Anthropic Just Banned Fake Sources Built to Sway AI Answers

The Big Picture

Google's September spam update finished on 8 October after nearly 14 days, and the same stretch of guidance (fact-check AI content, main content, fake authors) points to a larger core update before the holidays. At the same time, the week's strongest signals about AI visibility were about retrieval and trust, not scores: Mueller said AI crawlers fetch sitemaps and RSS feeds, Anthropic's new usage policy bans building networks of fake "independent" sources to sway AI answers, and Google appears to be tightening its stance on automated SERP scraping, which is the same kind of access visibility trackers depend on.

Google AI Mode & AI Overviews

No major new feature this week, but the plumbing is moving

There are no new AI Mode or AI Overviews events for 5 to 12 October. The branded-query jump reported last week (AI Overviews on over 80% of tracked brand searches from 26 September) has no follow-up data yet, and the effect on branded clicks is still unmeasured. Google also redesigned the Reader Revenue Manager page this week and it now lists AI Mode, AI Overviews and Gemini as surfaces, a small sign of how Google frames them as one product family.

Google appears to be pushing back on automated SERP scraping

Roger Montti (SEJ) argues Google is tightening its stance on automated search scraping, citing Cloudflare data that daily requests from AI agents on its network grew more than 1,700% in a year, and noting that AI Mode answers cost more to generate than classic results because they involve multiple model calls. SERPrecon's Ryan Jones says rank tracking is becoming hard to sustain. Treat the motive as the author's reading, not Google's statement. The article does not address AI visibility tools directly; the inference that tools querying AI interfaces face similar pressure is mine and unverified.

SEOmonitor separately reported that many rank trackers showed sudden ranking changes from 26 September that real users did not see, which it attributes to incomplete results returned to automated checkers. A tracker chart can show a movement that no buyer experienced.

Search Console reporting is lagging

Search Engine Roundtable reported on 11 October that the Search Console Performance report was more than 64 hours behind, which it links to the recent spam update. Do not read last week's numbers as final.

Key stat: Cloudflare reports daily AI agent requests on its network up more than 1,700% year on year, with over half of internet traffic now non-human.

• Annotate the Search Console delay in any weekly reporting before drawing conclusions.
• Search your brand name plus "pricing", "reviews" and "alternatives" and read what the AI Overview says. This is a yes-or-no accuracy check.
• Treat third-party rank charts for 26 September onward with caution and confirm in Search Console.

Sources: Google Is Fighting SERP Tracking. AI Agents May Make It Worse (SEJ), Google September 2026 spam update done rolling out (SEJ), Search Engine Roundtable

GEO / AEO / AIO

Anthropic bans manufactured "independent" sources aimed at AI answers

Anthropic published an updated usage policy on 8 October, effective 12 November. A new section, "Do Not Engage in Deceptive Campaigns or Artificial Activity," bars using Claude to manipulate the sources AI systems and search engines draw from, with the example of a network of sites posing as unaffiliated outlets that back the same claims. It also bars fake personas, outlets and reviews, and selling those activities as a service. Anthropic's September threat report described about 70 fake local news sites tied to a France-based ad agency, with at least 8,913 articles in about 20 languages, produced with Claude. The report does not say whether the network ever earned rankings or AI citations. Google's spam policies already treat attempts to manipulate generative AI responses in Search as spam. For B2B, the line to watch is any agency offering "citation networks" or pseudo-independent comparison sites.

GPT-6 and "Intelligent UI" in ChatGPT

OpenAI began rolling out GPT-6 in ChatGPT from 7 October, with answers that mix text with charts, forms, buttons and small in-chat tools, and OpenAI says GPT-6 decides better when to search. The announcement does not explain how sources appear inside these interactive layouts, and the speed claim (answers start 44% sooner on web-search questions) is OpenAI's own. SEJ's author argues simple calculator and converter pages may lose traffic when the tool can be built in chat. There is no data yet on citation rates or clicks.

Planning when AI will not attribute

Kevin Indig and Amanda Johnson (Search Engine Land, 8 October) argue for planning around pre-visit influence. Their cited data: US zero-click Google searches rose from about 60% in 2024 to about 68% in early 2026 (SparkToro and Similarweb); in one Similarweb clickstream study, direct visits rose after AI recommended a brand while search visits fell about 15%; AirOps found third-party signals drive about 85% of brand visibility in LLMs; and about 91% of AI citations appear in only one of ChatGPT, Perplexity or AI Overviews. Their advice: fund off-site work (digital PR, community, proprietary research) as its own budget line, and measure exposure, behaviour and business outcomes together. Note the disclosures: Search Engine Land is owned by Semrush, Indig sells his own planning products, and much of the data comes from SparkToro and Similarweb.

Where ViMi's position stands this week

ViMi's position is that AI ranking tools are not effective, because LLM output is probabilistic. An AI visibility score is a draw from a distribution rather than a measurement of a stable quantity, so averaging more runs narrows the interval around an average that still describes no individual buyer's experience, and it cannot support the before-and-after claims those dashboards are sold on.

This week's material fits that in three ways. The 91% single-platform citation figure says a brand's "visibility" in one engine tells you little about another, so any blended score averages unlike things. The SEOmonitor finding shows tools can report movement that users never saw. And Anthropic's policy and Google's spam rules mean the inputs a score rewards (citations) are being actively gamed, so a rising number may reflect manipulation by someone, not your work.

Contrary evidence deserves a fair statement. The Similarweb finding that direct visits rose after AI recommended a brand, and the claim that AI influence can be under attributed by up to 10x (Graphite), suggest AI answers do shape buyer behaviour, so ignoring AI exposure is also a mistake. That supports caring about what the engines say, not trusting a score of how often they say it. The data is also correlational and comes from vendor-adjacent sources. And Indig's three-layer framing includes "exposure (visibility, mentions)" as a measure.

Key stat: About 91% of AI citations appear in only one of ChatGPT, Perplexity or AI Overviews (SparkToro, via Search Engine Land).

• Replace any "visibility rate" in reporting with checks that have answers: can engines retrieve your key pages, are the claims they make about you accurate, and did a real buyer conversation change.
• Ask any agency selling "citation building" whether the sources are genuinely independent. From 12 November, Claude-assisted fake-source networks breach Anthropic's policy.
• Add a line to your sales call notes or intake form: "Did you use ChatGPT, Gemini or Perplexity before contacting us, and what did it say?"

Sources: Claude's New Rules Target Fake Sources Built To Sway AI Answers (SEJ), ChatGPT Gets GPT-6 And Intelligent UI (SEJ), How to plan 2027 when traffic is down and AI won't attribute (Search Engine Land), Google Is Fighting SERP Tracking (SEJ)

Traditional SEO & Organic Search

September spam update complete after 13 days 16 hours

Google confirmed the rollout finished on 8 October. It used nearly all of the two weeks it warned about, versus 19.5 hours (March), about 2 days (June) and about 2.7 days (August) for earlier 2026 spam updates. Google has not said what it targeted. Diib's early read of 32,569 sites found the median share of visits from unpaid Google search fell from 34% to 31% between 24 to 26 September and the same days a week earlier, and sites losing half or more of those visits doubled from 5% to 10%. Diib calls it an early read, not a verdict, and cannot tell which sites were flagged.

A core update before the holidays looks likely (speculation)

Roger Montti (SEJ) expects a significant core update before year end, citing Google's pattern of one or two updates around the US holiday season, six core updates since 2021 preceded or accompanied by a spam update, and recent guidance changes on main content, fake authors and AI content. Lily Ray and others have pointed the same way. This is inference, not announcement.

How long things take

A recap of Gary Illyes' 2 October session at Search Central Live Barcelona (not an official Google publication, and "typical" is undefined) gives these ranges: new URL discovery about 20 hours typical but potentially weeks or never; sitemap processing about 24 hours up to 14 days; title and snippet changes 1 to 2 days up to months; manual action removal 1 to 2 weeks; core update recovery 3 to 6 months, sometimes until the next core update.

Other items

Google disclosed a UGC Fresh Data Program (platforms with high-volume public forum or social content can apply, response in about 6 to 8 weeks, no guarantee of appearing in Search); most B2B brands will not qualify unless they run a large public community. Search Console now lets you filter by multiple countries. John Mueller and Martin Splitt said HTML sitemaps can still aid discovery and navigation but do not replace XML sitemaps.

Key stat: Diib found sites losing half or more of their Google organic visits doubled from 5% to 10% across the first days of the spam update (32,569 sites, early read).

• Compare Search Console performance across the fixed 24 September to 8 October window, and check impressions and clicks, not rank trackers.
• If you lost traffic, audit thin, templated or scaled pages and fake or unverifiable author profiles. Expect months, not weeks, for recovery.
• Avoid large site changes right before a likely core update unless they fix a known problem.

Sources: Google Completes September 2026 Spam Update After Nearly 14 Days (SEJ), Google September 2026 spam update done rolling out (Search Engine Land), What To Expect From Google's Next Search Ranking Update (SEJ), Google Shows How Long Crawling, Indexing & Recovery Can Take (SEJ), Google Search disclosed its UGC Fresh Data Program (Search Engine Land), Google Explains How To Use HTML Sitemaps For SEO (SEJ)

New Protocols & Standards

Mueller: AI crawlers request sitemaps and RSS, and llms.txt is not a replacement

On the 1 October Search Off the Record podcast, John Mueller said AI training crawlers usually offer no way to submit a sitemap, so sites that want AI systems to find content should use the default sitemap.xml name or an RSS feed, which is easier to discover because it is usually linked in the page head. He has seen AI crawlers request his sitemap and RSS files in his logs, though he did not name them or know whether operators document this. On llms.txt he said "the hope is bigger than the reality." He also explained that a valid sitemap can show "Couldn't fetch" because of server load or low crawl demand, which is often tied to perceived site quality.

Anthropic usage policy: search and AI answer manipulation named explicitly

The policy change described above is the other standards-type development of the week. It is a platform rule, not a technical protocol, but it is the first major model provider policy I have seen that names search engines and AI answers as sources that must not be manipulated.

SynthID and watermarking

Google opened its SynthID watermark detector to the public, and OpenAI said it will watermark ChatGPT text in the EU with an API opt-in. Neither affects ranking or citation today, but provenance signals are becoming part of the AI content conversation.

Key stat: Mueller on llms.txt: "I think the hope is bigger than the reality."

• Confirm sitemap.xml sits at the default location and is referenced in robots.txt, and expose a blog or news RSS feed linked in your HTML head.
• Check server logs for AI crawler requests to your sitemap and feed files.
• Investigate any "Couldn't fetch" sitemap errors for server capacity first, then crawl demand and quality.

Sources: Google's Mueller Says AI Crawlers Access Sitemaps & RSS In His Logs (SEJ), Google Explains The "Couldn't Fetch" Sitemap Errors (SEJ), Claude's New Rules Target Fake Sources (SEJ), Google Opens Its SynthID Watermark Detector To The Public (SEJ)

Google Ads Updates

Preview AI Max generated ad copy before opting in

Published 8 October (Search Engine Land) and confirmed in Google's help documentation on 9 October: advertisers can preview up to 10 sample headlines and descriptions that Google generates from the landing page, in about 30 seconds, before turning on AI Max text customization. Previews appear in the Asset optimization section for new and existing Search campaigns. Google says these are examples, not necessarily what will be served. For B2B, this is a chance to check brand voice, positioning and claims (compliance, pricing language, certifications) before automation writes ads for you.

AI summaries in experiment reporting

Google Ads added an AI-generated panel to the experiments Summary tab covering goals, findings and next steps, and clarified how results apply: migration experiments move all traffic to Performance Max, optimization experiments apply tested settings permanently, and custom experiments leave the choice to you. With low B2B conversion volumes, treat the AI summary as a starting point and judge significance yourself.

Elsewhere

Google introduced commerce audience sharing for retail media (low B2B relevance). Last week's items still apply: automated promotions are on by default from 12 October for eligible Search and Performance Max campaigns, and business names that differ from the destination domain become eligible from 19 October. OpenAI is also testing visual ads in ChatGPT image generation.

Key stat: Google generates up to 10 sample headlines and descriptions per AI Max text customization preview, in about 30 seconds.

• Run the AI Max preview on your top B2B landing pages and list any claim you would not approve in a human-written ad.
• If you run experiments, record the significance threshold and minimum conversions before you read the AI summary.
• Check whether automated promotions apply to your account today.

Sources: Google Ads lets advertisers preview AI Max-generated ad copy before opting in (Search Engine Land), Google Ads adds AI-powered insights to experiment reporting (Search Engine Land), Search Engine Roundtable, OpenAI To Test Visual Ads In ChatGPT Image Generation (SEJ)

Suggested Actions This Week

• Check retrieval, not rates. Confirm your top 20 commercial pages are crawlable, sitemap.xml is at the default path and in robots.txt, and an RSS feed is linked in your HTML head. Look for AI crawler hits in server logs.
• Read what AI says about you. Search your brand plus "pricing", "reviews" and "alternatives" in Google and the main AI engines. Log inaccurate claims and trace each to a source page you can fix.
• Challenge visibility scores. If a tool or agency reports an AI visibility rate, ask what decision it changed and how they separate your work from platform changes. Ask instead for buyer evidence: sales call notes and "how did you hear about us" answers.
• Vet any "citation building" offer. Ask whether sources are genuinely independent. Network-style tactics are spam under Google's policy and, from 12 November, a breach of Anthropic's.
• Audit spam-update exposure. Use the 24 September to 8 October window in Search Console; review thin and scaled pages and any unverifiable author profiles.
• Test AI Max previews on your best landing pages before enabling text customization.
• Plan for a possible core update before the holidays: avoid big untested changes and keep annotated baselines.