Hey all,
It has been a relatively quiet week for Google Search.
By “quiet,” I mean Google hasn’t publicly admitted to changing anything important. The SERPs may still be doing cartwheels behind the curtains while the engineers whistle innocently.
Two AI automation developments are worth a look, though. And authority piggybacking is still very much alive and kicking.
OpenAI Has Released a Managed Agents API
Status: Confirmed public beta
On 10 September, OpenAI released its new Agents API. It handles session management, context compaction and recovery for long-running automated jobs.
Agents can retain sessions, report their progress and connect to external tools or MCP servers. They can run inside OpenAI-hosted sandboxes or on infrastructure controlled by the developer. OpenAI’s API changelog contains the announcement.
Why It Matters
This could be useful for publishing workflows where one job needs to:
- Research a product
- Analyse competing reviews
- Produce structured content
- Check the claims and links
- Build the pages
- Pick itself up after falling over
At the moment, we generally have to manage all the progress records, retry information and intermediate files ourselves.
The Agents API could remove some of that faff.
The words public beta are worth remembering, though. They mean “ready for careful testing,” not “point it at 1,000 ClickBank reviews, nip down the pub and assume it’ll all be tickety-boo.”
What I’d Test
Give one agent a five-product batch and deliberately interrupt it halfway through the research or drafting stage.
Check whether it:
- Resumes from the right place
- Avoids repeating paid API calls
- Preserves the original sources
- Produces the expected file structure
- Leaves enough information to diagnose any cock-ups
I’d still keep the existing database as the authoritative record. The agent’s memory should not be the only place storing batch IDs, research or publishing status.
AI memory is useful.
A proper audit trail is more useful - especially when the AI confidently insists it completed something that plainly isn’t there.
OpenAI Added Prompt Cache Diagnostics
Status: Confirmed
OpenAI also released Prompt Cache Diagnostics for supported GPT-5.6 and later models on 8 September. The feature is listed in the API changelog.
This sounds painfully geeky.
The useful version is much simpler: if you keep sending the same large block of instructions, OpenAI can reuse some of the earlier work and charge less for processing it.
If you keep changing bits near the top of the prompt, the cache may miss and you end up paying for the whole thing again.
Marvellous.
Why It Matters
Our automated article jobs repeatedly send the same material:
- Writing rules
- Review templates
- Banned phrases
- Output formats
- Validation instructions
- Safety rules
Across ten articles, a little inefficiency is neither here nor there.
Across 1,000 articles, that little inefficiency puts on a suit and starts demanding its own budget.
What I’d Test
Keep all the shared instructions together at the beginning of the prompt. Put the changing product information at the end.
Then run a small batch twice and compare the cost, cache-hit rate and finished articles.
If the second run is cheaper and the content remains consistent, lovely jubbly. Lock that shared prompt, give it a version number and stop tinkering with it every five minutes.
OpenAI has also added expiry dates for project API keys. Each unattended publishing tool should have its own key rather than one all-powerful master key quietly controlling everything like a tiny digital Bond villain.
A bit of belt and braces here could save a proper headache later.
Google, Reddit and Parasite SEO Watch
Status: No new official Google update; plenty happening in the wild
Google’s official ranking dashboard still lists the August spam update as its latest confirmed ranking event.
That does not mean authority piggybacking has stopped working.
Far from it.
From what I’m seeing, Reddit is still going strong in Google, while Steemit is getting some decent traction too.
The Ninja Forms PDF exploit is also still killing it. Those pages can borrow the authority of an established domain and rank very quickly—right up until the webmaster discovers what has happened and knocks the whole thing on the head.
So the lifespan may be uncertain, but the rankings can be very real while they last.
This is observed behaviour rather than a confirmed Google change. I certainly wouldn’t build the family pension around it.
Still, authority piggybacking is alive and well in Google land.
The platforms and methods change. The basic idea remains the same: find somewhere with trust, get a page onto it and see how much of that trust Google allows the page to borrow.
Sometimes Google closes the door.
Sometimes it appears to leave the door open, put the kettle on and offer everyone a biscuit.
What I’d Track
For every authority-hosted page, record:
- Time taken to index
- Highest ranking reached
- Traffic or affiliate clicks
- How long the page survives
- Whether it drops before or after removal
- Which platforms keep working after an update
Reddit, Steemit and PDF pages should be treated as separate cohorts. Their survival patterns and ranking behaviour are not the same.
That will tell us whether we have a repeatable channel or merely a brief window before somebody notices the curtains are missing.
My Priority This Week
Test recovery before testing scale.
A publishing agent that produces 100 articles is useful. One that produces 73, falls over, repeats the research bill and then announces “Task completed successfully” is less impressive.
Start with five products, interrupt the workflow deliberately and confirm that every stage remains traceable.
Meanwhile, keep logging those parasite pages …
Boring engineering underneath.
A bit of mischief on top.
That sounds about right 
Onwards!
- Rohan