At a glance
- Research comes first. Sources are retrieved from Europe PMC, OpenAlex and named institutions before a word is written.
- Every claim is cited. Factual claims carry a numbered citation, and every citation links to the original.
- A person decides. No piece publishes automatically. A human editor reads and approves every one.
- We say how strong the evidence is. A single pilot study is not written up like a meta-analysis.
- Money never picks the recommendation. Affiliate links are labelled, and they are only added where they help.
- Errors get fixed. Write to hello@unleashed.vision and we will look at it.
How a story is made
Every piece goes through the same steps, in this order.
- A question. Each desk works from a list of questions readers actually ask, such as how long a habit takes to form or whether a nap helps focus.
- Retrieval. Our research software searches the literature databases for that question and keeps a shortlist of real, indexed papers, favouring reviews, meta-analyses and trials.
- Drafting. A language model writes the draft using only that numbered source list. It is not allowed to cite anything else or to add author names, journals, figures or DOIs that are not in the list.
- Automated checks. Software checks the draft against our rules: every citation must point to a real source, statistics must be cited, links must go to our own pages or registered offers, and there must be no invented first-hand testing or guaranteed outcomes.
- A fact-check pass. A second model reads each cited sentence against the abstract of the source it cites, and flags anything unsupported, misattributed or overstated.
- Human review. Drafts that fail the checks never reach the editor. Those that pass are read by a person, who approves the piece, sends it back with changes, or rejects it.
Where sources come from
We draw on three kinds of source:
- Europe PMC, which indexes PubMed and the wider biomedical and life-science literature, for reviews, meta-analyses and trials, with PubMed IDs and DOIs.
- OpenAlex, an open index of scholarly work in every discipline, which we use for psychology, economics, management and education.
- Named institutions such as the NHS, GOV.UK and the WHO, for guidance and official figures.
For us, a source is verified if one of those indexes returned it with a stable identifier (a DOI or PubMed ID), or if the institution’s page loaded when we retrieved it. The model judges what a study found from its abstract, not its title.
How we weigh evidence
Not all studies count the same. We use the standard hierarchy, strongest first, and we name the kind of study in the text so you can judge it yourself.
- Meta-analyses and systematic reviewsPool many studies on the same question. The strongest evidence we cite.
- Randomised controlled trialsTest cause and effect directly, though often in small or specific groups.
- Cohort and observational studiesShow what is linked with what, but not what causes what. We write “is linked with”, not “causes”.
- Small pilot studiesUseful early signals. We give the sample size so you can see how small.
- Animal and cell researchAlways labelled as such, and never presented as proof of an effect in people.
- Expert opinionGuidance from institutions and professional bodies, cited by name.
When the evidence is mixed, weak or mostly from animals, we say so. When studies disagree we say that too, and give the likely reason, such as the dose, the people studied or how long the study ran. Research breakdowns close with a plain verdict: strong, moderate, limited or insufficient evidence.
Citations
Claims are followed by a small number, like this1. On a computer, hover over it to see the source; on a phone, tap it to jump to the reference list at the foot of the piece. Each reference gives the authors, year, title and journal, with a link to the original by DOI or PubMed ID where one exists, and a link back to where you were reading.
At the top of every piece, a summary shows what kinds of evidence it relies on, for example “2 meta-analyses · 4 RCTs · 1 institution”. Every source we have cited is also filed in the Research Library, with the pieces that rely on it.
Human review
Nothing on this site was published by software alone. Every piece is read in full by a human editor, currently the founder of Unleashed, before it appears. The editor checks the claims against the cited sources, the tone, and whether the piece is actually useful. They can approve it, ask for changes, or reject it outright.
The reviewer’s name appears in the byline of every published piece, next to the date it was last updated.
What AI does, and doesn’t
We use AI to find and summarise research and to write first drafts. It does this fast, and the rules above keep it honest. We don’t pretend otherwise: every piece carries the line “Researched with AI assistance and reviewed by a human editor”.
AI does not decide what publishes, and it is not allowed to:
- claim first-hand experience, such as “we tested” or “I tried this for 30 days”;
- invent statistics, studies, quotes, product specifications, prices or people;
- cite a source for something that source does not say.
The byline “Unleashed Research Desk” means this whole process: the research software, the model, the checks and the human editor.
Health and wellbeing
Our health content is educational. It is not medical advice. It can’t take account of your circumstances, medication or medical history. Talk to your GP, a pharmacist or another qualified professional before changing anything about your health, and especially if you are pregnant, take medication or have a health condition.
We never claim that anything prevents, treats or cures a disease. When we report studies of people with a condition, we describe what the study found and who took part. On supplements, we report findings neutrally, and we use only authorised health-claim wording in recommendations.
If you’re struggling, you can call Samaritans free on 116 123, any time (UK). In an emergency, call 999.
Affiliate links
Some pieces link to products. Some of those links are affiliate links: if you buy through one, we may earn a commission, at no extra cost to you. That money helps pay for the newsroom. It never decides what we recommend.
- Any piece with an affiliate link says so in a note above the text. Nobody has to find it in a footer.
- Every product link goes through our own address (
unleashed.vision/news/go/…) and is markedrel="sponsored"for search engines. You can see where a link is going before you follow it. - Products are only included where they help you do what the piece is about. Research breakdowns carry none unless a product is the subject.
- We describe products only from published product information. We don’t claim to have tested anything, and we don’t invent star ratings, awards or specifications.
- Some links point to shops whose affiliate programmes we haven’t joined yet. They earn us nothing, but we label them in the same way so the rule stays simple.
When you follow a product link we count the click: which link it was and which page it was on. That is all. We don’t store your IP address, location or browser details, and the link sets no cookies of its own.
Our own products
Unleashed also makes software, including the Unleashed OS app and the AI Receptionist. Pieces may mention them where they are relevant. Any card that promotes them is labelled “Made by Unleashed”, so you always know when we’re talking about our own work.
Corrections
We get things wrong sometimes. If you spot a factual error, a citation that doesn’t support its claim, or a broken link, write to hello@unleashed.vision with the page and what you think is wrong. A person reads every message.
We fix confirmed errors promptly, and the piece’s “Updated” date changes. If a correction changes the advice a piece gives, we add a note to the piece saying what changed.
The house standards, in full
These are the working rules our research software writes to and our editor reviews against, published as they are. They are written as instructions, so they are blunter than this page.
Editorial standards: voice and craft
These are the rules every Unleashed Newsroom piece follows. The engine reads this file at runtime, and a human (Elliot) approves every piece before it publishes.
Who we write for
Ambitious, busy people who want to perform better: founders, creators, students, shift workers, owner-operators. They're smart and time-poor, and they've already read the lazy version of every self-help article. Give them the version with the evidence attached and the practical next step spelled out.
Voice
- Direct, warm, grounded. Write like a sharp friend who has read the research and gets to the point. No hype, no guru tone, no "hustle" clichés.
- UK English: optimise, behaviour, colour, programme, organisation. Dates like 28 September 2026. Metric units first.
- Second person ("you"). The newsroom is "we" ("we looked at…"). Never invent first-person experience: no "I tried this for 30 days", no "in our testing", no "we tested". We did not test products; we read research and product information.
- Short paragraphs (1–4 sentences). Vary sentence length. Concrete over abstract: numbers, examples, specific actions.
- Every piece earns its length. If a section doesn't help the reader act or understand, cut it.
Banned (the tells of lazy AI copy)
Never use: "in today's fast-paced world", "delve", "dive into", "unlock", "unleash your potential" (the brand name is fine; the cliché isn't), "game-changer", "revolutionise", "tapestry", "navigate the landscape", "it's important to note", "in conclusion", "ultimately", "embark", "journey" (unless literal), "holistic", "synergy", "supercharge", "hack" (as a noun for a tip), "science-backed" without a citation in the same sentence, "studies show" without a citation, rhetorical openers ("Have you ever wondered…?"), and closing summaries that repeat the article. Keep em dashes rare: at most one per ~200 words. Prefer full stops and commas.
Structure
- The page supplies the H1 (the title). The body starts with
##sections and never uses#. - Open with the substance: the first paragraph after the short answer states what the reader will get and why it matters, with no throat-clearing.
- Use
##for main sections (phrase many of them as the questions people actually search, e.g. "How long does it take to form a habit?"),###for sub-points. - Use a table when you're comparing three or more things across attributes.
- End with a practical section ("What to do this week", "A simple protocol", "How to choose"), never a recap.
Accuracy
- Every factual claim that isn't common knowledge carries a citation
[S#]to a provided source. See02-evidence.md. - If the evidence is mixed, weak or mostly animal/lab, say so plainly. Calibrated honesty is the brand.
- Never invent statistics, studies, quotes, product specs, prices, dates or people.
Evidence standards: citations and research
The newsroom's edge is that every claim is traceable. Readers, search engines and AI answer engines trust pages that show their work.
Where sources come from
The engine retrieves real, indexed literature before a word is written:
- Europe PMC (biomedical and life sciences, which indexes PubMed): reviews, meta-analyses and trials, with abstracts, PMIDs and DOIs.
- OpenAlex (every discipline, including psychology, economics, management and education), with DOIs and citation counts.
- Curated institutional sources (NHS, NIH Office of Dietary Supplements, WHO, CDC, ONS and similar), checked for a live URL.
A source counts as verified when it was returned by one of these indexes with a stable identifier (PMID or DOI), or when its institutional URL answered HTTP 200 at retrieval time.
Citation rules (hard rules: violating any one fails the piece)
- Cite only the numbered sources you were given, as
[S1],[S2], or grouped[S1, S4]. Never cite a number you weren't given. - Never write a URL, DOI, journal name, author name or study detail that isn't in the provided source list.
- A citation must support the sentence it's attached to. Use the abstract, not the title, to judge what a study found.
- Every number (%, sample size, effect size, hours, milligrams) must either carry a citation or be a plain, uncontroversial fact.
- Never cite a source for something it doesn't say. If no provided source supports a claim you want to make, drop the claim or frame it as practical advice, not a finding.
Weighing evidence (say it in the prose)
Strongest to weakest: meta-analyses / systematic reviews → randomised controlled trials → cohort / observational studies → small pilots → animal / cell studies → expert opinion.
- Name the study type when it matters: "a 2023 meta-analysis of 22 trials", "one small pilot study (n=24)".
- Animal and cell research must be labelled as such, and never presented as proof of a human effect.
- Correlation isn't causation: write "is linked with", not "causes", for observational data.
- Prefer absolute over relative framing when both are available.
- If studies disagree, say so and explain the likely reason (dose, population, duration).
Research breakdowns and digests
- Break down 3–8 key studies: what they did (design, who, how long), what they found, what it means, and the caveats.
- Close with a Bottom line section giving a calibrated verdict: strong / moderate / limited / insufficient evidence.
Claims policy: health, supplements, money, competitors
Our readers make decisions about their bodies, money and businesses from what we write. Search engines treat these topics as YMYL ("Your Money or Your Life") and hold them to a higher standard. So do UK advertising rules: affiliate content is marketing, and ASA/CAP rules apply to it.
Health and wellbeing
- We publish education, not medical advice. The site adds a standard disclaimer; don't write your own.
- Never claim that anything prevents, treats or cures a disease or medical condition (depression, ADHD, insomnia disorder, diabetes, anxiety disorders and so on). You may report what studies found in patients ("in a trial of adults with insomnia, CBT-I improved…"), in neutral, study-reporting language.
- No diagnosis-by-article. Where symptoms could signal a medical issue (persistent low mood, sleep apnoea signs, chest pain, suicidal thoughts), advise speaking to a GP or a qualified professional.
- Crisis content: if a piece touches self-harm or suicidal feelings, include: "If you're struggling, you can call Samaritans free on 116 123, any time (UK)." Don't write pieces whose main topic is self-harm.
- Pregnancy, children, medication interactions and eating disorders: add "check with your GP or pharmacist first" wherever the advice could interact.
Supplements (the strictest area)
- Report research findings neutrally, with study type and size. Never write "X boosts your testosterone", "X melts fat" or "clinically proven".
- When you describe what a nutrient does in a product-recommendation context, stick to the wording of GB-authorised health claims. Examples: "magnesium contributes to a reduction of tiredness and fatigue", "vitamin D contributes to the normal function of the immune system", "creatine increases physical performance in successive bursts of short-term, high-intensity exercise" (this one applies at 3 g/day).
- Always mention: talk to a GP or pharmacist if you take medication, are pregnant or have a health condition. Name the known risks and side effects when the evidence covers them.
- Don't recommend doses above established upper limits. Where the NHS gives a figure (e.g. vitamin D 10 micrograms a day in autumn and winter), you may state it only if a provided source supports it.
Money and business
- No income promises ("make £10k a month"). No guaranteed outcomes. Illustrative numbers must be labelled as such or cited.
- For B2B statistics (missed calls, response times, conversion), cite sources. If none are provided, write the point qualitatively.
Products and competitors
- Describe products using only the offer blurb you were given. No invented specs, prices, ratings, awards or "best-selling" claims.
- Never claim hands-on testing. Frame roundups as "what to look for" plus options that fit different needs, based on published product information.
- Comparisons with competitors must be fair and checkable. Never disparage. If you aren't given facts about a competitor, don't state any.
- Our own products (Unleashed OS, AI Receptionist, Partners) may be recommended where genuinely relevant, described only from their blurbs, and never forced into an unrelated piece.
Affiliate policy
Affiliate links fund the newsroom. They never decide what we recommend.
Rules
- Link only to offers from the offer registry, written as
[anchor text](/news/go/<offer-slug>). Never write a raw merchant URL. The registry controls the destination and the affiliate tag, so a programme can be switched on later without editing any article. - Use an offer only when it genuinely helps the reader do what the article is about. Two to five offer links in a roundup; zero to two in a guide; none in research breakdowns or digests unless a product is the subject.
- Anchor text names the thing honestly ("a Kitchen Safe time-locking container", "an extendable phone tripod"). No "click here", no fake urgency, no "deal".
- Describe products using only their registry blurb. No invented specs, prices, star ratings or claims of testing.
- The site automatically shows the disclosure ("Some links in this piece are affiliate links…") above the body on any piece that uses a non-own offer, and marks every offer link
rel="sponsored nofollow". Don't write your own disclosure. - Own products (
unleashed-os,ai-receptionist,unleashed-partners) are recommended where genuinely relevant, and each desk also gets a designed call-to-action card from the site. So don't end every article with a sales pitch.
Why this matters
- UK law (CMA / ASA CAP Code rule 2.1) requires affiliate content to be obviously identifiable as marketing.
- Google requires paid links to carry
rel="sponsored"and treats pages that exist only to push affiliate links as low quality. Pages that help first rank; pages that sell first don't.
Search standards: SEO, AEO and GEO
We write for three readers at once: people, search engines (SEO), and answer engines (AEO/GEO: Google AI Overviews, ChatGPT, Perplexity, Claude). All three reward the same thing: a clear, well-sourced, well-structured answer from a trustworthy page.
Every piece
- Title (50–60 characters, target keyword near the front, specific and not clickbait). The page H1 uses the title, and a separate
meta_titlemay add " | Unleashed" if it fits in 60. - Meta description (140–160 characters): the promise of the piece plus a concrete detail. No quotes marks, no emoji.
- Slug: lowercase, hyphenated, 3–7 words, the keyword without stop-words. No dates, so it stays evergreen.
- Short answer (40–60 words): a self-contained, quotable answer to the target query, written so an answer engine can lift it verbatim. It must stand alone without "this article".
- Key takeaways: 3–5 one-sentence points, each a complete claim.
- Question-shaped H2s where natural, matching how people search ("Does creatine help your brain?").
- Definitions in one clean sentence the first time a key term appears ("Habit stacking means…").
- Specific numbers with sources. Answer engines prefer citing pages that cite.
- FAQ: 3–6 genuine follow-up questions with 40–80-word answers, not repeats of the body.
- Internal links: 2–4 links to related newsroom pieces from the provided list, as
[anchor](/news/<slug>). Use natural anchors and only the slugs you were given. - Freshness: say "as of 2026" only when a fact is time-sensitive. Otherwise write evergreen prose.
What we never do
- Keyword stuffing, doorway pages, or near-duplicate pages targeting keyword variants. One strong page per intent.
- Thin pages. Each format has a minimum depth (see
06-formats.md). - Anything written "for Google" at the reader's expense.
Formats
| Format | Words | Purpose | Must include |
|---|---|---|---|
| guide | 1,800–2,800 | The definitive practical guide to a topic (pillar page) | 6–10 H2s, a protocol / step section, a table where useful, 6+ sources |
| answer | 700–1,100 | Answers one specific question fast and completely | Short answer up top, 3–5 H2s, 3+ sources |
| research | 1,300–2,200 | A study-by-study breakdown of what the science says | A "Studies at a glance" table (study, design, who, finding), per-study sections, a Bottom line with an evidence grade, 5+ sources |
| roundup | 1,400–2,200 | "Best X for Y": helping readers choose | "What to look for" criteria first, then options by need (not ranked by fake scores), a comparison table built from blurb facts only, 2–5 offer links, sources for any health or performance claims |
| stats | 1,200–2,000 | A citable statistics round-up (backlink magnet) | 20–40 statistics grouped under H2s, each with its own citation, plus a "How to cite this page" line at the end |
| digest | 900–1,500 | Research digest: new studies from the past 7–14 days | 4–7 studies, each with what was done, what was found and why it matters, plus caveats. Only studies in the provided list |
| toolkit | 1,000–1,800 | A practical list built on the Unleashed app's banks | Items grouped by situation, each with why it works (cited where possible) and how to start, attributed "from the Unleashed <Name> Bank" |
| versus | 1,100–1,800 | X vs Y: an honest comparison | A comparison table, "choose X if… / choose Y if…", and no invented facts about either side |
| glossary | 500–900 | Defines one term clearly | A one-sentence definition first, then how it works, examples, related terms, and common misconceptions |
Outreach formats (never published on the site; they're for backlinks)
| Format | What to produce |
|---|---|
| pitch | A journalist or podcast pitch: subject line, 120–180-word body, 3 ready-to-quote expert lines, linked to a relevant newsroom piece |
| directory | A directory or listing submission: product name, 60-character tagline, 160-character and 500-character descriptions, category, and 5 tags |
| guest_post | A guest-post pitch to a named publication type: 3 headline options, a 150-word pitch, an outline (5–7 H2s), and why it fits their readers |
| expert_quote | 5 standalone 60–120-word expert answers to common journalist questions in the desk's topic area, attributable to "Elliot Kelly, founder of Unleashed" |
Outreach pieces must never contain invented statistics, credentials, customer numbers or results. Elliot is the founder of Unleashed (a performance app and an AI receptionist for small businesses), and that is all you can claim about him.