Changelog

AI Data Stream now runs a second kind of model alongside the frontier LLMs on your own keys: Jev, the first System One model from TypeSafe AI. Jev never generates text. You send it a state and typed questions, and it returns a choice from your options or a yes/no probability, with a calibrated confidence, in about a second. It runs on our key, not your BYOK LLM key, and the blog post explains what a System One model is, how it differs from an LLM, and how we validated it.

Its first job is an experimental, switch-it-on-yourself beta for Pro and Business teams. GA4’s new “AI Assistant” channel only catches visits that arrive with an ai-assistant medium. Perplexity, Copilot and untagged ChatGPT links land in Unassigned instead, next to your own subdomains, link scanners and referrer spam. When the beta is on, every traffic-source lookup gives those rows a second opinion from Jev.

• A real channel for Referral and Unassigned rows: AI assistant, self-referral, link scanner or tool, referrer spam, industry referral, search, social, email, offline or campaign tag — with a confidence score on every row
• An is_ai_assistant flag your AI can total, so “how many sessions came from AI” finally includes the ones GA4 missed
• GA4 stays the source of truth: below a confidence threshold the GA4 grouping is kept and the model’s guess is only noted; if the model is unavailable you get plain GA4 results, never an error
• Judged once, cached 30 days per source, so the extra second only applies to the first lookup of each source
• On/off in Team Settings → Beta Features (team owner, paid plans). Off by default. With it off, your AI mentions it once when your data has rows it could help with, with a link to the toggle

Cloudflare is now a connected data source on Pro and Business plans. If your site sits behind Cloudflare, one read-only API token gives your AI the view from the edge: which AI crawlers are reading your site, what they fetch, and how much of your total traffic never reaches Google Analytics.

• AI crawler visibility: Requests and bandwidth per bot per day — GPTBot, ClaudeBot, PerplexityBot, CCBot, Bytespider and more — plus the pages they fetch most and the URLs they keep requesting that no longer exist
• AI referrals: Human visitors arriving from ChatGPT, Perplexity, Claude, Copilot and Gemini, by day or landing page (needs a paid Cloudflare plan on the zone)
• Edge traffic overview: Whole-site requests, page views, unique visitors, bandwidth, cache hit rate and blocked threats — every request counted, including the ones ad blockers and consent banners hide from GA4
• Dashboard widget: Daily requests and page views, the AI crawlers seen at the edge, top countries and status codes
• 2 new skills: AI Crawler Visibility, and Edge vs Analytics Gap (Cloudflare + GA4)
• Read-only: A token with a single permission, Zone → Analytics → Read, scoped to one zone

Google Ads connects with Connect with Google again: sign in, pick the account, done. Google retired API developer tokens on 9 September 2026, which removes the manager account, the token application and the wait for Google’s review that made Google Ads our most involved setup. Existing connections keep working unchanged, and service accounts remain available for agencies. See Connecting Data Sources.

Google’s Gemini 3.8 Flash is now the recommended Google model in AI Data Stream. It keeps the same introductory pricing as 3.7 Flash ($0.75 / $3.75 per million tokens through 31 December 2026, then $1.50 / $7.50) and brings better multi-step tool use, which is exactly what our data analysis relies on. Gemini 3.7 Flash stays in the picker as the previous-generation option. Gemini 2.5 Flash has been retired and anyone still using it is migrated to 3.8 Flash automatically. Gemini 2.5 Flash Lite remains the cheapest option and is not migrated.

Claude Fable 5.1 is now in the model picker at the same price as Fable 5 ($10 / $50 per million tokens) — but cache reads drop from $1 to $0.25 per million, which is where most of the cost of a long conversation goes. Anthropic reports better long-session reasoning and notably better reading of dense charts and tables in PDFs. Fable 5 stays available as the previous generation, and Sonnet 5 is still the default. Fable 5.1 needs 30-day data retention on your Anthropic account, same as Fable 5.

Meta Ads is now available as a connected data source on Pro and Business plans. Connect an ad account and ask your AI about Facebook and Instagram campaign performance — spend, results, audiences, placements, and individual creatives, with GA4 as the verification baseline when both are connected.

• 4 data tools: Campaign performance, performance over time, audience and placement breakdowns, and per-ad performance
• 4 new skills: Meta Ads Account Audit, Meta Tracking Health Check, Meta Lead Attribution Check, and Meta vs Google Ads Channel Mix
• Dashboard widget: Spend, impressions, clicks, reach, and top campaigns at a glance
• Read-only: OAuth with the read-only ads_read permission — we never create or modify campaigns

Connect your own AI assistant to your AIdata properties. Every team member gets a personal MCP URL that works in claude.ai, Claude Desktop, and any MCP-compatible client - covered by the AI subscription you already have, no API key needed.

• One personal URL, all your properties: spans every property you can access, across all your teams, with your permissions enforced per property
• Full data toolkit: the same analytics, ads, CRM, and annotation tools as the built-in chat, plus your skills - in the prompts menu and runnable by the AI itself
• Writes where you’re allowed: your AI can create and update annotations and save team memories, attributed to you - everything else is read-only
• Off by default: team owners enable External AI Access in Team Settings (Pro and Business); disabling it instantly cuts external access, and you can regenerate your personal URL at any time

Core Web Vitals is now a connected data source on every plan, including Free — real-user speed data from Google’s Chrome UX Report, with no account to link and no API key to create. Because the data is public, you can track competitors’ vitals right next to your own.

• Dashboard widget: All four metrics (LCP, INP, CLS, TTFB) at once — six months of weekly trends, phone/desktop toggle, and green “good zone” bands
• Competitor comparison: Track up to 5 competitor sites and see whose line is lower
• AI analysis: Ask your AI about any site’s vitals history, trends, and how you stack up against the competition
• Zero setup: Enable the connection, confirm your website address, done

Google’s Gemini lineup in AI Data Stream has been refreshed: Gemini 3.7 Flash replaces the Gemini 3 Flash preview as the recommended model, Gemini 3.1 Pro replaces the retired Gemini 3 Pro preview, and Gemini 3.5 Flash Lite joins as the fastest option. Anyone on the old preview models is migrated automatically. Gemini 3.7 Flash has introductory pricing of $0.75 / $3.75 per million tokens through 31 December 2026 (then $1.50 / $7.50). The Gemini 2.5 models stay available, including 2.5 Flash Lite as the cheapest option.

Good news for your API bills. Anthropic has made Claude Sonnet 5’s $2 / $10 per million token launch price permanent, cancelling the increase to Sonnet 4.6 rates planned for 1 September 2026. OpenAI cut GPT-5.6 prices on 30 July: Terra is 20% cheaper ($2 / $12 per million tokens) and Luna 80% cheaper ($0.20 / $1.20); Sol is unchanged. Because AI Data Stream is bring-your-own-key, these savings reach you directly - nothing to change on your account.

LinkedIn Ads is now available as a connected data source on Pro and Business plans. Connect an ad account and ask your AI about B2B campaign performance — including LinkedIn’s unique professional demographics: which job functions, seniorities, industries, and companies your ads actually reach.

• 4 data tools: Campaign performance, performance over time, professional audience breakdowns, and per-ad performance
• 3 new skills: LinkedIn Ads Account Audit, LinkedIn B2B Audience Insights, and LinkedIn Lead Attribution Check
• Read-only: OAuth with read-only reporting scopes — we never create or modify campaigns

Google Tag Manager is now available as a connected data source on Pro and Business plans. Connect a container and ask your AI to audit tags, triggers, and variables — and debug why tracking broke by pairing GTM with GA4.

• 6 data tools: Containers, tags, triggers, variables, entity detail, and workspace status
• 3 new skills: GTM Container Audit, GTM Change Monitor, and Conversion Tracking Debugger
• Read-only: OAuth with a read-only scope — we never modify or publish your container

Anthropic’s new frontier model, Claude Opus 5, is now the flagship Anthropic option — same price as Opus 4.8, with major gains on agentic and data-analysis tasks. Opus 4.8 stays available as the previous generation so you can compare. Opus 4.7 has been retired; if it was your default, you move to Opus 5 automatically.

Report templates have a new name: skills. A skill is a reusable analysis recipe - expert methodology built in - and running one produces a report, exactly as before. Nothing about how they work has changed; existing templates, schedules, and report history all carry over untouched.

• Skills catalog: 19 ready-made skills across audits, monitors, analysis, CRM, and ecommerce
• Same reports: Running a skill still produces a shareable, schedulable report
• Scheduled Reports unchanged: Schedules keep their name and keep delivering reports by email

Support can now only sign in to your account with your explicit permission. The new Settings → Support Access page puts you in control:

• Off by default: Nobody can access your account until you grant it, confirmed with your password
• Auto-expiry: A grant lasts 7 days and can be revoked at any time
• Full transparency: Every support sign-in and sign-out is recorded and shown to you on the page
• Secrets stay secret: API keys and connection credentials are never visible in support sessions

Reports can now run themselves. Put any report template on a schedule - weekly, monthly, or every N days - and the finished report lands in your team’s inbox, ready to read. The monitor templates that always promised “weekly” finally deliver on it, automatically.

• Any template, any cadence: Weekly on a chosen day, monthly on a chosen date, or every N days - at the time and timezone you pick
• Email delivery: Recipients get a link to the finished report the moment it’s ready
• Team-visible: Scheduled runs are shared automatically and badged in the reports list; template cards show their cadence
• Self-healing: A schedule that fails repeatedly pauses itself and emails you why, instead of burning tokens forever
• Paid plans: Active schedule limits scale with your tier; free plans can still run every report manually

The AI now remembers. When you tell it something worth keeping - your fiscal year, a campaign naming rule, which KPI actually matters - it saves the fact and references it in every future conversation on that property. Memories are shared with your whole team, fully visible, and editable at any time.

• Automatic saving: The AI saves durable facts you state in chat, and you can ask it to remember (or forget) anything
• Team-shared: One memory pool per property, so everyone gets consistent answers
• Fully editable: View, edit, and delete every memory from the new Memories page
• Incognito chats: Toggle incognito on any conversation for a fresh take that neither reads nor saves memories
• Per-property control: Turn memory off entirely in property settings
• Free plan: 20 memories per property; unlimited on paid plans

DataForSEO Backlinks is now available as a connected data source. Connect it with your own DataForSEO API key and ask your AI about your backlink profile — referring domains, anchor text, spam scores, and broken links — an independent, external view of your link graph alongside Search Console and Analytics.

• 4 data tools: Backlink summary, backlinks list, referring domains, and anchors
• Auth: Bring your own DataForSEO API key (login + password); requests are billed to your own account
• Cost control: Set a maximum rows-per-request ceiling when you connect
• Use case: Audit link quality, find toxic or broken backlinks, and analyse anchor text distribution

Claude Fable 5, Anthropic’s most capable model, is now in the picker as a premium option ($10 / $50 per million tokens). It’s not the default — Sonnet 5 still is — but it’s there for the hardest questions, now that OpenAI’s top-tier Sol sits alongside it. Note Fable 5 needs 30-day data retention on your Anthropic account.

GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna are now available as OpenAI model options, replacing GPT-5.4 and GPT-5.4 Pro. If you used GPT-5.4 you move to the identically priced GPT-5.6 Terra automatically - no cost change. Select a model any time from your AI settings. Update: OpenAI cut GPT-5.6 prices on 30 July 2026 - Terra is now 20% cheaper ($2 / $12 per million tokens) and Luna 80% cheaper ($0.20 / $1.20); Sol is unchanged.

The AI assistant now stores fetched data in private per-conversation SQL tables and computes sums, averages, rankings and comparisons with SQL — so every number in an answer is exact, not estimated. This also enables cross-source questions (join Search Console, Analytics and Ads data by date or page) and raises the per-fetch limit from 1,000 to 5,000 rows. Stored data is read-only for the AI, private to your team, and cleaned up automatically.

Anthropic’s new Claude Sonnet 5 is now the flagship Sonnet option — the best balance of speed and capability. Sonnet 4.6 stays available as the previous generation so you can compare the two on your own analytics. Update: the $2 / $10 per million token launch price is now permanent — Anthropic cancelled the planned increase to Sonnet 4.6 rates, so Sonnet 5 is a third cheaper than its predecessor.

Anthropic’s latest frontier model, Claude Opus 4.8, is now the flagship Anthropic option — same price as before, with sharper judgement on agentic tasks. Opus 4.7 stays available as the previous generation so you can compare. Opus 4.6 has been retired; if it was your default, you move to 4.8 automatically.

Give the AI always-on background for each property. In Property → Settings, add a context block covering what the site is and who it’s for, your primary conversion, traffic to exclude, and key competitors. It’s automatically included at the start of every conversation, with no lookup required.

• Always in context: Appended to every chat, alongside your chosen prompt preset
• Composes with personas: Keep shareable team presets like “SEO Specialist” and layer property-specific facts on top
• Markdown editor: The same editor as prompt presets, with live preview
• Pairs with the knowledge base: Property context for always-relevant facts, the knowledge base for searchable detail

OpenAI has deprecated the o3 and o4-mini reasoning models. Both have been removed from the model picker. If you had either set as your default, your AI provider will automatically use GPT-5.4 (for o3) or GPT-5 Mini (for o4-mini) instead — reasoning is now built into the GPT-5 family via the reasoning effort setting (low/medium/high).