What I Learned Publishing 167 Blog Posts

82 / 100

SEO Score

Notes from an unexpected experiment in content at scale

I didn’t set out to publish 167 blog posts. I set out to answer a question: what happens when you stop trusting a single AI and start asking five of them the same thing?

The posts were a side effect.

I’d built a small tool — a consensus engine — that queries five large language models simultaneously with the same question. OpenAI, Claude, Gemini, Mistral, Cohere. Each returns its own answer. The engine scores them, identifies where they agree and where they diverge, selects the strongest response, and synthesises the results into a single structured output.

The output needed to go somewhere. WordPress was already running. One thing led to another.

The accidental pipeline

I’m a retired industrial engineer. I spent forty years in manufacturing and metallurgy before I wrote my first line of Python. When I looked at the content creation problem, I didn’t see a writing challenge. I saw a production process with identifiable stations, quality gates, and measurable outputs.

The pipeline emerged over several months of iteration. A one-line query goes in. Five AI models process it simultaneously. A scoring system evaluates each response for completeness, relevance, and whether the model actually answered the question or politely declined. The best response is selected. The full set of results — agreements, disagreements, consensus percentage — gets formatted into structured HTML with schema markup, SEO metadata, and internal links. The finished post is packaged as a WordPress-importable XML file.

I batch the queries using calendar files. Each file contains a list of topics — sometimes ten, sometimes forty. The system processes them sequentially, generates the posts, and delivers them ready for WordPress import. No copy-pasting. No manual formatting. The XML file goes in, the published posts come out.

The 167 posts took six days of elapsed time. Not six days of writing. Six days of the pipeline running in batches while I did other things. I submitted a calendar file before bed on a Tuesday. By Wednesday morning, forty-two posts were sitting in WordPress, formatted, tagged, and scored.

Rank Math scored most of them at 81 or above, which I hadn’t expected from automated content. The SEO parameters — focus keywords, meta descriptions, heading structure, internal links — are set programmatically during the formatting stage, not added afterwards by hand. The pipeline treats SEO metadata the way a manufacturing line treats quality specifications: they’re built into the process, not inspected into the product after the fact.

The Re-Anchor Manager

Industrial Agentic Engineering from an Actual Industrial Engineer 13 chapters. 146 pages. 59 real sessions of proof. The first methodology for maintaining AI session continuity

Buy the Full Book — $19.95

What the numbers actually look like

I checked Google Search Console six weeks after the first posts went live. Here is what I found.

Google indexed 74 of my 168 pages. That’s 44%. More than half the site was invisible. Of the 94 pages Google chose not to index, 44 were pages it had visited, read, and decided weren’t worth including. Forty-one were excluded because of a technical tag. Seven returned errors. One had a server problem.

Total impressions across six weeks: 964. Total clicks: 4. All four clicks were on the homepage — people searching for the site by name.

The weekly pattern told its own story. In the third week, impressions were 110. Steady. In the fifth week, they spiked to 378 — Google was testing our content in higher positions, showing it to more searchers. By week seven, they’d dropped back to 113. Google had promoted us, measured the response, found zero engagement, and pulled back. That’s standard behaviour for a search engine evaluating a new domain. It gives you a window. If nobody clicks, the window closes.

The top-performing content page was a post about AI applications in veterinary medicine, ranking at position 8.2 on Google. Page one. Zero clicks. The search volume for “most promising AI applications in veterinary medicine 2025-2030” is apparently very small. But we were there, on page one, for a query nobody else had covered with structured, consensus-validated data.

The programming languages post ranked at position 7.8. Data analysts at 6.4. Pharmacists at 6.6. All page one. All zero clicks. All ultra-niche queries with minimal search volume.

These aren’t impressive numbers. They’re honest ones. A six-week-old domain with automated content, ranking on page one of Google for queries that almost nobody searches for — yet.

The 73 broken doors

The audit uncovered something I hadn’t thought about. Seventy-three URLs on the site were returning 404 errors. Not because pages had been deleted, but because WordPress generates pagination links for tag archives — /tag/optimize/page/5/ — when there aren’t enough posts to fill five pages. Crawlers follow those links, find nothing, and report the error.

Every one of those 404 responses was a wasted visit from Google’s crawler. On a young, low-authority domain, crawl budget matters. Every visit spent on a dead page is a visit not spent discovering and indexing a real post.

I fixed 54 of the 73 broken URLs with a single line. One regex redirect rule in Rank Math:

`/tag/(.+)/page/[0-9]+/?` → `/tag/$1/`

That one rule catches every tag pagination overflow and redirects it to the parent tag page. Four months of accumulating crawl waste, resolved in under a minute. The remaining 19 URLs needed individual redirects — old slugs pointing to renamed posts, trailing slash mismatches, a few genuine orphans.

Twenty minutes of work. Seventy-three broken doors closed.

Stop asking one AI. Ask five

Five AI models. One consensus answer. No hallucinations. Try free — validated results straight to your inbox in seconds.

Try Seekrates Free

What surprised me

Three things I did not anticipate.

First, the AI models disagree more than you would expect. On technical questions — “what programming languages will dominate in 2030?” — consensus runs high, often 85% to 98%. The models converge on similar lists, similar reasoning, similar timelines. On philosophical or speculative questions — “will AI agents have legal personhood by 2030?” — it drops to 60% or 70%. Four models might say no, while one offers a carefully hedged “it depends on jurisdiction.” The disagreement is itself useful data. It marks the boundary between what AI models collectively know and where they’re collectively guessing. When I started plotting consensus percentages across topics, a map emerged — high-confidence zones and low-confidence zones, with a surprisingly sharp boundary between them.

Second, Google indexed the automated posts faster and more consistently than the articles I wrote by hand. The structured posts include FAQ schema, article schema, keyword-optimised headings, and consistent HTML structure. The hand-written articles had better prose but worse machine-readability. Google’s crawler responded to structure, not style. That was a humbling observation for someone who spent time crafting sentences.

Third — and this one took months to discover — one of my five AI models had been misconfigured since launch. I had placed its system instructions in the wrong field of the API call. The model received my safety guidelines as if they were the user’s question, and the actual question as an afterthought. It scored zero on every query. For months, I assumed it was simply the most cautious model — refusing to speculate, declining to answer. It wasn’t cautious. It was confused.

Once I corrected the configuration — moving one parameter from one field to another — the model went from scoring zero to regularly producing the strongest response of the five. The “cautious” model was actually the most capable one in the group. It had been hobbled by a single misplaced instruction.

Nobody noticed. Not me, not the pipeline, not the readers. The consensus engine ran on four functioning models instead of five for months, and the output looked perfectly reasonable. That’s both reassuring and unsettling. The system was resilient enough to absorb the loss, but fragile enough that a silent misconfiguration persisted undetected.

The real question

I am not sure what 167 posts will become. The site is six weeks old. The domain authority is low. The search volume on our best-ranking queries is measured in single digits per month. Most of the content Google hasn’t bothered to index yet.

But the pipeline works. The consensus data is genuine. The structured posts rank on page one for queries nobody else has covered. And the research database is quietly recording every query, every response, every consensus score — building a dataset that will eventually tell us something about how five AI models collectively understand the world.

There is a question I cannot yet answer. The emerging generation of AI search engines — Perplexity, ChatGPT with browsing, Google’s AI Overviews — need authoritative sources to cite. When someone asks Perplexity about the future of veterinary medicine and AI, will it find our structured, consensus-validated post at position 8 on Google and cite it? Or will it prefer the WHO report, the McKinsey analysis, the university paper with thirty years of institutional authority behind it?

I tested this. I searched Perplexity for five topics we’ve published on. Seekrates was cited in zero of them. Every citation went to established institutions. That was not surprising, but it was clarifying. Domain authority is not something you earn in six weeks. It is something you build over six months, or twelve, or longer.

The content is the foundation. The citations will come — if they come — because the content was there first, structured correctly, answering questions nobody else had thought to ask five AI models at once.

What I do know is that asking five AIs the same question produces more interesting answers than asking one AI five questions. The disagreements are where the insight lives. The consensus is where the confidence lives. The gap between the two is where the work is.

The posts were a side effect. The data might be the point.

*Mohan Iyer is a retired industrial engineer based in New Zealand who builds AI-native development tools. His consensus engine, Seekrates AI, queries five leading language models simultaneously and publishes the results at seekrates-ai.comseekrates-ai.com.*