We published 421 AI-written articles in five months. They earned 10 clicks

I build a marketing automation tool. To prove it worked, I pointed it at four of my own websites and let it run unattended from March 2026: pick keywords, write the article, pass a quality gate, publish to WordPress. No human in the loop. Five months later: 421 published articles. 1,274,609 words. 10 clicks. Here is the data, and the reason — which is not the one I expected. What the engine actually did Articles published 421 Total words 1,274,609 Average length 3,027 words Period 12 March – 19 August 2026 Sites 4 The engine was not broken. It ran four times a day, every day. Articles averaged three thousand words and cleared a quality gate scoring structure, depth and factual grounding. Nothing crashed, nothing looped, nothing published garbage. Search Console over the measured window: Impressions 7,572 Clicks 10 Click-through rate 0.13% Median position 25 One honest caveat before anyone reuses these numbers: that window covers 28 days of Search Console data, not the full five months. My own analytics job had a bug that dropped days, which I only found while writing this. The click figure is what was actually measured, not an extrapolation. The thing I got wrong My first assumption was that the writing was bad. It wasn’t — or at least, that wasn’t what was costing me traffic. I pulled every keyword the sites had a measured ranking for and joined it against real search volume: Keywords with a measured ranking 109 Ranking on page one 40 Page-one keywords with ZERO monthly searches 16 (40%) In the top three 8 Top-three keywords with zero searches 3 Forty percent of my page-one rankings were for search terms that nobody types. Not low volume. Zero. The keyword research step had been inventing phrases that read like keywords — grammatical, plausible, on-topic, the sort of thing a human would nod at in a spreadsheet — and then the engine wrote three thousand words targeting each one and ranked first, because there was no competition, because there was no demand. Across all ranked keywords, 30% had zero volume. Why this is worth writing down Ranking #1 feels like success. Every dashboard I had was green. Impressions were climbing — they roughly doubled over the period. Positions were improving. The quality gate was passing. Every proxy metric said the system worked. The only metric that mattered said otherwise, and it took me three months to look at the join between “what we rank for” and “what people search for”, because no tool puts those two columns next to each other by default. The failure mode generalises beyond my setup: an LLM asked to produce keywords will produce plausible strings, and plausibility is not demand. If your pipeline does not validate every keyword against a real volume source before committing writing effort to it, you will get this outcome. The content can be excellent and it will not matter. The second problem, which is not solvable by writing Median position 25 is page three. Even for the keywords with genuine demand, the pages sit too deep to earn clicks. Every one of the four sites has zero referring domains. Zero. An automated content engine can produce unlimited pages, and unlimited pages with no authority plateau at roughly position 20–25 regardless of how good they are. Volume of content is not a substitute for anyone linking to you — which is, with some irony, why this article exists. What I changed A hard demand gate before writing. Every keyword is checked against real volume, and the floor adapts to the pool: sites with plenty of high-volume options stop writing filler automatically, while thin sites keep publishing rather than starving. Bottom-of-funnel targeting. Instead of informational topics, the engine now derives comparison and alternative queries, where buying intent lives and where vendors cannot rank for their own competitors’ terms. Harvesting from Search Console. The most reliable keyword source turned out to be queries the sites already got impressions for but had never targeted — proven demand and proven relevance, free. Impressions have since grown about 2.4× on the same sites. Clicks have not moved yet, and I would rather say so than publish a case study with a triumphant ending five weeks in. The dataset Every number here came from Search Console and DataForSEO. If you want the raw data — keywords, volumes, positions, impressions, publish dates across all 421 posts — email me and I will send the spreadsheet. No form, no signup. It is more useful to other people than it is sitting in my database. Chetan Sroay, Techno Believe Solutions Ltd See how Marketingsohigh can help Put these ideas to work with Marketingsohigh. Learn more Ready to get started? Marketing So High writes, optimizes, and publishes across 39 platforms. Your growth compounds while you build. Start Free