers. Fresh conversation does not prove personalization is absent. US wording is a prompt context, not a verified US connection location.

Retain exact answer text after completion and all visible source URLs that can be extracted from the answer/source UI. Record answer and source availability limits, timestamps, whether web-search behavior was visible, and private conversation locator separately. No private user material is submitted. Source-fidelity review checks only claims about SEO Informatica against public sources; other provider claims are not audited in full.

## Records

Fields: id, questionId, question, platform, attempt, status (planned/completed/blocked/failed/not-run), collectedAtUtc, mode, context, rawAnswer, sourceUrls, sourceCapture, metrics, notes, evidenceFile. Metrics are mention, recommendation, ownSiteCitation, thirdPartyBrandCitation, factualAccuracy; each value is yes/no/not-assessed with its evidence note. A mention is an explicit SEO Informatica/seoinformatica.com name in the answer body, not the user's question. A recommendation is affirmative inclusion for consideration, not a mere description. An own-site citation is an answer/source link whose host is seoinformatica.com or a subdomain. A third-party brand citation requires a cited third-party page to have been opened and confirmed to mention this business. Factual accuracy is assessed only for identifiable brand claims, and only when those claims have been checked against retained public evidence; a not-mentioned answer is not-assessed. Failure is never a zero.

Summaries show each yes count divided by that metric's assessed yes+no denominator, alongside not-assessed. Multiple source links are one answer's evidence, not extra attempts. Completed counts mean an answer was retained, not that every metric is assessed. One period cannot show change or uplift.

## Pre-query context clarification, 07:07 UTC

The current ChatGPT interface offers Temporary chat with a per-conversation Unpersonalized option. Before the first query, select this for every ChatGPT attempt; its visible explanation says this chat ignores memory, plugins and custom instructions. This improves the context control compared with a fresh personalized chat. Do not treat it as proof about server-side routing or geolocation. Keep the model at its default and record its visible label. Perplexity remains the default public interface and its actual context will be recorded.