Principles: evidence-based website, product and agent-facing copy

Status: canonical research synthesis for this organization; the Research log holds its passes. These principles guide copy for websites, product interfaces, emails, sales material and text written for AI agents. They do not replace research with the real audience, and they never carry a product's own words: a product's voice, glossary, message house and claims live with that product's owner.

Use Conflicts and open questions where a decision depends on context, and check Myths before adopting a memorable number or a named effect. The method that produces positioning and a message house is From product strategy to page copy; the review procedure, checklists and voice template are in Copy review playbook; source coverage is in the Research log.

How to read the evidence

Each principle names its evidence class. Experiment: a controlled study or large field experiment; it supports the finding under its tested conditions. Observational: large correlational data, for example benchmark reports; it shows association, not cause. Practitioner: canon and single reported tests; an inspection tool, not a law. Standard or law: binding within its scope. Synthesis: our application of the above; a hypothesis until tested on the real audience. Published effects shrink by about half on replication (replication); test anything a source note rates mixed or weak (evidence ratings, replication scorecard) before relying on it.

Rules at a glance

Terms such as awareness stage, hero, fold, marketese and agent reader are explained in Words used here at the end.

Principles

Before the words

1. Position before you write. Decide the competitive alternatives, the unique attributes, the value themes with proof, the best-fit customer and the market category before drafting. Copy that keeps failing is a positioning problem. Default to an existing category plus a qualifier; a new category name adds two jobs before differentiation can start. Evidence: practitioner consensus across positioning, narrative and conversion schools (positioning); category-creation claims rest on survivor data (Play Bigger critique). Limit: positioning is a decision, not research; it needs an owner who rules.

2. Take the words from the customer. Mine reviews of the alternatives, interviews, support tickets and sales calls; keep phrases verbatim and reuse them in headlines and problem statements. Only past behaviour becomes problem copy; hypothetical enthusiasm never becomes a claim. Evidence: practitioner method with reported tests (a review-mined headline got over 400 percent more clicks in one reported test) (Wiebe); interview method (Mom Test, switch interview). Limit: single reported tests; the method, not the number, is the rule.

3. Channel desire; do not manufacture it. Copy focuses a desire that already exists onto the product. If nobody is looking for the outcome, better copy will not fix it. Evidence: practitioner canon (Schwartz, Halbert).

4. Write the story before the page. A strategy that cannot survive a narrative memo or a PR/FAQ is not ready for a homepage. Evidence: practitioner (Working Backwards, six-pager).

The page

5. Match each surface to one awareness stage. Most aware readers need the offer; product aware, what is new; solution aware, why this approach wins against named alternatives; problem aware, the problem in their words; unaware, an identification line or the shift. Give each stage and segment its own door; keep the root for one stage. Evidence: practitioner framework from long direct-response experience (Schwartz); separate entry pages per buying state (Intercom). Limit: needs research to place the reader; guessing defeats it.

6. Clarity first. Answer what it is, whether it is for me, whether I want it, and why you, in that order. Buyers cannot judge value until they understand the offer. The hero passes the "read only this" test. Evidence: practitioner model built from B2B message tests (Wynter, Shapiro); visitors often leave within 10 to 20 seconds, and pages with a clear value proposition hold them far longer (analysis of more than 2 billion dwell times) (reading). Limit: the dwell-time curve is strong; stating the value within 10 seconds is an inference from it.

7. Front-load and write for scanning. People have time to read about 20 percent of a page's words, at most 28 percent, and see roughly the first two words of list items and links. Put the message in the first two paragraphs; start headings, bullets and links with the information-carrying word; make headings tell the story alone. Evidence: a model fitted to 2005 browsing logs of 25 users for the percentages; a link test with 80 participants for the first two words; eye tracking with 500+ participants across studies for the scanning pattern (reading). Limit: the percentages are model estimates; the direction is robust.

8. Put important content above the fold and avoid false floors. 57 percent of viewing time falls above the fold and 74 percent in the first two screens; a sparse hero can make readers think the page has ended. Evidence: eye tracking, 120 participants (reading).

9. Put proof beside the claim. Use specific, close social proof (people like the reader, named, with titles and photos), recognisable logos with permission, and numbers with their source. Unknown logos and follower counts are not remembered. Never state a bad norm ("most users never finish setup"). Evidence: recall experiments on social proof types (homepage); norm effects real but smaller than popular books claim (Cialdini). Limit: recall, not conversion.

10. Show the price, or at least a starting price. B2B users ranked pricing first among 28 information types; extra costs (shipping, tax, fees) are the top reason shoppers give for abandoning an e-commerce checkout (40 percent), and 12 percent leave because they cannot see the total up front. Evidence: usability research with 79 participants; abandonment surveys (homepage). Limit: e-commerce data transfers by analogy.

11. Keep key messages from looking like ads. Boxed, bright or animated promos in banner and side positions are ignored, and so is content next to them. Evidence: eye tracking replicated across decades (reading).

Sentences

12. Replace every superlative with a fact. A number, a name, a mechanism or a customer result. Precise figures read as measured, but only if true. Evidence: practitioner canon across a century (Hopkins, Caples); concrete language raised satisfaction and spending in large field data (Berger). Limit: a false number is worse than a vague one.

13. Run the three tests on every claim. Can the reader picture it? Could it be proven false? Could a competitor sign it? If a competitor could, rewrite it. Evidence: practitioner (Harry Dry).

14. Plain words for every reader. Needless complexity lowers judged intelligence; preference for plain versions rose with education, to 86 percent among lawyers; easy-to-process statements are judged more true. Aim for grade 7 to 9 for general readers and no higher than about grade 12 for specialists. Evidence: experiments on fluency and truth judgments (robust) (fluency); a 376-person paired-passage study (plain); landing pages at grade 5 to 7 converted best in benchmark data (observational) (landing); published reading-level targets from NN/g, Shopify and Hemingway (targets, Hemingway). Limit: readability formulas are smoke alarms, not targets.

15. Cut marketese. Objective copy measured 27 percent more usable than promotional copy, concise copy 58 percent, scannable 47 percent, all three 124 percent. Readers must filter hype to find facts, and a superlative leads readers to discount everything else you say (Hopkins). Evidence: controlled study, one site, 51 users; direction confirmed by later work (reading); Hopkins on superlatives, practitioner canon (Hopkins). Limit: quote the direction, not the 124 percent.

16. Length follows awareness and stakes. A short, complete hero, then as much as a hesitant buyer needs; every section answers a real objection. Evidence: practitioner canon on both sides, resolved by awareness (disagreements); word count correlates negatively with conversion (observational) (landing).

17. Frame equivalent facts positively, and test loss framing rather than assuming it. "99.9 percent uptime", not "down 0.1 percent of the time". Loss aversion is real on average but weaker for small stakes and experienced buyers. Evidence: framing robust in a 19-country replication; loss aversion mixed (Kahneman, loss aversion).

18. Repeat one core claim consistently. Repetition raises perceived truth, so the same true claim should appear in hero, sections and call to action, in the same words. Evidence: meta-analysis of 51 studies (fluency). Limit: it works for false claims too; this is why claims need owners.

Persuasion that holds up

19. Use the robust effects honestly. Anchoring (a legitimate first number, such as the annual total or the cost of the status quo), defaults (pre-select what is genuinely best for most people and label it recommended), picture superiority (show the product), left-digit pricing. Evidence: robust across replications and meta-analyses (scorecard).

20. Treat mixed effects as hypotheses. Centre-stage and compromise effects on pricing pages, decoys, choice-overload moderators, scarcity cues and identity wording are context-dependent or fragile. Evidence: mixed or failed replications (scorecard, Myths).

21. Never use persuasion you would not show the reader. No fake countdowns, fake stock, unverified reviews, invented reference prices, confirmshaming, pre-ticked paid add-ons or cancellation harder than sign-up. They convert, which is why they are regulated. Evidence: dark patterns more than doubled sign-ups in an experiment (dark patterns); German and EU law: UWG blacklist, section 5 UWG, section 11 PAngV for goods, section 312k BGB, withdrawal button from 19 June 2026, DSA article 25 for platforms (legal context). Limit: legal notes are orientation, not legal advice.

Product and interface copy

22. Interface text is copy. Buttons are verb plus object and predict the result; errors say what happened and how to fix it, next to the problem, without blame, keeping the user's input; empty states say what state the system is in and offer the next step; confirmations only for the irreversible, naming the consequence. Example (illustrative): "Save draft", not "Submit"; "Enter a date after today", not "Invalid date"; "No invoices yet. Create your first invoice." in place of an empty table. Evidence: practitioner guidance and usability research (microcopy patterns); inline validation of hard fields improved success in a 22-person study (homepage); labels above fields, never only in placeholders (homepage).

23. Voice is constant, tone flexes. Default to casual, serious and respectful, with moderate enthusiasm; no jokes in errors, billing or data loss. Trustworthiness explained far more of desirability than friendliness. Evidence: NN/g tone experiments (voice and tone).

24. One term per concept everywhere. One glossary for marketing, interface, docs and code. When the homepage noun and the interface noun differ, users pay the translation cost, and the model itself is often unclear. Evidence: practitioner guidance (working rules); ubiquitous language (Evans).

Language and localization

25. Choose one German address form per product and enforce it everywhere. Du suits consumer, developer and peer tools; Sie suits enterprise, public sector, finance, health and older audiences. Mixing forms is the real mistake. Buttons use the infinitive, which sidesteps the choice. Evidence: Duden etiquette, Microsoft and Apple German practice (German specifics).

26. Handle gender with direct address and neutral wording first. Direct address, then rephrasing around actions, then neutral nouns and participles, then pair forms; typographic forms only where the audience expects them. Evidence: Rat für deutsche Rechtschreibung criteria, Microsoft German guide (German specifics).

27. Write for translation from the start, and write German natively where register matters. Externalized strings, full-sentence messages with named placeholders, no concatenation, room for 40 percent expansion and far more for short labels. Tone and address form do not survive machine translation. Evidence: W3C, ICU, Microsoft localization guidance (localization, length expansion).

Mentioning AI and trust

28. Lead with the outcome, not the technology. Keep "AI" out of headlines and product names unless the reader wants the mechanism; where they do, make AI the subject ("your agents", "the agent you choose"), not an adjective ("AI-powered", "intelligent"). Words alone shift trust: "AI", "algorithm" and "computer program" change perceived complexity. Evidence: adding "artificial intelligence" to descriptions lowered purchase intent, more for high-risk products (experiments); a meta-analysis of 163 studies finds appreciation only when AI is seen as more capable and personalization is unnecessary, aversion otherwise (evidence digest, homepage).

29. When you disclose AI, state roles, not a label. "AI draft from your notes, checked by you" instead of "AI-generated". A bare label makes readers assume full automation; naming the human step shrinks the penalty. Disclose before someone else reveals it; exposure costs more trust than disclosure. Evidence: preregistered experiments on labels and disclosure (evidence digest). Limit: the trust cost to the person who discloses remains under every framing tested (pooled d = 0.81, Schilke and Reimann) (bottom line); accept it rather than hide.

30. Put control inside the message. Let the reader decide how much the AI does, and say so. People use an imperfect algorithm far more when they may adjust its output, and hybrid human and AI decisions remove most resistance. Evidence: incentivized experiments (evidence digest).

31. Keep AI out of the lead in high-risk, personal and emotional contexts. Hiring, CVs, welfare, health, and emotional brand messages carry the largest aversion. Avoid terms that imply the machine decides or authors ("our AI", "AI matching", "automatically generated"). Evidence: meta-analysis and experiments (principles).

32. Meet the legal floor where it applies. Since 2 August 2026: a chat that a person could mistake for a human says it is AI at the first turn, in plain words, near the input, with reminders for vulnerable readers and no human name or face; public-interest text generated by AI gets substantive human review with a named editor, or a visible label; AI image, audio or video that resembles real people, objects, places or events and would falsely appear authentic says it is AI-generated or manipulated, and a misleading AI product or food photo can count (deep fakes); a generator provider marks output machine-readably, and a provider whose generator was on the market before 2 August 2026 has until 2 December 2026 (Digital Omnibus on AI). Terms-and-conditions notices and site-wide "we use AI" lines are not enough. Evidence: AI Act article 50 as amended by the Digital Omnibus on AI, the Commission's article 50 guidelines and the code of practice, kept current in the legal floor; the 2 December 2026 transition for article 50(2) on the Commission's AI Act Service Desk timeline (https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act, read 29 October 2026). Limit: orientation, not legal advice; dates move.

33. Answer who can see it before what it costs. Trust copy answers, in order: who can read this, which AI vendor sees it, what leaves with me. Name only verified export and deletion behaviour. Never "secure", "private" alone, "DSGVO-konform" without a named basis, or "your data never leaves" when connected agents send content to their vendor. Evidence: synthesis of transparency duties and trust research (trust and privacy).

Writing for an agent reader

34. Write so each claim survives one summary. The first reader is often a small model summarizing for another model. Put fit-relevant facts on the first screen (what it is, who it is for, who it is not for, prerequisites, price model), then mechanics; use absolute links, one term per concept and one owning page per fact. Evidence: vendor and specification guidance (source synthesis, specs).

35. Every sentence an agent may repeat must be checkable. Agents repeat adjectives as facts. No superlatives or time claims ("in a minute"); separate what is guaranteed from what is currently observed; state unknowns explicitly so the agent does not guess. Evidence: MCP server-instructions guidance against marketing language (specs).

36. Make "no" reachable. Give disqualifiers the same weight and structure as use cases, place them before or beside the use cases, state the prerequisites the product does not control (client plan, admin approval, device), and tell the agent to ask when it does not know the user. A prompt that asks "where would this fit" presupposes a fit; ask "whether" first. Test it with the fit-check protocol. Example (illustrative): "Not for you if you need on-premise hosting. Before you start: your AI client must be on a plan that allows connectors; ask your admin if you are not sure." Evidence: synthesis; protocol in fit check.

37. Ask a visiting agent only what the user would endorse. A request to a visiting agent is legitimate only if the user would endorse the exact sentence, the benefit to the user is stated and checkable, it is visible and reversible, and nothing breaks if the agent declines. Memory requests route through the user ("if your user agrees, keep this; tell them what you saved"). Never steer the agent between tools or vendors. Imperatives gain nothing: models trained to treat tool output as untrusted give it no authority (inference). Evidence: prompt-injection norms in model specifications and OWASP; vendor submission rules (ecosystem prompt engineering).

38. Tool descriptions and errors are copy. Purpose first, then when to use and when not, one canonical example, side effects and failure semantics, within the client's size cap with margin; errors say what failed, why, and the next step, with the limit as a number. Evidence: vendor guidance and the MCP specification (checklist).

39. Discovery by answer engines comes mostly from others' pages. AI answers cite earned media and community threads more than brand pages, and engines differ: Google's AI Overviews draw partly on pages that already rank in the top 10, ChatGPT search mostly on pages ranked 21 or lower. On owned pages, quotable statistics, attributed sources and exact claims help, keyword stuffing does not. Structured data must match the visible page. Evidence: one causal lab study plus observational vendor data (citation evidence, SEO and AI answer engines). Limit: thin, fast-moving evidence.

Plans, capacity and prices

40. Lead each plan with the number people read first, then translate units into actions. Say what fits in a day in words for actions ("about N edits a day"), with "about" and a footnote stating the conversion and what makes it vary. A system's own names for its limits, and raw counts such as tokens, never reach the page; the AI chat plans in the register state limits as multiples of a usage window or as message counts per window, not as token counts. Evidence: NN/g comparison-table research; dated vendor pages (plan principles, evidence register).

41. Say what happens at the limit on the plan card, before anyone buys. Three parts: what stops, what keeps working, the ways out (wait for refill, upgrade, a published opt-in overage). Extra costs are the most named reason for abandoning a checkout. Evidence: vendor exemplars that state failure behaviour, and one that does not; abandonment surveys (plan principles, homepage).

42. Promise only deployed behaviour, and keep separate promises separate. "Reaching a limit never deletes anything" and "what happens when a paid plan ends" are different promises; each is present tense only once the behaviour ships. No surprise bills is stated as a mechanism ("a limit pauses, it never charges"), never as a slogan. Evidence: synthesis; UWG section 5b on price calculation (plan principles).

43. One total price, with its unit and tax status beside it; never "free forever". In Germany consumer-facing prices state the total including VAT and the unit ("per month, incl. VAT"); a price reduction starts from a price really charged, not one charged only for an unreasonably short time, and a reduction on goods also shows the lowest price of the previous 30 days; "free" only where nothing is charged. Business-only offers may lead with net prices when only businesses can buy. Evidence: PAngV sections 1, 3, 6 and 11, where section 11 covers reductions for goods only (https://www.gesetze-im-internet.de/pangv_2022/__11.html, read 29 October 2026); UWG section 5, which presumes a reduction from a price charged only for an unreasonably short time misleading (https://www.gesetze-im-internet.de/uwg_2004/__5.html, read 29 October 2026); UWG annex number 20; BGH I ZR 99/08 (German price display). Limit: orientation, not legal advice.

44. Limit messages have four parts: what is paused, what still works, when or by what it resumes, and who can act. One number per sentence at most, no internal nouns, and name which limit binds when several exist. For agents, add a machine-readable reason and a retry time; when no reliable time exists, say so honestly. Example (illustrative): "Imports are paused: this workspace used its 500 imports for November. Reading and editing still work. Imports resume on 1 December, or an admin can upgrade the plan now." Evidence: Microsoft and Google error guidance; MCP tool error semantics (activation principles, tensions).

Activation copy for a bring-your-own-agent product

45. The first screen carries a job, not a tour. Every first-run surface names a job and the artifact it produces; setup is the way to that job. Test: can the reader say what they will use this for tomorrow without inventing a workflow? Evidence: tutorials did not improve task success and made tasks feel harder in a 70-user study; practitioner onboarding research (activation principles).

46. When the customer's own agent does the work, the agent is the actor and the person is the principal. Verbs of doing take "your agent" (named when known); verbs of deciding take "you"; the product is the subject only of keeping, showing and enforcing, never of thinking or suggesting. State the boundary positively ("Use the agent you already work with. You keep everything it makes.") and use the negation once, where an expectation would mislead. Evidence: MCP human-in-the-loop guidance; synthesis (activation principles).

47. Explain access beside the action it enables. Name the client, the scopes, where it applies and how to revoke it at the moment of connecting; lead with the benefit and put the limit beside it; show the redirect destination prominently. Reasons raise permission grants; excessive scopes cause abandonment. Evidence: permission research cited by NN/g; Google, GitHub and Slack consent guidance; MCP security best practices (activation principles, setup and consent).

48. Empty states say which empty they are. No agent connected, connected but nothing yet, filtered or no access: each teaches what fills the area and offers one direct path, ideally a ready prompt that runs on a fresh account. Evidence: NN/g empty-state guidance (activation principles, empty states).

49. Setup copy ends with a check and a first job, and version pins cannot rot quietly. One pasteable block per client; say what sign-in will show; state version floors, not exact versions; every third-party click path carries "verified on date against client and version" and an owner who re-walks it; one home per pinned fact. Evidence: dated developer exemplars and timeless-documentation guidance (developer copy).

Testing and learning

50. Test comprehension with real readers before traffic allows A/B tests. Below tens of thousands of visitors a month, conversion A/B tests on copy are mostly noise. Use five-second tests, blind tagline tests, message-testing panels, first-click tests and moderated sessions; grade open answers, not confidence scales. Evidence: sample-size arithmetic and experimentation research; a comprehension study where confidence scores missed misunderstanding (testing, playbook).

51. Hold every copy opinion loosely. Even at Microsoft, Google, Booking and Airbnb only 8 to 33 percent of ideas win; micro-copy tweaks such as button colour and call-to-action wording average near zero across about 2,600 categorized retail and travel experiments. Spend effort on the offer, clarity, proof and price, not on button colour. Evidence: experimentation programmes (experiments) and a vendor meta-analysis that covers all its tests, not only winners (landing).

52. Learn from every correction. A founder's, reviewer's or test's correction becomes a case in the product's organization and, when it generalizes, a rewritten rule here (learn from corrections).

Words used here