[3/3]
"Semantic search" has mathematical blind spots. If you abandon exact keywords for pure vector search, you lose visibility on logic-based queries.

arxiv.org/pdf/2508.21038

[2/n] The results were wild:
❌ SOTA Embedding Models (Google, OpenAI) failed hard, scoring <20% recall.
✅ Old-school BM25 (Keyword matching) scored nearly 100%.

Embeddings are great at concepts, but bad at precise boolean logic.

Welp... Now I know what I'm doing for the next 37 minutes. Thanks Lily for sharing.

@wilreynolds.bsky.social I saw you liked my message 👀

Would be awesome to chat when you feel the timing’s right. Here's my calendar link if and when you feel up to it! meetnicolas.com.

Hey Wil. Big fan of yours. I built something similar. Maybe an opportunity to collab and build a single, stronger solution instead of chasing the same goal from two sides? I have few killer features in dev that could make it worth a chat.

The Forbes Advisors story is maybe the most unethical case of Search Engine Manipulation and they finally got burned after who knows how much money they made from it
larslofgren.com/forbes-marke...

Long read but a good one.
written by: @larslofgren.bsky.social

#SEO #SEOTools

🧵 (4 of ?)
3: The breakdown is for a single industry and results may vary for other industries (though I doubt it).

4: Similarity scores on page title aren't the *BEST* way to determine content relevance. But it's not a bad way either. More factors at play around content quality, etc...

🧵 (3 of ?)
Limitation 2: The analysis was done on a mix of non-branded head and tail terms more geared towards information queries.

* PS - I hate this character count limit!

🧵 (2 of ?)

Limitation 1: based on one of the "big 3" SEO tools provider so it could very well the other tools KW Difficulty Scores are more accurate. I will not name the company because the purpose of this is not to "throw any shade". The assumption is they all are equally as problematic.

Really enjoyed your talk! Great idea recording it.

Hi!