← Back to Case Studies

Optix Own product
Case Study · Research · Own product

Optix: verified AI content, from a thesis to my own site

AI writes fast and cites badly. For my Master’s thesis in Data Science I designed a method where every source is checked before the model writes a word, evaluated it, and built it into Optix. Then I ran it on this site.

Project
Optix · an AI article writer with verified citations
Role
Research, method design, evaluation, and the full build
Origin
Master’s thesis in Data Science on verified AI content
Proof
Run on nadiamohamed.me, measured in Google Search Console

The problem

Language models invent sources. Walters and Wilder tested the citations ChatGPT produces and found that 18% of the citations from GPT-4 were fabricated, and that 24% of the real ones contained substantive errors such as the wrong journal, year or authors (Scientific Reports, 2023).

A fake citation is worse than a weak sentence, because it looks right. It has the formatting, a plausible author and a plausible date, and nobody catches it by reading. For a company publishing under its own name, one of those is a credibility problem. For AI search, where answers quote sources, it is a reason not to be cited at all.

The thesis

The usual fix is to check citations after the article is written. My thesis started from a different question: what if the model never gets the chance to invent one?

The principle

Verify first, write second. The model only ever receives sources that already checked out.

A fabricated citation is not caught after the fact. It is never created, because the writing step has nothing unverified to work from.

The method

The pipeline is built so that verification is structural, not a prompt asking the model to be careful.

  • Research the live results

    For a keyword, the pipeline reads the pages that rank today and extracts the statistics they actually publish.

  • Check every claim at its source

    Each statistic is matched against the live source page, verbatim, and traced to the original rather than a blog quoting a blog. If it can’t be confirmed, it is dropped.

  • Write only from the verified set

    The brief and the writing step receive the confirmed claims and nothing else, with each citation already formatted as a link.

  • Gate the result

    A quality gate scores every article on citations, SEO and readiness for AI answers. Nothing ships below the bar.

  • Keep a claims ledger

    Every article carries a record of each statistic with its source URL, the quote that backs it, the HTTP status and a timestamp, so it can be checked again months later.

The evaluation

The method was evaluated empirically as part of the thesis, by checking the citations in the articles it produced against their sources.

100%citation integrity in evaluation: every source still live checked out
18%of GPT-4 citations fabricated without a check, the baseline this removes

I wrote up how the pipeline works for the Athens SEO Knowledge Hub: an AI content pipeline with verified citations.

The people using it see the same thing from the other side. Ricardo Batista, CEO of cloro, put it this way:

“The most valuable part was the source verification harness. It forces you to get credible sources, or it doesn’t let you proceed at all.”

Ricardo Batista, CEO, cloro

The results on this site

I started publishing Optix articles on nadiamohamed.me on 7 July 2026, into SEO and GEO topics where the competition writes about search for a living. No ads were involved. This is the shape in Google Search Console.

Weekly impressions in Google Search Console2026
7 Jul · Optix starts
AprJulSep

Weekly impressions, Google Search Console, 6 April to 13 September 2026. About 840 a week before 7 July, about 8,800 a week after.

Before July the line was flat. Once the articles started going out, impressions rose sharply within a few weeks and have stayed well above where they were: about 840 a week in the 13 weeks before 7 July, about 8,800 a week in the nine weeks to mid-September. Clicks followed the same direction, from 26 in the 13 weeks before to 118 in the nine weeks after. Average position got worse along the way, which is what happens when a site starts appearing for many more, harder queries, and it is the reason I keep publishing.

What this case shows

AI content does not have to mean trusting a model with your name. When verification is part of the architecture rather than a final read through, the speed stays and the fabricated sources go. The method came out of research, was measured, and is now the product other teams use.

Want articles you can put your name on?

Optix is available by request. See the plans, what every article includes, and how access works.