TL;DR
If you have access to BigQuery, make sure you're granted permission as a 'Vertex AI User' so you can experiment straight away.
The main caveat is sadly to scale it as computing resources can still be heavy.
Using Gemini 3.0 to capture entities and mine sentiment with AI.Generate may be powerful
Ofc, it depends on what you run it against.
Is it title tags? ❌
Is it excerpts from your content? ✅
I tried it on queries and the extracted entities went rogue – how is Google ADS api even related? LOL
✅ AI.similarity() works well if computed on your pre-processed datasets (e.g you ingest title tags, headings, custom extractions)
Upload an SF file to BQ with your custom extraction and use Gemini's text-embedding model to identify gaps w/ on-page SEO and your content.
AI.Classify uses Gemini 3.0 to map out intents based on rule-based classification.
The trouble is that search intent is too nuanced for LLMs to take the lead, making it error-prone
❌ Intent classification should be a manual task for your keyword research. Consider rule-based automation at worse
🔔 BigQuery supports AI similarity and embedding generation
Basically, they released new functions you can embed in your SQL queries, including entity extraction, intent classification and sentiment analysis of unstructured text with Gemini 3.0.
I ran a few tests 👉🏻
This is a fine example of stochastic mirroring within #AI
The story was fabricated to quickly get karma on Reddit.
AI bots emulated human negative emotions to tell a story.
The point is use LLMs to brainstorm, refine and create the code.
You do the analysis.
www.linkedin.com/posts/benluo...
This is actually a made-up story - quite sure it was manufactured to get karma on Reddit.
www.linkedin.com/posts/benluo...
Admittedly, I did test it on my own website's content, but had to remove it due to generating infinite URL parameters and polluting my GSC page profile – couldn't be arsed to add a line at the robots.txt for that
In turn, seodepths.com leverages a lean plug-in called Effi Share – no prompt injection
Microsoft detected AI memory poisoning subtly used for brand promotion.
Prompt injection is not a matter of black vs white SEO.
Instructing LLMs to “remember” a company can introduce serious cybersecurity and open to geopolitical cases when dramatically scaled.
www.microsoft.com/en-us/securi...
Never actually managed to touch the AI brand vs non brand split 🤔
I've seen this often. Interesting that their models are considerably skewed towards Python.
Although one can argue it's one of the most versatile and easiest languages to build programs and automations – if ChatGPT aims at mirroring human skills, it is probably just fair😆
AI agents are contemporary Narcissuses.
Language models and chatbots are progress in disguise as they can't help but mirror mankind's in-depth need for validation and self-comfort.
RLHF and stochastic mirroring are the reasons LLMs seem so real
Genius
julianajackson.substack.com/p/what-is-re...
I don't know if they ended up in men's rectums, but in Italy there have been several unexploded ordnance, especially from American allies near Milan or along the River Po.
As AI platforms emerge as new traffic channels, attribution becomes a crapshoot.
Sharing an advanced framework to track zero-click searches and guide your decisions on AI visibility by integrating Google Search Console and Bing data via BigQuery incremental marts.
seodepths.com/seo-research...
I'm not sure we should normalise 404 pages to redirect to the HP.
it's not the end of the world ofc, but IME should always be treated as a last resort
I've developed a framework that combines authority metrics with semantic relevance to return much deeper insights into link value.
seodepths.com/python-for-s...
Log file analysis shows llms.txt has limited impact on AI crawl behaviour.
Learn why robots.txt still matters most and how to deploy llms.txt safely.
seodepths.com/seo-research...
🔔 Did you know #OpenAI may keep a cached index of webpages?
When a page is fresh or heavy on #JavaScript, #ChatGPT can generate answers from related cached pages.
I built a small app using your OpenAI API to check the external_web_access endpoint
Tests and context -> github.com/simodepth96/...
According to my server logs they're all over the place hitting our robots.txt - and guess what? they've been around for so long, training ML models and LLMs.
Hold on a sec, where does llms.txt sit within all of that ahah
I was gonna catch up with the Godfather saga, but time is ticking ahaha
I'm not surprised at all.
If you're based in the EU and have a proper look into your server logs, this is what you get.
The idea per se is smart as LLMs may slice .md faster than .html but if it doesn't become a standard in the W3C then it lives in thin air
You're right, it was someone desperately trying to upsell a plug-in from their CMS.
this goes right along with my theory that the AI obsession in *users* is often bc…
1. they were never educated (esp "ipad generation") & don't know how computers work
2. our tools and services are all SO AWFUL NOW, degraded at best, actively user-hostile at worst
Head to Skyline to attend an insightful session on ‘Technical SEO nightmares on client-side React web apps’ by @simodepth.bsky.social
#BrightonSEO
Test built on stiff methodologies from another more traditional SEO test
will removing meta description lead to traffic decrease?
seodepths.com/seo-research...
A test on 7 URLs with missing meta descriptions showed:
- Google and ChatGPT favoured H1S and structured text
- Perplexity relied on cached metas.
LLMs and AI search engines tend to mimic Google, preferring structured, self-contained content during retrieval.
👉🏻 seodepths.com/seo-research...
"The schema markup (itemprop, itemscope, etc.) was there to label the data, but didn’t need it to understand or summarise the profile."
IMO, schema markup as data labelling is the best definition ever --- I'd borrow it from Jarno Van Driel from a recent podcast.
From the conversation, ChatGPT 5 went on:
"I relied on the plain visible content in the HTML page — the text inside headings, paragraphs, and lists."
I tested Sarah Chen
"""
✅ I used the human-readable content: “Sarah Chen — Chief Executive Officer & Co-Founder”
🔎 The schema markup was a signal about structure (helpful for machines, search engines, or AI testing) but not required for me to extract the details.
"""
chatgpt.com/share/68ac1c...
/3
It's probably our best bet to start optimising information retrieval with #AI engines
Lots of testing in sight though - AI mode vs Grounding Gemini API are based on different models, but to what extent is all to be found
