What we’re seeing
Google doesn’t like local seo pages that read the same. Ie a local page that is the same as all other pages on your city, just with the city name replaced,
Local pages rank when they include facts that are only true of that locality.
More of content generation by LLM has to be actually useful, the days of relying purely on LLM memory for hallucinated article writes are numbered.
Why the old model breaks
SCM started with basic keyword > single prompt > AI article,
Then it was single prompt > write section > combine to make 1 article.
Both of them had problems, the AI would hallucinate and splitting writing by sections caused the AI to repeat itself, creating a long mini repeating junk article..
Next try. Pass pre scraped serp content from the primary keyword and use AI memory by passing the output of one prompt to the next. This is the chain prompt + SERP modelling idea.
This fixed hallucinations and memory meant 1 article was long and sections flowed.
Here is the break
- The old chain prompt template gets one keyword and nothing else.
- Without local facts, every town page is the same page with the city name swapped.
- Google treats those pages as near-duplicates and doesn’t reward them.
The operational problem
To get actual researched facts, you need to add extra keyword searches.
There is no way to pre generate what those searches should be ahead of time.
Chain prompts only did 1 keyword SERP modelling on the primary keyword.
Concrete example.
Vet clinic for a town.
That town might have pet registration and regulations specific to that town, and that information might be on a website, a pdf or hidden somewhere requiring digging outside the original ‘vet clinic in box hill’ search term.
The new approach
Its all about the agents now.
Just like coding is all agents, LLM generated text requires agents now.
The new approach:
- An agent page still gets one keyword.
- We ask the agent to search the web and track down the research it needs.
- You don’t need to provide any extra keyword searches.
- We then spawn extra agents after the content is written to do 2 things.
- Check article is written to be helpful
- Check facts are properly sourced and accurate
How it works
- Page Pipeline, select the new page type from Agent list.
- Pick Local SEO Service, Local SEO Hub, or Article to let the agent choose the page type.
- Use medium to high effort for the deepest research.
This is how typical article writing step looks like now.
-
Agent searches the web
-
Agent opens results and reads content
-
Agent searches again if needed
-
Agent writes article
-
A review agent comes back to fix it
I started with only 1 keyword “vet clinic, box hill, mel, aus” but Agent found very specific information to the location and started reading the council website @ whitehorse.
The old method in SCM and most likely any tool you use, would never have found and read the actual council website for real facts.
What this means going forward
- Articles written with agent research, starting from Google to opening websites.
- The model effort level decides on how much research and checking it will do.
- More token cost
- Agents have full access to tools, they can read pdfs, make searches and even compile stats if required because they have access to scripting and files
As models get better and their speed increases, this will be the new normal.
Right now a well researched run will burn through 200K new tokens, and make about 1Mil in token reads. So maybe 5x what a simple chain prompt will do. However the quality of articles jumps. It is also the future and presents what Agents are capable of now.
Next update to SCM ships with page agents.
All older chain prompts are now deprecated and no longer show up in the prompt selector.
Model selection will also be cut down slightly to only support services that work with agents and tool calls.