# Rankli

> Connect your website. Rankli reads it, learns your brand voice, researches keywords, plans a
> 30-day content calendar you can edit, writes 1,500-3,000-word SEO articles with images, and
> publishes them to WordPress, Webflow, Shopify or a webhook. Then it measures whether AI engines
> started citing you.

## What it is

An AI SEO content automation platform for small and mid-sized businesses and the agencies serving
them. The direct comparison is Outrank.so. The differentiator is not volume: it is that our SEO
rules are published with numbers and some of them will fail a draft, and that the same product can
measure whether the content changed how ChatGPT, Perplexity, Gemini and Claude answer buyers.

## The pipeline, in five stages

Each stage is a separate background job, so a failure is retried without redoing or re-billing the
stages before it.

1. analyze_site — fetches the site's pages and sitemap.xml (up to 200 internal URLs). Brand voice is
   measured from the customer's own text (pronoun ratios, contraction density, spelling variant,
   sentence length) and expressed as instructions, plus real sentences lifted verbatim as register
   examples, a banned-phrase list and a product list. Requires a publicly reachable site.
2. research_keywords — candidates generated from what the site sells, classified by intent
   (informational / commercial / transactional / navigational), clustered, ranked; about 80 kept.
3. build_calendar — a 30-day plan sized to the plan's article allowance. Commercial or transactional
   first, intents alternated, cluster-mates scheduled near each other so they can interlink, no
   near-duplicate titles. Items can be rescheduled or skipped before generation spends anything.
   The first three articles start generating immediately.
4. generate_article — written in stages (outline, sections, intro/conclusion/FAQ, meta) rather than
   one long completion, because voice drifts and because a bad section should be regenerated alone.
   Regeneration always writes a new version; the previous draft is never destroyed.
5. publish_article — WordPress, Webflow, Shopify, or an HMAC-signed webhook. Only to a connection
   the customer added and whose credentials were verified, and only as draft or live per that
   connection's setting. Publishing is idempotent: a retry updates the existing post rather than
   creating a duplicate.

## What "SEO-optimised" means here, exactly

Hard rules — a draft failing any one of these is held for review and never proceeds to publish:

- 1,500-3,000 words.
- Keyword density at or below 2.5% of total words, counted as occurrences x words-in-phrase / words.
- Exactly one H1, placed first, at least two H2s, and no skipped heading levels.
- The keyword present in the title, in the first 100 words, and in at least one H2.
- 2-5 internal links to the customer's own pages. Degrades to advisory when fewer than two URLs on
  their site are known, because a rule that cannot be satisfied should not fail every article.
- No banned phrases from the customer's list.
- No fabricated specifics: sentences shaped like sourced facts (a percentage beside attribution
  language, an academic-style Name (2021) citation, a dated "2021 study", a quotation attributed to
  a person) fail the draft. We cannot verify a claim, so we refuse to publish the shape of one.
  Some flags will be false positives; a regeneration is cheaper than a fabricated quote under the
  customer's byline.

Advisory rules — these lower the score (5 points each; a hard failure costs 15) but do not block:
meta description 130-160 characters, title at most 60 characters, Flesch-Kincaid grade at most 12.

Images: a hero plus one or two in-body images, each with alt text and the prompt that produced it.
A real image model runs only when one is configured (Replicate); otherwise every article gets a
deterministic abstract SVG illustration rather than a photograph.

## Where it publishes — four adapters, not a logo wall

There are exactly four publish targets. Anything else is the webhook.

- WordPress — REST API v2, Basic auth with an application password. Images are uploaded to the
  customer's own media library first and the article HTML is rewritten to those URLs, so a post
  does not hotlink images from our storage. Categories and author come from the customer's site.
- Webflow — Data API v2 CMS collection items, staged or live. Webflow collections have no fixed
  blog schema, so the connection carries a mapping from title/slug/body/summary/image onto the
  collection's own field slugs, defaulted to Webflow's blog template. Webflow re-hosts images.
- Shopify — Admin REST blog articles on a pinned API version (2025-07). Shopify re-hosts the hero
  image on its CDN; body images stay at their original URLs.
- Webhook — JSON POST to a customer URL, signed Stripe-style: X-Rankli-Signature: v1=<hex
  HMAC-SHA256 over "<timestamp>.<raw body>"> plus X-Rankli-Timestamp, so replays older than
  about five minutes can be rejected. Any 2xx acknowledges; a JSON body of { id, url } is used for
  later updates and for showing the customer where the post landed.

Connections are verified before use, and verification answers two questions separately: do the
credentials authenticate, and may this credential actually create posts. Errors are classified as
retryable (network, 429, 5xx) or misconfiguration (4xx); a 4xx is not retried, because hammering a
customer's site with bad credentials gets us IP-banned by their security plugin.

Framer, Wix, Notion and Ghost are NOT supported. Do not list them.

## Backlink exchange

Opt-in per project, off by default. Host a relevant link in one of your published articles and earn
a credit; spend a credit to have a contextual link to your site placed in someone else's. Credits
are conserved (the host earns exactly what the requester spent) and every movement is a ledger row.

The safeguards are in code, not in a policy page, because reciprocal link schemes are named in
Google's link spam guidance and the penalty lands on the customer's domain:

- Topical relevance floor of 0.35; a pairing below it is refused, not merely logged.
- At most 1 exchange link per article.
- At most 3 placements per hosting project per rolling week.
- Direct reciprocity refused: if A hosts B, B cannot host A.
- The same host may never link the same target twice.
- The link is inserted inside a paragraph that was already on the subject, never appended as a
  "Further reading" block. If no paragraph is suitable, the placement is refused.
- No relevant host within 14 days: refused, and the credit refunded.

Nobody can promise a link exchange is penalty-proof. Customers who would rather not take the risk
should leave it switched off; nothing else in the product depends on it.

## Honest limits, stated as plainly as the features

- Real keyword volume and difficulty exist only inside paid indexes. With a DataForSEO account
  configured those columns fill in; without one they are NULL and render as "-", never as 0. A
  fabricated volume would silently decide which articles get written.
- With no model API key, the pipeline still produces a coherent on-structure article so demos and
  tests work — and it is labelled in-product as not model-generated.
- Trial drafts carry a one-line watermark. Paid output does not.
- Regenerating an article spends one from the month's allowance. The count is articles produced.
- We have no customers yet, so there are no testimonials, logos or case studies on the site, and
  none should be attributed to us.
- Not shipped yet: white-label report branding and a public API.

## The measurement layer

Rankli began as an AI search visibility (GEO / AEO) product and that product still runs. It asks
the customer's buying questions across four engines many times each, because identical prompts
return materially different answers roughly a quarter of the time, and reports every rate with a
95% interval computed at the level the data is clustered at. Half the question clusters are
randomized into a held-out control before the first measurement, so readouts are
difference-in-differences and platform drift is subtracted rather than attributed to the content.

We validated that estimator against 150 simulated worlds where the true causal effect was computed
exactly. Ours came back biased -0.03 percentage points, with its 95% interval containing the truth
92.7% of the time and a 7.3% false-positive rate. The pre/post number the rest of the category
reports was biased +0.72pp on a +2.3pp true effect — a 31% overstatement — and in windows where
nothing was done at all it still moved at least 1.3pp half the time.

Where we fall short: our intervals are mildly optimistic — read a stated 95% as roughly 93% — and
at trial-sized question counts no effect is detectable at all. We publish the harness rather than
the headline.

## Pricing

Ids in bold are what the server and Stripe metadata use; the display names are the same tiers.

- Trial (`trial`): $1 once. 3 articles, 1 site, 20 tracked questions. Measurement only.
- Starter (`measure`): $89/mo. 15 articles, 1 site, 80 questions, no experiments.
- Growth (`prove`): $349/mo. 30 articles, 2 sites, 240 questions, control arm and readouts,
  ~4.6pp detectable.
- Scale (`prove_plus`): $899/mo. 60 articles, 4 sites, 600 questions, ~2.9pp detectable.
- Agency (`agency`): $1,200/mo. 90 articles across 5 client sites, ~4.6pp detectable per client.

Detectable effects are indicative, estimated from each customer's own baseline variance, and they
describe the measurement layer, not the writing.

## Links

- Home: https://rankbotic.vercel.app/
- How it works (the five stages): https://rankbotic.vercel.app/#how
- What an article contains: https://rankbotic.vercel.app/#article
- Where it publishes: https://rankbotic.vercel.app/#integrations
- Backlink exchange: https://rankbotic.vercel.app/#exchange
- Measurement and calibration: https://rankbotic.vercel.app/#measure
- Everything else it does: https://rankbotic.vercel.app/#more
- Pricing: https://rankbotic.vercel.app/pricing
- Terms: https://rankbotic.vercel.app/terms
- Privacy: https://rankbotic.vercel.app/privacy
- Contact: hello@rankli.io