Confirmed
Zepto halves P90 search latency by migrating to OpenSearch's OR2 instances
Zepto migrated its OpenSearch Service clusters from r7g.12xlarge to the new or2.12xlarge instance type, AWS and Zepto detailed in a joint Big Data Blog post published August 17, 2026. OR2 instances store primary data on local EBS storage with Lucene segments replicated to S3, using segment replication instead of document replication. Load testing showed indexing throughput doubled, P90 search latency dropped 52% (187ms to 89.1ms), and Zepto cut its data node count by a third while saving 30% in cost.
In plain termsOpenSearch is the search engine many e-commerce sites use to power on-site product search, and OR2 is a newer, cheaper way to run it that answers searches faster by storing the underlying data differently.
Why it mattersA quick-commerce retailer proving OR2 cuts both latency and cost at production scale gives any e-commerce team running OpenSearch a concrete migration case to weigh for its own product-search infrastructure.
Our takeWorth watching whether other high-traffic retailers follow Zepto's lead, since a 52% latency cut plus a 30% cost drop is the kind of production case study that tends to accelerate migrations once real numbers replace a vendor benchmark.
Confirmed
Google adds Gemini 3.7 Flash as a selectable model in AI Mode
Google Search VP of Product Robby Stein announced on X on August 14, 2026 that Gemini 3.7 Flash, released by Google DeepMind the day before, is now rolling out globally in AI Mode as a selectable model for Google AI Pro and Ultra subscribers in English. Stein said the model is better at following instructions and understanding user intent. Users select it via the plus icon in AI Mode's picker. It isn't the default for all users yet.
In plain termsAI Mode is Google's chatbot-style search experience for paying subscribers, and this update lets those subscribers manually pick a newer, smarter underlying AI model for their searches instead of being stuck with whatever model Google set as default.
Why it mattersA newly released model reaching AI Mode within a day signals Google now ships model upgrades to its AI-search surface on the same cadence as its consumer chat products, not a slower search-specific release cycle.
Our takeThe one-day turnaround from model launch to AI Mode availability is the detail worth tracking. It suggests Google now treats AI Mode as a standard rollout surface for new Gemini releases rather than something that lags weeks behind the consumer app.
Confirmed
OpenAI documents that robots.txt may not bind ChatGPT's user-triggered fetch bot
OpenAI's crawler documentation states that ChatGPT-User, the bot that fetches a page when a live ChatGPT user asks about it, is not used for crawling the web in an automatic fashion, and that because these actions are initiated by a user, robots.txt rules may not apply. This is separate from OAI-SearchBot, the crawler that determines ChatGPT search-result visibility, and from GPTBot, which crawls for model training. Both of those bots still respect their own robots.txt disallow rule.
In plain termsRobots.txt is the file a website uses to tell automated bots which pages they can visit, and OpenAI is saying that rule was built for bots crawling on their own schedule, not for the one that fetches a page only because a person using ChatGPT just asked about it.
Why it mattersA site can disallow every OpenAI bot in robots.txt and still have ChatGPT-User fetch its pages on a live user's behalf, so robots.txt alone can no longer be treated as a complete access control against ChatGPT.
Our takeI've written about how a Disallow rule controls what a crawler fetches, not what ends up indexed. This is the same gap from the opposite direction: OpenAI is arguing Disallow may not even control the fetch itself once a live user's question triggered it, narrowing what robots.txt can promise even further than the indexing side already showed.
Deep dive: Disallow Stops the Crawl. It Doesn't Stop the Index.
Confirmed
The llms.txt spec adds formal link tags for discovering a page's Markdown version
Jeremy Howard's llms.txt specification, modified August 10, 2026 per its own changelog, adds two standard link relations so AI agents can find a page's Markdown version without guessing a URL pattern: rel="alternate" type="text/markdown" points to the Markdown file, and rel="describedby" points to the site's llms.txt. Sites can implement either via an HTML link element or an HTTP header, and the spec author notes OpenAI, Anthropic, and Google's Gemini already publish llms.txt files for their own developer docs.
In plain termsllms.txt is a proposed standard file, similar in spirit to robots.txt, meant to tell AI systems which pages on a site matter and where to find a clean, text-only version of each one; v2 adds a formal way to point from a normal web page straight to that clean version.
Why it mattersThe original llms.txt spec told agents which pages exist but not where a machine-readable version of each page actually lived, so v2 closes the exact discovery gap that made most implementations a link directory instead of something an agent could parse directly.
Our takeWhether this gets adopted matters more than the spec update itself. Nobody has confirmed a live AI assistant actually fetches a stranger's llms.txt before answering a question about that site, so the formal linking mechanism only pays off once retrieval pipelines are confirmed to use it, not just publish it.