Why I Removed the Scores From My GEO Formatting Test

I originally published this page with scores for four GEO content formats. I removed them because I could not find the unedited model answers behind those numbers, and the examples did not hold the wording constant. A score for “caveat retained” means little if the reader cannot compare the source text with the answer that was scored.

The old page called itself a formatting test. The more accurate description is an editorial correction: what I claimed, why the comparison could not support that claim, and what I would need to save before making it again. Removing the scores made the article less impressive, but more accurate.

What I could not substantiate

The previous version assigned the paragraph a 0/2 for caveat retention and the source block a 2/2. It also described the definition list as the best compact format and the source block as the highest-scoring option. Those were presented as results from a run. Without the exact model responses, prompt context, and scoring notes, I cannot check whether either answer actually lost or preserved the caveat. I no longer treat those numbers or winners as evidence.

There was a second problem that is visible in the old article itself. The four HTML samples were not simply four wrappers around identical text. One version named screenshots and a reproducible test; another added an “Evidence” field with a source note and date. Some caveats said “no ranking, indexing, AI answer, or citation guarantee”; another only said “no visibility or citation guarantee.” If a generated answer handled those versions differently, the wording might explain the difference. The format alone would not.

Two excerpts that exposed the problem

These are constructed examples from the earlier comparison, not results from a saved model run. I have kept just two to show the problem. The paragraph and definition list both concern source blocks, but they do not give a summarizer exactly the same facts.

<p>
  Use a source block when a GEO recommendation depends on platform behavior,
  crawler access, structured data output, screenshots, or a reproducible test.
  This is useful for technical articles and documentation pages, but it does
  not guarantee ranking, indexing, inclusion in an AI answer, or citation.
</p>

<dl class="geo-extraction-facts">
  <dt>Recommendation</dt>
  <dd>Use source blocks for GEO claims that depend on platform behavior,
  crawler access, structured data output, screenshots, or reproducible tests.</dd>
  <dt>Applies to</dt>
  <dd>Technical articles, documentation pages, and implementation notes.</dd>
  <dt>Evidence</dt>
  <dd>Keep the source note, date, limitation, and verification step together.</dd>
  <dt>Limit</dt>
  <dd>A source block improves verifiability; it does not guarantee visibility or citation.</dd>
</dl>

For instance, only the definition list mentions a source note and date. Only the paragraph explicitly names indexing. A summary that includes one of those details and misses the other would not tell me whether a <dl> outperformed a <p>. These snippets can still help an editor think about readability; they cannot establish an answer-extraction winner.

What I would record before scoring again

First I would write one recommendation with the same scope, caveat, evidence, and action in every version. Then I would change only its HTML presentation. I would keep the prompt identical, save each exact input and every unedited output, record the model and date, and explain each judgment next to the sentence that supports it. That would let someone else inspect a claimed omission instead of taking a score on trust.

I have not done that controlled comparison for this page, so I am not replacing the old ranking with a new one. The separate answer extraction review describes how to keep one source section and its generated summary together. If the reader’s question is which layout helps a human use a page, the content structure examples show actual SeoGeo Tech pages and the editorial reasons behind their formats.

That is the distinction I missed in the first draft. I can explain why a table, paragraph, or labeled list serves a reader. I cannot claim one performs better in AI answers without the output that would demonstrate it.