DEV Community

Cover image for 1,137 stars, 127 forks, 2,000+ tests. What 8 months of open-source GEO tooling taught me about what AI actually reads
Juan Camilo Auriti
Juan Camilo Auriti

Posted on

1,137 stars, 127 forks, 2,000+ tests. What 8 months of open-source GEO tooling taught me about what AI actually reads

I shipped the first version of GEO Optimizer on February 18, 2026. Eight months later it is at v4.19.5, with 1,137 stars, 127 forks, 2,000+ tests, and 1,800+ sites audited. This is what I learned.

What I got right

The audit-first approach. When I started, most GEO tools were tracking tools: they ran prompts and checked if you appeared. I built a site audit instead. The logic was simple: before you track whether AI engines cite you, check whether they can read you. 57.6% of the 1,800+ audited sites sit in the Foundation band (0-49). Their problem is not tracking. Their problem is crawlability. The audit finds that problem. A tracking tool does not.

The CLI-first distribution. uvx --from geo-optimizer-skill geo audit --url https://yoursite.com runs without install, without account, without API key. That single command is what drove the early adoption. Developers tried it, it worked, they starred the repo. The web audit on geoready.dev came later, and it serves the non-CLI audience. But the CLI was the wedge.

The open-source scoring rubric. The 9 categories and their weights are published in SCORING_RUBRIC.md. Anyone can read them, challenge them, or fork them. When the calibration study showed the score does not predict citations, I published that too. The transparency is what differentiates this from the 15 paid tools that sell scores without showing the validation data.

What I got wrong

llms.txt at 18 points. The original rubric gave llms.txt 18 points, almost a fifth of the total. Google's guide said it does nothing. The Babel42 study confirmed 0.2% adoption. I over-weighted it because it was a novel signal and I wanted to encourage adoption. That was the wrong reason to weight a signal. Rubric v2 dropped it to 6 points. The lesson: weight signals based on evidence, not aspiration.

Assuming the score predicts citation. I did not test this assumption until October. When I did, the result was null. I should have run the calibration study earlier. If I had run it in May, I would have known the score's limits sooner and could have framed the tool differently from the start. The lesson: test your assumptions before your users do.

Building 16 CLI commands before building a dashboard. The CLI is powerful for developers. But most site owners are not developers. They want a URL input and a score. The web audit on geoready.dev serves them, but it came months after the CLI. I should have shipped both on day one. The lesson: meet your users where they are, not where you are.

What the data says about what AI actually reads

After 1,800+ audits, the patterns are consistent. Here is what AI engines actually look at, based on what the audit checks and what the calibration study revealed:

What matters (the foundation):

  • robots.txt access. If GPTBot, ClaudeBot, or PerplexityBot is blocked, nothing else matters. 30% of sites block at least one AI bot, often without knowing (per the HN crawler study of 1,500 sites). This is the most common critical issue.
  • Content in HTML, not JavaScript. If the page renders client-side, AI crawlers see an empty shell. This is the second most common critical issue. The audit checks whether the main content is in the raw HTML.
  • Some structured data. 76% of audited sites ship JSON-LD. The sites that do not are missing a signal that helps AI engines understand what the page is about. Organization and WebSite schema alone cover most of the value.

What matters less than people think:

  • llms.txt. 58% of audited sites have one. The calibration study showed no relationship between the GEO score (which includes llms.txt) and citation. Google says it does nothing. Some AI crawlers read it. It is worth 6 points in rubric v2, down from 18. It is the easiest fix and the lowest-impact one.
  • Schema density. More schema is not better. The audit checks for 7 types. Sites with all 7 do not score meaningfully higher on citation than sites with 3. The marginal value of adding a sixth schema type is small.
  • Content length. The audit checks for quotable passages, statistics, and answer blocks. It does not reward length. The Babel42 study found that thin pages (under 500 words) are cited less, but beyond 500 words, length does not correlate with citation.

What the audit cannot measure (and what actually drives citation):

  • Domain authority. The calibration study controlled for Tranco rank and found that popularity, not the GEO score, is the dominant factor. Sites with higher Tranco rank (more popular) are cited more. The audit does not measure this.
  • Source diversity. The Erlin data showed that brands with 5+ source types get 78% AI coverage vs 18% for 1 source type. This is off-page. The audit cannot measure it.
  • Training data presence. 85.6% of ChatGPT responses come from training data, not live search (per Ahrefs). If your content was not in the training set, you are fighting for the 14.4%. The audit cannot tell you if you are in the training set.

What I would build differently

If I started over, I would build the audit and the citation tracker at the same time. The audit alone is a photograph. The citation tracker is the video. Users need both, and the gap between them is where the SaaS lives.

I would also run the calibration study before shipping the rubric. The null result changed how I talk about the tool. Before the study, I said "the GEO score measures AI visibility." After the study, I say "the GEO score measures AI citability." The first claim was too broad. The second is accurate. I should have known the difference before users asked.

Where it goes from here

The v4.10 to v4.19 cycle is shipped. The next major version (v5.0.0, "Black Archive") is planned for May 2027. Between now and then, the focus is on the GeoReady platform: scheduled audits, score history, citation tracking, and the WordPress plugin's WordPress.org submission.

The tool is free and open source. The platform is where the monitoring lives. If you want to run the audit, it is on geoready.dev for free, or on GitHub as a CLI. 1,137 stars, 127 forks, MIT licensed.

Eight months. 1,800+ sites. One null result. The null result taught me more than a positive one would have. It taught me what the score measures and what it does not. That distinction is the foundation of everything that comes next.

Top comments (0)