Can AI run SEO autonomously?

Partly. Over 90 days our autonomous pipeline handled keyword research, drafting, technical auditing and refresh prioritisation without a human writer in the drafting step — indexing 127 URLs and producing 8,804 impressions. Publication still requires human review. Autonomy got the site indexed, not ranked: clicks lagged and required a human-guided layer of snippet and authority work. AI can run the production line; it cannot yet replace strategy and link-earning.

What did 90 days of autonomous SEO produce?

On a roughly three-month-old domain, the archived experiment generated 8,804 Google Search Console impressions, 21 clicks (CTR 0.24%) at an average position of 48.5, with 127 URLs submitted and indexed via sitemap and zero indexing errors. Around 45 articles were drafted; public publication now remains behind human QA. The best single page reached 435 impressions at position 12.8.

Does AI-generated content rank?

It gets indexed reliably and can earn impressions quickly, but ranking is harder. Our AI-generated articles were indexed at scale yet mostly sat in positions 40-50 because a young domain lacks backlink authority. AI-generated content ranks when it has genuine commercial relevance, avoids cannibalising itself, and sits on a domain with earned authority — not on volume alone.

Is autonomous SEO right for B2B SaaS?

It is a strong production engine, not a complete strategy. For B2B SaaS it works best as an always-on layer that handles indexing, drafting and refresh signals, freeing humans to focus on commercial-intent targeting, link acquisition and conversion. Used alone it risks chasing vanity impressions; paired with human strategy it compounds.

What is the 'Four Systems' pipeline?

It is iVanHub's autonomous SEO pipeline of five coordinated systems — deterministic keyword opportunities, content writer, local onsite audit, refresh recommender and orchestrator — running on the protected Contabo runtime. Local Ollama is used only where generation materially helps, and every publication remains behind human QA.

Build in public · 90-day experiment

What an archived 90-day SEO experiment actually produced

We pointed an autonomous pipeline — five coordinated systems we call Four Systems — at a brand-new domain and let it run SEO with no human writer in the drafting step for 90 days. It researched keywords, drafted articles, audited pages, and decided what to refresh. The experiment predates the current human-QA publication gate; today generated content remains a draft until reviewed. The archived run used external datasets, but the current iVanHub runtime is self-hosted and makes no paid search, hosted-AI, or hosted-email requests.

This is the honest result, straight from Google Search Console for ivanhub.co.uk over 8 March – 6 June 2026. No rounding up, no cherry-picked wins. The numbers are modest because the domain is three months old — and that is exactly what makes the lessons useful.

The results, with nothing hidden

90 days. Google Search Console, domain property ivanhub.co.uk. Every figure below is the real recorded value.

These figures describe an archived experiment, not a forecast or a promise. The current pipeline keeps generated content behind human QA.

8804

GSC impressions

127

URLs indexed

Articles drafted

48.5

Average position

Total impressions	8,804
Total clicks	21
Average CTR	0.24%
Average position	48.5
URLs submitted & indexed	127 (0 errors)
Best single page	/insights/schema-markup-b2b-saas-structured-data — 435 impressions, position 12.8

How impressions ramped

GSC impressions per day. Effectively nothing through mid-March, first real movement in late April, a peak of 457 on 13 May, then settling into the low hundreds through early June.

Mid-Mar

Late-Apr

457

13 May (peak)

235

Late-May

200

Early-Jun

Representative points from the daily GSC impressions curve, 8 Mar – 6 Jun 2026. Not every day is plotted.

How the five-system pipeline works

Each system does one job and hands off to the next. The only system allowed to make content live is the last one — everything else feeds it.

Keyword research

The archived experiment fanned out from seed topics into candidate queries. The current runtime keeps this step deterministic and reports on-page opportunities without inventing search-volume or CPC estimates.

Content writer

For each cluster the system drafts an outline, expands it into a full article, then appends an FAQ block and key-takeaways. Drafts land in the database with status 'draft' — never auto-live.

Onsite audit

Local Chromium/Lighthouse runs against safe public URLs to catch performance and technical regressions before and after publishing. Results are logged for trend analysis.

Refresh recommender

Combines page age with captured page evidence to queue pages that need rewriting rather than endlessly producing new ones. Hosted search datasets are not required.

Orchestrator

Dequeues the refresh queue and leaves generated drafts behind the human QA gate. Publication is an explicit reviewed action; no outbound digest is promised.

Four honest lessons

These are systemic, not cosmetic. They are the real reason the numbers look the way they do — and the part most worth linking to.

Content velocity created keyword cannibalization

Generating quickly meant the system produced near-duplicate articles on overlapping topics that ended up competing against each other in the SERPs. The fix was structural, not cosmetic: consolidate overlapping pages and 301-redirect to a single canonical article per topic. Lesson — fast generation needs a topic-ownership layer, or your own pages split your authority.

Big impressions are not the same as commercial intent

Auto-generated tool-comparison pages pulled large impression counts but ranked on pages 6-8 and converted almost nothing. High volume, low commercial value. Lesson — optimise for queries with buying intent, not for the raw impression numbers that look impressive on a dashboard.

Generation gets you indexed, not ranked

Impressions ramped fast, but clicks lagged badly. A three-month-old domain simply lacks the backlink authority to convert rankings into clicks. Lesson — content velocity is not authority. Automation can get you indexed at scale; earning the trust that moves you up the page is a separate, slower job.

CTR needs a second layer on top of generation

Generation alone produces functional but weak SERP snippets. Turning impressions into clicks required deliberate work on titles, meta descriptions and FAQ schema. Lesson — the autonomous pipeline is the first stage; click-through optimisation is a deliberate second pass, not an emergent property of writing more.

What's next: authority and links

The 90-day takeaway is simple: generation gets you indexed; it does not get you ranked. The pipeline proved it can index at scale, audit itself and prioritise refreshes without supervision. What it cannot manufacture is authority.

So the next phase moves from volume to trust: consolidating cannibalising pages behind canonicals, sharpening titles, meta and answer-engine-optimised snippets, chasing commercial-intent queries instead of vanity impressions, and — the slow part — earning the backlinks a three-month-old domain has not had time to build. Automation handles production; people still handle authority.

We build and run this for clients. See how the pieces fit in our SEO service and growth systems, read the deeper write-up on autonomous SEO agents, or meet the team behind it as an SEO agency in London.

Book a strategy call Run a free SEO audit

Cite this experiment

Found this useful? You're welcome to reference or link to it. Here is a ready-made citation.

iVanHub (2026). Autonomous SEO: What 90 Days of AI-Run SEO Actually Produced. Retrieved from https://ivanhub.co.uk/autonomous-seo

Canonical URL: https://ivanhub.co.uk/autonomous-seo