Hi everyone, this is Neo.

No algorithms today. Something more dangerous: where the numbers in your dashboard actually come from.

The trigger was an unusual piece on Search Engine Journal. Duane Forrester (who builds measurement tooling in this space and disclosed that interest up front) laid out three predictions for 2027, then spent the entire back half explaining why the fourth one — “you won’t be able to check whether the first three were right” — matters far more.

My takeaway: the value isn’t the predictions. It’s a framework for judging whether a number can be trusted, which is exactly what independent site owners are missing.

One: the three predictions (briefly)

Prediction 1: search ad money keeps growing, but the share earned by sending people to your site keeps shrinking.

In Alphabet’s Q2 2026 report (July 22), Google Search revenue came in at roughly $63.3 billion, up 17% year over year — more than a year after AI Overviews went wide and AI Mode rolled out globally. The cannibalization thesis did not show up in that number.

Read the next line, though: that 17% was the first deceleration in six quarters, and the CFO flagged a tougher comparison in Q3.

The author’s framing: both things stay true at once — total search ad revenue keeps growing, while the portion earned by handing a user off to someone else’s site keeps shrinking. The money didn’t leave Google. It left the part of the arrangement that used to pay you.

(What would disprove it: two consecutive quarters of negative search revenue growth with no obvious macro cause.)

Prediction 2: ranking and earning stopped being the same job.

The old arrangement was clean: Google ranked you, the ranking sent traffic, and Google sold ads against the intent that traffic represented. Everyone in that chain needed the click, which is exactly why the click got measured.

The answer layer broke the arrangement without breaking the revenue: most searches now end without a click to an independent site, and the ad business grew anyway. Those two facts side by side are the whole story.

So: by 2027, the surface that ranks and the surface that earns are no longer the same surface. The ranked results page probably persists as a legacy interface — advertisers and two decades of habit require it — while the consequential product decisions happen elsewhere.

And the operational consequence: nobody wins outright, so most organizations end up running parallel — optimizing a ranked surface and an answer surface at the same time, with different assumptions and one shared budget. That’s not a transitional state. For most companies, it’s the steady state.

Prediction 3 / 4: by the end of 2027, this industry will make more decisions on inferred data than at any point in twenty years — and report those decisions with the confidence it used to reserve for data it actually counted.

The first three are parlor talk. The fourth is cold water.

Two: the core distinction — counted versus inferred

The definitions are restrained and genuinely useful:

	A counted number	An inferred number
Source	An event observed on infrastructure you or your vendor control	A sample extrapolated to a population nobody can fully enumerate
Can you check it?	Yes — you can go find the underlying events	No — you can only ask the vendor about their method
What you can do when it surprises you	Locate, fix, recompute	Doubt, negotiate, or accept

The crucial part is the second and third rows: this is not a good-data/bad-data ranking.

The author is explicit: panel work answers questions server logs cannot touch. It’s a difference of method, not of quality.

But the consequence is hard: when a number surprises you, its class determines what you can do about it.

• Counted: open the logs, find the events, recompute it yourself;
• Inferred: ask about the method. And if no method was published, you can’t even determine whether it’s wrong.

Three: example one — GA4’s two most important numbers look identical

His example is Google Analytics, precisely because it contains both classes and displays them identically.

• Session counts — close to counted;
• Channel attribution — inferred;
• Direct — where that inference fails quietly rather than loudly.

Nothing in the interface tells you which column you’re looking at. Same font, same chart, same export, same two decimal places.

Then two things happened in 2026.

First: on May 13, GA4 added an “AI Assistant” channel to the default channel group.

A useful change, long requested. It also arrived with:

• the full list of recognized AI sources unpublished;
• no statement about how that list would be maintained;
• the channel definitions page not yet updated to describe it.

Within weeks, the set of platforms being reported diverged from the set named at launch. The announcement named ChatGPT, Gemini, and Claude; by June, the live documentation listed ChatGPT, Gemini, Deepseek, Copilot, and Grok. Claude — named at launch — isn’t in the current published definition. Perplexity was never in it (its visits still land in Referral). And Google’s own AI Overviews and AI Mode clicks are routed to Organic Search.

More importantly: the channel counts forward only. Nothing before May 13 gets reclassified, ever.

The author’s line on it is the one to keep: this is a counted number whose definition moved without anybody being told, inside a tool that reports the result to two decimal places.

Second: on September 1, standard reports showed zero traffic across a very large number of properties, while Realtime kept showing users the whole time.

Unpack that: collection was working. Reporting was not. Those are separate systems that fail separately.

• If you understand that, you can explain the zero to an executive in one sentence;
• If you don’t, you spend a day re-checking tags that were never broken.

A week later, there was still no public confirmation that the affected data would be restored or backfilled.

The author’s assessment is polite and precise: neither of these is a scandal. Both are ordinary. That is what makes them worth learning from.

Four: example two — a claim accurate for one week, inherited for a year

This is the part that got me.

When AI Mode launched in May 2025, its citation links carried a noreferrer attribute, which stripped the referrer and dropped those clicks into Direct. Practitioners caught it fast, John Mueller said publicly it looked like a bug on Google’s side, the attribute came off within days, and named practitioners confirmed traffic classifying as organic again.

Then, through 2026, a substantial volume of vendor content has asserted that AI Mode strips attribution by architectural design, that it’s deliberate rather than accidental, and that no workaround exists. Almost none of it references the correction.

In his words, roughly: a claim that was accurate for about a week was inherited forward for a year and hardened into permanent design intent along the way.

Then this line, which I’d print and tape to the wall:

Whether the attribute sits on those links today is beside the point. The point is that almost nobody repeating the claim knows either way — because settling it means opening the page and reading the markup, which takes about ninety seconds.

They are not reporting an observation. They are repeating one.

Neo’s take: this describes the most common error of the AI era — mistaking someone else’s conclusion for your own data. You open ten vendor articles, eight repeat the same sentence, and in your head it graduates from “one person’s claim” to “industry fact.” It may never have been checked once.

Five: supporting evidence — two things can be true at the same time

Prediction 1 (money decoupling from traffic) has a good companion data point, and it also shows how hard “methodology” is.

SparkToro’s report, built on Similarweb’s U.S. clickstream panel: for every 1,000 U.S. Google searches, only 232 clicks reach what it calls the open web. 68% of searches ended without a click; 32% produced one — of which 66% went to the open web, 27% to Alphabet properties (YouTube, Maps, AI Mode), and 6% to paid ads.

Against the 2024 figure (360 clicks per 1,000 searches, from the Datos panel), the share of searches producing at least one click fell from 41% to 32%.

But notice the layer underneath: those numbers come from different panels in different years — 2016 and 2019 from the defunct Jumpshot panel, 2024 from Datos, 2026 from Similarweb. The author himself calls the cross-year chart “a bit of apples and oranges.”

Meanwhile Google’s framing is entirely different: Liz Reid has said organic click volume is “relatively stable” and that AI Overviews mostly remove “bounce clicks.” She’s talking about total volume. The panel measures clicks per search. Query growth can offset a falling click rate, which means both claims can hold.

That spot — where two contradictory numbers are both technically true — is the most dangerous real estate in any report.

Six: so what do you do? He refuses to give a checklist, and gives four questions instead

This section is the most practical and the most counterintuitive: he considered publishing a validation test, and the honest conclusion was that no definitive version can be built.

Four solid reasons:

1. There’s no ground truth to score against, which means a test you run can never fail;
2. Provenance is frequently undisclosed, so any step that says “identify the population” fails at the vendor’s discretion, not your competence;
3. Most real numbers are hybrids — counted, then modeled inside the same figure;
4. Any classification you make expires the moment a vendor changes method, usually in release notes nobody reads.

So, in his words: publishing a checklist here would reproduce the exact error the rest of the piece is describing.

What he gives instead is four questions to put to any number before it goes into a deck:

Question	Why it matters
What population does this describe, and can it be named?	If you can’t enumerate it, you can’t judge how representative it is
Was the value observed or extrapolated?	Determines what you can do when it’s wrong
What would move this number if the world outside stayed exactly the same?	Surfaces definition changes and pipeline edits
When the system can’t determine an answer, where does that session go?	This is where Direct lives

That last one deserves repeating: if the answer is “it drops out of the dataset,” that’s clean handling. If the answer is “it gets filed somewhere quiet,” someone will later read that bucket as a finding. Direct traffic has spent the past two years carrying exactly that role, and it was never built to.

Neo’s take: the beauty of these four questions is that they don’t produce a score. They produce either an answer or a silence — and the silence is itself the finding. Even a badly applied version forces a vendor to answer something they ought to be able to answer.

Seven: three things independent site owners can start today

You don’t need a full measurement audit. These three are doable this week.

1. Build your own counted layer

Only data you control and can verify belongs here:

• Server / CDN logs — the only complete event record that’s yours;
• Your own Search Console properties — impressions, clicks, queries, unfiltered by a third party;
• Your own landing page parameters and UTM conventions — you decide when the definition changes.

Put these side by side with GA4 and your third-party tools instead of substituting one for the other. The discrepancies will surface on their own.

2. Label every number in your report

It doesn’t need to be sophisticated. Add one column to your internal reports: is this number observed or extrapolated?

The moment you do, you’ll notice something immediately: a lot of the numbers that walk into decision meetings are extrapolated, but presented as counted.

3. When a number looks wrong, ask the three questions first (the September 1 lesson)

Is the collection broken? Is the reporting broken? Or did the definition change?

The diagnostic paths are completely different:

• Collection broken: fix tags, tracking, consent management;
• Reporting broken: change nothing, wait for the backfill (and confirm there will be one);
• Definition changed: find the release notes and log it in your change register.

Most people start by fixing tags — and burn a day on a pipeline that was never broken.

One small habit while you’re at it: annotate every definition change in GA4 (or whatever tool you use). Two minutes today saves an entire day next year — and stops the next person from reading a channel’s sudden appearance as a performance spike.

The end

I pulled this piece out because of one thing: it draws a distinction that rarely gets stated clearly. Whether a number can be checked matters more than what the number says.

The most valuable skill for 2027 probably isn’t GEO or prompt engineering. It’s being able to tell, at a glance, which class of number you’re looking at.

Because over the next year you’ll see more and more beautiful, two-decimal-place, chart-ready numbers that cannot be checked.

And they look exactly like the ones that can.