# Batch 5 retained exercises

Method version: 2026-10-06. Two bounded teaching exercises were commissioned for the approved review package. They are not a model comparison, an accuracy benchmark, a controlled experiment or an SEO performance study.

## Shared conditions

- Product: ChatGPT web interface, signed-in Chrome session, fresh conversation for each exercise.
- The model menu inspected before the keyword run displayed `6`, `Latest` selected, and `Pro` effort. The composer for both runs displayed `Pro`. These are observed interface labels, not an independently verified backend model identifier.
- Backend snapshot, temperature, sampling seed, system prompt and hidden account-level memory/customization were not exposed or independently audited. Both prompts explicitly requested no browsing, tools, other-conversation context or personal memory. That instruction is retained; it does not establish an isolated laboratory environment.
- Only the public prompt text was submitted. No private analytics, customer records or attachments were supplied. The Google searches used to review topics were performed separately by the researcher, not represented as model browsing.
- One initial request per exercise. No selected best-of-N response and no follow-up correction is part of the initial response. Any interruption or second attempt must be recorded rather than silently replaced.
- Raw response content is retained separately from review decisions. Copy-button output or rendered response text preserves the returned wording; serialization details are documented with the retained files. No hidden reasoning is included.
- Source/inventory checking and package preparation were assisted by Codex. The package is prepared for owner review; no final human or subject-matter sign-off has been fabricated. The proposed author must accept responsibility before publication.

## Keyword exercise

Input: six researcher-selected editorial seeds and summaries of 11 existing public pages. The supplied set is deliberately bounded, not a complete inventory or an observed demand dataset. The retained `page-inventory.json` records those source paths, headings and hashes at the inspected repository commit.

Task: return 30 distinct candidates, including six supplied seeds, with plausible task, provisional action, existing owner and the next evidence check. Retain raw generated candidates before looking at review decisions. Classify provenance as supplied seed or generated suggestion. No numeric keyword metrics are solicited.

Review: compare each candidate with actual page content and preserve the two separately observed exact-query Google samples for `ai keyword research` and `ai generated content seo`. Do not imply that other variants were searched. Use inventory-only or related-query evidence labels where appropriate. A suggested owner is an editorial judgment, not proof of ranking overlap or search demand. New-page decisions remain proposals.

Fields should distinguish the original task/action/page from the reviewed task/action/page, evidence scope/date/market, rationale, next action and reviewer status. Missing metrics remain not collected. The model's proposed evidence check is a task to do; it must not be mistaken for completed evidence.

## Content exercise

Input: the eight facts and A/B/C references in frozen pack `B05-KEYWORD-METRICS-v1`, checked 6 October 2026. The full source-pack file is retained in `research/content-source-pack.md`. The exact submitted prompt paraphrases no additional factual claims and is kept in `content-prompt.txt`.

Task: produce a short beginner explanation plus title, description, proposed alt text for an unmade teaching image and a headline/description JSON fragment. The JSON is not requested as complete schema. No fictitious business result, author, measured metric or intentional error is supplied.

Review: inventory every substantive assertion and metadata element, map to the linked primary source, then keep, qualify, correct, remove or leave unresolved. An accurate response is a valid outcome. Do not invent errors merely to dramatize the guide. Image alternative text remains proposed until checked against a real matching image. Two metadata strings do not establish complete, valid, eligible structured data.

The reviewed version must preserve supported qualifications, identify changes and separate factual corrections from clarity edits. This exercise's source pack is narrow and carefully bounded; it cannot establish how well the product would perform without sources, on other topics or at scale.

## Search sample context

The accepted comparison sets were observed 6 October 2026 through Valentin in Chrome: US English desktop, visible `10007, New York, NY` footer and `Results are not personalized` indicator, signed-in browser. These are dated location samples, not universal national positions. The first five standalone standard organic destinations exclude ads, AI Overview references, People Also Ask, sitelinks and carousels. A standalone Reddit result counts. Public tool shells and visible discussion content were readable; paid features, generated product outputs and all collapsed replies were not tested.

See `research/competitor-gap-map.md` for the actual destinations and retained review. No measured demand, traffic lift, citation rate, lead gain or ranking effect follows from these exercises.