Chris Green SEO

Here is another good glimpse of a future compensation model for AI. Paying publishers for using their content. For everyone who thinks Publishers are facing an existential crisis, this may suggest a potential alternate future. 🧵 1/3

This raises many, many more questions than it does answers for me. How do you calculate the worth of being part of an answer? Is this worth enough for a publisher? How do you stop these mechanics from being manipulated by scrapers? The list goes on. 🧵 2/3

At the heart of understanding AI search is understanding where the data is coming from. With ChatGPT, we’d suspected it was Google and Bing. But that has become MUCH more complex recently. 🧵 1/3

Peec has done great work exposing more of this processing and helping shed light on the sources that could really play a part in generating AI answers. 🧵 2/3

What if an SEO extension could harness AI to help you run a better audit without delegating the entire thing to an Agent? I wanted to see how doable this is. 🧵 1/4

I’m not trying to replace SEO audit tools or change your existing workflow, just provide a set of tools that have a relatively light footprint, but will be genuinely useful where it’s needed. 🧵 3/4

What if an SEO extension could harness AI to help you run a better audit without delegating the entire thing to an Agent? I wanted to see how doable this is. 🧵 1/4

AI can render content in a browser - it’s pretty good at doing it now. It can also read text in an image and can interact with JavaScript-rendered content. Does that mean AI Chatbots or other search surfaces will do this every time? That’s an assumption that’ll create risks! 🧵 1/2

ICYMI - if you want an approach (and tool) to help you understand if your pages are likely part of AI retrieval, then this could be quite useful! 🧵 1/2

Sometimes someone builds a tool I wish existed a long time ago - this time it is Mersudin who has stepped forward. Migration Risk uses your domain and GSC account data to profile the most important pages on your site and the pages that pose the greatest risk if not migrated correctly. 🧵 1/2

Bot control in 2026 isn’t as easy - it’s not as simple as just using robots.txt anymore. It also needs to involve more of the business than ever, which is both good and bad. 🧵 1/2

When you are searching for something, where would you rather spend the effort: researching or verifying returned results? Duane’s latest Substack explores the idea that AI search hasn't removed the effort needed, but has merely shifted where that effort is required. 🧵 1/5

and synthesis process that produced it.” I think there’s a quality issue here, but Duane’s conclusion is more helpful for SEOs and search professionals now: “Our information may be extracted in pieces, combined with other sources, compressed into a smaller answer, and presented… 🧵 4/5

When you are searching for something, where would you rather spend the effort: researching or verifying returned results? Duane’s latest Substack explores the idea that AI search hasn't removed the effort needed, but has merely shifted where that effort is required. 🧵 1/5

For those who don’t want to research and don’t care about verification, AI search may be a net gain. Those who've been using it for a while may feel that sometimes the process is not much quicker. That’s what I feel at times, at least! Duane does suggest that this shift is better (overall): 🧵 2/5

What’s more important: prompt wording or prompt intent? “I don’t think users would search for that” is a common response to outputs of prompt generation or prompt selection. We don’t have a source of data (like AdWords Keyword Planner) to help us adjudicate/validate our prompt selection. 🧵 1/7

Brand inclusion (for this brand) is robust to wording changes, while citations are more dependent on retrieval. Changes in prompt wording can alter the search queries(sources) or query fan-outs generated, which can then surface a different set of sources. 🧵 6/7

What’s more important: prompt wording or prompt intent? “I don’t think users would search for that” is a common response to outputs of prompt generation or prompt selection. We don’t have a source of data (like AdWords Keyword Planner) to help us adjudicate/validate our prompt selection. 🧵 1/7

However, when looking at one established brand within the test set, the pattern was notably different: - Brand mentions: ~99% stability - Brand-domain citations: ~35–45% stability 🧵 4/7

What’s more important: prompt wording or prompt intent? “I don’t think users would search for that” is a common response to outputs of prompt generation or prompt selection. We don’t have a source of data (like AdWords Keyword Planner) to help us adjudicate/validate our prompt selection. 🧵 1/7

I believe this is one of the larger issues of prompt tracking. I took 32 prompts, rewrote each prompt four different ways while preserving the underlying intent, and tested each version five times (to make a small account for “noise” whilst making it achievable in the timeframe I needed). 🧵 2/7

This has only gotten truer over time. I love a good Venn, and Cyrus has created a corker. I know I promise to stop re-litigating the SEO vs GEO debate, but it’s refusing to die. 🧵 1/3

Every time someone sees a difference between ChatGPT’s results & “traditional” results, a tier of company/consultant/tool provider steps in to expose what “no one else is talking about” or “SEOs are in denial of” Let’s be honest, that’s someone trying to mislead you to sell more through fear 🧵 2/3

Myriam’s work on the changing role of images in AI search could be huge. But there’s an itch I need to scratch: - We know some models can understand images. - We know multimodal embeddings can put text and images into shared semantic spaces. 🧵 1/6

There’s a big jump from “this technology exists” to “this signal participates in retrieval” and then another jump to “we should optimise for it.” I think this is one of the most interesting conversations in “AI Search” right now, and I don’t think people are paying enough attention. 🧵 5/6

Myriam’s work on the changing role of images in AI search could be huge. But there’s an itch I need to scratch: - We know some models can understand images. - We know multimodal embeddings can put text and images into shared semantic spaces. 🧵 1/6

but I think for people to take this approach seriously (because it takes a lot of effort and isn’t something that SEOs were historically that strong in), some experimentation here is crucial. 🧵 3/6

Myriam’s work on the changing role of images in AI search could be huge. But there’s an itch I need to scratch: - We know some models can understand images. - We know multimodal embeddings can put text and images into shared semantic spaces. 🧵 1/6

There are two key schools of thought with AI trackers at the moment - ones that I think will continue to develop further, but are critical to consider. - Fidelity to “real” results - “more accurate”. Usually through scraping the interface of the AI Chatbot 🧵 1/9

visibility data from it can be as useful as a market representation; it can still be useful. To be clear, API + Tool calling is not the same as scraping the UI. Scraping the UI is not the same as a “real user” searching. 🧵 8/9

There are two key schools of thought with AI trackers at the moment - ones that I think will continue to develop further, but are critical to consider. - Fidelity to “real” results - “more accurate”. Usually through scraping the interface of the AI Chatbot 🧵 1/9

- Wait, “traditional” rank tracking has become less and less accurate, but we (as an industry) don’t seem to be as bothered. Surely the most important thing is how we interpret and present the data? I have come to the view that which method you should use depends on what you need to convey 🧵 6/9

There are two key schools of thought with AI trackers at the moment - ones that I think will continue to develop further, but are critical to consider. - Fidelity to “real” results - “more accurate”. Usually through scraping the interface of the AI Chatbot 🧵 1/9

- But the API (with tool calling) is nowhere near as close to the UI experience; is it close enough to be valuable? - Personalisation means that UI fidelity is a bit of a fiction - so this can be most misleading 🧵 4/9

There are two key schools of thought with AI trackers at the moment - ones that I think will continue to develop further, but are critical to consider. - Fidelity to “real” results - “more accurate”. Usually through scraping the interface of the AI Chatbot 🧵 1/9

- Deep/probing - a more complete view of parametric memory. Usually via the API, calling the “search” tooling if required Initially you’d think the merits of each method were quite easy to understand. More accurate data is better for a low number of prompts, and... 🧵 2/9