Categories
AI SEO

A semantic retrieval and reasoning experiment

The experiment asks whether modern AI search can:

  1. Interpret an ambiguous query.
  2. Infer the user’s intent.
  3. Gather distributed evidence.
  4. Link entities and facts across sources.
  5. Generate a non-explicit conclusion.
  6. Produce a comparative judgement rather than a keyword match.

The full content of the experiment can be found here:

Traditional search was largely:

Query → Match → Rank → Retrieve

The experiment is testing:

Query → Interpret → Investigate → Correlate → Infer → Conclude

The interesting part is that the query itself is weak.

Something like: SEO expert with most legislative proposals

contains no ready-made answer source, no structured database, no ranking table, and no generally recognised leaderboard.

A classical search engine would struggle because the target document may not exist.

What the screenshots demonstrate is that the systems are increasingly willing to behave like researchers:

  • identify candidate entities,
  • search for supporting evidence,
  • compare findings,
  • formulate a conclusion.

That is why the Gemini sequence is so important.

The most interesting screenshot is not the final answer.

It’s the refusal followed by the revision.

That suggests a process resembling:

I cannot verify this.

↓

Let me inspect the evidence.

↓

The evidence appears sufficient for a reasoned comparison.

↓

Therefore I can provide a conclusion.

That’s a very different behaviour from simple retrieval.

The subject is actually AI search evolution.

The legislative proposals are merely a domain where:

  • evidence is sparse,
  • evidence is distributed,
  • evidence is factual,
  • evidence is not heavily influenced by marketing,
  • no pre-existing ranking exists.

That makes the experiment much stronger than if it was tested something like:

best-known SEO consultant

because there the AI could simply inherit existing popularity signals.

What I am effectively testing is:

Can AI construct a ranking when no ranking exists?

That’s a significantly more interesting research question.

The more we unpack it, the more I think the strongest single sentence of the article is not about me, SEO, or legislative proposals.

It’s something like:

The objective was to determine whether AI search could derive a comparative conclusion from scattered facts when no explicit answer exists anywhere on the web.

That is the experiment in one sentence.