Brandon Lazovic

Pulse · August 19, 2026

Confirmed

Google releases the August 2026 spam update globally, across all languages

Google's Search Status Dashboard confirmed the release of the August 2026 spam update, starting August 18 at 9:27 a.m. Pacific and applying globally across all languages. Google said on LinkedIn this is a normal spam update and expects the rollout to finish within a few days, with effects possible across all locations and languages during that window.

In plain termsA spam update is a scheduled tune-up to the systems Google uses to detect and demote content built to game search results instead of help readers, and it runs everywhere at once, not gradually by country.

Why it mattersThe multi-day rollout window means ranking volatility this week does not by itself indicate a penalty, so the useful response is watching Search Console over the next several days, not reacting to day-one swings.

Our takeThe multi-day rollout window Google names here is worth taking literally. Ranking swings that show up mid-week and settle by the weekend are the update finishing, not a second signal stacking on top of it.

Confirmed

OpenAI announces ChatGPT Ads expansion to 31 European countries, its largest yet

OpenAI said August 18 that ChatGPT Ads will expand to 31 European countries next week, including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands, and Austria. Advertisers get initial access only through OpenAI's Ads Solutions team, agency partners, and technology partners, with self-service Ads Manager access following later this summer. Ads remain limited to Free and Go plan users, and paid subscribers stay ad-free.

In plain termsChatGPT Ads is the sponsored-placement system OpenAI is testing inside chat answers, and this expansion means people asking ChatGPT questions in Germany, France, and 29 other European countries will start seeing paid results mixed into their answers, the way Google mixes ads into search results today.

Why it mattersAdvertisers running or planning ChatGPT campaigns now have a firm date for European market access, though self-serve tooling lags the initial managed rollout by weeks.

Our takeThe sequencing matters as much as the country list. Managed access through OpenAI's own ads team and partners goes live next week, while self-service Ads Manager, the tool that lets a marketer run a campaign without a middleman, does not arrive until later in the summer. That gap is where OpenAI keeps control over quality and volume before opening the platform to everyone.

Confirmed

Snowflake makes per-user AI cost quotas generally available in CoCo, lists MCP server controls as 'GA soon'

Snowflake made per-user daily and monthly AI credit quotas generally available across every CoCo surface on August 18, enforced automatically with no custom code required. The same post previews three more governance layers due 'generally available soon,' including an organization-wide policy that restricts which MCP servers, models, and tools a CoCo installation can reach, plus team- and query-level controls. Usage stays queryable through Snowflake's account-usage views.

In plain termsCoCo is Snowflake's AI coding and work agent, and a governance control here is a rule an administrator sets ahead of time, like a spending cap or an approved list of outside tools, so the agent cannot go past those limits even if a user or a prompt tries to push it there.

Why it mattersCost quotas ship today, but the layer that actually restricts which MCP servers, models, and tools an agent can reach is listed as 'generally available soon,' not shipped yet.

Our takeI wrote in August that Snowflake's prior MCP governance launch shipped nine of twelve tagged capabilities in private preview, not general availability. The pattern holds here. Per-user cost quotas are GA today, but the control that actually bounds which MCP servers an agent can reach is again listed as 'GA soon,' not shipped.

Deep dive: An MCP Server Is a Supply-Chain Dependency With Reach Into Your Data

Confirmed

Databricks adds Precision Mode to AI Extract, beats frontier models by seven points on accuracy

Databricks launched Precision Mode inside its ai_extract document-extraction API on August 18, pairing custom-trained extraction models with an agent harness that splits large documents into parallel subtasks before merging results. Evaluated across roughly 9,000 documents, including files up to 2,000 pages and schemas with more than 300 nested fields, Precision Mode reached 94.7% accuracy, beating the strongest frontier-model baseline, GPT-5.6 Sol, by seven points.

In plain termsDocument extraction means pulling specific facts, like a price, a date, or a line item, out of an unstructured file such as a PDF contract or invoice, and Precision Mode is built for the hard cases, like a 2,000-page filing or an invoice with thousands of line items, where a single request to an AI model runs out of room to hold the whole document at once.

Why it mattersContracts running hundreds of pages and invoices with thousands of line items are exactly the documents that break single-call LLM extraction, and Precision Mode targets that specific failure instead of competing on general model size.

Our takeThe comparison baseline matters as much as the headline number. Chunk-and-merge, the workaround most engineering teams already reach for once a document exceeds a model's context window, was the benchmark Databricks beat, and clearing that real-world workaround is a harder test than beating a single-call baseline nobody serious still uses.