The Two Scoreboards: Why Your Content Looks Dead in Analytics and Alive in ChatGPT

The Two Scoreboards: Why Your Content Looks Dead in Analytics and Alive in ChatGPT

Data Study #1. The Starr Conspiracy. Published 2026-10-03.

Measurement window 2026-07-03 to 2026-10-01, 90 days. Sources: Google Search Console (domain property), Google Analytics 4, Ahrefs Brand Radar weekly snapshots, and our own content database. Every figure comes from one script that reads those sources directly. Methodology and limits are stated in full at the end, including the places where our own earlier drafts were wrong.

---

The short version

We pointed an autonomous content engine at our own domain and let it publish. Over ninety days that corpus earned 684,474 impressions in Google and 995 clicks. A click-through rate of about one tenth of one percent. Of 10,247 queries where we hold a page-one position, 10,225 earned no clicks at all.

For about a week we read that as a failure and discussed shutting the program down.

We were reading the wrong scoreboard. On the second one, the same corpus was the most-cited domain in our category across ChatGPT, Gemini, Perplexity and Copilot in nine of the last thirteen weeks, ahead of LinkedIn and Reddit. A channel that did not exist in March now sends people straight to the answer pages rather than the front door.

This study reports both, including the parts that do not flatter us, and one we found after we first wrote it: many of the statistics on those cited pages were wrong.

---

What we actually built

An autonomous engine that generates, validates, internally links and publishes across twelve content formats, grounded in a structured go-to-market model rather than free-running generation. No human wrote the bodies.

Published documents live today: 1,362.

Publishing velocity, all time:

Two things to note, because both cut against the easy version of this story.

The sprint ended. Nearly all of the volume landed in a sixty day window in April and May. What followed is maintenance, not acceleration.

And the corpus is smaller than its peak. We consolidated duplicate and near-duplicate pages and archived the losers, retired a news feed that earned almost no readers, and in October retired eighteen more pages that were broken or left with nothing to say by the cleanup described below. That is deliberate. A library that competes with itself for the same query is not an asset. Any study telling you its content corpus only ever grows is describing a decision nobody made.

---

Scoreboard 1: traditional search. The rankings arrived. The clicks did not.

The first thing to kill is the idea that machine-written content cannot rank. Ours ranks.

Twenty-two clicks, spread across 10,247 page-one rankings.

A page-one position is conventionally worth somewhere between five and thirty percent click-through depending on rank. Ours returns close to nothing. The five pages that earned the most impressions in the window:

The third row is the study in one line. Thirty-two thousand impressions. No clicks. That page is being shown to people constantly and read by almost none of them, because the answer is delivered above the link.

Where the impressions actually go

Search Console reports the query for roughly 51% of impressions and 26% of clicks. Google anonymises the rest. Inside the visible slice:

Three findings sit in that table, and they are the ones we would want a competitor to miss.

More than a quarter of our query-visible impressions are not people. 99,655 impressions came from machine queries and produced zero clicks. Counting those as audience would have inflated every engagement metric we report.

Question-format queries are the AI fan-out surface, measured directly. 5,418 question queries, 50,110 impressions, two clicks. This is what it looks like when a retrieval layer is reading you: constant retrieval, essentially no click-through.

Of the human clicks we can see, 82% are branded. 214 of 261. Google is mostly delivering people who already know our name. Non-brand discovery, the thing content marketing is supposed to buy, returned 47 clicks in ninety days.

If this were the only scoreboard, the rational move would be to kill the program. We came close.

---

The turn

Here is what changed our minds, and it is not a sentiment. It is that the clicks did not vanish. They changed form, and the form they changed into is measurable if you go looking for it in places a standard SEO report does not cover.

Three independent sources, same corpus, same ninety days.

---

Scoreboard 2, receipt 1: machines are reading the site at scale

Of 28,705 direct-traffic sessions in the window, 11,101 came from data-centre locations with engagement rates under 20%, most of them far under:

That is the signature of an automated agent rendering a page, not a person reading one. A word on method, because this is the sort of number that deserves suspicion: we filtered on engagement, not on city name alone. Des Moines showed 2,092 direct sessions at 80.3% engagement. Those are people, and counting them would have inflated this figure by almost a fifth. They are excluded. Boardman, a data-centre town that met the test in our first draft, sat just above the line this time at 20.2% engagement, so it is excluded too. The rule decides, not the result.

This is a floor, not a count. Our analytics only sees agents that execute JavaScript. The declared crawlers, GPTBot and ClaudeBot and PerplexityBot and their peers, never appear in analytics at all and are ingesting the corpus on top of this.

You do not get cited by a model that has not read you. In 2026 being read looks like direct traffic with terrible engagement, which is exactly the pattern most analytics training teaches people to dismiss.

---

Scoreboard 2, receipt 2: we are the most-cited domain in our category

Ahrefs Brand Radar tracks a fixed set of category prompts across ChatGPT, Gemini, Perplexity and Copilot, and records which domains the assistants cite. We have twenty-one weekly snapshots.

The Starr Conspiracy ranked first by responses in nine of the last thirteen weekly snapshots, including seven in a row from 2026-08-03 to 2026-09-14 and the most recent, 2026-09-28. On 2026-09-21 redbranchmedia.com took first and we were second. In early June we ranked seventh. At the most recent snapshot the nearest non-platform domain is redbranchmedia.com at 8 cited pages and 22 responses, against our 23 and 25. LinkedIn and Reddit both sit below us.

Here is the full series, including the parts that do not trend:

We are publishing the whole series rather than two endpoints for a specific reason. The cited-page count is noise. It runs 15, then 33, then 18 inside a fortnight. An earlier draft of this study claimed a "50 percent increase in citation footprint in four weeks" by comparing 16 to 24. That was reading variance as a trend, and we cut it.

Rank is the durable measure. It moved from seventh to first and has mostly held there, losing first place for one week in September. If you take one number from this study, take that one, and note that we showed you the noisy one next to it.

---

Scoreboard 2, receipt 3: assistant referrals, with the claim we had to retract

A channel that did not exist before April now sends real people:

433 assistant sessions across the window. Before April, zero. Not small. Zero, every month, for the life of the domain.

Now the part where our earlier draft was wrong, which we are including because a study that only reports its wins is an advertisement.

That draft claimed assistant referrals were "on pace to deliver more visitors to our content pages than Google search does." They are not. Google sent 775 sessions to our content pages in this window. The assistants sent 318. Google is ahead by a factor of 2.4. Assistant referrals did grow, from 116 sessions in July to 161 in September, while Google held roughly level, but Google still sends more than twice the content traffic. We only caught the original claim by re-deriving the numbers instead of updating the date on them.

What does hold is the shape, and it is the more interesting finding anyway:

Three quarters of assistant sessions land directly on the page that answers the question. Google sends a quarter of its sessions to our homepage, and 82% of the human clicks it delivers are branded searches from people who already know who we are.

So the honest statement is not that the assistants have overtaken Google. It is that the two channels are doing different jobs. Google is largely returning our own audience to our front door. The assistants are delivering people who did not know us to the exact page that answers what they asked.

---

What we found after the first draft: the cited pages were not all right

In September we audited the research claims on the pages the assistants were citing. The engine had attributed statistics to named research firms across hundreds of pages. When we checked a sample against the publishers' own pages, 43 of the 45 claims we could verify were defective: misattributed, misquoted, or not traceable to any report at all. The prompt that asked the engine for sourced statistics was the cause, and we rewrote it.

So on 2026-09-22 we wiped the slate. Every citation on the site to a report dated before 2026, from any publisher, and every research-firm claim carrying no year, came out: 9,675 removals across roughly 700 pages in the first pass, 376 more edits on 122 pages in a second, and a further round in October that also retired pages the cleanup left empty. Every edit is recorded and reversible. New figures come back only through a rewrite, highest-traffic page first, with each number checked on the publisher's own page.

This changes how to read scoreboard 2. Citation rank measures being retrieved, not being right. The assistants cited pages carrying statistics that did not survive a check against their sources, and nothing in the citation data showed it. If you track AI visibility, track accuracy next to it. Being the source an answer is built from is a liability when the source is wrong.

It is too early to say what the cleanup does to our rank. The one snapshot taken after it, 2026-09-28, still shows us first. One week is not a trend, and we will report the series either way.

---

What we think this means

You are probably reading the wrong scoreboard. If we had judged this program on clicks we would have killed the only channel that was growing. Before concluding your content is failing, check the four signals a standard report does not show you: impressions decoupled from clicks, question-format queries with no click-through, data-centre traffic with dead engagement, and citation tracking.

The unit of visibility is becoming the citation, not the link. The win is being the source an answer is assembled from. We now watch our rank in cited domains the way we used to watch keyword positions, and, after September, we check what those cited pages actually say.

The lag between publishing and citation is long. We published 1,060 documents in May. Citation rank did not reach first until 2026-06-29, roughly six weeks after the sprint ended, which is about what indexing and retrieval lag would predict. If you are measuring an AI-visibility program on a four-week cycle you will conclude it failed before it has had a chance to register.

We are deliberately not claiming that any single thing we did caused the move. We consolidated duplicate pages starting 2026-07-21, three weeks after rank first hit one, so consolidation cannot be the explanation for the move itself, though rank has largely held through it. With one domain and no control we can report the sequence and not the mechanism. Treat anyone who tells you otherwise from a sample of one with suspicion, including us.

And the traditional scoreboard still matters, because it still pays. Google sends us more than twice the content traffic the assistants do. Anyone telling you to abandon search for answer engines is selling you the opposite of the mistake we nearly made.

---

Methodology and limits

Window. 2026-07-03 to 2026-10-01, 90 days, ending three days before measurement to allow for Search Console finalisation.

Sources. Search Console domain property for impressions, clicks, positions and query segmentation. Google Analytics 4 for sessions, landing pages, engagement and source. Ahrefs Brand Radar weekly snapshots for citations, stored locally, 21 of them. Our own content database for corpus counts.

Reproducibility. Every figure comes from one script. It calls the same measurement library our internal reporting uses rather than reimplementing the classification, after an earlier version of that script disagreed with the shipped one about which queries are branded (68 non-branded clicks versus 41) and which are questions (3,152 versus 3,902). Two instruments producing two answers for one quantity is how a study gets pulled apart, so there is now one instrument.

Refreshed before publication. The first version measured 2026-06-10 to 2026-09-08 and was due out on 2026-09-29. Rather than publish September's numbers in October, we re-ran the same script for the latest ninety days. What changed: our run of first-place citation weeks broke once, on 2026-09-21; machine queries fell from about a third of visible impressions to just over a quarter; and the citation cleanup described above happened in between.

Query-dimension coverage. Search Console reports the query for 50.6% of impressions and 26.2% of clicks. Google anonymises the remainder. Every segmented figure describes that visible slice and is labelled as such. Site totals are the only complete numbers, and we never compare a segment figure to a site total and call the difference a change.

Assistant referrals are undercounted. Answers read in-app, copied links and most desktop flows pass no referrer and land in direct traffic. Our 433 is a floor. This undercount only makes scoreboard 2 look worse than it is, so no directional claim here depends on it.

Machine traffic is inferred, not declared. We classify a session as automated from a data-centre location combined with an engagement rate under 20%. That is inference. We excluded Des Moines, which met the location test at 80.3% engagement and is plainly people.

Brand Radar measures a fixed prompt set, not the whole of AI search. Rank first in that set means first among the domains cited for those category prompts. It does not mean first everywhere, and the prompt set is ours.

Corrections to earlier drafts of this study, all found by re-deriving rather than re-dating:

1. The claim that assistant referrals exceed Google traffic to content pages. Wrong. Google leads 775 to 318.
2. The claim of a 50% four-week increase in citation footprint. That was variance, not growth.
3. The claim that the corpus had grown to roughly 1,600 pages. It was 1,380 when we wrote the first version and is 1,362 now, deliberately smaller than its peak.
4. A zero-click example built on the query "the b2b growth engine," presented as a US commercial term absorbed by an AI Overview. Its impressions are predominantly outside the US and its volume has collapsed to single digits. Removed.

We are a B2B marketing agency and this study is about our own website, so treat it as an interested party reporting on itself. That is exactly why the methodology, the noisy series, our own retracted claims and the accuracy problem are printed here rather than summarised.