On September 1, 2026, GA4 standard reports showed zero traffic across a large number of properties. Realtime kept showing active users the entire time. Collection was working. Reporting wasn't — and nothing in the interface told anyone which system had failed.
That gap is the problem this piece solves. GA4 mixes two different kinds of numbers — ones observed directly from your own tracking, and ones estimated by extrapolating from a sample — and displays them in the same font, the same chart, the same export, rounded to the same two decimal places. If you're presenting AI-referral or Direct traffic numbers to a client or executive, you need to know which kind you're looking at before you put it in a deck.
This is a workflow, not a diagnosis of Google's roadmap. It won't tell you whether AI Overviews are cannibalizing your traffic or whether your rankings are healthy. It will tell you which of your own numbers are checkable and which aren't — and what to ask before you cite either one.
Why Your GA4 Numbers Got Less Trustworthy This Year
Three things happened to GA4 in 2026 that practitioners are still reporting around, often without realizing the ground shifted.
May 13: The AI Assistant channel arrived quietly changed. Google added an AI Assistant channel to the default channel group — a long-requested update. But it shipped with the full list of recognized AI referrers unpublished, no statement on how that list would be maintained, and the channel definitions page not yet updated to describe it. Within weeks, the set of platforms being reported differed from the set named at launch. The channel counts forward only; there's no backfill of history, so you can't retroactively compare this quarter to last.
September 1: Reporting broke, collection didn't. Standard reports zeroed out across many properties while Realtime kept working. A practitioner who understands that collection and reporting are separate systems can explain the zero in one sentence. A practitioner who doesn't spends a day re-checking tags that were never broken. As of the following week, there was no public confirmation that the affected data would be restored or backfilled.
May 2025 (still being repeated in 2026): a one-week bug became permanent folklore. When Google launched AI Mode, its citation links carried a noreferrer attribute that stripped the referrer and dropped those clicks into Direct traffic. Practitioners caught it fast. John Mueller said publicly that it looked like a bug on Google's side, and the attribute came off within days — named practitioners confirmed traffic classifying as organic again shortly after.
A year later, a substantial amount of vendor content still asserts that AI Mode strips attribution by architectural design, that it's deliberate, and that no workaround exists — without referencing the correction. An observation that was accurate for about a week got inherited forward and hardened into permanent design intent. Whether the attribute is on those links today is a separate question from whether the people repeating the claim have checked. Most haven't, because checking takes about 90 seconds of reading page source and repeating a claim takes zero.
The Distinction That Actually Matters: Counted vs. Inferred vs. Hybrid
A counted number comes from an event observed on infrastructure you or your vendor control. A inferred number comes from a sample extrapolated to a population nobody can fully enumerate. Neither is automatically better — panel data answers questions server logs can't touch — but only one of them can be independently checked when it surprises you.
Most real-world GA4 metrics are hybrid: counted data that gets modeled or bucketed inside the same figure. Channel attribution is the clearest example. The underlying session may be real and counted; which channel it gets attributed to is inferred, and Direct is where that inference fails quietly rather than loudly.
Use this three-way classification, not a binary "trust it / don't trust it" label:
- Counted — tied to a specific observed event you can trace back to raw data.
- Inferred — extrapolated from a model, a sample, or an unpublished matching list.
- Hybrid — a counted event wrapped in an inferred label (most channel and source/medium data falls here).
Step-by-Step: Audit Your Own GA4 Property Before It Goes in a Deck
Run this against the property you're about to report on. It takes 20-30 minutes and produces a classification you can defend if someone asks "how do you know?"
Step 1: Pull raw session counts, not channel-grouped totals, first. In Explore or the API, pull total sessions/users for the date range without any channel breakdown. This number is close to counted — it comes from events your tag actually fired. Write it down separately from anything that follows. This is your anchor number.
Step 2: Check whether the reporting outage window overlaps your date range. If any part of your reporting period falls in a window when standard reports were known to zero out, cross-check that segment against Realtime exports or BigQuery raw event data if you have it linked. Don't average an outage window into a trend line without flagging it.
Step 3: Inspect your channel definitions against the current default group. Go to Admin > Data display > Channel groups and look at what's actually defined for AI Assistant / AI referral traffic. Compare it against what you assumed was included last quarter. If the definitions page doesn't list every platform you think is being captured, assume the list has changed and note the date you checked.
Step 4: Segment Direct traffic and ask what's hiding in it. Pull Direct as its own segment and look at landing pages, not just volume. A spike in Direct sessions landing on pages that also rank in AI Overviews or get cited in AI Mode is a signal — not proof — that referral-stripped AI traffic is being misclassified as Direct. This is your first hybrid-metric flag.
Step 5: Spot-check AI referral links for the noreferrer attribute yourself.
Don't cite the claim secondhand. Open a live AI Mode or AI Overview result that cites your site, right-click the citation link, and choose "Inspect" or view page source. Search the anchor tag for rel="noreferrer" or rel="noopener noreferrer". Do this on three to five separate citations, on different days if you can, since Google can change this without announcement. Record what you find with a screenshot and a date — that's your evidence, not a vendor's blog post.
Step 6: Classify each metric before it leaves your spreadsheet. For every number going into the deck, tag it counted, inferred, or hybrid. If you can't confidently tag it, that's information too — mark it "unverified" rather than guessing.
Step 7: Note the collection date on every classification. Channel definitions, referrer lists, and attribute behavior can change without notice. A classification from three months ago may no longer be accurate. Date-stamp your audit so you know when it needs to be redone.
The Question Script: Run This Before Citing Any Vendor or AI-Traffic Number
Before a number from GA4, a third-party tool, or a Google rep goes into a client-facing report, ask:
- What population does this describe, and can it be named? If the answer is vague ("AI platforms" without a list), treat the number as inferred, not counted.
- Was this value observed or extrapolated? Ask directly. If they can't answer, assume extrapolated.
- What would move this number if nothing in the real world changed? A method change, a new referrer added to a list, a backend reclassification — any of these can shift a number with zero underlying behavior change.
- What happens to sessions the system can't classify? Do they drop out of the dataset, or do they get filed somewhere — like Direct — where someone later reads them as a finding?
None of these produce a score or a pass/fail grade. They produce either an answer or a silence, and the silence is itself useful information: it tells you the number came from a black box, not a checkable process.
What This Audit Can't Do
Be upfront about the limits when you present this workflow internally or to a client:
- There's no ground truth to check against. No independent, authoritative baseline exists for AI-referral traffic, so this audit can confirm internal consistency, not external accuracy.
- Methods change without notice. The May 13 channel launch and the noreferrer attribute's disappearance both happened without advance warning. Any classification you make today can be invalidated by a change you won't hear about until you check again.
- This doesn't tell you anything about rankings or business outcomes. It tells you whether a traffic number is checkable. It says nothing about whether your content is ranking well, whether AI Overviews are helping or hurting you, or whether the underlying strategy is working.
- Most numbers are hybrids, and hybrids resist clean labels. Expect to mark a good portion of your metrics "hybrid, use with caveat" rather than a clean counted or inferred. That's an accurate answer, not a failure to finish the audit.
Before You Present the Number
The gap between what GA4 displays and what it actually measured isn't a bug you can file a ticket against — it's the current state of the tool. Treat every AI-referral and Direct-traffic figure as unverified until you've run it through the classification above, and say so out loud when you present it. "This number is counted" and "this number is our best inference" are both defensible things to tell a client. Presenting an inference as a count, because the interface didn't tell you the difference, is the thing that gets a report torn apart later.
Source: 3 Predictions For 2027, And Why You Won't Be Able To Check Them, Search Engine Journal, originally published on Duane Forrester Decodes, September 17, 2026.
