# ChatGPT capture protocol

The second half of the flagship. Google side captured 2026-09-02; this is the comparison.

## Setup (once)

**Session state — decide and record which you used.** The Google captures were logged out, so
logged-out is the symmetric choice. But ChatGPT Shopping features may require an account. Check
first:
- If shopping/product cards work logged out → use logged out. Cleanest.
- If they need an account → log in and use **Temporary Chat** for every run. It disables memory
  and personalisation. Record that you did this; it is a methodology-block line, and it is a
  weaker control than logged-out, which must be stated honestly.

**Use the web app, not mobile.** Easier to screenshot, and the model name is visible.

**Record before you start:** date, time + timezone, model shown in the UI, session state
(logged out / Temporary Chat), and your city.

## The loop

For each of the 10 queries, 5 times each. **50 runs total.**

1. **New chat every single run.** Not a follow-up in the same thread. Context from run 1
   contaminates run 2, and that is the single easiest way to ruin this dataset.
2. Paste the query **exactly** as written in the query list. Same string every time.
3. **Do not steer.** No follow-ups, no "what about X", no clarifying. Take the first answer as
   given. If it asks you a clarifying question, record that as the outcome and move on — a
   clarifying question instead of a recommendation is itself data.
4. Screenshot the full answer.
5. Record the row (see template).

**Watch for whether ChatGPT shows its search queries** — a "Searching for…" line or a sources
panel. If it does, **capture that string verbatim.** That is the actual fanout query, and it
would let us replace the guessed Google reductions with the real ones. That single detail would
materially strengthen the study.

## What to record per run

| Field | Notes |
|---|---|
| Products named, in order | Verbatim as ChatGPT wrote them |
| Product card rendered? | Y/N — the shopping surface vs prose |
| Sources cited? | Y/N, and which domains if shown |
| Fanout query shown? | The literal search string, if visible |
| Refusal / hedge / clarifying question | Instead of products. This is data, not a failed run |
| Timestamp | |

## The 10 queries — ask these verbatim

**Single-constraint**
1. What's the best moisturizer for dry skin?
2. Best face sunscreen for daily use
3. What vitamin C serum should I buy?
4. Best cleanser for acne-prone skin
5. What's a good retinol to start with?

**Multi-constraint**
6. What moisturizer should I use for sensitive skin under $30?
7. Best fragrance-free sunscreen for oily skin that doesn't pill under makeup
8. What should I use for hyperpigmentation if I'm pregnant?
9. Best affordable skincare routine for combination skin in your 40s
10. Gentle retinol alternative for sensitive skin that won't cause purging

**Note the asymmetry, and state it in the methodology block.** ChatGPT gets the conversational
question. Google got the short product-focused reduction. That is deliberate — it mirrors how
each surface is actually used — but it is a real difference and hiding it would be a hole a
critic could put a fist through.

## Expect these, and record rather than discard

- **Q8 (pregnancy) may hedge or refuse.** A medical-adjacent query producing a "talk to your
  doctor" answer instead of a shelf is a finding, and arguably a good one.
- **Q9 asks for a routine, not a product.** It may return a multi-step answer rather than a
  ranked list. Record the products named anywhere in it.
- **Prose instead of cards.** Published figures put ChatGPT's product-card rate around 87% and
  Perplexity's near 1%. If you see prose, note it.

## Do query 1 first, all 5 runs, then stop

Look at the shape of the data before doing the other 45. If the answers come back in a form the
template does not fit, better to find out after 5 runs than after 50.

## Then

Overlap gets computed at top-10, top-40 and full-grid depth against the Google shelves, and
reported as a curve. Membership, not rank — rank is unstable between sessions on both sides.

---

## CAPTURE METHOD — REVISED 2026-09-02 (screenshots were insufficient)

Viewport screenshots truncated the answers on q01 and q02, making every brand list a lower
bound and the overlap percentages an upper bound. Do not rely on screenshots for the data.

**Primary capture is TEXT, not images.**

1. **Use ChatGPT's copy button** on each response (the copy icon under the answer). It copies
   the full response text regardless of scroll position. Paste that.
2. **Screenshot separately, for evidence and for the things text misses:** product cards (price
   and merchant), ad units, and the sources chips. These are UI, not prose, and will not appear
   in the copied text.
3. **Full-page screenshot does NOT work on ChatGPT.** Confirmed 2026-09-02. Google Shopping is
   a normal document whose body scrolls, so DevTools captures the full scroll height. ChatGPT
   is a fixed-height app that scrolls inside an inner container, so the document body never
   grows and both Chrome's *Capture full size screenshot* and Firefox's *Save full page* return
   a single viewport. Do not waste time on it.

   Use instead:
   - **Cmd+P → Save as PDF** — print rendering walks the whole conversation, not the viewport.
     This is the reliable full-thread evidence capture.
   - **Zoom out to 50-67% (Cmd+minus) then screenshot** — at that zoom the answer, cards and
     ad usually fit one frame. Fine for capturing cards/ads specifically.

**So per run, send:** (a) the copied response text, (b) a note of card price·merchant, (c) any
ad brand, (d) the source chip names.

Text is the dataset. Images are the evidence.

## Re-do required

q01 and q02 need re-capturing as text before any number from them is published.
