Skip to content
Agile Snowglobe

Grounded AI blog posts, on a schedule

One grounded post per site, every morning.

Agile Snowglobe reads your website, learns what you sell and how you sound, then writes, checks, illustrates and publishes one post per scheduled day. You approve. It writes.

No account needed for the sample. The free plan has no card.

About 9 minutes
from a pasted URL to a finished sample post, measured
Two cents
of AI per post, metered on every model call
Up to 30 pages
of your site read before the first word is written
Two models
one writes, a second judges quality, depth and style

How it works

The same six steps, every morning.

Nothing is written from a blank prompt. Every post starts from what your site already says and ends with a second model checking the first.

Watch the six steps run on your own site before you sign up.

  1. 01

    Reads your website

    Crawls your pages, products and articles first, and indexes what they actually say. Pages sent by the embed script count only after the crawler verifies them.

  2. 02

    Learns your voice

    Builds a knowledge pack: brand facts, offers, audience and tone. Brand facts in a post can only come from here.

  3. 03

    Chooses a topic

    Picks the next search your customers make from the keyword map, and checks it does not compete with a page you already rank for.

  4. 04

    Writes grounded

    Drafts the post from your pages plus verified external facts, each statistic cited as a link. No invented numbers.

  5. 05

    Checks facts and depth

    A second model reviews every claim, scores quality, depth and style, and deepens thin sections before you see it.

  6. 06

    Links, images, publishes

    Adds links to your money pages, a banner and an infographic, meta and schema, then publishes to WordPress or your API at the scheduled time.

Inside one post

What happens in those nine minutes.

Ten stages, each with a rule it cannot break. This is the same pipeline for the free sample, the daily post and a refresh.

  1. 01

    Reads your website

    The crawl comes before any writing: up to 30 pages for a sample, the whole site once you connect, re-crawled weekly.

    • Products, services and articles are fetched first; navigation, footers and boilerplate are stripped.
    • Every page is passed through a sanitiser that removes instruction-like text and marks the rest as untrusted data before a model sees it.
    • Text is split into passages and embedded on our own hardware (bge-small), so retrieval costs nothing per post.
    • Pages sent by your embed script are used only after the crawler has verified them.
    crawlersanitiserbge-small embeddings
  2. 02

    Learns your voice

    A knowledge pack is assembled for each post from the passages that matter for its topic.

    • Hybrid retrieval: keyword search (BM25) and vector similarity, fused by reciprocal rank, so exact product names and paraphrases both surface.
    • The pack carries brand facts, offers, audience and tone; a site profile is written once per site and refreshed on re-index.
    • Edits you make to posts become voice signals, learned weekly into a voice profile the writer reads.
    • The rule: a brand fact that is not in the pack cannot appear in the post.
    hybrid retrievalsite-profilevoice-learn
  3. 03

    Chooses a topic

    The keyword map decides what to write next, and checks it will not compete with what you already rank for.

    • Seeds come from your offer pages, your own seed list and competitor topics; they are expanded into candidate searches.
    • Volume and difficulty are added from a keyword data provider when one is configured; intent is classified; candidates are clustered into topics with a primary term, secondaries and a priority score.
    • A keyword ownership registry compares the term with pages you own: write new, refresh the existing post, or link to your page instead.
    seed-keywordsintent-classifycannibalisation check
  4. 04

    Grounds it in the search results

    What ranks today defines what the post must cover and where its facts come from.

    • The top-ranking pages for the term are fetched (a search data provider, with a scraper fallback); their headings become the coverage checklist.
    • A fact harvest pulls statistics from those pages together with their source URLs.
    • Facts enter the prompt as untrusted data with the citation attached, so the writer can quote them and the judge can verify them.
    SERP groundingfact-extract
  5. 05

    Writes the brief

    A plan before prose: title, outline, FAQ questions, a differentiating angle and the money page to link.

    • The brief is built from the knowledge pack and the coverage checklist, not from the model's memory.
    • Pillar posts get a longer brief that maps the cluster of child topics they will link to.
    topic-briefpillar-brief
  6. 06

    Writes the draft

    The writing chain drafts the post inside the brief, in your house style and output language.

    • Brand facts only from the pack; external statistics only from the harvested facts, at least three of them cited as links.
    • Today's date is in every prompt, so a post never calls last year "this year".
    • The writing and judging chains are different model families, so the judge never grades its own prose.
    article-writewrite chain
  7. 07

    Checks depth, facts and style

    A second pass measures the draft and fixes it before you see it.

    • A depth scorer measures section length, answer-first openings, specifics and subtopics missed against the checklist; below 7, thin sections are rewritten.
    • A judge model scores quality and lists concrete issues; a revise pass fixes them and is not allowed to drop a citation.
    • A style lint and a humanising pass strip the usual AI tells; an originality check compares the draft with your recent posts.
    section-deepenquality-judgequality-revisehumanize
  8. 08

    Adds the links

    Internal links are placed by meaning, not by keyword match.

    • Each section is embedded and matched against your indexed pages: money pages first (up to two), the pillar, then articles (up to three).
    • A hub page is skipped when one of its children is already linked; anchors are chosen by a skill that never uses the post's own keyword.
    • Every post opens with a short overview and closes with key takeaways, so assistants and search engines can quote it correctly.
    linkeranchor-pickkey takeaways
  9. 09

    Packages it for search

    Meta, structure, schema and images, ready to ship.

    • Meta title and description trimmed at word boundaries, a clean slug, canonical handled; table of contents from the headings.
    • The FAQ is rendered as a table and emitted as FAQPage JSON-LD beside the Article schema; OG tags carry the banner.
    • An image plan produces a 1680×882 banner and an infographic in your colours with your logo, saved as JPEG and WebP with alt text.
    meta-packageimage-planschema
  10. 10

    Publishes, then keeps learning

    You approve, it ships, and every post teaches the next one.

    • Approve-first by default, with signed review links that work from email for 72 hours and an optional auto-approve window; full auto unlocks after 30 days at an average quality of 8 or better.
    • Publishes to WordPress through the REST API, or your CMS pulls it with a per-site key or receives a signed webhook; search engines get an IndexNow ping.
    • The keyword is registered as owned, the analyst turns the report into suggestions, Search Console feeds performance, striking distance and decay into the refresh queue.
    • Every model call is metered against the plan budget with a hard stop, so cost can never run away.
    lifecyclewebhooksanalystbudget guard

Why it does not read like AI

Grounded, cited, reviewed.

Three rules the pipeline cannot break, so the post says what your site says and nothing your site does not.

1

Brand facts only from your pages

The writer sees a knowledge pack built from your own site. If a fact about you is not in it, the post cannot say it.

2

Statistics only from verified sources

External numbers are harvested from the pages that already rank, checked, and cited as links. Unsupported claims are cut, not softened.

3

An editor model, then you

Every draft is judged for quality, depth and style by a second model. New sites run approve-first until the 30-day average earns full auto.

On every post

Everything a post needs to ship.

A banner and an infographic

Generated for the post in your colours, 1680×882, served as WebP and JPEG, no watermark, alt text written.

Key takeaways for AI answers

An answer box at the top of every post so assistants and search engines can quote you correctly.

Schema, table of contents, FAQ

Article and FAQPage JSON-LD, a linked table of contents, and an FAQ written as questions people actually ask.

Links to your money pages

Contextual internal links to the products, services and guides you already have, never to a hub that adds nothing.

Meta that fits

Title and description trimmed at word boundaries, a clean slug, and canonical handled for you.

Publish where you already are

A WordPress and WooCommerce plugin pulls approved posts with featured image, SEO meta, schema and product cards for the products each post links to. Any other CMS pulls them with a per-site API key or receives a signed webhook with HTML, Markdown, meta, schema and image URLs.

Approve first, then trust

Posts wait for you until the 30-day quality average clears the bar. Then full auto, if you want it.

Keyword map and calendar

See the next two weeks, drag a slot to another day, ban a topic, add a seed. Search Console shows what moved.

Free sample

See it on your own site first.

Paste your address. In about five to ten minutes you get a real post written from your pages, with its scores, the links it added and the sources it cited. No account, no card.

  1. Reading your website
  2. Learning your voice
  3. Choosing a topic
  4. Writing
  5. Checking facts and depth
  6. Adding links and images

Pricing

Pay for posts, not for seats.

Every plan includes the full pipeline. The AI cost of each post is metered and included.

Free

Free
  • 1 site
  • Up to 1 post a week per site
  • 1 site
  • 1 AI-written post per week
  • Sitemap-driven content gap analysis
Start free

Pro

$99 / month
  • 3 sites
  • Unlimited posts
  • 3 sites
  • Unlimited posts (priority generation queue)
  • Everything in Starter
Choose Pro

Agency

$299 / month
  • 10 sites
  • Unlimited posts
  • 10 sites
  • Unlimited posts
  • Everything in Pro
Choose Agency

Compare every plan

Questions

Before you paste a URL.

Will it publish without my approval?

Not at first. New sites run in approve-first mode: every post waits for you, and you can set an auto-approve window. Full auto unlocks once the 30-day average quality clears the bar, and you can keep approve-first forever.

Where do the facts come from?

Two places only. Facts about you come from a knowledge pack built from your own pages. External statistics come from sources that already rank for the topic, are checked, and are cited as links in the post.

Does it work with WordPress?

Yes. Connect with an application password and approved posts publish through the REST API. Anything else pulls posts with a per-site API key or receives signed webhooks.

What does the AI cost me?

Every model call is metered. A post has cost about two cents in AI so far, and that is included in every plan. You can see the spend per site on your dashboard.

Can I edit a post?

Yes. Edit the title, Markdown body, meta title and description in the review screen; saving re-renders the HTML, table of contents, FAQ schema and meta. The pipeline learns from your edits.

Start with a sample, not a signup.

One real post for your site, written from your pages, in minutes.