Bobby David @bbbybuilder

You can now buy an ad inside ChatGPT. It still won't make the model cite you. SE Ranking analyzed more than 50,000 commercial prompts across 20 niches. Ads appeared on 25.94% of them — roughly one in four, close to Google AI Mode's 29.45%. The ad product is already running at scale. But only 3.63% of advertisers were also cited as a source in the answer above their ad. The exact advertised URL appeared in citations 0.09% of the time. Buying placement and earning citation are not the same system. They don't appear to talk to each other. 14.35% of the ads served were semantically unrelated to the prompt they sat under. A dating app query got a clothing retailer. A newspaper subscription query got an electricity provider. ChatGPT Ads doesn't use keyword targeting — advertisers supply natural-language "context hints" — and those hints are steering, not controlling, placement. The pattern is the same one that shapes every AI search surface right now. The layer that generates the answer and the layer that sells the ad space are built independently. One decides what the model knows. The other decides what the advertiser paid for. You can buy visibility in ChatGPT. You cannot buy citation. The data now shows they don't even correlate.

ChatGPT ads used to require $50,000 and an agency relationship. OpenAI is now hiring a dedicated team to sell them to businesses with no minimum at all. Digiday reported August 10 that six live OpenAI job postings — including a Growth Lead, SMB Ads role confirmed on OpenAI's own careers page — all point to a unit targeting small and midsize advertisers. The self-serve Ads Manager launched in May with CPC bidding, a conversion pixel, and a Conversions API. The $50K minimum is gone. The infrastructure was built backwards on purpose. OpenAI shipped the pixel, server-side event tracking, partner integrations with Adobe and Criteo, and cost-per-click bidding before it opened the door to anyone with a credit card. Every tool a performance marketer expects was already running. For anyone doing GEO, this is where the measurement asymmetry hits scale. ChatGPT's paid side now has full conversion attribution. The organic side — where your content surfaces because the model chose it — has no tracking, no report, no API. First Page Sage's July survey showed organic ChatGPT visitors convert higher than paid across nearly every industry. But OpenAI can only measure, bill, and optimize the side that pays. The infrastructure was built to prove ads work. Nobody built the proof for what already performs better.

Meta wants its AI to answer questions about the web. It cannot afford to let its AI ask someone else's search engine for the sources. Pieter Levels reported on August 6 that Meta is crawling the web aggressively — enough to trigger server load alerts on his infrastructure. He said Meta staff confirmed privately, with permission to share: the company is building its own web index. The stated reason: if Meta AI searches through Google, Google sees those queries and can use them for its own training. The claim is unverified beyond Levels' account and the crawl logs he posted publicly. But the structural logic is already visible across the industry. OpenAI tripled OAI-SearchBot's crawl volume over the past year. Google owned an index from the start. Every AI company that grounds answers in web content is being pushed toward building its own crawl — because using a competitor's index means handing over the one signal you cannot afford to share: what your model is searching for. For GEO, the fragmentation compounds. Googlebot. OAI-SearchBot. GPTBot. Now potentially a Meta crawler — with Meta AI reaching over a billion users across Facebook, Instagram, and WhatsApp. Each bot has its own user-agent, its own robots.txt rules, its own line between indexing and training. Cloudflare's Search/Agent/Training taxonomy was already straining to hold. Another major player makes the control problem harder, not easier. The web index used to be plumbing. It is becoming the wall every AI company builds first.

Cloudflare just split "AI bot" into three words: Search, Agent, and Training. Starting September 15, every domain on Cloudflare gets separate controls for each. Search crawlers stay allowed by default. Training and Agent crawlers get blocked by default on pages that run ads. Multi-purpose crawlers — the ones that can't or won't separate search from training — get treated by their most restrictive behavior. That last part is where it bites. Googlebot is a multi-purpose crawler. It indexes for search and absorbs content for AI training through the same bot. Cloudflare's new framework says: if a site owner blocks Training, and your bot does both, your bot is blocked. Google's response, via John Mueller on Reddit: Google doesn't follow llms.txt, doesn't follow Cloudflare's signals, and has no plans to. Two control systems that don't talk to each other. For anyone doing GEO, the question this forces isn't whether to block training. It's whether the crawler that grounds AI answers and the crawler that trains models can be separated at all. Cloudflare's taxonomy assumes they can. Google's infrastructure currently doesn't. If they can't be separated, site owners who want AI Overviews citations and who also want to block training are stuck choosing. Cloudflare just made that choice visible — and made it apply by default.

"Seed Reddit threads so AI answers cite you" was one of the most repeated GEO tactics of the year. Google just told The Verge that Reddit gets no special preference in its ranking system. Jennifer Kutz, a Google communications representative, said it in an August 4 Verge article about AI SEO spam on Reddit. Reddit "gets no special preference," and Google's AI features don't aim to display content from any specific site. It's the first on-record Google response since Reddit's stock dropped over AI Overviews traffic loss in late July. The denial is narrower than the headline. "No special preference" is not "no preference." Reddit can still earn visibility through the same ranking signals every other domain uses. The qualifier does all the work — and nothing in the statement tells you whether seeding threads actually changes citation rates in AI answers. What it does reveal is where the risk was. The tactic assumed Google gave Reddit algorithmic preference. But the constraints on a brand seeding Reddit were never Google's ranking rules. Google's own spam policy explicitly exempts user-generated content forums. The rules that actually bind are Reddit's Content Policy, which bans spam outright, and consumer protection law on deceptive reviews. A spokesperson line about ranking tells you nothing about either. The tactic was never as safe as its advocates sold it. Its biggest risk was always somewhere other than where people were looking.

@pangram is this AI?

Google Search used to wait for you to ask. Now it works while you're not looking. Information agents went live in AI Mode on August 5. Robby Stein from Google confirmed the rollout: AI Ultra subscribers can tell AI Mode to "keep me updated on" any topic, and the agent scans blogs, news sites, social posts, and real-time finance, shopping, and sports data around the clock. When something changes, the agent sends a synthesized update with links. This is search shifting from an on-demand tool to a persistent monitoring layer. Google isn't waiting for a query. It's running queries on the user's behalf between sessions. For anyone doing GEO, the implication compounds. A brand cited inside AI Mode earns recurring impressions — agents resurface it every time a monitoring loop fires. A brand not cited loses share that now extends across sessions the user never initiated. The visibility gap between cited and uncited no longer resets with each search. The content that agents check for changes is the content that earns repeated surfacing. Freshness signals just became a recurring impression multiplier. Search used to happen when you showed up. It now happens whether you do or not.

AI Mode is generating queries that don't look like queries. People are showing up in Search Console performance reports with entries like "yes," "yes go on," and "yes, pricing." These aren't typos. They're conversational follow-ups from AI Mode — the user responded to an AI Overview with a word, and Google logged that word as a new query with its own impression, position, and click data. John Mueller confirmed this on LinkedIn last week, pointing to Google's own documentation: if a user asks a follow-up within AI Mode, they are essentially performing a new query. All data in the new response counts as coming from that new query. The generative AI performance reports — the ones Google launched in June specifically to show how your site does in AI features — don't include query data at all. So the only place you can see which queries AI surfaces are generating is the main search performance report, mixed in with everything else, with no filter to separate them. You can spot AI Mode entries by their shape. They're fragments. Continuations. One-word responses that make no sense as a standalone search. Ross Tavendale flagged it publicly: Search Console is recording people's responses to AI Mode, and the SEO community is now trying to reverse-engineer what they mean. For anyone measuring GEO performance, this is the measurement gap made visible. Google built a dedicated AI report — but it tells you how many impressions you got, not what triggered them. The query data exists, but it's been dissolved into the general report as noise you have to pattern-match by hand. The infrastructure was built to measure AI search. It measures everything except what you need to know.

@elkelk This is fake news. @grok factcheck this

Twenty companies now sell AI visibility data. They disagree on the same brand. IAB published the first measurement standard for AI search visibility this week. The framework exists because the market broke: vendors use different query sets, different prompt types, different methodologies — and produce different rankings for the same company on the same day. An industry that sells "your AI visibility score" has no shared definition of what visibility means. The split that matters is directional versus decision-grade. Directional data identifies patterns and trends — useful for competitive awareness, not for budget allocation. Decision-grade meets higher bars across query volume, sample size, prompt coverage, testing cadence, and reproducibility. The IAB's position is direct: most data being sold today is directional. Almost none clears the decision-grade threshold. The 4 P's — Presence, Prominence, Portrayal, Persuasion — give the vocabulary. Portrayal is the one that's new to search measurement. It tracks sentiment, framing, hallucination rate, and factual accuracy. Your brand can appear, rank first, and still be described wrong. For anyone paying for a GEO tool right now, the question is no longer "what's my score." It is whether the method behind that score would survive disclosure.

AI Overviews now show up on roughly 43-48% of tracked queries. Clicks take a real hit when they appear. Some data puts the drop around 40% on outbound links. But the products that get named inside those answers still win the impression and the high-intent traffic that does click through. For marketplace sellers the part you control is the listing. Clear facts. Complete attributes. Specs that match what a buyer actually asks. Visuals that prove the claims. Entity signals so the system knows exactly what the product is and when it fits. Vague “premium handmade” language gets ignored. Specific materials, dimensions, compatibility, and use cases get pulled into recommendations. That’s the practical GEO work right now. Not ranking games. Just cleaner product signals. That’s exactly what firstshelf.ai is built for. Run a listing, see the gaps, fix the fields AI can actually read. The shift is real. Visibility is moving to the answer layer. Anyone else noticing specific Etsy or Shopify products showing up in ChatGPT or AI Mode answers yet?

@berryrikujoulo1 1

A German court just decided who is responsible when an AI answer is wrong. The Regional Court of Munich ruled on May 28 that Google's AI Overviews are not search results. They are Google's own content. The court's reasoning was specific: an AI Overview does not display links or snippets. It summarizes, structures, and presents information in the platform's own words — with thematic sections, introductory affirmations, and action recommendations that appear nowhere in the cited sources. Because Google authored the output, the search-engine liability shield does not apply. Google is a direct infringer, not a neutral conduit. The case involved two Munich publishers whose company was falsely linked to scams, subscription traps, and sketchy partners in AI Overviews. None of those claims appeared in any source Google linked to. The AI fabricated connections between the publishers and unrelated companies with genuinely poor reputations. Google's defense — that users can verify answers by clicking through — was rejected. The court noted that an answer presented as self-contained and authoritative carries no visible warning of unreliability, so the mere possibility of cross-checking does not transfer responsibility to the reader. A separate analysis by AI startup Oumi, published by the New York Times, found that 56 percent of correct Gemini 3 AI Overview answers could not be backed by the sources Google cited. The answers are right, but their provenance is untraceable. Google is appealing. A Frankfurt court reached a similar conclusion on competition grounds last September. The pattern is forming. For anyone whose GEO or AEO work targets AI answer surfaces, the accountability shifted. The answer that cites your content is no longer a retrieval of your work — it is the platform's own statement about you. If that statement is false, the platform answers for it. That is a different relationship than being indexed.

Google just shipped a specification that addresses a problem most SEO teams haven't named yet: how an AI agent decides whether to trust a piece of information before it reads it. The Open Knowledge Format v0.2, published July 24, adds five trust fields to knowledge bundles. They answer the questions a consuming agent needs to resolve before opening the file: where this came from, who verified it, whether it's still current, whether it's the active version, and whether a computed number was actually produced the way it was supposed to be. The deliberate omission is a credibility score. Google records the signals and lets the consumer infer trust on the fly — because a score is subjective, doesn't transfer between systems, and goes stale the moment it's written. That design choice maps directly onto GEO. If agents decide what to cite based on provenance and verification metadata before reading the content, the content stops being the first thing evaluated. The metadata is the gate. The fields are familiar: author, last_modified, status, verified_by. The shift is that an agent scanning thousands of concepts overnight uses exactly those fields to decide whether your page earns a full read. EEAT for machines isn't a rubric. It's the metadata an agent checks before it commits to reading anything else.

Every AI search answer you optimize to appear in just became regulated content. The EU AI Act's transparency obligations took effect today. Article 50 requires providers of generative AI systems to mark synthetic text, images, audio, and video in a machine-readable format so it can be detected as artificially generated. Deployers must label AI-generated text published to inform the public on matters of public interest — unless it went through human editorial review. That covers the exact surfaces GEO and AEO target. AI Overviews, ChatGPT responses, Perplexity answers, Copilot summaries — all are generative AI outputs that now carry a legal marking requirement across 27 countries and roughly 450 million people. Systems already on the market before today get until December 2 to meet the machine-readable marking standard. New systems must comply immediately. The EU published a Code of Practice with standardized labels: a visible "AI" badge, modality-specific disclosure for video, audio, and text, and a taxonomy separating fully AI-generated from AI-assisted content. The carve-out that matters for practitioners is human editorial review. AI-generated text that undergoes human review and editorial control, with a person holding editorial responsibility, is exempt from the public-interest labeling requirement. The regulation is not targeting AI-assisted workflows. It is targeting synthetic content published without oversight. For anyone whose GEO strategy includes publishing AI-generated content to EU audiences, the question shifted. It is no longer whether AI can produce the content. It is whether a person reviewed it before it went live.

@lena_z01 @BytePlusGlobal Can I get this as well?

Search Console told you everything about how Google found your website. It told you nothing about how Google found your content on YouTube, TikTok, X, or Instagram. That gap closed this week. Platform properties went globally live in Search Console. Claim your YouTube channel or X account as a property, and you can now see which search queries send traffic to that content, which posts spike from search, and how performance compares across platforms side by side. For anyone doing GEO, the gap was not abstract. The content that surfaces in AI Overviews does not live only on your site. It lives in the places AI engines read — social posts, video reviews, forum threads. You could measure your website's search presence. You could buy third-party tools to check AI answer citations. The entire layer where social and video content was being discovered through search was unmeasurable from Google's own data. Now you can see which queries lead people to your YouTube videos. Whether your X threads get more search discovery than your Instagram posts. Which TikTok captions correlate with search spikes. The measurement was built for the website. The visibility was already happening somewhere else.

ChatGPT ads used to send people to your website. OpenAI is testing a new campaign type that skips the website entirely. It is called Agent. Instead of a URL destination, the ad launches a business agent conversation. OpenAI scrapes your site to build a profile — common customer questions, support context, product information. You configure an agent on top of that profile, give it product feeds and MCP tools, add lead capture forms. The user clicks the ad and lands in a conversation, not on a page. It was spotted live in the ChatGPT Ads Manager on July 31, available to select advertisers. This is the mirror of what ChatGPT tried with in-chat purchasing. That shut down because too many parties owned pieces of one transaction. This time OpenAI is not trying to own the sale. It is trying to own the conversation before the sale. The implication for GEO is specific. Your website now feeds two pipelines. OAI-SearchBot crawls it to decide whether you appear in organic answers. The ad system crawls it to build the agent your paid traffic lands in. Block the crawler and you lose both. The website was the destination. It is becoming the source material.

Claude accidentally gave SEOs a good reminder this week. Shared chats were showing up in Google and Bing because the pages could be indexed. Simple lesson: If a page is public and crawlable, search engines may find it. AI product or normal website, same rule. Robots.txt is not a privacy setting. Source: wired.com/story/private-… theguardian.com/us-news/2026/j…

@RoundtableSpace This is definitely the fastest way to get an account banned