Every three months someone asks us to "10x their content with AI". Every three months the same conversation happens. We say no, we explain why, and we offer the thing that actually works. Sometimes they listen. Sometimes they go to someone else who says yes. Those ones come back a year later asking what happened to their rankings.
The short version: Google has already done the sorting. The Helpful Content Updates, starting in 2022 and rolled into the core ranking system in the March 2024 core update, have quietly moved the floor. Sites pumping unedited AI content at volume are down. Sites using AI as a research and drafting assistant, with a human editor in the loop, are up.
What we have seen
Across six twelve-month engagements in FinTech, legal, B2B SaaS, and Australian e-commerce, here is what we have observed.
What moved rankings.
- Topical authority at the cluster level, not the page level. Ten well-linked pages on one tight topic beat fifty scattered pages.
- Original research and proprietary data. A survey, a benchmark, a dataset. This is the strongest signal we can create, and an AI model cannot generate it for you.
- Real author entities. Knowledge panels, verified LinkedIn profiles, published work elsewhere. The byline is a ranking factor now.
- Schema.org markup that describes what is actually on the page. Not every page needs Article schema, but everything that can have it, should.
- Fast sites. Green Core Web Vitals. No hero image at 2.4 MB. This is still table stakes, and the majority of sites we audit fail it.
What wasted time.
- Programmatic SEO. Ten thousand auto-generated city pages are the 2020 playbook. Google deindexes them now.
- AI "pillar" pages with no new information. Models synthesise what is already out there. If you publish what is already out there, there is no reason to rank you over the existing page.
- Mass internal linking with anchor text stuffing. The link graph is a signal, but not a blunt one.
- Chasing top-of-funnel vanity keywords. "What is AI" gets you traffic, but not traffic that converts. We target mid-funnel intent instead.
The workflow that works
This is the one we run on every engagement.
-
Topic cluster map. Human strategy, AI-assisted research. We pick one pillar topic and fifteen to twenty supporting pieces. Clusters are tight. Each cluster has a clear audience and a clear conversion path.
-
Editorial brief. H2 outline, target queries, entity coverage, required sources, internal link plan. One page, approved by the client before any draft is written. This is where most agencies cheat. We do not.
-
AI first draft. Claude for structure and voice. GPT for alternate takes. We use both because they fail in different ways, and the variation gives us better drafts to edit from. Not a paragraph of AI copy ships without a human touching it.
-
Human edit pass. A real editor. Voice, accuracy, specificity, Australian English, removes AI tells ("dive into", "seamless", "robust", em dashes). Adds original observations, a specific example, a direct quote. Adds the author's opinion where the brief calls for one.
-
Schema markup. Article, Person, Organization where relevant. Fills in the knowledge graph signals that Google needs to understand who wrote what.
-
Internal linking within cluster. Every piece links to every other piece in its cluster. Anchor text varies. The link graph reinforces the topical authority signal.
-
Measure, iterate in 90 days, prune the dead. Pages that are not pulling their weight after a full quarter are merged, rewritten, or removed. Thin pages actively hurt the whole site, since site-wide quality is now baked into core ranking. We are not precious about culling.
The numbers
Across the six engagements: non-brand organic up 180 per cent to 420 per cent year on year. Revenue attributed to organic up two to five times. None of this is magic. It is the workflow, run consistently, for twelve months.
No client ever tried the mass-AI approach and beat this. Two tried and we had to rescue them twelve months in. Both recoveries took nine months.
The shortcut that works
If you want to rank faster than twelve months, the shortcut is proprietary research. Run a survey of a thousand people in your space. Publish the results with a methodology page, a downloadable dataset, and real quotes. Every outlet that writes about your topic will cite you, because they have nothing else to cite.
We have done this for two clients this year. Both ranked for the core query within ninety days. Both still rank. The research took about six weeks to produce and a tenth of the cost of the content programme they had before.
AI helped analyse the survey data. AI did not pretend to be the research.
Where AI genuinely helped
For the record, because we are not anti-AI, here is where AI earned its keep in those twelve-month engagements:
- Keyword expansion from a seed list
- Entity suggestions for a brief
- Alternate H2 proposals
- First-pass drafts against a tight brief
- Schema markup generation
- Meta title and description variations for A/B testing
- Editorial QA: "does this draft cover every point in the brief?"
AI is a research assistant and a drafting assistant. It is not a publisher.
The bar went up
Google's position is that content should be helpful. Site-wide quality (the signal the standalone Helpful Content classifier used to track, before it was folded into core ranking in March 2024) evaluates the site as a whole, not page by page. That means one mediocre AI-spun page can drag down pages that would otherwise rank.
Which means the old content marketing maths, where you publish a lot and accept that most of it will not perform, no longer works. The underperforming pages hurt the performers.
Treat that as a gift. Publish less. Publish better. Measure ruthlessly. Edit the pages that are already ranking up, not out. The sites that do this in 2026 will have quieter content calendars and better numbers than the sites that do not.