Richard Hearne

@glenngabe On a more serious note, they must be supremely confident that this won't hallucinate in any way. The legal jeopardy around news reporting by AI isn't something their legal people would baulk at.

@glenngabe You'd almost think Google was looking for a fight...

More on SER about AIOs showing up for breaking news. I included a quote from Google from May 2025 where they said "AI Overviews are purposefully not triggered for hard news queries." Well, I guess that changed... Again, a big heads-up for news publishers: seroundtable.com/aios-breaking-…

@dejanseo 6 profiles. Free, Ent paid, Pro. Base model is there, but no /optimization_guide_prediction_model_downloads/ subfolders.

Anyone sees these LoRA files for Gemini Nano in their local Chrome folders? / ├── OptGuideOnDeviceModel/ ← Base Gemini Nano (component updater) │ └── / │ └── weights.bin ← ~4GB base model │ └── / ← e.g., "Default/" └── optimization_guide_prediction_model_downloads/ └── / ← one folder per OptimizationTarget └── / ├── model.tflite ← config/model for that target └── ← LoRA weights live here # macOS find ~/Library/Application\ Support/Google/Chrome/Default/optimization_guide_prediction_model_downloads -type f # Windows dir /s "%LOCALAPPDATA%\Google\Chrome\User Data\Default\optimization_guide_prediction_model_downloads" # Linux find ~/.config/google-chrome/Default/optimization_guide_prediction_model_downloads -type f

@dejanseo @Hobo_Web @aduttya I'm thinking more about variable results for fixed queries TBH. But your point is yet another pain point to consider.

@RedCardinal @Hobo_Web @aduttya It keeps shifting if query fanout backend returns variable queries leading to variable results.

Great question by @aduttya about tackling the fuzzy probabilistic nature of AI search. Arrows point at optimization steps where five instances of the same model disagreed on what should be the result #2. The black line is the median value of five parallel searches representing model consensus. I couldn't secure top spot as Wikipedia appears to be super-glued to the top.

@Goog_Enough I wonder will Brightedge admit how daft their css is at 360px? The problem with their analysis is that they only check the serp results for the current query. If (and it's an if) Google has extended fan outs for AIOs, then cited urls may rank for a related query?

AIOs are showing up 58% more often than last year, but it’s still a toss up whether you’ll find any cited sites in the top 100 organic results. Something is clearly out of whack, but don’t expect Goog Enough to admit there’s an issue...

@DotWeekly @Sedo Do you track domains listed for sale?

Well, lets use real domain data from Zone Files to help answer this "chat" question and separate it from "search" data. 📊 Domain Saturation ~ Registered as an exact-match domain in 686 of 23,312 TLDs (2.9% saturation). ~ 443.6K domains across the zone file contain "chat" in their name. 🌐 Web Presence ~ 1,680 of the world's top 1M websites contain "chat" — 255 of those are on the .chat TLD. ~ Top sites: chatgpt .com (#40), snapchat .com (#116), lc .chat (#1,264), livechatinc .com (#1,799), wechat .com (#2,103) 📈 Registration Activity ~ Ranked #38 among 10,000 tracked keywords by total domain registrations. ~ EMD NRD 7d: 1 new exact-match registrations (↑98% vs 30d avg). ~ EMD NRD 30d: 200 new exact-match registrations. ~ Broad presence: 443.6K zone file domains contain "chat" in their name.

Is the #Chat keyword about to become the hottest domain #trend? 👀 We’re in the middle of a genuine technological revolution: #ArtificialIntelligence. AI now touches every industry, and its ripple effect has also an impact on #domainname searches. Why “Chat” is gaining more traction now: ☑️ AI progress fuels the rise of ChatGPT, chat‑bots, and AI‑chat #solutions. ☑️ As these tools become mainstream, the demand for chat‑related domain names could spike more and more. ☑️ “Chat” domain names can become more valuable for commercial use, brand building, or simply as a high‑value asset. Will we see another surge in interest for chat‑domains as #AI continues to evolve? 💬 Drop your thoughts in the comments! Note about the chart: : The chart shows how search interest changes during a chosen timeframe, using a relative index that runs from 0 – 100 on the Y‑axis. A score of 100 marks the peak interest within the displayed period, 50 represents half of that peak, and 0 means there isn’t enough data to calculate a value. Source: https://trends .google. com/explore?geo=Worldwide&gprop=web&date=today%205-y&q=chat

@GuideTwit It's been mental all year so far. I can't remember extended volatility like this since we got into this game...

Calm down already G

@dejanseo @Adoubleagent @dataforseo The code that ships with Chrome might be useful if you decide to grab pages directly. It's not beyond the pale to imagine they might use the same extraction patterns. I remember thinking it was unusual that they chunked on <br>. I went digging when you shared that small model.

@RedCardinal @Adoubleagent I use @dataforseo to return clean MD per page.

On popular demand, you can now set the target URL and its queries to see which parts will represent it in Google's AI search: gs.dejan.ai

@dejanseo @Adoubleagent Interesting. Is this using the chunking pattern from the small model bundled with Chrome by any chance? I presume this is part of your pipeline and not the model API itself?

@RedCardinal @Adoubleagent A segment of text followed by newline:

@gaganghotra_ Not weird if they don't segment this process for desktop/mobile. A good test will be looking for other sites which run m. to see if citations are for mobile version of site.

Noticing a lot of mis citations like on desktop Google is citing URLs of Yelp mobile pages which are on m .yelp .com subdomain of course it redirects to the desktop version as Yelp detects user is on desktop but bit weird from Google side to cite mobile version of URL in answers when user is using AI Mode on desktop #SEO

@DotWeekly As someone who used to be a very active auctions snipper in the past, but only a seller in the past few years, your tooling looks great. I kinda wish I could get into the buy side again. But seeing numion at $1k bid makes me wonder how much margin is left in wholesale market?

💎 Hidden gems in today’s auctions (Mar 6, 2026) 79,442+ domains analyzed across 8 marketplaces. Here’s what stood out: ▶️ EmailDefense.com — 74 bids @ $426 ▶️ Numion.com — 55 bids @ $1,066 ▶️ LoveDoll.ai — 40 bids @ $239 Full list: dotweekly.com/blog/daily-hig… #ad

Did anyone really expect Google not to leverage citations to generate more revenue? At a minimum, they needed to offset lost revenue from lower ctr on organic prop results/features. Always going to happen...

@DInvesting @DomainNameWire Gotta bank that $69 fee though...

@DomainNameWire I’m talking about an unsolicited offer made on a domain name not listed for sale (so no asking price or listing). Imagine getting an unsolicited $7,000 offer for DNW .com from a broker. Broker should tell client that’s well under a 3L valuation and owner is a pro.

Domain brokers: you are not *obligated* to present a lowball offer to the owner of a domain name. Either educate your client about domain name values or fire him. Don't waste time with an unrealistic offer that makes you look foolish.

@pedrodias Just you wait. Everything will be synthetic soon.

Some people vibe coding AI prompt trackers have no clue about the difference between user interaction with an interface and synthetic prompting. As for me, I’m all for enjoying great moments of laughter and appreciation of human ignorance

@forgebitz What was the purpose of them reporting these in the frontend to begin with? Were they ever accessible to the user? We're these just properties that were sent by the app for no good reason? Genuinely curious?

since most people don't know; nothing changed to the API scraping (but you can't track brand visibility using that, you don't have reddit for example) the logged in version of chatgpt still shows the queries the logged out version (used to track AI visibility) stopped sharing them

looks like chatgpt query fanouts are gone don't know if it is directly related to gpt 5.4 since we don't see any of them also on 5.3 as predicted, the only way to track citations/prompts is directly on CDN level final nail in the coffin for the vibe coded prompt trackers

@GeorgeKirikos You're very likely to run into serious hallucinations here. Where do you think the model got either the data or the equivalent tasks in its training data to provide useful responses? Even if it uses tools to search, what do you think it will find? Beware of such prompts/tasks.

By adding one's domain name portfolio to an AI tool's memory, e.g. help.openai.com/en/articles/85… one can run prompts like: "Analyze my portfolio list. Which do you consider to be the top 25 in terms of valuation? Valuation should be considered in terms of average liquid value that can be reasonably achieved over a period of 5 years." ChatGPT is getting better and better at this....its latest answer (using version 5.4 Thinking) started with: "Using a 5-year liquid end-user lens rather than a heroic holdout price, I would rank your portfolio’s top 25 like this......." and ended with: "If you want, I can next turn this into a full top-50 table with three columns for each domain: wholesale floor, 5-year liquid value, and strategic ceiling."

@cyberandy @Kevin_Indig @dawnieando Very interesting. I suspect that passages selected for citation are not always used in the synthesis process. Theory: Google is under intense pressure to link out, and *may* cite docs aligned with the synthesised answer, but not used directly in its generation.

👉 Article: wordlift.io/blog/en/embedd… 👉 @Kevin_Indig’s research: growth-memo.com/p/the-science-… 👉 @dawnieando’s latest piece: womenintechseo.com/knowledge/gene… Read Dawn’s piece first. It will make it clear that I’m not offering a GEO hack, but a closer look at how embeddings are evolving.

Embeddings are becoming a key layer of AI search retrieval. After @Kevin_Indig showed that 44.2% of ChatGPT citations come from the intro of articles, I dug into how embeddings from Google, OpenAI and Perplexity influence what gets retrieved and cited. wordlift.io/blog/en/embedd…

@Marie_Haynes Does it manage if the display:none is attached via a class style? It would need to parse the stylesheet to figure that out.

This is wild. Me: Run this prompt to give us ideas for improving freshness. My tool: Change this line here. Me: That line isn't there. Please don't hallucinate. ...I keep modifying the prompt. Nothing is working. Me: Hey AntiGravity, can you help me figure out why Gemini keeps hallucinating content that is not on this page? AG opens up a browser and confirms those words aren't on the page. But why is Gemini hallucinating? It gets an idea...check the source code. And sure enough the words we thought Gemini was hallucinating were in the source code commented out with display:none. So AG rewrote the prompt to tell Gemini to ignore words within a "display:none" or other hidden text. Problem solved.