I’m watching OpenAI discontinue ChatGPT Atlas, its standalone desktop browser, and move its browser-based AI features into the new ChatGPT desktop app. That app brings together ChatGPT Work, OpenAI’s work-focused agent, and ChatGPT Codex.
The end of Atlas. I’m taking note of an Aug. 9 retirement date after OpenAI’s James Sun confirmed the plan on X.
I’m also noting Sun’s exact wording: “The current targeted date for deprecation is 8/9, and we’ll share more information in the upcoming days both in-app and via email.”
One desktop app. I see the new ChatGPT desktop app becoming OpenAI’s primary desktop product, complete with built-in browser capabilities. Instead of maintaining a separate AI browser, OpenAI is combining browsing, work-agent features, and Codex in one place.
Chrome users can keep Chrome. If I prefer using Chrome, I can access ChatGPT and Codex through OpenAI’s Chrome extension without switching to a dedicated OpenAI browser.
As AI reshapes search, every glowing path to discovery carries commercial value—turning SEO investment into a conversation about pipeline, risk, and customer acquisition costs.
Why I care. I see this as an important shift because OpenAI is moving AI browsing into the main ChatGPT experience, where more people can ask questions, research brands, and complete tasks. In my view, that gives ChatGPT another opportunity to influence discovery beyond traditional search results.
I first saw ChatGPT Atlas launch on Mac in October. OpenAI later released a dedicated Codex app and added an in-app browser in April. Now, I’m watching those capabilities move into the new unified ChatGPT desktop app.
I’m watching Google add a new layer of AI transparency to ads across Search, YouTube, and Discover. The company said its new How this ad was made section will appear inside My Ad Center, giving people a clearer view of whether AI played a role in the ad creative they see.
The panel will show whether an ad was created or modified with AI. I see this as a meaningful expansion of Google’s ad transparency tools, especially as more advertisers rely on generative AI to produce images, copy, and other campaign assets at scale.
What it looks like. I’ll be able to access the disclosure from the three-dot menu or the info icon on an ad. In the screenshot Google shared with Search Engine Land, the My Ad Center panel includes a dedicated section explaining how the ad was made.
Google will handle some disclosures. When advertisers use Google’s own generative AI ad tools, Google will automatically add the disclosure inside My Ad Center.
Google’s My Ad Center adds a clear AI disclosure, helping users see when ad creative may have been created or edited with generative AI.
For advertisers using third-party AI tools, Google said they will have control over whether to disclose AI use. Depending on local requirements, an AI label may also appear directly on the ad, either automatically or after the advertiser uses that control.
Why I care. AI-generated ads are getting easier and faster to create, so disclosure matters more than ever. I want to know when creative was made or changed with AI because requirements can vary by market, platform, and ad format.
Existing ad rules still apply. Google said its ad policies still prohibit misleading or deceptive advertising, whether AI was involved or not. This update adds more visibility into how an ad was made, but it does not change the requirement that advertisers clearly identify who they are and what they are promoting.
A bold Google logo sits atop layered, colorful digital documents, evoking the fast-moving world of search marketing, ad formats, campaign assets, and platform updates.
Earlier AI safeguards. Google already embeds imperceptible signals, including SynthID, into content created with its generative AI tools. Election advertisers are also required to disclose synthetic or digitally altered content in political ads, under a policy Google introduced in 2023.
I’m watching YouTube take a bigger step into conversational search by expanding Ask YouTube to signed-in U.S. desktop viewers who are 13 and older. What started as a Premium-only experiment is now reaching a much broader audience.
What is Ask YouTube? I see Ask YouTube as YouTube’s AI-powered search layer. Instead of typing a traditional keyword query and scanning a list of videos, I can ask a natural-language question in the YouTube search bar and get an AI response that may include text, video clips, long-form videos, Shorts, and suggested follow-up prompts.
Access is expanding. When YouTube announced the test in April, Ask YouTube was limited to U.S. YouTube Premium members who were 18 and older and opted in through youtube.com/new. On July 6, YouTube expanded it to signed-in U.S. viewers 13 and older using English-language searches on desktop.
Signed-out viewers and supervised accounts are still excluded for now. YouTube also said it plans to bring the feature to more devices, languages, and users worldwide in the coming months.
Standard YouTube Search is not going away. If I land on an Ask YouTube results page and want the usual video results, I can click All or return to the Home page. That means Ask YouTube remains a separate search option, not a full replacement for traditional YouTube Search.
Views still count for creators. YouTube said videos featured inside Ask YouTube responses can give creators another path to discovery. Views from Shorts, videos, and previews shown in Ask YouTube responses count toward total view metrics and YouTube Partner Program eligibility.
I also noticed that featured videos display the video title and channel name, which matters for attribution and visibility. For creators, YouTube’s guidance is clear: publish unique, high-quality content with descriptive titles and clear chapters so its systems can better match video segments to viewer questions.
A bold Google logo sits atop layered, colorful digital documents, evoking the fast-moving world of search marketing, ad formats, campaign assets, and platform updates.
Why I care. YouTube is putting conversational AI search in front of a much larger group of U.S. desktop users. If I’m creating or optimizing video content, this raises the value of clear titles, useful chapters, and segments that directly answer specific questions.
For SEO and content teams, this is another reminder that discovery is shifting from simple keyword matching toward answer-based experiences. The videos most likely to benefit are the ones that make it easy for YouTube to understand what each section covers and which viewer questions it solves.
What it looks like. YouTube shared a GIF showing Ask YouTube in action, where users can ask a question, review AI-assisted results, and continue with follow-up prompts.
I see these two new analyses as an important reminder that ChatGPT citations are not as fixed or transparent as they may look. The sources shown in an answer can change when ChatGPT routes search traffic through different hidden retrieval pipelines.
Research from Chris Green and Suganthan Mohanadasan adds a new wrinkle to AI visibility tracking: the final answer does not reveal how ChatGPT selected its sources. Both researchers found internal source-selection labels, including Labrador, Bright, Oxylabs, and SERP, but those labels sit behind the answer rather than inside the citation cards users see.
Green tested 1,000 prompts up to 10 times each and captured 9,946 completed search runs. In most cases, prompts stayed on one retrieval source. Labrador accounted for 88.1% of primary search sources in his dataset, followed by Bright at 9.9%, Oxylabs at 1.7%, and SERP at 0.3%.
What stands out to me is that 11.6% of prompts changed their primary search source across repeated runs. When that happened, URL overlap dropped from 0.273 to 0.149, and domain overlap fell from 0.265 to 0.155. Green calculated that as roughly 45% lower URL overlap and 42% lower domain overlap.
Mohanadasan looked at the issue from another angle. He inspected two days of raw ChatGPT network traffic from one logged-in Pro account and logged about 1,240 source records across a few dozen searches. He found a result_source field attached to web results, with four observed values: SERP, Labrador, Bright, and Oxylabs.
He described Labrador as including established publishers and reference sites, Bright as tied to Bright Data, Oxylabs as tied to Oxylabs, and SERP as an open-web baseline that appeared mostly in news-style results. While Green’s repeated-prompt test found Labrador dominating his dataset, Mohanadasan saw Bright play a larger role in his sample, especially for commercial, shopping, finance, weather, and local queries.
I also think the skipped-search finding matters. Mohanadasan found that ChatGPT classified some queries before searching, using a turn_use_case field. Some prompts were filed as text and skipped web search entirely, even when they sounded current. In those cases, no page could be fetched, cited, or used as evidence.
Rows of illuminated data cabinets and paper files stretch into the distance, capturing the pressure on marketers to turn fragmented customer data into a smarter performance engine.
More complex “thinking” queries behaved differently. Mohanadasan found that ChatGPT could branch into many searches, including site: probes, pricing checks, and searches for unnamed competitors. That changes which pages can enter the answer process because ChatGPT may search rewritten queries, direct site probes, or follow-up checks instead of the exact phrase a user typed.
Another useful distinction is that fetched does not always mean cited. Mohanadasan separated three outcomes: fetched, cited, and mentioned. A page can be pulled into ChatGPT’s context without being shown to users, cited as support for a specific sentence, or skipped as a source even when a brand is mentioned in the answer.
In his small commercial-query sample, Reddit and YouTube were both fetched often, but Reddit was cited and YouTube was not. He attributed that gap to text availability: Reddit threads expose text, while YouTube search results often provide metadata rather than full video transcripts. Vendor pages were cited for their own facts, such as prices and specs, while third-party pages were more likely to support broader recommendation claims.
The practical takeaway for me is that there is no single ChatGPT visibility result to measure. A page may never be considered if ChatGPT skips search, uses another retrieval source, or finds a clearer third-party page to support the claim.
Both analyses also point back to readability. ChatGPT’s source selection depends partly on what it can retrieve and understand. Mohanadasan found cases where ChatGPT appeared to prefer official pricing pages, then fell back to third-party sources when prices were hidden behind JavaScript or otherwise hard to parse.
Green’s results showed that source routing can change which URLs and domains enter the answer set. That makes plain HTML, crawlable facts, clear pricing and specs, strong third-party coverage, and text-heavy pages more important when source selection depends on retrieval and readability.
I see advanced architecture as much more than a technical framework now. It shapes whether my content can be found, understood, and surfaced by search engines and AI systems.
That is why I am paying close attention to the next SMX Now on July 15, featuring Shari Thurow, co-founder, information scientist, and search director at the Information Architecture Gateway. She will explain how advanced architecture really works and where many AI, SEO, and site development workflows tend to fall short.
In this session, I will explore a five-phase framework Thurow has tested through decades of client work with organizations including Microsoft, Google Cloud, Abbott Laboratories, CVS Pharmacy, WebMD, Sony Music, the Library of Congress, Best Buy, and Merriam-Webster. I will learn how architecture decisions influence labeling systems, wayfinding networks, taxonomy, wireframes, and AI access to valuable content.
I also expect the session to challenge some long-standing assumptions, including the three-click rule, the idea that taxonomy is only a hierarchy, and the belief that AI can create effective wireframes without a deeper architectural model behind them.
Old search marketing tools give way to a faster, connected future, with data streams, AI icons, and a glowing search hub symbolizing SEO innovation and community growth.
By the end, I will have a practical framework for building sites that communicate more clearly with users, search engines, and human-centered AI systems.
I see ChatGPT’s high-reasoning mode acting like a very different search surface for brand visibility. In a Semrush analysis with Kevin Indig, ChatGPT cited different domains than it did in minimal reasoning mode and ran nearly five times as many web searches before answering.
By the numbers, the shift is hard to ignore. Only 25.6% of cited domains overlapped between minimal and high reasoning for the same prompts. That means nearly three in four sources changed when ChatGPT moved from Instant-style answers to Thinking-style answers.
I also noticed that Thinking mode used more sources overall. Citation rates rose from 50% in minimal reasoning to 68% in high reasoning. When ChatGPT did cite sources, it used more of them too, increasing from 2.6 to 4.5 citations per response. Across the test set, high reasoning ran 1,130 web searches, compared with 245 for minimal reasoning.
Reddit lost ground in high-reasoning answers. Reddit’s citation share dropped from 15% to 7% when high reasoning was turned on. User-generated content and review sites also declined, falling from 14.3% to 6%.
At the same time, I saw more weight shift toward institutional and official sources. Government and academic sources rose from 1.9% to 8.8%, while official documentation and support pages grew from 12.4% to 17.5%.
Comparison prompts drove the most search activity. At the comparison stage, high reasoning averaged 24 sub-queries per prompt, compared with 5.5 for minimal reasoning. Average citations also peaked there, reaching 9.8 per high-reasoning response versus 5.8 for minimal reasoning.
For example, I would expect a CRM comparison to trigger separate searches for pricing, integrations, security, support pages, and documentation before ChatGPT forms its final answer.
Early citations also appeared to last longer. High reasoning was more likely to carry a brand from early research into later buying questions. In four of the 20 journeys tested, a brand cited at the problem stage still appeared at the selection stage. Minimal reasoning showed no full-journey persistence, meaning no brand cited at the Problem stage survived through to the Selection stage of the same journey.
I also found the domain reuse pattern important. High reasoning reused the same domains more often within a single answer, with the same domain appearing multiple times in 51 of 100 high-reasoning responses. Minimal reasoning did this in 26 of 100 responses.
Finance saw the biggest citation jump. The lift varied by category, but finance had the largest increase, with citation rates rising 28 percentage points in high reasoning. Health and lifestyle rose 24 points, while B2B SaaS gained 16 points.
Consumer tech barely moved, rising only 4 points. Even though high reasoning ran more sub-queries for consumer tech prompts than for any other category, it often landed on the same brands and sources as minimal reasoning.
Why I care about this: content can appear in fast ChatGPT answers but disappear when users ask more complex questions. Visibility depends on whether my pages, documentation, and third-party references can surface across the smaller searches ChatGPT runs before it answers.
About the data: Semrush and Indig tested 100 prompts across 20 buyer journeys in B2B SaaS, finance, consumer tech, and health and lifestyle. Each prompt ran once in minimal reasoning and once in high reasoning. The analysis tracked citation rate, cited sources, and fan-out queries.
I’m reading this Cornell Tech research as a clear warning: deep-research AI agents can be steered by surprisingly small edits on public, user-generated pages. In the study, a single injected Reddit-style comment could become a cited recommendation for fake products, services, or entities.
The researchers described these altered pages as “poisoned” because the added text was written to influence what an AI system cites and repeats. The weakness appears in systems that search the web, collect sources, and produce cited reports. The paper calls the attack WARP, short for Web Agent Retrieval Poisoning.
How I see injected text reaching reports. The attack does not require access to the model, prompts, search engine, or retrieval system. Instead, an attacker edits or appends text to a page the agent already tends to retrieve, such as a Reddit thread, Wikipedia page, or forum post.
When the agent later searches related topics, it may pull in that page, cite it, and repeat the attacker’s chosen message as part of an otherwise normal-looking answer.
That matters because deep-research tools often run many related searches for a single user request. The paper found that the same user-generated pages surfaced across related queries, giving poisoned content more chances to appear.
Reddit stood out as the biggest opening. Across STORM, Co-STORM, and OmniThink, 17% to 23% of retrieved URLs came from user-generated platforms, including Reddit, YouTube, Facebook, and Wikipedia.
Reddit made up the largest share of those pages. It accounted for 54% to 71% of the user-generated URLs retrieved by the three open-source systems.
The researchers did not alter live websites. Instead, they used a simulation framework called GeoStorm to insert manipulated text into retrieved content during testing.
A few words were enough. What stood out to me most is how little text the attack needed. The researchers found that snippets as short as about 13 words could influence what these systems recommended.
In one test, a 15-word sentence pushed a fake cryptocurrency, BananaCoin, into a Co-STORM report as an “emerging” long-term investment option. The report cited the altered source alongside legitimate crypto sources.
When the manipulated page was retrieved, the fake entity appeared in 38% to 51% of reports across systems. When the researchers targeted multiple pages, that range increased to 42% to 62%.
The attack still worked when systems retrieved full Reddit threads, although mention rates were lower. When injected text was added to complete Reddit threads and represented less than 4% of the retrieved content, the fake entity still appeared in 30% to 53% of reports when the page was retrieved.
The defenses struggled. Blocking user-generated domains stopped this attack path, but I see the tradeoff immediately: it also removes useful sources such as firsthand product experiences and local recommendations.
The tested text filters also failed to reliably separate injected passages from normal user content. Because the manipulated passages were fluent and written by an AI model, perplexity-based filters were more likely to flag normal user content than the injected text.
Report-level checks missed the manipulation too. The altered reports looked similar to clean reports because the agent itself folded the fake recommendation into an answer that otherwise appeared normal.
Why I care. A small edit to a public page can become part of a cited AI answer, even when the underlying source is user-generated. Misinformation planted on sites like Reddit or in forums can move from discussion threads into AI recommendations that look credible to users.
About the research. The paper, Deep-Research Agents Can Be Poisoned via User-Generated Content, was written by Tingwei Zhang, Harold Triedman, and Vitaly Shmatikov of Cornell Tech and posted to arXiv on May 22. The researchers tested the full attack on three open-source systems: STORM, Co-STORM, and OmniThink.
They also analyzed OpenAI Deep Research and Gemini Deep Research for user-generated citations, but they did not run live manipulation tests because doing so would require publishing altered content to the open web.
When I look at Similarweb’s findings, the message is clear: users who saw a brand recommended by ChatGPT were much more likely to visit that brand’s website within a week.
What happened. I found the biggest takeaway in the behavior shift. On average, users were 2.5 times more likely to visit an AI-recommended brand than a direct competitor, based on Similarweb’s study of U.S. desktop activity across finance, travel, and beauty.
Similarweb tracked users who asked ChatGPT industry-relevant questions, received a specific brand recommendation, and then visited either that recommended brand’s website or a competitor’s site within seven days.
To keep the data focused, the study excluded users who had visited the brand’s site in the prior four weeks or had named the brand directly in their prompt.
Recommendations shifted traffic. I saw the same pattern appear across all three industries Similarweb analyzed, which makes this more than a one-category trend.
In finance, after an American Express recommendation, 7.2% of users visited American Express, compared with 3.1% who visited Capital One. After a Capital One recommendation, 14.2% visited Capital One, compared with 3.8% who visited American Express.
In travel, after a Skyscanner recommendation, 9.5% visited Skyscanner, compared with 7.6% who visited Kayak. After a Kayak recommendation, 12% visited Kayak, compared with 3.4% who visited Skyscanner.
In beauty, after a Sephora recommendation, 7.9% visited Sephora, compared with 3.3% who visited Ulta. After an Ulta recommendation, 7.6% visited Ulta, compared with 4.6% who visited Sephora.
AI demand showed up in search. What stands out to me is that most AI-influenced visits did not appear as AI referral traffic. ChatGPT may shape the user’s brand choice, but the later website visit often shows up in analytics as search traffic instead.
Similarweb found that 55.9% of AI-influenced visits came through search, compared with 40.4% of non-AI-influenced visits.
Direct traffic told a different story. It accounted for 19.9% of AI-influenced visits, compared with 38.8% of standard visits.
Recommended users stayed longer. I also think the engagement data matters. AI-influenced visitors viewed 12 pages and spent 11.8 minutes on site, on average, compared with 6.5 pages and 5.6 minutes for non-AI-influenced visitors.
That deeper engagement suggests these users may have already narrowed their options during the AI conversation before they ever reached the brand’s website, Similarweb said.
Why I care. AI visibility can drive meaningful visits even when referral reports miss the original source of influence. I need to understand whether ChatGPT is creating demand for my brand or sending that demand to a competitor.
About the data. Similarweb used its opted-in U.S. desktop web panel to track user journeys from July through December 2025. The report focused on finance, travel, and beauty brand pairs with competitive overlap.
