Observed
Rank-Tracking Vendors Report Google Swapping In Fake Search Results To Block Scraping
Nozzle's October 10 research report says Google began blocking rank-tracking scrapers on September 13, then from September 20 started returning normal-looking pages with swapped rankings and no error pages, a pattern Nozzle calls poisoning rather than blocking. Nozzle measured 8.0 percent of tracked SERPs poisoned since September 13, up from a 0.17 percent baseline, with the single worst day (September 29) hitting 27 percent and poisoned pages keeping only 20.7 percent of normal top-10 results.
In plain terms'Poisoning' here means Google serves the scraping tool a fake but normal-looking results page instead of blocking it outright, so the rank-tracking software has no error to catch, just wrong numbers that look real.
Why it mattersAnyone reporting rank-tracking numbers to a client right now should flag the same window Nozzle measured, since a poisoned SERP can show a tracked site falling out of the top 10 with no actual ranking change behind it.
Our takeIf Google keeps swapping results rather than blocking requests outright, a tool's anti-detection method stops being the deciding factor, since the fake page looks exactly like a real one. Watch whether tracking vendors converge on cross-checking SERPs against each other rather than trusting a single scrape.
Confirmed
Google Merchant Center: Poor Image Quality Alone No Longer Blocks Product Listings
Google's Merchant Center help page now reads 'Your product will continue to show in ads and free listings' for poor image quality, replacing stricter wording that said a low-quality image 'may be prevented from showing in ads and free listings until the image quality improves,' per a September 4, 2025 Wayback Machine capture Search Engine Roundtable screenshotted and linked. The removed exclusion covered subjective image quality; a separate, objective 500x500-pixel minimum still applies, with warnings since July 2026 and exclusion enforcement starting January 31, 2027.
In plain termsMerchant Center is the feed Google pulls product listings from for Shopping ads and free listings; this change means a blurry but correctly sized photo no longer gets a listing hidden, though an image literally too small in pixels still will, starting in 2027.
Why it mattersMerchants holding back weak product images out of fear of suppression can treat image quality as a conversion lever again rather than a pass/fail gate, but the real floor to watch is the January 2027 pixel-size enforcement date, a harder resolution-based cutoff than the quality language it replaced.
Our takeGoogle swapping a subjective quality judgment for a hard pixel-count floor is the same trade search ranking made with Core Web Vitals, a fuzzy editorial standard replaced by a number a script can check at scale. Expect the July 2026 warnings to be where most merchants first notice their image dimensions matter more than how the photo looks.
Model & AI releases
AI model launches, model API price changes and major assistant features, kept separate from the search news above. Every model story Pulse has covered is collected under LLM Models & Agents, and the monthly model comparison shows how they stack up.
Confirmed
Microsoft Launches Decision-1, A Decision-Scoring Model Priced Under 5 Cents Per Million Tokens
Microsoft released Decision-1 on October 9, a decision-scoring model built by post-training Qwen3.5-9B that outputs structured yes/no, multiple-choice, rating or rubric-based judgments for routing, classification, prioritization and workflow control rather than open-ended text. Microsoft reports the highest accuracy across 36 benchmarks spanning roughly 150,000 questions, latency about 35 times faster than GPT-6 Sol, and pricing of $0.042 per million input tokens with free output tokens. The Copilot team is using it internally to quality-check response outputs.
In plain termsA 'decision-scoring model' here is a stripped-down model that returns a short structured verdict, like yes/no or a 1-5 rating, instead of writing paragraphs, which is why it can run far cheaper and faster than a general chat model.
Why it mattersFor anyone routing agent decisions or classification tasks through a general-purpose LLM, a purpose-built decision model priced at roughly four cents per million input tokens is worth benchmarking against whatever it currently costs to get a yes/no or rubric score out of a larger model.
Our takeSub-cent pricing on a narrow decision model is the kind of release that gets skipped past the general-purpose leaderboards and then quietly replaces a bigger model wherever an agent pipeline only needed a verdict, not prose. Worth testing anywhere a workflow currently burns a full LLM call on a yes/no or routing decision.