🌐

​AI Mode expands to over 180 countries. With this expansion come new agentic features for restaurant reservations and the ability to share search results. Users can then pick up the shared search with follow up questions.

📉

​Google’s Knowledge Graph saw its largest contraction in a decade in June: a two-stage, one-week drop of 6.26% – over 3 billion entities deleted, trading sheer volume for clarity and confidence.

⛓️‍💥

​Referral traffic from ChatGPT to websites is down 52% since July 21, according to Profound’s Josh Blyskal, the analysis of over 1 billion citations showed a consolidation to sites like Wikipedia and Reddit.

🧍

​A new study shows which sources LLMs use when searching for well-known people. Different 'types' of people trigger different sources, with Wikipedia, YouTube, Britannica, and top news sites being commonalities.

📄

​ChatGPT is using Google search results via SerpAPI for grounding, after OpenAI reported 'significant quality issues' with results they were previously fetching from Bing. Ouch.

🕵️

​Exposed ChatGPT configuration files hint at how website citation may work. Some indications seem to confirm that as suspected, when performing RAG, LLM-based systems are favouring fresh content.

Need more leads to find your business on ChatGPT?

• Figure out why ChatGPT is not mentioning your brand
• Receive a personalized action plan to close your visibility gaps
• Win long term by doubling down on initiative's that move the needle

llms.txt is not the secret ingredient

llms.txt isn't a thing. No major LLMs have documented support. No major LLMs have announced they use it. The file appears to not get consistently crawled. Spend your time elsewhere 😉

This isn’t chemistry—It’s tokenisation

This is a visual explanation of why your favourite LLM does not use schema in their core training data (ignoring the fact it's likely stripped out during pre-training) ⤵️
​
LLMs work by "tokenising" content. That means taking common sequences of characters found in text and minting a unique "token" for that set. The LLM then takes billions of sample "windows" of sets of these tokens to build a prediction on what comes next.
​
The image below is some example schema that has a colour change applied which represents that set of characters is a unique token as made by the GPT-4o model.
​
What you will notice is that the schema gets "destroyed". For instance, the schema "@type": "Organization", gets broken down so there are separate tokens for "type" and "Organization", which means that in terms of tokenisation the regular words "type" and "Organization" are not distinguishable from schema.
​
If schema was included in this training data, all it would do in reality is say there is a slightly (likely insignificant) probability of tokens such as "@ appearing before the word "content".
​
Schema is useful because it is explicit. This explicity is lost during tokenisation.
​
Of course, LLMs could be using schema in some kind of RAG process, but this would likely be very similar to what we are seeing with search engines.

They’re still cooking with the same recipe

DOJ documents reveal that Google use many of their 'traditional' metrics to decide which sites to surface for Gemini... Just like Gary Illyes told us 🙃

This is my territory: Exact Match Domains

One of the main reasons that Exact Match Domains (EMDs) rank well is they naturally generate a high Site Quality Score, as defined by Google ⤵️
​
How Is Site Quality Score Calculated? 🧮
​
Site Quality Score = S / U
​
For a given website:
​
S: The count of unique queries that directly refer to the site (for example, explicit branded searches mentioning the site's name).
​
U: The count of unique queries that result in a visit or are associated with the site (all queries that led to the user clicking into the site).
​
The theory that a higher ratio means the site is being specifically sought after more often, which could suggest higher quality or reputation.
​
Check out the Site Quality Score patent 🤓 📜

Say my name

Google automatically generates what it believes is the best name to describe a website, but it can get it wrong sometimes! 🤦‍♂️
​
While you can't force or manually set the displayed site name, one of the strongest ways you can indicate a preference is by specifying one with WebSite structured data.
​
Outside of this, Google takes into account content from a site's home page and references to it that appear on the web.

Top Core Updates referrer leaderboard

A big thank you to our top referrers, who have signed up over 10 people to the Core Updates newsletter, go follow them!

🥇 MJ Cachón Yáñez (LinkedIn / Bluesky)

🏅 Nikki Pilkington (LinkedIn / Bluesky)

🏅 Lidia Infante (LinkedIn / Bluesky)

🏅 Saijo George (LinkedIn / Bluesky)

🏅Alice Rowan (LinkedIn)

🏅Martijn Jeurissen (LinkedIn)

🏅Chester Beard (LinkedIn)