> Current revision, September 20: **Family 07, three focused question ranges**.
> Shopify: returning cohorts, discount share, products bought together.
> Recharge: signup retention, churn, new recurring revenue.
> Attribution: discovery, unknown marketing sources, purchase journeys.
> Each is an 11-second sequence: three questions at three seconds each, then a
> two-second invitation. ChatGPT, Big Idea, layout and destination stay constant.
> The matching static shows all three questions. These are supported question
> invitations, not recorded answers. Native copy matches each subject.
> The 13-second real Shopify workflow remains the proof comparison, unchanged.
> Twelve motion exports, three statics and 27 exact draft units across five channels.
> Family 06 exports and immutable handoff remain the previous baseline.

# Talk to your data: matched topic cohort

Status: local creative exploration. No uploads, campaign changes or spend authorized.

For ecommerce operators who know their business question but do not realize they
can ask it in their existing AI. SourceMedium connects and maintains the data;
ChatGPT, Claude and compatible agents supply the workspace. The creative invites
an actual question, rather than inventing an answer or promising an outcome.

Current output: `.traffic/creatives/question-family-07/`.
Review: https://sourcemedium-demo-directions.lordhumunguz.chatgpt.site/talk-to-your-data/.
Local: http://127.0.0.1:4404/question-family-07/paid-copy.html.

The earlier Family 04 addition was a matched **Shopify question** cohort in ChatGPT:

| Question | Website prompt ID | Source |
| --- | --- | --- |
| Which customer cohorts keep coming back? | `retention-review` | `src/data/use-cases.ts` |
| How much revenue goes to discounts? | `discount-review` | `src/data/mcp-showcase.ts` |
| What products are bought together? | `bundle-affinity` | `src/data/mcp-showcase.ts` |

Each has a six-second 16:9, 4:5 and 9:16 invitation plus a landscape static.
Source logo, AI client, layout, timing, music, native copy and destination are
held constant; only the question changes. These are designed invitations, not
new recorded answers. `questions.json` retains each website prompt ID, reviewed
source, product documentation and account-coverage conditions. Website wording
is a starting point; narrower Shopify coverage is intentional.

Family 04 retains the prior topic cohort and real Shopify proof: 21 motion
exports and six statics, selected into 48 draft units across five channels.
Three new ideas add 21 placement units, not 21 independent experiments.
Family 03 remains the proof-study baseline; Family 02 remains the original matched
three-topic baseline. Earlier outputs and immutable handoffs are preserved.

Family 04 native ad copy invited viewers to browse the upcoming MCP prompt library.
Its tracked destinations retained `/mcp#prompts`; the displayed brand URL remains
SourceMedium.com/mcp. The updated page was confirmed live in Chrome on September 20:
“Your commerce data. In your favorite AI.”, the film, prompt library and maintained
connections/models/definitions message are present. The walkthrough CTA reaches
`#demo`. The preview at https://com-dev.sourcemedium.com/mcp remains historical
review evidence. The user confirmed the live form/CRM flow works and declined
a repeat QA submission on September 20. No test lead was created; the confirmation
is operator-provided rather than an agent-observed CRM receipt.

## Creative structure

- **Concept:** talk to a recognizable source or business domain.
- **Variation:** topic + documented question selection + compatible AI workspace.
- **Format:** the active question ranges use 11 seconds / three questions plus an
  ending, or a static with those same three questions. Compose motion independently
  at 16:9, 4:5 and 9:16. Earlier single-question formats remain historical.
- **Ad unit:** an exact export + native copy + CTA + placement + experiment.
- **Campaign:** objective, audience, budget, bid strategy, schedule and exclusions.
- **Distribution:** approved package, actual provider IDs and read-back receipt.
- **Learning:** join delivery to the exact unit revision, then revise the question
  or creative hypothesis. A new export is not a new concept.

Source of truth: [question catalog](../tools/demo-video/projects/questions/questions.json),
[format registration](../tools/demo-video/projects/questions/project.json), and
[exact ad selections](talk-to-your-data-sep-2026.json). The existing Creative OS
owns approval and packaging. Campaign specs retain ownership of money and targeting.

## Earlier topic comparison (retained)

Shopify / ChatGPT is the template anchor. The selected questions are:

| Topic | Question | Comparison |
| --- | --- | --- |
| Shopify | How much revenue comes from repeat customers? | Repeat-revenue relevance |
| Recharge | Which signup months have the best retention? | Subscription retention relevance |
| Attribution | Where do customers first find us? | Observed first-touch discovery relevance |

All three hold the workspace, six-second duration, layout, soundtrack, native
copy and CTA fixed. Topic, source logos and question change together. Compare
within the same audience and placement; this is a topic test, not a pure headline
test. Source-specific targeting would also change the audience and must be treated
as a separate test. Static and motion results stay separate.

Keep the earlier Claude Code, Muse and Grok variants for a later workspace test.
Do not launch every topic/workspace/format combination together.

Start with one relevant audience and a small matched set, not every combination.
A source-specific ad should reach operators using that source. Verify actual
platform targeting options before implementing this audience intent; do not assume
platform-user targeting exists or that native AI placements identify product users.

## Shopify proof study

The 19-second cut shows the actual ChatGPT question (0–4s), SourceMedium retrieval
(4–8s), matching DemoCo percentage answer (8–16s), and a short CTA (16–19s).
Native app motion comes from timestamped browser captures; long waits are removed.
No earlier LTV/promo answer or generated screenshot is substituted. Landscape,
feed and portrait use the same sequence with inherited camera overrides.

Compare the complete treatment against the Shopify invitation in matched audience,
placement and native copy. Duration and proof change together. Keep Recharge and
attribution as topic invitations until their own matching workflows are verified.
The SourceMedium activity and answer were observed in the real app, and the
arithmetic reconciles. Raw tool-result verification remains a release check;
ChatGPT's own assertion of verification is not an independent receipt.

## Coverage and proof

The short invitations are explicitly branded **Try asking** cards, not simulated app screenshots.
Those cards assert no answer, customer metric, tool execution or result. Questions carry
source documentation and coverage notes in the catalog. Read those before capture.

- Shopify questions use valid orders, product lines and the reporting timezone.
- Recharge subscription reporting requires the relevant Pro tables. Check bundle
  component grain and mature cohort age. Geography/channel fields are placeholders.
- Fulfillment examples concern configured fulfillment **costs**, not shipment
  tracking, late deliveries or carrier scans. Confirm configured cost inputs.
- Attribution is observed credit, not proof of causal lift. Match first/last-touch
  definitions and coverage; never sum attribution dimensions.
- Ecommerce questions require the selected sources and relevant models to be
  available for the specific demo or customer's account.

The source catalog is an invitation library. Before an idea becomes recorded
proof, run it in the real selected app, inspect SourceMedium context/schema calls,
check the answer, then capture the question, evidence and matching output.

## Channel and campaign intent

All references below are **planned intent**, not provider resource IDs. Budget,
bid amount, start/end dates and live status are unset. The account specifications
remain unchanged. Prepare the exact operator-reviewed change before distributing.
Do not clone these drafts into an enabled campaign merely because the assets exist.

### talk-to-your-data-meta

Goal: qualified demo requests. Candidate placements: feed 4:5, Reels 9:16.
Keep topic tests within matched audiences and placements. Separate prospecting
from any consented, eligible remarketing audience. Select the validated conversion
and optimization event from the account doctrine. Allocate from an explicitly
approved test budget; do not silently add spend. Bid strategy and caps belong in
`meta-campaigns.yaml`, with exclusions, schedule, attribution settings and pixel
verification included in the proposed change.

### talk-to-your-data-reddit

Goal: qualified visits and downstream demo requests. Candidate media: 4:5 feed
video and image. Use a relevant ecommerce/operator audience with verified targeting
availability. A title change needs a new promoted post; retain the old post/ad
mapping for rollback. Budget and bid settings belong in the reviewed campaign/ad
group plan and `reddit-campaigns.yaml`. Traffic has no Reddit apply workflow.

### talk-to-your-data-x

Goal: qualified visits and demo requests. Candidate media: 16:9 motion and image.
Use identical audience and placement when comparing the three topics. Verify account
credentials before provider preparation; warehouse delivery is a separate surface.
Funding instrument, budget, bid, audience, schedule and placements belong in
`x-campaigns.yaml` and the operator-reviewed change. Traffic has no X apply workflow.

### talk-to-your-data-chatgpt

Image cards with native title/body. Video exports are not assigned to ChatGPT Ads.
Existing account contract, review status and context hints govern eligibility;
no audience identity is inferred from a card placement. Keep budget/bid/targeting
in `chatgpt-campaigns.yaml`. A creative update can trigger review again. Read back
provider state before recording any update as complete.

### talk-to-your-data-google-video

YouTube: the selected six-second landscape candidates are for bumper evaluation;
vertical candidates are for a compatible reach configuration. The longer twelve-
second invitation cuts remain in the library, outside this matched cohort. The handoff
uses Traffic's `instream`/`shorts` placement names. It does **not** configure a
bumper campaign merely by exporting a six-second file. Match the campaign subtype,
objective, inventory and bid strategy in the explicit provider plan.

Google's current [format guide](https://support.google.com/google-ads/answer/2375464)
describes bumper ads as 5–6 seconds. Review the exact export duration and account
format eligibility again before upload. The [Shorts guidance](https://support.google.com/google-ads/answer/16041697)
recommends 10–30 seconds for action-oriented creative. Reserve the six-second
Shorts candidate for a compatible reach configuration; a longer, verified workflow
is the next action-oriented candidate. Do not lengthen a six-second card merely
to fill a duration target. Multi-format campaigns may optimize asset
delivery across formats, so a handoff placement label is intent, not a platform
restriction. YouTube hosting/channel ownership,
video IDs, campaign subtype, budget/bids, audience, exclusions and conversion
measurement are still distribution inputs. No YouTube upload is authorized here.

References checked 2026-09-19: [Reddit video](https://www.business.reddit.com/advertise/ad-types/video-ads),
[X creative specifications](https://business.x.com/en/help/campaign-setup/creative-ad-specifications).
These are creative candidates, not a claim that every provider accepts every
combination. Verify the exact account/placement at release. Two- or three-second
teasers are optional brand reminders; do not compress analytical questions into them.

## Historical release checks, 2026-09-19

At the September 19 check after preparing the proof cut, `https://sourcemedium.com/mcp` displayed
**Page Not Found / ERROR 404**. That dated observation is superseded for creative
planning and availability by the September 20 production check above. Production
measurement remains unverified. A form embedded in
the 404 page does not establish message match. No form was submitted. The prepared website `/mcp` page renders locally on the
clean `rev-mcp` branch; `main` lacks that route. Its walkthrough CTA correctly
reaches `#demo`. Fixing production requires the planned website release, not a
second landing page in Traffic. No website deployment was performed.

The website attribution suite passed 30 tests, covering capture/persistence and
metadata handling with mocked external calls. The browser URL retained all test
UTMs, but the hidden input readout was empty despite the metadata binding being
present. Submission-time hydration has not been exercised through the browser;
do not claim an end-to-end conversion pass. Verify the deployed form payload and
confirmed lead event before launch.

UTM campaign names now identify this family and channel (`talk-to-your-data-meta`,
etc.), with the exact unit ID in `utm_content`. This corrects the original generic
channel-only campaign labels without changing historical handoffs. Verify these
values survive the live destination and reach the successful `lead_signup` event
and CRM metadata; a source-code check is not an end-to-end conversion test.

Campaign references below name planned intent. Objective, event, audience,
exclusions, schedule, budget/bid strategy and caps still need a concrete,
operator-reviewed provider change. The selected media does not choose those
financial inputs or create a serving campaign. Choose one channel/cohort to start;
keep the remaining placement exports ready without fragmenting an unknown budget.

## Release and measurement

Select exact units, verify the live `/mcp` destination and message match, listen
to the Big Idea mixes, inspect placement UI overlays, confirm music/logo usage,
complete Creative OS reviews, then package the approved cohort. Do not reinterpret
an earlier film's approval as approval of these new prompts or assets.

Only after an explicit distribution instruction should an operator prepare/upload
approved media, record actual provider campaign/ad/media IDs, confirm status and
attach the read-back receipt. Join dated bindings to the immutable handoff with
`paid_results.py`. Unknown outcomes remain unknown. Qualified outcomes, not views
alone, decide which questions graduate into deeper demo creatives.