# Consumer citation repeatability wave — 14 September 2026

## Scope

This package records a second 48-answer observation of the same 12 unbranded
US-English shopping questions used on 9 September 2026. Each question received
two retained answers in ChatGPT's consumer interface and two in Claude's
consumer interface. The frozen questions and suffix are in
consumer-repeatability-panel-v1.json.

This is the DTC shopper repeatability panel. It is separate from Citelift's
20-question, 80-answer commercial discovery panel and from the newer Shopify
merchant diagnostic panel. Neither is included in these files or denominators.

Collection used existing consumer subscriptions on 14 September 2026. No API
answers or paid API calls were used. Every canonical run records its UTC capture
time, interface, displayed model and effort when exposed, web-search setting,
panel and prompt versions, and submitted prompt in the private archive.

ChatGPT used Temporary Chat with Unpersonalized visibly selected, High effort,
and web search selected in the composer. ChatGPT did not expose a model name.
Claude used fresh incognito conversations with Opus 5 and High displayed, and
each retained answer visibly reported web search.

## Prompt and attempt audit

Fourteen ChatGPT prompts matched the frozen prompt bytes. Ten used the same
question and suffix with one space between them instead of two newlines. Their
private records preserve the actual input and runs.csv labels them
whitespace_normalized. The analyzer accepts whitespace-only differences while
rejecting changed words and missing fresh prompts. All 24 Claude canonical
prompts match the frozen wording and whitespace exactly.

Seven answered Claude attempts were excluded for protocol compliance before
analysis, without regard to citation outcomes. Six used changed wording for
both repeats of FOO-01, FOO-02, and PET-01; six exact-prompt replacements were
then collected. One first APP-01 repeat-2 attempt followed a reload that retained
the preceding incognito conversation, and Claude's response recognized that
prior question; it was replaced with a fresh-conversation answer. The private
archive preserves all seven excluded attempts and their audit metadata. They do
not appear in this public package. ChatGPT had no personalized or non-temporary
submissions.

## Offline analysis

The analyzer makes no network calls. It accepts visible HTTP(S) text links,
excludes ChatGPT and Claude interface/help domains, removes URL fragments and
common tracking parameters, and deduplicates normalized URLs within an answer.
Paths, other query parameters, schemes, and subdomains remain distinct.

Answer–URL occurrences count a normalized URL once within each answer and again
when it appears in another answer. Repeat overlap is the Jaccard intersection
divided by union for the two URL sets from the same prompt and engine. Empty
complete sets would be not applicable. Missing or unavailable answers and
unreviewed fields remain null; an observed zero is emitted only after inspecting
a completed retained answer.

The 48 canonical answers all passed schema, prompt, metadata, and identity
validation. Both engines completed 24/24 answers and 12/12 repeat pairs, and all
48 answers had at least one qualifying visible link.

- ChatGPT: 145 answer–URL occurrences, 116 unique normalized URLs, 55 unique
  normalized domains, mean Jaccard 0.231886.
- Claude: 189 answer–URL occurrences, 166 unique normalized URLs, 147 unique
  normalized domains, mean Jaccard 0.162852.
- Exact Citelift mentions: 0/24 in each engine.
- Owned-domain links: 0/24 in each engine.

These are descriptive observations from one date and fixed prompt set. They do
not measure search demand, buyer demand, revenue, ranking, or a platform trend.

## Source-support spot check

The raw capture annotations remain null, so summary.json correctly reports 0/334
embedded source-class and support-review coverage. A separate deterministic
sidecar, source-sample-review.json, reviews the first qualifying link occurrence
in repeat 1 for APP-01, HOM-01, BEA-01, and SUP-01 in each engine.

The sidecar covers 8/334 selected answer-link occurrences (0.023952 occurrence
coverage): five supported the nearby claim tested, one partially supported it,
one did not support it, and one was unverifiable. Seven occurrence checks reached
accessible text. These are occurrence counts, not unique-page counts: the same
American Academy of Dermatology URL was selected independently in the ChatGPT
and Claude answers. Six selected labels contained a collapsed +1 or +2 group;
those additional destinations were not reviewed.

This fixed spot check is not representative, not an error rate, and not a review
of the full answers. A publisher ownership class describes the page owner, not
source quality or independence.

## Files

- summary.json — panel, engine, and category aggregates.
- runs.csv — one canonical row per retained answer, with no answer text.
- visible-links.csv — normalized observed answer-link occurrences.
- repeat-overlap.csv — the 24 within-engine repeat comparisons.
- consumer-repeatability-panel-v1.json — frozen questions and suffix.
- source-sample-review.json — fixed eight-occurrence manual audit sidecar.

Raw answer text is not included. The public 9 September study package remains a
separate immutable archive.