# September 17, 2026 answer-share study

Original research by Marketing Agent. Collection used an existing Chrome profile, English UI and a US location, on September 17 Pacific time (timestamps use September 18 UTC). This is a single-profile snapshot. The questions match the previously published twenty-question instrument. No claim of representative sampling, independent generations, market-wide share, causal effects or trend is made.

## Files

- `observations.csv`: all 120 planned question/pass/engine cells. Empty presence means unknown, never zero.
- `questions.json`: exact twenty questions, unchanged.
- `summary.json`: coverage and allowed per-engine rates.
- `google-evidence.json`: the forty completed Google Overview captures, including full visible answer text, expanded related-source links, capture times and search URLs; one capture-interruption event is retained as well.

The raw Google text includes AI-generated claims and third-party snippets that have not been fact-checked as advice. They are evidence of what appeared on screen only. Search URLs are dynamic; these retained captures, rather than a future revisit, document the observation. A capture timestamp is the time the answer was saved, not necessarily the time its generation started.

## Scoring

Presence equals 1 only for a marketingagent.com or subdomain link attached to a completed composed answer. Unlinked names and links appearing only in retrieval are separate. Position is the ordinal location among answer citations; framing describes recommendation, neutral mention or mischaracterization. With presence 0, position is not applicable and framing is absent.

For each engine, a question is positive if either of its two passes contains a qualifying citation. A question is negative only if both passes can be verified negative. The rate is positive questions divided by the frozen set of twenty, multiplied by 100. With unresolved questions, the engine's rate is withheld. Repeated links do not create additional positive questions.

For negative Google scores, the entire expanded Overview and its related-source list contained no marketingagent.com link. This conservative superset is sufficient to establish absence; it would not suffice on its own to establish a positive answer citation. Every Google question has two such retained observations.

## Actual collection conditions and deviations

Google was signed in, with personalized results, in an existing browser profile. Searches were opened separately for each exact question. Repeating searches in one profile may reuse cached answers; two passes are not necessarily independent draws. Q20 pass 2 had a capture interruption before retention. Its subsequent completed capture is retained along with the failure; the interrupted response is unknown and not scored.

ChatGPT used signed-in Temporary chats in a personalized account, the visible High setting, and default automatic search. Fresh chats did not remove account context. Twenty-five complete answers were retained. Seventeen observations have verified zero citation presence; eight have unresolved citation coverage because grouped sources were not fully preserved. Fifteen planned observations were not submitted after a request-limit notice. Five completed responses already in progress were recovered after dismissing that notice, without regeneration. No ChatGPT engine-wide percentage is certified. Personalized answer text is excluded from this public export.

Perplexity used signed-out Search with the Best label. One full answer and its expanded ten-source panel were retained with zero target presence. The next question was sign-up gated. The remaining 38 planned observations were not submitted. No Perplexity percentage is certified.

This run intentionally used the available browser after the researcher authorized that context. It departs from the earlier published logged-out/fresh-profile protocol and must not be compared to it as a trend. The September 7 records failed a separate evidence audit and are not a validated earlier baseline.

## Result

Google AI Overviews: zero questions with a qualifying citation out of twenty; forty completed, verified negative observations; 0% question-level answer citation coverage. ChatGPT and Perplexity: unavailable at the planned twenty-question/two-pass scope.