Assisted, edited, then published

Publish faster without publishing anything you'd be embarrassed by.

There is a version of this service that generates four hundred articles a month, floods your site with confident nonsense, and quietly damages the search visibility you already had. We do not build that one, and we will argue with you if you ask for it.

What we do build is the machinery around publishing: briefs assembled from real data, drafts prepared so a writer starts at sixty per cent rather than zero, metadata and alt text generated at scale, translation into other markets, and the tedious multi-channel distribution that eats an afternoon every week. A person still decides what gets published. That gate is the product.

Hold the gate with your cursor
drafted 0awaiting review 0published 0
DraftsAssisted
PublishingNever automatic
ClaimsSourced or cut
BylineA real person

Position

Six things we will not build

Stated up front, because these are the requests we get most often and it saves everybody a meeting.

None of this is squeamishness about the technology — we use it daily. It is that these particular applications reliably produce a worse outcome than doing nothing, and we would rather lose the work than build something that damages a site we helped make.

01

Bulk article generation

Hundreds of posts targeting keyword variations. Search engines have spent years getting good at recognising this, and the downside is site-wide rather than page-specific.

02

Fabricated authors

Invented names, stock-photo headshots and made-up credentials attached to generated articles. It is deceptive, and it is the first thing that unravels when anyone checks.

03

Competitor rewriting

Scraping the top-ranking pages and paraphrasing them. It produces content with nothing new in it, which is precisely what ranking systems are built to filter out.

04

Publishing with no human gate

Straight from model to live site. Everything else on this page is defensible; removing the person is what turns it into a liability.

05

Invented statistics and studies

Models produce plausible numbers and citations that do not exist. On a professional services site, one fabricated figure is a credibility problem you cannot un-publish.

06

Fake reviews or testimonials

Generated social proof. Illegal in several jurisdictions, against every major platform's terms, and genuinely damaging when discovered.

Division of labour

Who should do what

Twelve tasks in a typical content operation, and where the line sits on each. Filter the columns to see the shape of it.

The pattern is consistent once you look: automation earns its place on the mechanical, high-volume, low-judgement work — the parts nobody enjoys and everybody skips. It has no business anywhere near the decisions that carry your name.

TaskAI draftsEditor decidesNever
AI does the first pass A person owns the decision Not automated at all

Interactive

What the edit pass actually removes

The same opening paragraph for a commercial HVAC company, before and after a person worked on it. Hover a note to see the change highlighted in both versions.

The left column is not a bad model or a lazy prompt — it is what competent generation produces by default. It is fluent, confident, and says almost nothing. Note four is the one that matters: a specific, credible-sounding statistic that has no source and no basis, which is how a model damages a professional reputation without anyone noticing at the time.

Raw generation
After an editor

What changed, and why

The stack

Into the CMS you already publish from

Drafts land as drafts in your existing system, in your existing workflow, for your existing approver. Nobody learns a new tool, and nothing goes live without somebody pressing publish.

Content management

WordPress
Webflow
Contentful
Sanity
Strapi
Ghost

Commerce & other platforms

Shopify
WooCommerce
Squarespace
Drupal
Storyblok
Prismic

Models & language

OpenAI
Anthropic Claude
Google Gemini
Hugging Face
Ollama (self-hosted)
LangChain

Translation & multi-market

DeepL
Crowdin
Weblate
Directus
Markdown
Git versioning

SEO, analytics & delivery

Search Console
Google Analytics
Semrush
Yoast SEO
Cloudflare
Vercel

Distribution — social

LinkedIn
X
Instagram
Facebook
YouTube
TikTok

Applications

Six jobs worth automating

Notice what these have in common. None of them is "write the article". They are the surrounding work that never gets done because it is boring and endless.

Job 1 of 6
01 — Metadata

Titles, descriptions and alt text

The fields everyone skips. Meta descriptions on four hundred pages, alt text on every image, structured data kept current — drafted automatically, reviewed in bulk.

  • Alt text that describes, not keyword-stuffs
  • Descriptions within length limits
  • Flags pages missing them entirely

Highest return for the least risk

02 — Commerce

Product descriptions at scale

Thousands of SKUs where the manufacturer supplied a spec sheet and nothing else. Generated from real attributes, in your voice, then reviewed by category.

  • Built from actual specifications
  • Consistent structure across the range
  • Never invents a feature

Catalogues too large to write by hand

03 — Languages

Translation and localisation

Your site in French for the Quebec market, or any other language, with terminology held consistent and a native speaker reviewing before anything goes live.

  • Glossary keeps product names fixed
  • Adapts units, dates and currency
  • Re-translates only what changed

Genuinely useful in a bilingual market

04 — Repurposing

One piece into six formats

A long article becoming a newsletter section, three social posts, a summary and a set of talking points — drafted from the source rather than rewritten from scratch.

  • Keeps the substance, changes the shape
  • Per-channel length and tone
  • Queued for approval, not auto-posted

Where the afternoon actually goes

05 — Briefs

Research and outlines

Before a writer starts: what already ranks, what your existing pages cover, what questions customers actually ask, and which internal links belong.

  • Assembled from real data, not guesses
  • Flags where you would cannibalise yourself
  • Saves the writer the first two hours

The writer still writes

06 — Maintenance

Keeping old content honest

Finding pages with stale prices, dead links, retired products or last year's dates — the slow rot that quietly undermines a site nobody has audited since launch.

  • Flags outdated claims for review
  • Suggests internal links that now exist
  • Prioritises by traffic, not age

Nobody's job, until it is

Honesty

The risks, plainly

Content automation carries a particular kind of danger: the failures are invisible for months and then arrive all at once.

01

Search visibility is site-wide

Ranking systems assess sites, not only pages. A large volume of thin generated content can drag down pages that were performing perfectly well on their own.

02

Everything sounds the same

Generated copy converges on a house style that belongs to the model, not to you. Across a whole site the effect is a business with no personality.

03

Fabrication is confident

Invented figures, misattributed quotes and citations to papers that do not exist — all delivered in the same assured tone as the accurate parts.

04

Review fatigue is real

Approving forty drafts a week degrades into rubber-stamping within a month. Volume must be set to what a person can genuinely read.

05

Translation fails quietly

Fluent output that is subtly wrong reads fine to anyone who does not speak the language. Native review is not optional, it is the control.

06

Regulated claims

Health, legal, financial and safety statements carry consequences beyond embarrassment. In those sectors the draft is a starting point for a qualified person, full stop.

07

Disclosure expectations shift

What platforms and regulators expect you to declare about AI-assisted content is still moving. Keeping a record of what was assisted is cheap insurance.

08

Your own data leaks

Pasting unreleased material into a consumer chat tool is a common and invisible risk. Business-tier configuration and clear staff rules both matter.

09

Volume is not the goal

Twelve genuinely useful pages will outperform two hundred adequate ones, and cost less to maintain. If a proposal is measured in output count, be suspicious.

Scope

What a content automation build includes

Most of the effort goes into the constraints — the voice guide, the fact rules and the review step — rather than the generation itself.

01

Content audit

What you already have, what performs, what duplicates itself, and what is quietly out of date. Frequently the audit alone changes the plan.

02

Voice definition

A written guide built from your best existing pages — what you sound like, what you never say, and the words you use for your own products.

03

Source grounding

Connecting real material — specifications, policies, past work, live data — so drafts are assembled from your facts rather than the model's recollection.

04

Fact rules

Explicit instruction that statistics, quotes and claims must come from supplied sources or be omitted. Then checked, because instruction alone is not enough.

05

CMS integration

Drafts arriving as drafts in your own system, with the right fields, categories and internal links already populated.

06

The review step

A screen that makes approving or rejecting fast, shows what the draft was built from, and records who signed it off.

07

Distribution

Once approved, the piece reshaped for each channel and queued — still with a person on the send button where it matters.

08

Measurement

Which pieces earned traffic, links and enquiries, so the pipeline gets pointed at what works rather than at what is easy to produce.

09

Guardrail testing

Deliberately trying to make it fabricate, drift off-voice or exceed its remit — before it is anywhere near your live site.

Questions

AI content, answered

Will Google penalise us for using AI?

Google's stated position is that it rewards helpful, original content regardless of how it was produced, and targets content made primarily to game rankings. In practice that means the method matters less than whether the page is genuinely useful.

Where sites get hurt is volume without substance. A hundred near-identical pages built around keyword variations is the pattern that causes trouble — and it is the pattern we decline to build.

Can it write in our voice?

Closely, given enough of your existing material to work from and a written voice guide. Grounding it in your real pages does most of the work.

It will still drift, which is one of several reasons a person reads everything before it publishes. Voice is easier to maintain than to establish.

How much time does this actually save?

On mechanical work — metadata, alt text, product descriptions, first-pass translation — a great deal, because those jobs are currently either not done or done grudgingly.

On genuine writing, less than vendors claim. A draft saves the blank-page problem; a good editor still spends real time on it. The honest framing is faster and more consistent, not free.

Do we have to disclose that AI was involved?

Expectations are still forming and differ by platform and sector. Our recommendation is to keep an internal record of what was AI-assisted regardless, because reconstructing it later is impossible.

Public disclosure is your call, and worth taking advice on in regulated industries.

What about the byline?

It belongs to a real person who reviewed and stands behind the piece. We will not build invented author profiles, and we would push back hard on being asked to.

If nobody is willing to put their name on it, that is useful information about whether it should be published.

Can it publish automatically once approved?

Yes — scheduling and distribution after human approval is entirely reasonable, and it is where a lot of the time saving lives.

What we will not wire up is generation straight through to a live page with nobody in between.

Will this work for our e-commerce catalogue?

Product descriptions from real attribute data are one of the strongest applications, particularly for large catalogues where the alternative is a manufacturer's spec dump or nothing at all.

Review happens by category rather than per product, which keeps it practical at a few thousand SKUs.

What if we want the high-volume approach anyway?

Then we are not the right firm, and we would rather say that now than take the work and hedge.

We would still be glad to help with the parts of your content operation that do not depend on it.