The 2026 attribution reset: four changes that break your Q4 numbers

Four separate measurement changes landed in 2026, and all four come due in the same fortnight of Q4. Your Google Ads attribution model was migrated without you choosing it. GA4 grew a new channel for traffic arriving from chat assistants. Shopify switched on storefronts that take orders inside someone else's product. And the share of revenue from AI-referred sessions stopped being a rounding error.

Each change on its own is survivable. Together they mean the year-over-year comparison you are about to run in November is measuring a different world from the one it is comparing against. Here is what actually changed, from the primary documentation, and what a Q4 dashboard should show instead.

Google Ads: your attribution model was changed for you

Google's attribution model documentation now states that "the first click, linear, time decay, and position-based attribution models are no longer supported by Google", and that "conversion actions that used the deprecated attribution models have been upgraded to use data-driven attribution".

If any conversion action in your account ran on one of those four models, the numbers on either side of the migration are not comparable. Per Google's data-driven attribution page, DDA is now the default for most conversion actions, and Google recommends "at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period" for it to work well. Small accounts scraping that threshold should treat channel-level credit as directional.

GA4: the AI Assistant channel, and the hole it punches in year-over-year

GA4 now documents a channel for assistant traffic. Per Google's default channel group definitions, "AI Assistant is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok", matched where the medium is exactly ai-assistant.

Two consequences for Q4. First, sessions that previously fell into Referral or Direct now appear in a channel that did not exist in your comparison period, so a Referral decline this November may be a reclassification rather than a loss. Second, Google's list does not include every assistant — traffic from anything it does not recognise lands in Unassigned, which the same page defines as "the value Analytics uses when there are no other channel rules that match the event data". A rising Unassigned line is a measurement gap, not noise.

Which of your GA4 numbers are estimates

Worth knowing before a number goes in front of a board. GA4's behavioural modelling documentation sets the thresholds for modelling to kick in: "at least 1,000 events per day with analytics_storage='denied' for at least 7 days" and "at least 1,000 daily users… 'granted' for at least 7 of the previous 28 days".

Modelled data is also excluded from "audiences, user explorer explorations, segments with sequences, retention reports, predictive metrics, or data exports like BigQuery" — which is why a BigQuery total and a GA4 report total legitimately disagree. And attribution keeps moving after the fact: Google's attribution documentation notes that "attributed conversion data for each channel can still be updated for up to 12 days after the conversion is recorded". Reading BFCM channel performance on the Monday is reading a draft.

Meta: the one gap you can still close yourself

Platform-versus-store discrepancies are mostly structural, but duplicate counting is not. Meta's deduplication documentation states that a Meta Pixel's eventID must match the Conversions API's event_id, and that where events do not differ, "we generally prefer the event that is received first".

If browser and server events do not carry matching IDs, purchases are counted twice and your platform ROAS is inflated at exactly the moment you are deciding where to put Q4 budget. Checking this takes an afternoon and is the highest-value measurement fix available before November.

Orders that never touch your storefront analytics

The newest gap is structural. Shopify announced on 24 March 2026 that Agentic Storefronts went live, with the detail that "orders flow into the admin with ChatGPT referral attribution, so merchants can see exactly where sales came from". Its explainer adds that the feature "is automatically activated for eligible merchants", with per-channel toggles for direct checkout.

Read that twice. The order reaches your admin. It may never produce a session on your storefront, so your on-site analytics can be structurally blind to a growing revenue line even while the orders arrive.

The volume is no longer trivial. Shopify's AI search insights, published 11 May 2026, report that in Q1 2026 AI chatbot referral sessions "grew more than 8x year-over-year" and AI-referred orders "grew nearly 13x", converting at "nearly 50% higher rates than organic search" with "14% higher average order values". The behaviour differs too: "more than half of AI-referred sessions start on a product detail page, compared to about 20% for organic search".

On the protocol side, OpenAI's commerce documentation is explicit that "OpenAI is not the merchant of record in the Agentic Commerce Protocol" — you keep your payment provider, supply a product feed with daily snapshots, and publish order.created and order.updated webhooks. The reconciliation duty stays with you.

Reconcile in the right order

The fix is not a better attribution tool. It is an order of operations. Start from store records — orders, refunds, net revenue from the platform that took the money. Then bring in platform claims as claims, labelled by source and attribution model, and expect them to overstate. Then reconcile blended: total net revenue over total ad spend, which is the only figure nobody can inflate.

Then state the uncertainty explicitly. Which numbers are observed, which are modelled, and which will still move for twelve days. A dashboard that hides that distinction produces confident decisions on soft numbers.

The Q4 dashboard that survives all four changes

Six rows, refreshed daily through BFCM week. Net revenue and orders from store records, not platform reporting. Blended MER, spend divided into total revenue. Channel claims with the attribution model named next to each. AI Assistant sessions and agentic orders as their own line, so the blind spot is visible rather than absorbed into Direct. Unassigned as a health metric, because a rise means the channel map is out of date. And a plain note on each panel saying whether the figure is observed, modelled, or still settling.

What to fix this week, in order

Ranked by how much the number moves against how long the fix takes.

• Verify pixel and server deduplication. Confirm eventID and event_id match on purchase events. Inflated ROAS misdirects real budget.
• List every conversion action that was migrated to data-driven attribution and annotate the date in your reporting, so November is not compared against a differently modelled October.
• Add AI Assistant and Unassigned to the Q4 dashboard before the volume arrives, not after someone asks why Referral fell.
• Check whether Agentic Storefronts is active on your store and decide the per-channel toggles deliberately. It activates automatically for eligible merchants.
• Confirm your product feed and order webhooks are publishing if you are selling through the Agentic Commerce Protocol; the order data you reconcile against depends on them.
• Write the word "modelled" on every panel that is modelled. It takes five minutes and prevents the worst category of Q4 decision.

Why the assistant should explain the numbers and never calculate them

The obvious shortcut is to hand the whole mess to a language model. The published evidence says not to hand it the arithmetic. FinSheet-Bench, posted on 7 March 2026, found the best model scored 82.4% accuracy, with per-file performance ranging from 48.6% to 86.2%, and concluded that "no standalone model achieves error rates low enough for unsupervised use in professional finance". Its recommendation is to build architectures that "separate document understanding from deterministic computation".

That is the rule BriefSieve is built on: code computes every figure from the synced data and the assistant explains it, cites the sync it came from, and is not allowed to do arithmetic on money or metrics. It is the same discipline we apply everywhere — the reasoning is in why unowned AI output costs more than it saves.

Four changes, one quarter. The teams that come out of BFCM with usable numbers will be the ones that fixed deduplication, added the AI lines to the dashboard, and wrote "modelled" next to the figures that are modelled.