Confirmed
Google confirms AI Mode follow-up questions register as brand-new queries in Search Console
Google's Search Console help documentation states that when a user asks a follow-up question inside AI Mode, it counts as a brand-new query, with all of that response's impressions, clicks, and position data attributed to the new query rather than the original search. Google Search Advocate John Mueller pointed a site owner confused by short, conversational strings like 'yes' and 'ok' appearing in their Performance report straight to that documentation section.
In plain termsSearch Console is the free Google dashboard site owners use to see which searches bring people to their pages, and this confirms that when someone keeps chatting with Google's AI Mode, each follow-up gets logged as its own separate, brand-new search, even though the person never typed anything into a normal search box.
Why it mattersAnyone treating Search Console's query list as a clean keyword report now has to filter out AI Mode's conversational fragments before drawing any conclusion from query-level data.
Our takeI wrote yesterday about how Google's generative-AI reporting in Search Console still leaves real gaps for tracking AI-driven traffic. This is the same reporting surface producing a different kind of confusion: raw follow-up fragments landing in the query list with no way to tell, from the report alone, that they came from one continuous AI Mode conversation rather than three unrelated searches.
Confirmed
Google confirms AI-generated recipe images in AI Overviews were a paused experiment
Recipe blog Inspired Taste posted on X that Google's AI Overviews began showing AI-generated step illustrations in its content, a month after Google said it would bring image generation into AI Overviews. Google Search VP Robby Stein replied on X on August 17, 2026 that the images were a small experiment Google is no longer running, distinct from the Nano Banana image feature Google announced in July, which only generates images when a user explicitly asks.
In plain termsAI Overviews is the AI-written summary Google shows above normal search results, and this specific test replaced a recipe site's own step-by-step photos with Google's own AI-drawn pictures, which is exactly the kind of original content creators make a living producing.
Why it mattersA creator's screenshot forcing a same-day Google confirmation shows how fast an unannounced AI Overviews test now surfaces and gets walked back once it touches content creators depend on for income.
Our takeWorth watching whether this resurfaces once Google works out the creator backlash, since the underlying incentive, AI Overviews needing visuals for content it doesn't have images for, doesn't go away just because this particular test got paused.
Confirmed
Snowflake announces dynamic model routing for Cortex AI Gateway, adds DeepSeek and GLM open models
Snowflake said dynamic model routing is coming soon to Cortex AI Gateway, letting enterprises set approved models and cost/quality tradeoffs while the gateway picks the cheapest model that clears the bar per task. Snowflake also added DeepSeek-V4-Flash 0731 to Cortex AI in private preview, scoring 74.4% on the ADE-bench agent benchmark, and said GLM-5.3 is coming soon. In internal testing, routing cut token use up to three times on a dbt pipeline workload and about 25% on a coding workload.
In plain termsCortex AI Gateway is Snowflake's traffic-cop layer for AI requests, and dynamic model routing means it automatically sends each task to the cheapest AI model that can still do the job well, instead of a developer having to hardcode which model handles what.
Why it mattersTeams running AI agents on Snowflake now get automatic per-task model selection instead of hand-coding routing logic, which is the same cost-versus-capability tradeoff every team building on a metered LLM API already has to manage manually.
Our takeI've written about how a model's price-to-performance curve is what pulls a given AI use case into economic reach, not the model's raw capability alone. Snowflake automating that same match-the-task-to-the-cheapest-capable-model calculation inside its gateway is the enterprise-infrastructure version of the exact tradeoff that curve describes, just decided per request instead of once at build time.
Confirmed
Snowflake details the internal semantic layer it built to keep AI agents and dashboards from disagreeing
Snowflake published a case study on the internal semantic layer it built so AI agents and dashboards stop disagreeing about basic definitions like 'active customer.' The layer sits between raw tables and every consumer, translating business language into database schemas via semantic views that Snowflake version-controls through dbt. Snowflake said its product data science agent handled over 5,400 queries from more than 400 users, and cited an AtScale benchmark showing semantic context raised text-to-SQL accuracy from 20% to over 90%.
In plain termsA semantic layer is a translation step that sits between a company's raw database tables and anyone asking a question, whether that's a person building a dashboard or an AI agent, so both get the same answer to something like 'how many active customers do we have' instead of each one calculating it a different way.
Why it mattersThe same problem Snowflake describes, AI agents and dashboards disagreeing on what a metric means, is exactly what breaks any retrieval or agent pipeline built on inconsistent source data, making this a preview of a fix pattern any team feeding structured data to an LLM will eventually need.
Our takeI built a version of this problem at much smaller scale for my own knowledge base: a curated index that gives an AI consistent context to retrieve from beats pointing it at raw, inconsistent sources and hoping it sorts out which definition to trust. Snowflake is describing the same fix at enterprise scale, a governed layer between the agent and the raw tables, which tracks with what I found: the retrieval layer's quality matters more than the model answering from it.