Research log: copy and messaging passes

Status: append-only pass history for this organization. Each pass records date, agent and model, sources covered, earlier claims rechecked, changes, contradictions, actual verification, and remaining gaps. Queued passes below carry ready-to-run briefs; a pass that runs moves its brief into its log entry.

Pass 1: founding research, 29 September 2026

Principal instruction: the owner, 29 September 2026: research the best resources for website copy and copy in general, including the psychology behind it and how products get built and directed; have one agent review what else to research and where the blind spots are, so that a product's website and in-product copy fit its target audience.

Agents and models: one coordinating agent. Five parallel research lanes of the same model: copywriting canon, persuasion psychology, positioning and product direction, content design and UX writing, web reading evidence and testing. One read-only coverage review by a second model against a product organization's own audience, positioning and voice documents and its shipped copy. Second wave, after the coverage review: six research lanes (live copy audit, voice-of-customer desk research for one product's first audience, mentioning AI to skeptical and regulated readers, writing for an agent reader, plan and capacity copy, activation copy for a bring-your-own-agent product). Generalizable results became the source notes on mentioning AI, agent readers, and activation and plan copy; product-specific findings went to the product organization.

Sources covered: see the eight source notes, five from the first wave and three from the second; every title, author, year and URL was checked against a live source in its lane, with unverifiable items flagged in each note's verification section.

Rechecked in this pass: no earlier pass exists. Cross-lane checks: the Morkes and Nielsen 1997 figures appear identically in the canon and evidence lanes; the Unbounce 2024 readability figures appear in both with the same caveat; the Oppenheimer 2006 result appears in psychology and evidence lanes with the same direction.

Contradictions found: long against short copy and clarity against curiosity (resolved by awareness, see Conflicts); loss aversion (open); the value of llms.txt (Google says unnecessary for search; agent-mediated entry is a different job, see Principles).

Coverage review findings, stated generally:

Remaining gaps: the queued passes below.

Legal recheck, 27 October 2026

Agent and model: one agent, one model, one session, applying the owner's 27 October 2026 ruling that this organization's calendar pages carry a "Last checked" line and a monthly recheck. Scope: the EU and German legal-floor blocks in two source notes, persuasion psychology, section 3 and mentioning AI, section 2.

Sources read on 27 October 2026: the UWG annex and § 5 and § 5a UWG as amended by BGBl. 2026 I Nr. 43 (in force 27 September 2026), the AI Act Service Desk timeline and the Commission's notice on Article 50 applying from 2 August 2026, the Council's press release of 21 October 2026 on the data protection omnibus, and the European Parliament's legislative train and legal press on the Digital Fairness Act.

Earlier claims rechecked and held: Art. 50(1), (4) and (5) apply from 2 Aug 2026; the Art. 50(2) grace for systems already on the market runs to 2 Dec 2026; Annex III applies from 2 Dec 2027; UWG blacklist Nr. 5, 7, 23b and 23c are unchanged; § 5(2) Nr. 1 and § 5a(1) and (2) UWG still say what the AI note cites; the EU withdrawal button applies from 19 June 2026.

What changed: both blocks gained a "Last checked: 27 October 2026" line. The persuasion note gained the new UWG blacklist entries on environmental and durability claims (Nr. 2a, 4a to 4c, 10a, 23d and following) and § 5 Abs. 3 Nr. 4 UWG on future environmental claims, and its Digital Fairness Act line now says the proposal was still not tabled and was expected on 11 November 2026. The AI note's GDPR omnibus lines now record the Council mandate of 21 October 2026. AGENTS.md now links the sibling UX and UI organization's AGENTS.md and routes the new Loop: Recheck the legal blocks (monthly, first slot 1 December 2026). The organization had no open reference findings.

Not done: no claim outside the two legal blocks was rechecked. The first Loop cycle after 11 November 2026 checks whether the Digital Fairness Act was tabled.

Review pass, 28 and 29 October 2026

Principal instruction: the owner, 28 October 2026: run heavy review workflows over this organization and fix everywhere, public documents included, without changing settings, the public switch, members, connections, or what a product promises customers.

Agents and models: one coordinating session at a time (the pass ran across several sessions after plan limits stopped earlier ones), all Claude Opus 5.5. Workflows: understand (16 readers and one synthesizer), review (39 finder lanes), verification (a source skeptic and an editorial skeptic per batch, a third on a split), fix (one fixer and one checker per document), completeness (four critics, then a skeptic, a fixer and a checker per document).

Sources covered: every document in the organization. Fact lanes refetched the cited primary sources where they answered; 517 cited URLs were checked for reachability.

Rechecked in this pass: every numeric and legal claim the lanes sampled, among them the O'Donnell scarcity replication (four of 18 originally significant effects replicated, per the PNAS correction of November 2022; held), the Wynter panel figures, the NN/g reading figures, § 11 PAngV, § 312k and § 356a BGB, and AI Act Art. 50 with the Digital Omnibus grace.

Changes: 401 findings fixed in place; every change, its reason and its source is in Review log: 29 October 2026.

Remaining gaps: the open items and the unchecked sources listed in that review log.

Queued passes

Queued: multilingual and low-literacy end-user copy

Run when a product with vulnerable or low-literacy end users has its next user-facing copy task. Goal: rules for copy that readers at CEFR A1/A2 understand in the product's main language and in several translations (including a right-to-left script) on a small translation budget, plus a register rule per surface and reader. Already covered: Leichte against Einfache Sprache and DIN SPEC 33429:2025-03 (content design); ISO 24495-1 is not. Scope: public plain-language and easy-language standards, with German as the worked case (DIN SPEC 33429:2025 on Leichte Sprache, Netzwerk Leichte Sprache rules, Bundesfachstelle Barrierefreiheit guidance), ISO 24495-1 plain language, CEFR A1 to B1 descriptors, and published adult second-language teaching materials as register references; localization craft (text expansion, right-to-left scripts, gendered forms, formality levels, pseudo-localization, back-translation and native-review sampling on a budget) from Microsoft, Apple, Mozilla and W3C guidance; published research on second-language and low-literacy adult readers (for example, how readers react to text written on their behalf, and whether showing their own-language input next to the output builds trust). Return: a principle for A1/A2 readers and for choosing and checking translations, a register recommendation per surface, a translation QA procedure, and a Conflicts entry on Leichte Sprache against dignity for educated adult readers. Primary research, where a product runs this pass: five think-aloud sessions with participants in two languages, with informed consent and a clear rule for deleting the recordings.

Queued: German B2B buying and trust copy

Run before a product's next trust block, data-protection answer pack or pricing page. Goal: what a German B2B buyer reads to say yes, and which claims are legally safe. Already covered: trust and privacy copy rules, including "DSGVO-konform" without a named basis (mentioning AI, section 5). Scope: what a data protection officer checks (Art. 28 GDPR processing agreement, technical and organisational measures, subprocessor list, server location, transfer mechanism), what a works council checks (co-determination for technical systems that monitor employees' behaviour or performance, section 87(1) Nr. 6 BetrVG, https://www.gesetze-im-internet.de/betrvg/__87.html, read 29 October 2026); BfDI, state authorities and DSK guidance on AI; UWG sections 5 and 6 on misleading and comparative claims, testimonial consent, "DSGVO-konform" as a claim; price-unit wording that fits how the buyer budgets. Check the real provider chain (hosting, subprocessors, model and support vendors) before any claim about where data goes. Return: trust block rules, an answer pack outline, a claims register, price-unit wording. Data protection removes objections; it is never the headline. Primary research: two conversations with a data protection officer or works council contact at a buyer organization.

Queued: recipe and content page copy and search visibility

Run before a content or recipe site's next template or content push. Goal: recipe and content pages that readers finish and that search and AI answers surface, in the creator's own voice. Already covered: AI disclosure on recipe pages (mentioning AI) and structured data for AI Overviews (agent reader). Scope: Google Search Central recipe structured data and rich result requirements, schema.org Recipe and HowTo status, video and ingredient markup; evidence on the story-before-recipe pattern and jump links; Google's helpful content, site reputation abuse and AI-generated content policies for agent-produced articles; E-E-A-T through creator identity; when a translated edition of a content site earns its cost (search demand in the target language, translation and review cost, hreflang and duplicate-content handling). Return: a page template with copy slots, a structured-data checklist, a creator voice guide outline that each site fills with its creator's own words, a translation decision rule, and a measurement note using Search Console. Primary research, where a site runs this pass: its Search Console query and page export, plus a two-question on-page survey owned by its creator.

Queued: validity of simulated persona panels

Run before the next persona-panel round is used to decide copy. Goal: whether LLM-simulated readers predict real reader response, and a calibration rule. Scope: Argyle et al. 2023 (silicon samples), Aher et al. 2023, and 2024 to 2026 critiques on homogenization, sycophancy and demographic caricature; verify each. Return: a principle and a calibration protocol where every simulated round names the real-reader check it will be compared against, and a Conflicts update.

Queued: primary voice-of-customer research (product organizations)

Briefs live with each product organization: Mom-Test interviews with each product's first audience (no product shown; timeline of a recent switch or handoff; four forces), moderated first-run sessions from landing page to first useful artifact (score where copy, not software, stopped people), and conversations with each stakeholder who reads the product's output. Results become language banks in the product organization; generalizable findings return here as principle updates.

Queued: emails and sales material

Run before a product's next lifecycle, onboarding or sales email, or its next sales deck or proposal. Goal: principles and a review pass for transactional, lifecycle and sales emails and for sales decks and proposals. Already covered: subject lines and notifications (content design) and the sales pitch order (From product strategy to page copy). Scope: NN/g and public style guide research on email; the German legal floor for advertising email (prior express consent under § 7 Abs. 2 Nr. 2 UWG and the existing-customer exception in § 7 Abs. 3 UWG, https://www.gesetze-im-internet.de/uwg_2004/__7.html, read 29 October 2026); evidence on outbound sales copy and on decks read by several stakeholders. Return: principles, a playbook pass, and a legal line for the owner to place in the legal floor.