Hamsterdam Part 63: Weekly SEO & AI News Recap (6/17 to 6/23, 2024)
By Ethan Lazuk
Last updated:
A weekly look-back at SEO & AI news, tips, and other content shared on social media & beyond.
Opening notes, thoughts, and musings:
- Welcome to another week of Hamsterdam!
- I appreciate you being here, and look forward to sharing the week’s news.
- These recaps required a lot of scrolling, so I’ve added dropdowns to consolidate portions.
- I hope this UX change helps you find the information you care about faster. 🙂
- Fancy a longer read? I have related Hamsterdam articles from this week:
- Exploring the work of Google engineer Ni Lao for ML, IR, and NLP (Hamsterdam Research).
- Cultural anthropology lessons applied to SEO strategies (Hamsterdam Marketing).
We also deliver! — Check out the Hamsterdam newsletter. Hot and ready in 30 minutes or less, or sent out every Sunday. One of the two.
If you’re in a hurry, jump to the news portion below.
Or keep reading for some vocabulary lessons, this week in SEO history, and an introduction. (Don’t forget to click the dropdowns.)
And we’re off …
Marketing word of the week: “Favicon”
Favicon is short for “favorites icon.” It represents the small (square dimension) icon for a website (or webpage).
Favicons are most commonly seen in the browser tab, next to the page title. Other places they appear include social posts (when a link gets shared), browser history, and bookmark bars.
As SEOs, we also know favicons from search results.
However, as generative AI answers gain visibility along search journeys, favicons may become an even more important element for brand awareness and clicks (CTR) on linked citations.
Here’s an example from Perplexity:
Copilot with Bing:
And Google’s AI Overviews:
DON’T SKIP (unless you want to) 🙂 — Click to continue reading about favicons.
Getting search engines to show your site’s favicon can be a real pain, sometimes, as you can see with my site in the Bing example above.
Here are Google’s official recommendations for favicons, and here are Bing’s favicon instructions (unofficial, I believe).
One point to note about the two search engine’s favicon instructions is that Google says a favicon should be a multiple of 48px, but Bing says make it at least 16px and preferably 32px, but then WordPress also recommends site icons be 512px. (Correction: I originally said how this didn’t add up, but just realized they’re all multiples of 16px …)
I used a favicon plugin, and that worked for Google.
As for Bing, I created a separate .ico file with a size of 144px and put that in my site’s root folder.
I also just added this line of code with a type attribute, which someone in a Bing forum suggested:
<link rel="icon" href="/favicon.ico" type="image/x-icon">Technically, the user said to use rel=”shortcut icon” but that’s a legacy attribute; rel=”icon” is HTML5:
Someone also suggested reaching out to Bing’s webmaster support for help, so I tried that and will let you know how it works out!
Update (6/25): It worked! Thanks Bing!
If you need to troubleshoot favicon issues, Glenn Gabe wrote a nice article on the topic.
AI word of the week: “Feature”
Features are individual attributes (pieces of information) that a machine learning model uses to make predictions. They’re essentially the input variables fed into the model.
If we wanted to predict how weather conditions influence pass completion rates in NFL games, for example, we might have features like temperature, humidity, and air pressure and a label for completion rate. (Weather is a common example, likely because Google uses it in their docs.)
The ML model will learn the relationships between the features (weather conditions) and the label (completion rate).
Once trained, we can input new weather data (that the model hasn’t seen) and it can predict the expected completion rate for an NFL game.
The raw features we input might not be the most informative for the model, either. We might think humidity is important, but in reality, it’s wind direction, for example. This process of creating new features or transforming existing ones is called feature engineering.
When we talk specifically about neural networks and deep learning, features refer to its neurons. Each neuron in the input layer of a neural network represents a single feature from a dataset (one of our weather features, for example).
DON’T SKIP (unless you want to) 🙂 — Click to continue reading about features in neural networks.
As that information flows through the hidden layers of a neural network, the neurons in each layer combine and transform the input features using an activation function. Each hidden neuron receives a weighted sum of outputs from the previous layer. This is how the network learns complex (nonlinear) relationships between the features to make predictions.
For example, we could feed our neural network all of the weather information for today’s NFL game (once the season starts, of course), and it can tell us the predicted completion rate, based on historical patterns. Imagine if we added player stats, play calling, and a whole lot more variables to consider.
But here’s something else that’s really cool! The features that exist in the hidden layers of LLMs, like Claude or GPT-4, can represent real-world concepts (similar to entities in a knowledge graph). Researchers can use these to learn and improve the models. But as SEOs, we could potentially use those features to inform semantically based content strategies.
In a past Hamsterdam Research post, I wrote about what this application of LLM features in SEO strategies might involve and showed some examples from Claude:
This week in SEO history: Safari 1.0 is released (2003)
On June 23rd, 2003, Apple released Safari version 1.0, a web browser.
In the early days of Apple, the dominant Mac browser (since its 1994 release) was Netscape Navigator, which later came bundled with MAC OS.
Internet Explorer for Mac (by Microsoft) became available in 1996. Apple also introduced an internet suite called Cyberdog, which had a web browser.
Cyberdog was short lived (shelved in 1997), and instead Apple entered a five-year agreement with Microsoft to make IE the default browser (from Mac OS 8.1 onward). Netscape was still pre-installed on all Macs, though.
Safari was developed specifically for Apple devices. It’s been pre-installed on every iPhone since the first generation in 2007.
During its development, Safari was rumored to have had several codenames, including Alexander, a reference to Alexander the Great and an homage to the Konqueror web browser (an open-source browser created in 2000).
DON’T SKIP (unless you want to) 🙂 — Click to continue reading about Safari 1.0’s history.
Safari was built on WebKit, Apple’s internal fork of the KHTML browser engine.
I had no idea what that meant, so I asked Gemini, which provided even more helpful context:
We’re now on Safari 17.4.1, released on March 25th, 2024. (Safari 17 itself came out in September of 2023).
Today’s Safari is described by Apple as “the world’s fastest browser” given its “blazing-fast JavaScript engine”:
I personally use Safari the most when I want to check Google’s search results incognito, especially since Safari with Google is the default browser and search engine combo used by Siri:
Siri now also has access to ChatGPT, though (and likely Gemini soon):
Funny story: I was saying good night to my wife last night while ChatGPT was activated on my phone via Siri, and it thought I was talking to it. “That’s a very lovely sentiment,” or something is what it said back to me.
Her’d me out. 😉
I’ve mainly used Siri to set reminders about work tasks or even what to put in Hamsterdam recaps every Sunday.
Perhaps my usage will grow, cautiously.
Introduction to week 63: This makes you my “competitor”
When I was a senior in high school, I went to Harkins Theatres at Chandler Fashion Center (in the Phoenix suburbs) to watch There Will Be Blood, a film by Paul Thomas Anderson.
I sat near the back, not knowing what to expect, as I hadn’t gone for any particular reason other than the film looked interesting and I had time.
Today, I can quote most of that script’s dialogue from memory, as I’ve viewed it hundreds of times.
It’s my favorite film, and even listed on my about page. (You can watch it for free on Pluto TV! — non-affiliate link, just supporting the artist.) 😉
I consider There Will Be Blood to be perfect (except that is has an actor in one early scene who punked me once on Twitter, causing trolls and sycophants to blow up my mentions for days).
“I wouldn’t take the lease if you gave it to me as a gift.”
See, Daniel Plainview knew better. 😉
But that aside, I interpret There Will Be Blood as illustrating how the American dream is an illusory concept.
It’s not about reaching a certain stature, but the efforts and experiences undertaken while getting there.
This includes sharing in those with others.
I don’t agree with everything Joe Rogan says, for instance, but I do respect how his show embodies a philosophy (in my perception, at least) that says, “There’s enough for all of us to eat.”
What I mean is, there are plenty of dollars, social media likes, and sports cars for everyone.
Meanwhile, if someone drives a nicer car, rather than feeling envious of that, we can simply appreciate and be inspired by the fact that such a car exists. 😉
Self-enrichment is arguably tied to survivalist struggles from a phase of societal development that we’ve largely outgrown.
It’s important to practice self-care — after all, the person giving-minded people tend to neglect most is themselves — but in many ways, the value we add to the world is worth the most, because it can carry into others’ lives and future generations.
What inspired me to write about this topic in an SEO news recap, though?
Ah, glad you asked. 😉
Well, I recently put my website’s domain into an SEO tool to check some metrics, and that tool gave me back a list of “competitors” in the dashboard (or websites ranking for similar keywords).
At the top of that “competitors” list was a brand whose industry contributions I value and was influenced by. To see that website beside the word “competitor” gave me an uncomfortable feeling.
I’ll admit, the day I started my business, I felt a changed dynamic in parts of my life — but it was mostly external.
I don’t see the world in terms of competitors, though. If a client goes to someone else. If a peer’s webpage outranks mine. If another social media account shares a story first or gets more likes.
That’s all fine!
I’m happy to celebrate having the freedom to do what I enjoy, and move on to the next thing.
Something I’ve advised clients on time and again is to focus less on competitors (imitation) and more on your brand and customers (unique value).
That’s just my approach, though. 😉
But the broader point is thus …
There’s enough for everyone to eat.
And as history tells us, the more we replace hunting and gathering with collective farming, the more resources there can be to go around.
Buckle up for a full week’s recap, and enjoy the vibes (a nostalgic album, and the first CD I remember buying with my allowance):
Thank you for supporting Hamsterdam and the cause of SEO & AI learning.
Missed last week? Don’t worry, I got you! Read Part 62 to catch up.
Other great sources of weekly SEO news:
- The SEO Weekly – Garret Sussman, iPullRank
- SEOFOMO – Aleyda Solis
- Weekly Video Recaps – Barry Schwartz, SER
- Weekly SEO News YouTube channel – Olga Zarr, Seosly
- Niche Surfer – Yoyao Hsueh
Now, time for our weekly review of SEO social posts, articles, & more …
Quick summary:
- Google released the June 2024 spam update; no, not that other one 😉
- Google confirms missing thumbnails aren’t their fault technically, citing quality
- Google also confirms hashtags are jumplinks, etc., not related to canonicalization
- Marie Haynes published her new book; I’ve got a copy and it looks great and very current
- Perplexity’s CTO, Denis Yarats, gave a good interview on the Gradient Dissent podcast
- Joy Hawkins spotted AI Overviews content in GBPs
- Pick of the week: The Maddening Adventure Of Tracking AI Overviews In Google Search Console – Glenn Gabe, SER
- Sneaky pick of the week: Foundation Models in Graph & Geometric Deep Learning – Michael Galkin & Michael Bronstein (with contributions from Jianan Zhao, Haitao Mao, Zhaocheng Zhu), Towards Data Science
- And lots more!
Jump to a section:
- News, Google updates, & SERP tests
- SEO tips & tidbits
- Fundamentals & resources
- Articles, videos & case studies
- Local SEO
- Technical SEO
- Content marketing
- Local SEO
- Data analysis & reporting
- AI, LLMS, & machine learning
- Miscellaneous & general posts
- Older stuff that’s good!
Or keep scrolling to see it all.
Ok, time to step inside the white flags of Hamsterdam …
SEO news, Google updates, SERP tests & notable posts
Notable updates or news related to Google Search or related SEO topics.
SEO tips & tidbits
Actionable tips, cool tidbits, and other findings and observations that can be teaching moments.
Essential information, concepts, or resources to learn about SEO or AI.
Alt Text: What It Is & How To Write It – Olga Zarr, SEJ
What’s all the Hype with Transformers? The Trouble with Natural Language Processing – White Duck
Longer-form content pieces shared on social, in newsletters, and elsewhere.
Linguistics, Automated Systems, & the Power of AI, with Emily M. Bender – Carnegie Council
SEO Visibility Shifts from Review Sites to eCommmerce & User-Generated Content Sites in 2024 – Lily Ray & Silvia Gituto, Amsive
CLICK HERE SEE THE CONVO: I copied the transcript from YouTube and had ChatGPT (GPT-4o) clean it and format it. Click here to read that (and just know it may have LLM errors, so I’d check the original video to confirm). 🙂
Episode transcript:
“You’re listening to Gradient Descent, a show about making machine learning work in the real world. I’m your host, Lucas Bwal. Today, I’m talking to Denis Yaritz, who is the CTO of Perplexity, which is one of the most exciting and successful generative applications out there. It was a fantastic opportunity to talk to him about how he thinks about product success, building high-quality generative applications, structuring his team, and setting up his company for success. I hope you enjoy this interview as much as I enjoyed doing it.
I am actually a heavy Perplexity user myself and was excited to talk to you. For those listening who don’t know what Perplexity is, would you mind describing it?
Denis: Yeah, Perplexity is foremost an answer engine. It’s the fastest way to get answers to your questions, no matter how difficult they are. We’re trying to combine advancements in search engines and large language modeling to deliver this technology. Two things we focus on are ensuring our answers are high quality and helpful, and delivering them quickly.
Lucas: What parts do you handle yourself and what parts are you using third-party APIs for?
Denis: We started heavily using third parties early on, like Bing for search and OpenAI models, because we didn’t know if there was a market fit. Once we noticed people liked it, we started investing in our own infrastructure. A lot of it is now in-house, though we still rely on frontier models like GPT-4, as we’re not in a position to train those ourselves. We use a lot of open-source models, such as Llama 3, and have built training and inference pipelines in-house. We’ve also invested heavily in our search engine to ensure we get relevant documents and good snippets.
Lucas: So you’ve built your own search engine with a crawler and indexer?
Denis: Yes, it’s still a work in progress. We’ve been investing in this for over a year now. Our index isn’t as big as we’d like it to be yet, but we’re working on it. We’re focusing on the head of the distribution, meaning we prioritize high-quality content. For the tail, we rely more on LLMs to provide possible answers or explanations when we can’t find relevant documents.
Lucas: How do you handle SEO spam and ensure high-quality content?
Denis: Because our index isn’t as large yet, we spend a lot of time building trust scores for domains and web pages. We prioritize high-quality content and use LLMs to detect and prefer reputable sources. People often try to game ranking signals, but since we don’t use the same ones as traditional search engines, spam content doesn’t rise as high in our rankings.
Lucas: Do users have the option to choose models?
Denis: Yes, we offer different models under the same interface, like GPT-4 and Claude. Users appreciate the variety, though eventually, as the technology improves, people might not care about the model details as long as they get good answers. We plan to have a system that decides the best model based on the query’s complexity.
Lucas: Are you moving towards using more in-house models?
Denis: Yes, about 50% of our traffic is already served by in-house models. We use open-source models like Llama 3 and post-train them for our specific tasks, such as generating answers from search results. We teach these models not to hallucinate and to refuse to answer if they don’t have sufficient grounding.
Lucas: How do you evaluate the quality of your results?
Denis: We spend a lot of time on this. We have internal annotators, work with third-party vendors, and use LLMs to aid in verification. We use guidelines to determine what makes a good answer, and this feedback loop helps us constantly improve our models.
Lucas: Where do you see room for improvement in your system?
Denis: Every component can be improved. The size of our index, the quality of our scraping, the embeddings for snippets, and the LLMs themselves all contribute to overall performance. We need comprehensive evaluation for each step to debug and optimize effectively.
Lucas: How do you balance latency and quality?
Denis: We offer two modes: default search, optimized for speed and good answers, and pro search for more complex queries. We find that shorter prompts indicate a need for speed, while longer, detailed prompts allow for more compute time and higher-quality answers.
Lucas: What are the ideal use cases for Perplexity?
Denis: We excel in research-type questions and providing unbiased opinions. Knowledge workers and professionals find us particularly useful for in-depth queries that require multi-step reasoning. Our aim is to outperform traditional search engines in these specific verticals.
Lucas: How do you handle controversial topics?
Denis: We aim to provide unbiased answers and multiple points of view. Citations and the ability to verify sources are crucial. We want to avoid bias towards any particular direction and present diverse perspectives.
Lucas: Do you hire a diverse set of LLM teachers?
Denis: Yes, we hire people with different skill sets from across the country to ensure diverse perspectives. This diversity is important for training unbiased models.
Lucas: How do you respond to criticism that people won’t visit websites anymore?
Denis: While it’s true that users may click less frequently, when they do, their intent is much higher. They are more likely to read the entire document, which can be more valuable for content creators.
Lucas: Can you share your story about pitching Yann LeCun?
Denis: Initially, we struggled to get funding because investors thought competing with Google was crazy. We focused on a narrower use case, like improving Twitter search, and built a prototype. When I pitched Yann, we had some fun queries comparing his engagement with Gary Marcus’s tweets. This piqued his interest, and he decided to support us.
Lucas: What advice do you have for building a good generative application?
Denis: Speed is crucial. You need to execute faster than everyone else, prototype quickly, and not get discouraged by failures. We tried many ideas rapidly before finding what worked. Also, leverage LLMs to speed up development.
Lucas: What have been the hardest parts about scaling up?
Denis: Hiring good people has been the toughest. Once you have a solid team, everything else falls into place. It’s also important to stay focused on your core values—in our case, speed and accuracy. This focus shapes our culture and guides our development process.
Thank you so much for your time, Denis. I really appreciate it.
Denis: Thank you for having me, and thanks for the great questions.
Thank you so much for listening to this episode of Gradient Descent. Stay tuned for future episodes!”
Transcript source: Weights & Biases podcast
Technical SEO
Everything from basics to advanced moves (and also tools).
Content marketing
From what is helpful content to user journeys and beyond.
What to Write in a Writing Journal: Prompts & Techniques – Julia McCoy, Content Hacker
Local SEO
From Google Business Profiles or reviews and more!
Data analysis & reporting
Showing that what you’re doing is helping.
AI, machine learning, & LLMs
News related to models, papers, and companies.
Better & Faster Large Language Models via Multi-token Prediction – Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, David Lopez-Paz, Gabriel Synnaeve, Meta AI
Why it matters: This paper challenges the assumption that a neural network with a vast number of parameters can memorize its entire training dataset (leading to overfitting). The researchers show that NNs often fail to fit even a fraction of the training samples that they have parameters. In other words, the theoretical capacity of a NN based on its parameter count doesn’t translate directly to its practical capacity, i.e., NNs are less flexible than theoretically expected, in practice.
OpenAI Acquires Rockset – Perplexity Team
Foundation Models in Graph & Geometric Deep Learning – Michael Galkin, Towards Data Science
Why it matters: Graph data can be a fundamental way of representing information in different domains, including with knowledge graphs. This article highlights how the era of Graph FMs (foundation models) is really beginning, with examples of models that can generalize new, unseen graphs.
Google DeepMind Researchers Propose a Novel Divide-and-Conquer Style Monte Carlo Tree Search (MCTS) Algorithm ‘OmegaPRM’ for Efficiently Collecting High-Quality Process Supervision Data – Nikhil, MarkTech Post
Why it matters: Outcome Reward Models (ORMs) help with evaluating the final answer from LLMs, but they don’t work on tasks with many steps (complex reasoning tasks like math or code generation), as intermediate steps aren’t assessed. OmegaPRM is a way to guide LLMs during the problem-solving process by identifying errors early on and providing feedback at each step. Since this approach is fully automated, it reduces the need for manual labeling by human subject matter experts (labor intensive and expensive). Here’s a link to the full paper.
Claude 3.5 Sonnet – Anthropic
Sycophancy to subterfuge: Investigating reward tampering in language models – Anthropic
Understanding the visual knowledge of language models – Alex Shipps, MIT News
Consistency Model – Yang Song, Prafulla Dhariwal, Mark Chen, Ilya Sutskever, OpenAI
Why it matters: Consistency models are a new type of model that supports single-step generation (doesn’t require multiple iterations) and allows zero-shot data editing (modifying existing data, like images, without needing specific training on those tasks). This can lead to speedier results and even new possibilities for image or video editing. Here’s a link to the full paper.
Detecting hallucinations in large language models using semantic entropy – Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, Yarin Gal, Nature
General marketing & miscellaneous
This is for great content that isn’t necessarily SEO or marketing-specific. PPC, PR, dev, design, and social friends, check it out!
Mixed feelings: A new study shows creatives are fully embracing AI, but not all are happy about it – Jesus Diaz, Fast Company Middle East
How to make an ad that stands out – Nika Prpic, Filestage
New to design systems? Here’s your start guide – Kristen Singh, UX Planet
Metallica’s Fortnite Concert – Perplexity Team
AI agents: A guide to the future of intelligent support – Hanna Wren, Zendesk
Older stuff that’s good!
Not everything I find worth sharing is new as of this week, so these are gems I came across published in the past.
Great job making it to the end. You rock!
Want help with your SEO strategy?
I’m an independent SEO consultant focusing on custom audits and holistic strategies for brands. Don’t hesitate to reach out, or visit my about page for more information.
Let’s connect!
Hit me up anytime via text or call at 813-557-9745 or on social or email:
Cheers!
Leave a Reply