# Citation pilot collection record

Collection date: 6 October 2026. The protocol fixed four exact questions, two platforms and two fresh-context attempts per platform before the first query. This is one collection period. It cannot establish a trend, representative visibility, a national result or an effect of website work.

Timestamp correction: the frozen protocol's manually typed addendum heading says 07:07 UTC. That minute is incorrect and was not a clock reading. The retained file was created at 07:05:41 UTC and last modified at 07:06:19 UTC before collection. Tool order records the Unpersonalized selection, then that addendum write, then the first prompt submission. The first retained in-progress ChatGPT snapshot was written at 07:06:46 UTC. No exact prompt-submit timestamp was retained. The original protocol text is left intact so this correction is visible; the sequence is evidenced, but the addendum's typed minute must not be used as precise timing.

## Context

ChatGPT used the connected Chrome signed-in session, visible Pro mode (the selector showed Pro, 5 of 5), Temporary chat, and the per-conversation Unpersonalized option. The visible explanation said the conversation ignores memory, plugins and custom instructions. This was recorded before submission. A fresh conversation was opened for each of Q1-Q4. No files, private account data or previous conversation content were supplied. Prompts requested current public sources; visible search activity was observed on the identity attempt. These visible labels do not disclose the underlying model routing or prove a geographic location.

Perplexity used its signed-out public Search interface, with a model selector labelled Model but no underlying model identity exposed. Each question began at a fresh search rather than as a follow-up. Optional sign-in invitations were dismissed through normal controls; no account was created and no paid option or privacy setting was changed. Q2-Q4 supply United States as question context. Connection geography, personalization beyond the visible state, and backend model routing were not controlled.

## Collection and interruption

The protocol defines order within each platform: Q1-Q4, then Q1-Q4 again. Platform activity was interleaved. The first four ChatGPT prompts were submitted in separate contexts while responses were running. A shared browser-control interruption then made the original tabs blank and unavailable to the browser API. Native accessibility also showed no answer content. Existing conversation URLs were preserved privately. Reopening the original first temporary-conversation URL in a new tab returned a fresh blank temporary chat, not a recoverable answer. No prompt was resubmitted to replace those attempts.

C01-C04 therefore have status Failed: the request was submitted, but a completed answer could not be retained. This describes capture failure, not a provider's refusal or a proved failure to generate an answer. C05-C08 have status Not run: the repeat pass was not submitted after capture reliability was lost. Every metric for these records stays Not assessed. No missing response is counted as a zero.

P01 was copied through Perplexity's Copy control, with its ten-source panel expanded. P02's full original answer was recovered by opening its existing conversation URL in a fresh Chrome tab. This did not submit the question again. P02-P08 preserve the full rendered answer text with the displayed citation labels; layout and Markdown formatting differ from the first copy export. Their source panels were expanded and all visible source URLs retained separately. Rendered transcripts end before suggested follow-up questions and the source panel itself. None of the collected questions was edited or regenerated.

Original conversation locators and full browser snapshots remain private. Public files contain the fixed prompts, observed answer text, source links, dates, context, decision notes and limitations. A later reader cannot replay a private conversation URL from the public bundle. Public raw answers may contain statements we have not verified; read source-review.md before using them as business facts.

## Reading the measures

The target is SEO Informatica and seoinformatica.com. A name already present in Q1 or Q2 is a prompted branded check; it is not unaided provider discovery. Keep those questions separate from Q3 (criteria) and Q4 (provider discovery). Recommendation means affirmative inclusion for consideration, including a conditional fit statement. A neutral identity description does not qualify.

Each field retains Yes, No or Not assessed independently. Report Yes divided by Yes plus No for that field, plus the number not assessed. An own-site citation is one answer-level observation even if the answer links to several pages. Source lists include retrieval candidates and are not counts of independently verified endorsements. A third-party brand citation needs a confirmed citation and a destination read identifying the business. Company-controlled LinkedIn pages must not be presented as independent validation.

Clarification after collection: the frozen protocol's phrase answer/source link is too broad if a platform mixes citations with retrieval candidates. A Yes needs an explicit answer citation or a source clearly presented as cited, not mere presence in a retrieved-source list. The four own-site Yes rows were independently checked against explicit answer-body anchor URLs, with the source panel excluded, and are supported in citation-evidence.json. The original protocol remains unchanged. Candidate-only evidence must remain Not assessed. Where the full retained answer and displayed source list contain no own-domain URL, this capture supports No within that visible scope.

Whole-answer factual accuracy was not established. Some source identities and selected brand facts were checked, with limits in source-review.md. A future reviewer can perform a defined claim-level review; this pilot does not invent that sign-off.

Use this small retained example to inspect the method and decide a next investigation. Do not combine its differently worded questions into a market share, attach a significance claim to two repeats, or treat the two repeats as independent random samples. There is no second period, no before-and-after comparison, and no citation or ranking uplift claim.

## P05 transcript-boundary correction

Final review found that P05 originally began inside the displayed prompt, before the answer, because the business name occurs in both. Canonical rawAnswer and raw/P05-answer.md now exclude characters 0 through 128 (zero-based, end-exclusive offset 129; 129 UTF-8 bytes). The excluded prefix is the truncated question followed by the Researched UI label. raw/P05.md is preserved byte-for-byte as the original capture. transcript-corrections.json records the exact excluded prefix, offsets and both hashes. The assistant answer after that boundary is unchanged, and all P05 metric decisions are unchanged. The original capture and the answer-only file are both available; use the answer-only file for answer measures.