AI watermarks are live and impacting rankings, Meta is building a search index, and Google is hiding the search button

In this issue: Meta’s crawler traffic exploded this month in a pattern that strongly resembles a company building its own search index. Google tested a home page for logged-out users that replaces the classic search button with AI shortcuts, and the Generative AI Performance report in Search Console has been rolled out globally. Plus: three tips worth reading this week.

This week’s activity: The Generative AI report startup audit—how to correctly interpret your new Search Console data from day one.

AI Watermarks Are Live, and Watermarked Content Is Already Ranking Five Spots Lower

As of August 2nd, all major AI platforms have to label their AI-generated content under the EU AI Act’s transparency code. This week, Anthropic confirmed that Claude is now embedding an invisible watermark into generated text, TechCrunch reported. The technology uses Google DeepMind’s SynthID-Text method in conjunction with C2PA metadata for files. The watermark accompanies the text when it’s copied and pasted, and Anthropic says the technology will be rolled out to older models as well. Google already watermarks Gemini’s output in a similar way, and OpenAI appends C2PA data to ChatGPT images.

Does it affect SEO? Well… First Page Sage conducted a controlled study in which they published 1,682 pieces of content on 139 sites across four B2B industries after August 2nd. They then tracked how watermarked content performed compared to content produced without AI. Non-watermarked content ranked 6th on average, while watermarked content ranked 11th on average. Non-watermarked content was also cited almost twice as often in AI responses: 12% compared to 7%. The study can’t prove that the watermark is causing the drop, as the watermarked and non-watermarked texts were not otherwise checked for identicality. But the trend was the same across all four industries in the sample.

Why this matters: The industry has been debating for two years whether Google can identify AI content. Now there’s a watermark that does just that, active on all major platforms, and early numbers suggest it’s already showing up in rankings and AI citations.

What you can do:

Tools are emerging that claim to remove watermarks, but they haven’t been confirmed to work yet. They may also violate platform terms and conditions and, depending on the jurisdiction, transparency or disclosure laws. The requirement is not optional for platforms, and trends indicate that it won’t be optional for publishers either.

If you’re using AI tools for drafts: set aside time for proper editing. The study didn’t isolate what the watermarked texts looked like before publication, but the sheer size of the ranking difference is reason enough to stop treating a first AI draft as ready for publication.

Check your own rankings for AI-assisted content published after August 2. If you see the same pattern, it’s worth raising in conversations with clients now—before it becomes an accepted truth instead of an open question.

In short: Watermarking has gone from being a compliance box to a measurable ranking signal. It changes how much editing an AI draft needs before it is published.

Meta's Crawler Traffic Just Jumped From 2% to 38% of All AI Crawler Requests

Pieter Levels wrote on X this week that Meta employees, in his words, have said that Meta is “ALLEGEDLY building its own Google search engine” – allowing Meta AI to search the web without going through Google while also giving Google free visibility into what Meta’s users are searching for, according to a report by the Search Engine Roundtable. The claim is not just an unsupported rumor. Promptwatch crawler logs show that Meta-WebIndexer’s share of all tracked AI crawler traffic rose from about 2% in mid-July to 37.8% on August 9 – about 17x higher in less than a month from a single crawler.

Facebook has wanted its own search engine for over a decade. The company partnered with Bing for a long time, but then ended that partnership. Now the same ambition is back, with AI training as an additional reason. If Meta is going to build the answers, it also wants to own the index that feeds them.

Why this matters: Crawling at this scale and pace is exactly what you would expect from a company building a web index from the ground up. If Meta launches a real search product based on this crawl, whether your content was indexed during the build could determine whether it can be cited later.

What you can do:

• Check your server logs and robots.txt for Meta-WebIndexer, Meta-ExternalAgent, Meta-ExternalFetcher, and FacebookBot. Getting crawled now is what could make citation possible later, if a Meta search product actually launches.
• Start treating Meta as an emerging AI search space, not just a social platform. The same citation logic you already apply for ChatGPT and Perplexity applies here.
• Keep an eye on your crawl budget and server load if you see a spike from these user agents. Sites have already been hit with load alerts due to the intensity of this crawl.

In short: No one has yet confirmed that Meta is building a search engine. However, the crawler logs are confirmed – and they look exactly like a company is building one.

Google Tests Home Page Without a Search Button

Search Engine Watch discovered a Google home page for logged-out users where the classic Google Search button disappears entirely. Instead, there are three shortcuts: Create images, Ask about files, and Brainstorm, as well as an AI Mode button right in the search bar. I’m Feeling Lucky is the only survivor of the old layout, which seems fitting. Google confirmed the buttons two days after the discovery, but hasn’t said how many people will see the version or whether it will be permanent.

Ask about files, Brainstorm, and AI Mode worked without logging in when Search Engine Watch tested them in Chrome, Edge, and Perplexity’s Comet browser. Create images was the exception: Google still wants you to log in there. When you log back in, the three AI shortcuts disappear and you’re back to the home page you’ve known for two decades.

Why this matters: A home page test is still a change in visibility, and visibility changes affect behavior before any algorithm even has a chance to do so. Show a logged-out visitor a Brainstorm button instead of a search button, and a certain percentage will click there first – especially on exploratory, uncertain searches that your top-of-funnel content is built to capture.

What you can do:

• It's just a test currently, so there are no actions to take. Just take note of it and wait for a broader rollout before changing your strategy.
• If you have visibility into your own traffic mix by search intent, track exploratory and comparative searches in particular. They’re most likely to be redirected to a chat-based tool before the user even reaches a results page.
• Review whether your top-of-funnel content still deserves a click on its own merits, regardless of ranking position. That’s the content that’s most at risk if AI shortcuts continue to move earlier in the search journey.

In short: The search bar didn’t move. The button next to it did, at least in this version. See where Google continues to test before you react.

Google Search Console AI Report Rolled out Globally

Google’s Generative AI Performance report in Search Console, which launched in June for a limited number of properties, rolled out globally this week. Search Engine Watch reported that Search Console now displays a pop-up that takes site owners directly to the report, which feels like Google’s attempt to get more people to discover something that most people haven’t seen yet.

The report shows impressions in AI Overviews, AI Mode, and generative AI features in Discover. It’s filterable by URL, country, device, and time period. It doesn’t yet include click data or the actual prompts that triggered an impression, and Google hasn’t promised a timeline for when any of that will be added.

Why this matters: This is the first real, built-in metric Google has given you for how your content is performing in its AI features, although it’s still limited. Previously, you had to piece together the picture with guesswork and whatever third-party tools you were relying on that particular week. Now the data is in one place, even though impressions are still the only metric offered.

What you can do:

• Check to see if the report has appeared in your own Search Console properties and in the client accounts you manage. The rollout is described as global, but it could still happen incrementally per property.
• Set a benchmark now, even if the data is just impressions. When Google eventually adds clicks, you’ll want a before picture to compare against.
• Compare spikes in AI impressions with your regular Performance report. A page that gets more AI impressions while organic clicks decline tells you something specific about where the traffic is going instead.

In short: Google continues to add ways to measure AI visibility—on its own timeline, never yours. Get to know what’s out there now before the click data arrives and everyone else tries to catch up.

Worth Reading This Week

I just published a guide on crawling and AI bots. If you’ve ever wondered how to stop AI from training on your content for free while still being cited in AI search, this is for you.

Duane Forrester on why no one will pay for the AI ​​visibility data everyone says they need (Duane Forrester Decodes). Forrester surveyed 163 practitioners and found that price isn’t the issue: participants rated the underlying AI visibility data a 4.20 out of 5, but the willingness to pay for a platform to get the data a 3.19. The methodological transparency everyone is demanding is also the one thing no venture-funded vendor can offer without undermining their own product. Forrester has built such a platform himself, says it openly, and yet doesn’t shy away from the results.

Mark Williams-Cook on why the evidence everyone cites for llms.txt could also prove that a cat can rank (his Substack). Mark invented a fake specification called cats.txt, in which his office cats were officially given job titles and designated spin levels. Then he tested it against the exact four “proofs” the industry uses to justify llms.txt: bots crawled the file, Google indexed it, an LLM provided a detail that isn’t found anywhere else, and ChatGPT recommended using it. It passed all four. None of that proves that llms.txt does anything—and that’s the point: half of the GEO tactics billed to clients are based on the same level of proof that a fictional cat just passed.

Sarah Fitzpatrick on why your brand isn’t showing up in ChatGPT. Fitzpatrick’s main point is that most invisibility is because no one has yet properly checked, long before there’s any hidden penalty involved. A good reality check about AI visibility has started to feel like an unsolvable mystery to you or a customer: her point is that it's almost always diagnosable when someone takes the time to investigate it.

Free For Subscribers

The Generative AI Reprt Starter Audit Steps + Prompt