Google's 3rd Spam Update Roll-out, ChatGPT breaks up with Reddit, and local businesses get a boost

In this issue, get the latest news on the latest spam update, what happened with reddit, and the news for local business visibility + a free AI Disclosure Building skill and a prompt to help optimize your content for AI visibility

Google Spam update #3 Rolls Out:

Google just rolled out their latest 2026 spam update on Aug 18. This update is going to be applied globally across all languages. This is the third spam update shipped this year, showing that Google is really cracking down hard on Spam (and possible AI content maybe, given this interesting study by .. )While the update was officially announced to roll out Aug 18, many SEOs and rank trackers detected volatility spikes between Aug 1st and 6 - which, of course, Google never confirmed as an update. However that means that your tracking may be messy and you should look at both events to assess overall impact. 

Spam updates like this one target manipulative tactics, but Google’s own documentation does admit they make mistakes and classify good players as spammy even if they’re not engaging in these tactics. In these cases, you have to wait for the next update and hope Google catches it.

Here’s what to do:

• Compare Google Search Console impressions and clicks before and after the update across different sets of pages. It's often helpful to isolate key areas of your site as overall traffic numbers might mask changes.
• If your drop points to Aug 1-6, that points to an earlier, unconfirmed update. It may not be a spam update, so don’t misdiagnose the cause.

Bottom line:

Since this update targets manipulative web practices, legitimate sites shouldn’t be impacted or need to make any changes. If you were affected (fairly or unfairly) you should cross reference your site with Google’s official documentation on the specific tactics it's looking for to uncover if anything you’re doing could have been perceived as manipulative, and correct it ASAP. 

Reddit has gone missing from ChatGPT, have they broken up?

New data from Promptwatch shows that over the last week Reddit’s share of ChatGPT citations fell from an average of 3.83% to an average of 0.52 ( that’s a huge dip of 86% in just over 4 days!) Promptwatch also noticed an earlier (smaller) decline around Aug 8th, which aligns with when ChatGPT said they were changing their query fan-out behavior.  

Honestly? If Reddit made up less citations Id be pretty happy if that means less spam on my once-beloved site, but that’s a whole other post.Here’s what to do:

• Pull your own AI citation data by platform - make sure to look at the data individually rather than a rolled up/blended number 
• Set a recurring weekly check on your top referring AI sources, which will hep you spot shifts like this early. 
• Don’t build a whole strategy around getting cited in a single site. Reddit was the popular girl for a bit, but here’s the proof that things can shift and change on a whim.
• Watch for any comments by OpenAI before assuming this is permanent.

Bottom Line:

This shown us that things like losing 86% citation share in four days without any announcement do happen, and it’s important to always look at your citations sources and stay on top of news and trends so you are reporting the correct numbers. 

On the subject of citations, let’s talk about Gemini’s Local Citations

A new study from Steady Demand, found that business websites are currently making up nearly 60% of Gemini’s local citations. This is more than directories, review sites and forums combined and is honestly a really nice change for small business owners who can’t afford all of the crazy fees and PR needed to be listed everywhere. The second top source was Reedit at close to 14%, which was ahead of all the local service directory categories like Angi, Thumbtack and HomeAdvisor. 

However, we want to look at what Ben Fisher (Co-Founder of Steady Demand) calls “Grounding Drift” which is not as peaceful as it sounds. If you repeat the identical query multiple times, Gemini returns overlapping cited sources only about 40% of the time, and the same business only about 7% of the time. 

A control test on Google’s local pack was shown to return the same top-listings about 90% of the time. So to see if the same behavior was found in AI, they also ran those queries through ChatGPT, and found that the two rarely agreed and cited the same domains only 8% of the time, and recommended the same top business just about 4% of the time. Gemini skews toward business websites and ChatGPT leans more heavily toward Reddit (though maybe we will have to wait and see if the breakup mentioned above is more permanent)

Here’s what to do:

• Treat your local visibility efforts as platform specific as both Gemini and ChatGPT pull from different sources, and you want to be listed in both.
• Make sure to keep your business website optimized and maintained for Gemini specifically as it's the highest-weighted search type.
• Don’t treat a single ranking as your baseline. Use it as a trend, the same thing I would recommend when checking rankings, you can’t rely on 1-2 cases to show your true performance, check it at regular intervals and treat each like a snapshot to see how it's trending overall. 
• If Reddit threads are popular for your category, it’s worth considering how you can be more visible as a brand on Reddit,

Bottom Line: At 7% consistency with multiple queries, this makes your ranking more of a coinflip, so watch trends over time

Claude’s new text watermarks explained

If you have been under a rock and have missed the whole hullabaloo on Anthropic’s watermarks, we now have more data on how they will work. Anthropic published a blog post that walks through the mechanics of the watermark it will be rolling out to Claude.

As a reminder, this new watermark is a requirement under the EU AI Transparency Act, so while we are focusing in on Claude right now, other LLMs will be following suite - especially (as Anthropic explained) this watermarking is utilizing Google DeepMind SynthID Text.  Other major models have also agreed to take part in the same Code of Practice and will be shipping out their own versions of this watermarking feature. 

So how does it work? Text is not made up of pixels, so a watermark cannot be embedded directly in the text. Instead Claude will create a series of invisible choice word patterns in the content output, using a key. Then a detection key can analyze that pattern to see if it was used. Anthropic says that this is not impacting the quality or token-usage of content created. 

This means that heavy edits may strip out most of the key, while if you make only small adjustments here and there it will still contain Claude’s full signature. Even if you wrote it and passed it through Claude, even minor edits it makes will carry that signature and make it more likely to be identified as AI written. 

It honestly doesn’t sound incredibly clear at this point exactly how high the % of a present key needs to be discovered to label the entire piece as AI generated, and we cannot even check this ourselves at the moment as Anthropic has not made the key detection system available yet. It sounds like this may come closer to 2027. 

It is also unclear what the actual impact of this will be at the moment. I feel like if this will have any impact on things like SEO, areas like translated content should be treated a bit differently since if you use AI to simply translate articles pre-written by a human it will contain a key.

Here’s what to do:

• Map out all of the areas where AI is touching your content. Do you use it for translating content, proofreading, etc? It’s good to understand the risk of all areas, because even a watermark surviving light edits is not a reliable defense against AI detection, and we currently don’t know impact other than labeling. 
• If you are publishing content that also serves EU readers, make sure to check that your own toolset and vendors are in compliance with the EU’s full disclosure requirements. 
• Don’t get AI style detection tools mixed up with watermark detection tools, as these look at very different things. And at this point nobody can reliably do a check on a watermark since Claude has not even released its key checker.
• It's a good idea to add an AI disclosure policy to your site now, just to have one available.

Bottom Line:

Watermarking gives a definitive answer on whether or not content was written by AI, but leaves a lot of room to interpretation as to whether Claude just breathed on it or took a pen and rewrote the entire thing. 

Worth reading this week:

Duane Forrester wrote about how the keyword universe is getting smaller than we thought. Duane discusses how AI citation tools are measuring things that traditional rank tracking tools do not, and it's all about brand eligibility. His data shows the 3 major LLMs only agree on a brand around 41% of the time, but they did agree on who was eligible to show up in that shortlist of citations 91% of the time.

Tony Uphoff disusses how people complaining that AI killed their reach across Linkedin and Substack may be angry at the wrong target. Zoom out and organic click are down 41% since AI overviews and LinkedIn reach is down about 47% after their feed ranking was rebuilt. But these structural disruptions don't neccessarily represent real decline.

Vinicius Stanula on JavaScript links can make your pages invisible to AI search. Stanula ran his own 41-day crawler experiment, deliberately hiding half the internal links on a 2,400-page directory site behind JavaScript. Every AI crawler he tracked, including GPTBot, ClaudeBot, and PerplexityBot, found zero of the JavaScript-linked pages. Even Googlebot, the one crawler that did render the JavaScript, only reached 2% of them. When he converted the links back to plain HTML, GPTBot swept the entire blocked section within three hours. Google still hadn't fully caught up by day 41. The assumption that AI crawlers render JavaScript fine falls apart the moment you test it against real navigation.

For Subscribers:

With the new disclosure requirements for AI use, I have made you a free Dislosure Policy builder that you can upload to Claude or ChatGPT. This builder will interview you about you (or your team's) workflow, and write the policy for you based on what you need to disclose under the new laws and regulations. And yes, I do realize the irony of having AI help build your AI policy disclosure, but it works great - try it out by downloading the MD file below and uploading that to your fav AI platform.

I also have a great prompt for that helps you optimize your content for AI through a comparative analysis, check it out below