01One Network, Four Records
From April 2024 to September 2026, Google, Bing, Google’s AI answers and Copilot each kept an account of the Sustainability Directory, and when they are read side by side, those accounts disagree on almost all of it.
- Google removed the Directory with a manual action on 30 May 2026, after which the clicks it had from Google went from 370,755 in March 2026 to one in all of the four months that followed, and by 25 September there were none of its pages in Google’s index.
- The warning came about eight weeks early, when Google showed the network 134 million times in February 2026, more than twice as often as in January, while only 0.24 % of those who saw it clicked, and it was on 26 March that the clicks broke.
- Bing kept sending readers, with between 31,000 and 39,000 clicks in each month from February to September 2026, and once Analytics counted only people, 97 % of all the visits that came from search came through Bing’s index.
- Copilot cited the Directory more after the verdict than before it, about 24,800 times a day in November 2025 and about 124,000 times a day in September 2026, which made 17.5 million times in all.
- The engines quoted different pages, so that the pages for terms drew 50 % of Copilot’s citations and only 35 % of Google’s clicks, and 40 % of all the pages Copilot cited are not in the Google data we have at all.
- Each assistant followed its own index, so that Gemini sent 4,350 visits in March 2026 and only 43 in August, while Copilot, which answers from Bing, sent 2,179 in March and then 3,930 in August.
- Most of what Analytics counted was machines, which we saw when the tags began to wait for a real scroll, tap or key press on 17 September 2026 and recorded sessions fell from about 692,700 a day to 20,200 while visits from search barely moved.
Our Research Question
We built the Directory, we ran it for two and a half years, and in September 2026 we saved every account it had before the engines let them go. This paper asks what each engine recorded about it between April 2024 and September 2026, why those records disagree, and what reading them side by side tells a firm about how it is found. A firm that reads how it is found through Google alone would have read this property wrong at every turn.
Through the spring of 2026 the Google line said success; by the summer it said the site was dead, and both times another engine said something else.
In 11 months Copilot cited the Directory 17.5 million times, and it did so more after Google’s verdict than before it.
So the answer reaches well past one site. It goes to what any firm owns when it is found, and to how that firm would know.
Every figure below comes from the Directory’s own accounts, with its window and its source; the last part, Method and Sources, says how each was pulled and checked so that another firm could repeat the work on its own.
Twelve Sites, One Engine
Noo–Studio built the Sustainability Directory from April 2024 as an open reference on sustainability, which anyone could read for free and use under a CC BY 4.0 licence, with one page for each concept and each page linked to the next one. It ran on a root domain with eleven subject sites beside it: lifestyle, climate, energy, pollution, esg, product, prism, fashion, news, term and learn.
All of the subject sites went live between January and October 2025. Each carried the same parts, terms, questions, explainers and a page for every tag, so that one concept led to the next. Every page declared in structured data what it was, a defined term, a question or a collection.
It grew very large, and a count we made in September 2026 puts it at about 120,000 articles and about a million tag pages, each of them split into numbered pages of thirty. At the start of the three months that Search Console’s page report keeps, in late June 2026, Google listed about 26 million of its addresses as known and not indexed.
Those numbered pages were built to stay still. On most sites the second page of an archive shifts every time a post goes out, so that an engine finds a different page on each visit. The Directory counted its archive pages from the oldest post, so the first page held the oldest thirty and, once it was full, never changed again, while new posts gathered on the root.
That made every archive page an address that could be cached for good and indexed as itself, and by our reading it is a large part of why the reference ranked.
It is also how a million tags became millions of pages.

Written by a Machine
No person typed a page of the Directory. An engine ran one queue for each of the subjects on Google’s Gemini models, and each of them had its own prompt, schedule and daily limit. A question engine wrote pages that answer what people ask, and then it proposed the questions that follow.
Then a third engine wrote a description of several sections for each of about 1.1 million tags, and a glossary linked every page to the terms it mentioned. Before anything went out, a humanizing pass reworked the rhythm of the sentences. We engineered the prompts the way we write software: one for each subject, a word range for each section, a list of banned words, and a pass rate measured version by version.
No subject expert read the entries one by one, and each page was published and linked into the rest of the reference without a hand on it.
A reviewer at any engine could see that from the pages themselves. In May 2026 one did.
02Two Clocks Running Together
The Directory’s first pages went out in April 2024, a month before Google showed its first AI Overviews in the United States, so the reference and the answer engines grew up side by side.
Built on Google’s Models
Gemini wrote the pages from June 2024, through a gateway we ran ourselves, and each new release of it was in the engine within weeks. The archive holds 162 publishing scripts dated that month, most of them successive versions of one engine, rewritten as we learned what the model could hold.
In 2024 machine text had a name for being easy to spot. The first detectors judged writing on how predictable its words were and how much its sentences varied. Our publishing prompts of that year met the test head on.
They asked the model to put readability, perplexity and burstiness ahead of a rigid page, and to write with the experience, expertise and trust that Google asks of any page.
Google released Gemini 2.0 Flash in December 2024, and by 18 December the question engine ran on it. By February 2025 the question chain called on five of the Gemini models, from the light Flash-Lite up to the experimental Pro. Around it the answer engines arrived one by one.
AI Overviews came to the United States on 14 May 2024, and then to more than 100 countries on 28 October. ChatGPT search launched on 31 October, and AI Mode opened to everyone in the United States on 20 May 2025.
The assistants found the Directory early on. Its first visits from one of them came in May 2024, when Perplexity sent four. ChatGPT and Gemini were next, in August 2024, and in the month that AI Mode opened, ChatGPT sent 2,754 visits and Gemini 651.
Copilot was the last of them to come, in August 2025.

Time to a Thousand a Day
The subject sites opened across 2025, each on a curve of its own, and the first of them, esg, had its first visit from search on 24 January while the last, learn, had its first on 21 October. Five of the eleven reached a thousand visits from search in a day within 190 to 321 days of their first, roughly six to eleven months, while the other six never reached it.
Lifestyle was the fastest of them, at 190 days from its first visit from search on 17 February 2025. The slowest of the five was esg, which took 321 days from 24 January 2025.
Four of them, term, news, product and learn, only opened in September and October 2025, so they had eight and a half months or less before Google acted, and their curves stop short.
Term never reached a hundred visits from search in a day.
| Subject | First visit from search | Days to 100 a day | Days to 1,000 a day |
|---|---|---|---|
| lifestyle | 17 Feb 2025 | 98 | 190 |
| climate | 12 Feb 2025 | 114 | 236 |
| energy | 12 Feb 2025 | 83 | 238 |
| pollution | 17 Feb 2025 | 66 | 247 |
| esg | 24 Jan 2025 | 207 | 321 |
| prism, fashion, news, product, learn | 27 Jan to 21 Oct 2025 | 45 to 139 | not reached |
| term | 12 Sep 2025 | not reached | not reached |
Peak Month, March 2026
Search Console’s history begins on 13 May 2025, with 19,389 clicks in its first 19 days, and from there Google’s count climbed in all but two of the next ten. Its busiest was March 2026, with 370,755 clicks, the most the Directory ever had in a month.
Lifestyle carried the network, with 914,721 Google clicks in the sixteen months and 187,466 of them in March 2026 alone, and after it came pollution with 392,707, climate with 215,192 and energy with 186,018. Lifestyle’s prompt was the most human of the eleven, by our own reading, and it asked for an opening in the middle of a scene. Bing came to it later and more slowly than Google did, with 2,971 clicks in August 2025, 23,274 in October and 36,499 in March 2026, which was also the month in which Copilot cited the Directory 1.85 million times.
Where the two Google counts overlap, they agree.
In March 2026 Analytics recorded 424,954 sessions that had come from Google search, and Search Console recorded 370,755 clicks for the same month. That is close for two systems that each count a different thing.
So, seen through that one window, the spring was a success and the months after it read as a collapse.

The numbers
| Month | Visits from search (Analytics) | Google clicks (Search Console) |
|---|---|---|
| Apr 2024 | 237 | |
| May 2024 | 285 | |
| Jun 2024 | 628 | |
| Jul 2024 | 1,436 | |
| Aug 2024 | 957 | |
| Sep 2024 | 1,932 | |
| Oct 2024 | 3,402 | |
| Nov 2024 | 3,879 | |
| Dec 2024 | 5,793 | |
| Jan 2025 | 4,939 | |
| Feb 2025 | 4,283 | |
| Mar 2025 | 5,866 | |
| Apr 2025 | 7,433 | |
| May 2025 | 15,227 | 19,389 |
| Jun 2025 | 43,482 | 49,122 |
| Jul 2025 | 60,490 | 59,948 |
| Aug 2025 | 67,124 | 63,565 |
| Sep 2025 | 124,808 | 112,696 |
| Oct 2025 | 227,323 | 171,197 |
| Nov 2025 | 300,208 | 242,703 |
| Dec 2025 | 313,919 | 231,844 |
| Jan 2026 | 236,178 | 220,997 |
| Feb 2026 | 413,473 | 316,701 |
| Mar 2026 | 498,505 | 370,755 |
| Apr 2026 | 211,523 | 108,815 |
| May 2026 | 186,672 | 46,249 |
| Jun 2026 | 506,892 | 1 |
| Jul 2026 | 481,765 | 0 |
| Aug 2026 | 182,431 | 0 |
| Sep 2026 | 155,462 | 0 |
03Where the Answers Came From
On 10 February 2026 Bing put out a report that shows a site where Copilot cites it, on which searches and how often, with a history that reached back to 1 November 2025.
Eleven Months on Record
Copilot cited the Directory 743,409 times in November 2025, which is where Bing’s report begins, and it cited it 1.85 million times in March 2026. The count dipped in the spring, to 1.66 million in April and to 1.41 million in May. Then it went higher than ever, with 2.03 million in June, 1.93 million in July, 2.19 million in August and 2.72 million in the first 22 days of September.
Over all 326 days that makes 17,500,282 citations, which came on 15,839 of Copilot’s own searches and went to 48,040 of the Directory’s pages. The busiest day was the last one on record, 22 September 2026, with 164,855.
A few searches drew a great deal of it. Ethical sourcing practices alone took 61,289 citations, and almost all of them, 61,268, went to one page of the root site. Greenwashing had 45,217 of them and the IFC performance standards had 14,565.
And almost none of it was buying. Of the 1,571,175 citations in Bing’s sample of grounding queries, 86 % came on searches it labels informational, research, or learn and solve. Only 1,939, about 0.1 %, came on searches it labels commercial.

The numbers
| The search Copilot answered | Citations | Intent |
|---|---|---|
| ethical sourcing practices | 61,289 | Informational |
| greenwashing | 45,217 | Learn and Solve |
| explain the concept of genetic conservation and its importance. | 18,853 | Informational |
| ifc performance standards | 14,565 | Informational |
| eco ai | 10,035 | Informational |
| bbso sustainability | 7,880 | Informational |
| local supply chain | 7,504 | Informational |
| smart commute | 6,743 | Informational |
| authenticity maintenance | 6,464 | Learn and Solve |
| the influence of cultural emphasis on patience and long-term orientation on investment strategies and savings habits | 6,332 | Research |
| urban wildlife definition | 6,268 | Informational |
| decomissioning | 6,017 | Learn and Solve |
| water-efficient appliances | 6,010 | Informational |
| ikm | 5,910 | Others |
| global monitoring system | 5,304 | Informational |
Why Definitions Get Quoted
Copilot ran its searches in a few words at a time. Of its citations of the Directory, nearly two thirds came on grounding queries of one to three words. The length that came up most was two words, which had 32 % of them, as in smart commute.
Only 12 % of Copilot’s citations of the Directory came on searches put as questions.
But the rest named a thing, and Copilot went for the page that says what that thing is. Term pages took 50 % of its citations, where on Google they had taken 35 % of the clicks; the pages for questions took 20 % of Copilot’s citations and 35 % of Google’s clicks.
A term page in the Directory says what one thing means in one line, then opens each of its sections with the question a reader would type, what is the X of Y. We once tried a version that varied that wording; it ranked worse, and the plain one came back. By our reading this is the long tail rewritten short: a person may ask a long question, the assistant searches with the two or three words that carry it, and the page that defines those words is the page it quotes. Longer searches are in the record too, as 9 % of the citations came on grounding queries of eight words or more, and one of them, on how a culture’s patience bears on saving and investing, drew 6,332 citations; they are few.
The definitions, by our reading of the archive, carried the site. Each one was short, parsed at once, and stood on a page of its own.
We hold one more reading that these accounts cannot prove. An engine that answers millions of questions picks the page that gives a certain answer for the least work to fetch, read and quote. A static page that defines one term costs it almost nothing.
Copilot also cited pages that, as far as the Google data we hold can tell, Google’s search had never shown at all. Of the 138,256 pages that Bing lists as cited on the subject sites, there are 55,347, or 40 %, that are not in our Google data. Only 51 % of Copilot’s citations fell on pages to which Google had at some time sent a click.

Most of the Weight
Copilot leaned very hard on a few of the Directory’s pages. The top 1 % of the pages it cited drew 35 % of its citations, and the top 10 % drew 71 %. On Google the top 1 % of the pages it sent clicks to had 23 % of all the clicks, and the top 10 % had 60 % of them.
Were the cited pages put in a row by their count, the one in the middle would have been cited 9 times. On Google, the page in the middle of all those it sent a click to had only 1. And when Copilot used the Directory it often used a lot of it.
On 49 % of its citations the Directory’s pages made up from a quarter to a half of all the sources that were behind the answer, and on 17 % they made up more than half of them. They made up less than a tenth on only 5 %.

The numbers
| Measure | Share |
|---|---|
| Citations on grounding queries of 1 word | 11.0% |
| Citations on grounding queries of 2 words | 32.4% |
| Citations on grounding queries of 3 words | 21.3% |
| Citations on grounding queries of 4 words | 8.9% |
| Citations on grounding queries of 5 words | 7.3% |
| Citations on grounding queries of 6 words | 5.6% |
| Citations on grounding queries of 7 words | 4.5% |
| Citations on grounding queries of 8 or more words | 9.0% |
| Citations where the Directory made up under 10% of the sources | 5.1% |
| Citations where the Directory made up 10 to 25% of the sources | 28.8% |
| Citations where the Directory made up 25 to 50% of the sources | 48.9% |
| Citations where the Directory made up over 50% of the sources | 17.1% |
| Citations on grounding queries put as a question | 12.0% |
| Copilot’s citations on the top 1% of cited pages | 35.4% |
| Google’s clicks on its top 1% of pages | 23.5% |
| Copilot’s citations on the top 10% of cited pages | 70.6% |
| Google’s clicks on its top 10% of pages | 60.4% |
Office Hours
At weekends Copilot cited the Directory at 54 % of the rate it had on weekdays. Over the whole of the record Tuesday was the busiest day of the week, at 69,149 citations on average. Saturday was the quietest, at 26,202.
People who clicked through from Google and Bing had the same rhythm, though a less steep one. From November 2025 to March 2026 Google’s clicks at weekends ran at 69 % of what they were on weekdays, and Bing’s at 60 %. And when Google’s AI answers showed the Directory, it was on a desktop 78 % of the time, 135,787 times out of 173,290.
We read that as people at work, asking the question their job turns on between two meetings, and it is the reader a firm wants, and the record shows it by the day of the week, without a survey.

04Each Engine’s Own Map
Google and Copilot read the same pages and kept different lists of what mattered on them, by subject, by kind of page and by the searches that led there.
Which Subjects Each Engine Chose
Lifestyle led all three counts, with 45.9 % of Google’s clicks on the eleven sites beside the root, 42.8 % of its AI answers and 25.8 % of Copilot’s citations, and it is below lifestyle that the orders part.
Google’s AI answers leaned on product and news, which had 16.8 % and 7.0 % of all their appearances and only 1.5 % and 1.8 % of the clicks. Copilot, for its part, gave esg 20.8 % of its citations, which was 3.16 million of them on 7,196 of its searches, when esg had only 3.6 % of Google’s clicks and 0.9 % of its AI answers.
Climate and energy, taken together, had 24 % of Copilot’s citations, which was more than their 20 % of Google’s clicks and more than twice their 10 % of its AI answers. Learn shows it most plainly of all, since it had 38 Google clicks in its whole life, no AI answers, and 386,529 citations from Copilot.
We read the two maps this way. Google’s AI answered everyday choices, what to buy and how to live. Copilot treated the network as a reference for work, where reporting standards, climate and energy are the job, and both readings hold.

The numbers
| Subject | Google clicks | Google’s AI answers | Copilot’s citations |
|---|---|---|---|
| lifestyle | 914,721 (45.9%) | 75,600 (42.8%) | 3,921,217 (25.8%) |
| pollution | 392,707 (19.7%) | 39,006 (22.1%) | 1,620,597 (10.7%) |
| climate | 215,192 (10.8%) | 13,682 (7.8%) | 1,984,019 (13.1%) |
| energy | 186,018 (9.3%) | 3,258 (1.8%) | 1,661,375 (10.9%) |
| fashion | 75,899 (3.8%) | 650 (0.4%) | 233,087 (1.5%) |
| esg | 72,458 (3.6%) | 1,578 (0.9%) | 3,160,966 (20.8%) |
| prism | 67,229 (3.4%) | 604 (0.3%) | 666,843 (4.4%) |
| news | 35,952 (1.8%) | 12,270 (7.0%) | 498,786 (3.3%) |
| product | 29,967 (1.5%) | 29,726 (16.8%) | 691,702 (4.6%) |
| term | 1,608 (0.1%) | 157 (0.1%) | 354,545 (2.3%) |
| learn | 38 (0.0%) | 0 (0.0%) | 386,529 (2.5%) |
Demand Google Never Showed
The pages that Copilot lists and the pages that Search Console lists only overlap in part. Google’s own AI answers took 95 % of what they showed from pages to which Google’s search had at some time sent clicks, and of the 8,081 pages they used, there were 2,789 that Copilot also cited. But Copilot went further afield, and of its citations 66 % came on searches that also appear in the Google query data we hold, the other 34 % on searches that do not.
That 34 % is the figure in this paper we can least defend.
Google hides the rare searches from its tables, and its query table for March 2026 covered only 15 % of that month’s clicks. So a search that is missing from our copy may still have reached the Directory through Google.
A firm that reads its own accounts can still make use of the number. It should take the direction as real, since some of what an assistant answers is demand that search never showed it. Then it can test the size against its own Bing report, where the grounding queries carry its name.

The numbers
| Page type | Google clicks | Google’s AI answers | Copilot citations |
|---|---|---|---|
| Terms | 35% | 26% | 50% |
| Questions | 35% | 15% | 20% |
| Explainers (learn) | 8% | 14% | 16% |
| Tag pages | 16% | 29% | 7% |
| News | 1% | 10% | 2% |
| Paginated archives | 3% | 5% | 0% |
Many Positions at Once
Every engine builds its picture of a site from what its own users search for, and it cites what it keeps; for a firm, then, ranking well means holding a place for each question in each engine, all at the same time, and each engine decides which page answers.
By our reading the Directory held its places from the small end. Its terms answered searches of four and five words, too rare to show in Google’s tables, and the pages for one-word and two-word terms gathered on top of them. It is a mesh that grows upward, from the longest search to the shortest, and the 15,839 searches of Copilot’s that reached it fit that reading.
A reading that looks at one engine sees one map and takes it for the ground.
05Google’s Verdict
Well before the notice of 30 May 2026, the Directory’s clicks had begun to fall, and Search Console kept the order of every one of the steps.
Impressions Without Clicks
In February 2026 Google showed the network’s pages 134 million times, which was more than twice the 62.5 million it had shown them in January. The clicks did not keep up; they rose to 316,701, and so the click rate fell to 0.24 %.
Of the 134 million, lifestyle alone took 77 million. The extra showings went to the ordinary questions and terms. And the translated results, the ones Google translates for a reader who searched in another language, went up from about 922,000 in January to 4.1 million in February.
In March the showings fell back to 81 million while clicks peaked at 370,755, a click rate of 0.46 %, and then the clicks went too.
We now read showings that rise without clicks as the first warning that a firm gets. Here it came about eight weeks before the break and four months before the action, which is time enough to act if someone reads the account every week.

The numbers
| Month | Impressions | Clicks | Click rate |
|---|---|---|---|
| Nov 2025 | 37,393,609 | 242,703 | 0.65% |
| Dec 2025 | 46,649,675 | 231,844 | 0.50% |
| Jan 2026 | 62,526,714 | 220,997 | 0.35% |
| Feb 2026 | 134,034,690 | 316,701 | 0.24% |
| Mar 2026 | 81,189,377 | 370,755 | 0.46% |
| Apr 2026 | 31,990,876 | 108,815 | 0.34% |
| May 2026 | 16,010,135 | 46,249 | 0.29% |
| Jun 2026 | 404 | 1 | 0.25% |
From the Break to the Notice
On 26 March 2026 Google sent 12,384 clicks, which was about its March average of 11,960, and by 3 April it was sending 2,984 a day. April as a whole came to 108,815, which is 71 % less than March, and May came to 46,249.
The pages sank before they went. Their average position in Google’s results was 8.0 in March, 13.6 in April and 22.1 in May.
Google’s AI answers kept the Directory up to the last days. The Generative AI report in Search Console only has a history from 18 May 2026. It shows that in the eleven days up to 28 May they showed it 172,636 times, between 8,425 and 19,293 times a day.
Then 138 on 29 May, the day before the notice. From 29 May to 12 June there were 632 in all, and from 13 June to 24 September there were only 22.
The message from Google is dated 30 May 2026. Google sent one click in June and none after it.

Judged as One Pattern
Search Console’s notice of 30 May 2026 reads “Major spam problems”, and it affects all pages, the root and every one of the subject sites, and Google issues a manual action only when a person on its staff has reviewed a site. The description it gives names a family of techniques that it acts on, scaled content abuse, cloaking, scraping and repeated violations, without saying which of them it found.
Our reading of the record finds three things that match lines of it.
| What we see now | What the Directory did | The line it matches |
|---|---|---|
| Scale | About a million tag pages on about 120,000 articles, each tag page split into numbered pages with its own AI-written description and an ad, and a question engine that fed the questions it wrote back into its own queue | Scaled content abuse, pages made in bulk to rank |
| A reveal | To keep bots out of the ad auction, images waited for a visitor’s first real scroll or tap; Google’s crawler never scrolls, so it saw the pages without them | Cloaking, a different page for the crawler |
| Words | The site described itself as a network of satellites, programmatic at scale | The vocabulary a reviewer reads first |
A manual action is a person’s judgement, and a person looks at a site the way a buyer would, at what it is for and how much of it there is. An engine that scores pages singly might have passed most of the Directory’s, and a reviewer who asked why a reference on sustainability needed millions of pages would have found the answer in the engine.
That is our reading, and Google has not said.
Taken together, the pages looked like pages made to rank, whatever each one said.
A reviewer reads how a whole site was made, and judges it as one pattern.
We took the reveal out on 1 June 2026, and we took the ads out with it. No request for review had gone to Google by the end of September 2026.

Out of the Index
By 25 September 2026 there was nothing left of the Directory in Google’s results. Search Console had 0 of its pages in the index and 10.7 million out of it, for 14 reasons, and the largest of them was “Crawled, currently not indexed”, which held 10,595,075 pages.
Its record of the network’s links was entirely empty, 0 from outside and 0 from within. Googlebot still called, and in the 90 days up to the capture it made 3.31 million requests and fetched 77.4 billion bytes of the site. AdsBot, the crawler that checks pages for ads, made 85 % of those requests.
Lifestyle alone took 2,309,907.
For Google the property no longer existed, and its crawler still came to it every day.

06What Happens After a Manual Action?
Outside Google the same pages kept their readers through the summer of 2026, both the people that Bing sent to them and the assistants that answer from its index.
Second Index, Same Pages
Every answer engine in this record cited what its own search engine had kept. Google’s index is the one that feeds Google search, its AI Overviews and AI Mode, and Gemini, and all of them let the Directory go between April and June 2026.
Bing’s index feeds Bing and Copilot, and Yahoo, DuckDuckGo and Ecosia serve it to their own users as well. In the six days after the filter of 17 September, it was Bing that sent the most, 3,976 visits a day, then Yahoo with 847, DuckDuckGo with 109 and Ecosia with 91. Google sent 135.
Which index ChatGPT and Perplexity read is not in the record, so the drawing leaves them out.

Clicks That Held
Bing’s count was between 21,657 and 24,278 a month from November 2025 to January 2026, and from February to September it ran between 31,000 and 39,000 in every one of the months, with no break at all when Google acted. There were 32,440 in June, 31,579 in July, 38,790 in August and 35,180 in the first 22 days of September.
A manual action is one company’s decision about its own results, and it always stops at what that company shows.
And the pages stayed up through all of it. Readers found them through the other engines, and the assistants that read another index went on citing them.

The numbers
| Month | Google clicks | Google AI answers | Bing clicks | Copilot citations |
|---|---|---|---|---|
| Nov 2025 | 242,703 | 23,848 | 743,409 | |
| Dec 2025 | 231,844 | 21,657 | 787,568 | |
| Jan 2026 | 220,997 | 24,278 | 890,102 | |
| Feb 2026 | 316,701 | 32,243 | 1,299,961 | |
| Mar 2026 | 370,755 | 36,499 | 1,852,508 | |
| Apr 2026 | 108,815 | 31,489 | 1,655,612 | |
| May 2026 | 46,249 | 172,938 | 32,735 | 1,406,185 |
| Jun 2026 | 1 | 340 | 32,440 | 2,031,858 |
| Jul 2026 | 0 | 5 | 31,579 | 1,926,793 |
| Aug 2026 | 0 | 0 | 38,790 | 2,186,748 |
| Sep 2026 | 0 | 7 | 35,180 | 2,719,538 |
Citations Kept Rising
After a dip in the spring, Copilot went past its March high, with 2.03 million in June and then 2.72 million in the first 22 days of September. As a daily average, that was about 24,800 in November 2025 and about 124,000 in September. It also widened, from about 7,261 of the Directory’s pages cited in a day in November 2025 to about 22,861 in September 2026.
And it cited each of them more often, 3.41 times a day on average at the start and 5.41 at the end.
Those daily figures rest on Bing’s daily totals, which Bing gives complete, and they would move only if Bing changed how it counts a citation, which nothing in its report says it has done.
The answers in Google’s own results had already gone out with its index at the end of May.
Assistants Follow Their Index
Gemini, Google’s own answer engine, sent the Directory 4,350 visits in March 2026 and only 43 in August. Copilot, which answers from Bing’s index, sent 2,179 in March and 3,930 in August, more after the action than before it.
ChatGPT went down from 16,837 visits in February to 393 in August. It draws on several sources, and how much of that fall came from Google’s action is a question this record leaves open. But these counts carry a doubt of their own.
Up to 17 September 2026 some of the automated traffic also came in with an assistant as its referrer. Perplexity’s 18,191 visits in July rose with a wave of bots that began on 24 July. A buyer should trust which way each line goes and hold the exact counts loosely.
Being named in an AI answer is the visible end of an index, so a firm that exists in only one of them is one decision away from leaving half of the answers.

The numbers
| Month | Visits from Google (Analytics) | Visits from Bing (Analytics) |
|---|---|---|
| Jan 2026 | 203,558 | 24,049 |
| Feb 2026 | 358,225 | 42,487 |
| Mar 2026 | 424,954 | 57,098 |
| Apr 2026 | 141,628 | 54,401 |
| May 2026 | 69,905 | 94,878 |
| Jun 2026 | 21,560 | 268,989 |
| Jul 2026 | 1,566 | 251,220 |
| Aug 2026 | 2,228 | 114,687 |
| Sep 2026 | 1,905 | 104,923 |
07How Many Visitors Were Bots?
One window misleads for a second reason, which has nothing to do with Google, because most of what Analytics recorded on the Directory was never a person.
Sessions Before and After
On 17 September 2026 we set the analytics tags on all twelve sites so that they load only when there has been a real scroll, tap or key press, and we changed nothing else. Every visitor and every crawler still got the same page as before; it was only the counting that changed.
We call the filter a trusted interaction.
A person who reads a page scrolls it, taps it or presses a key, and the browser marks that event as the person’s own. But a script that fetches a page to copy it seldom does any of that, so its visit is never counted.
The sessions it recorded fell from about 692,700 a day, from 1 to 15 September, to about 20,200 a day from 18 to 23 September. So 97 % of the visits Analytics had been reporting had never touched the page. Site by site the share ran from 85 % to 99 %.
The visits that came from search barely moved, from 5,744 a day to 5,161. So the filter took out the machines and kept the people who searched, and three months after Google’s action about 5,000 people a day were still reaching the network from engines other than Google. Engagement fell with the bots as well.
With them in, between 45 % and 60 % of the sessions counted as engaged; without them, it was 17 %.

The numbers
| Channel | Sessions a day, 1 to 15 September | Sessions a day, 18 to 23 September |
|---|---|---|
| Direct | 680,399 | 14,584 |
| Unassigned | 6,330 | 88 |
| Organic Search | 5,744 | 5,161 |
| Referral | 99 | 171 |
| AI Assistant | 86 | 111 |
| Organic Social | 32 | 47 |
| 1 | 0 | |
| Organic Video | 0 | 0 |
Growth in the Direct Channel
The bots had bent every one of the rates in the monthly report. In the first half of September, before the filter, 40 % of all the sessions came from Singapore. And 62 % ran Chrome on Windows and 38 % Chrome on a Mac, a mix we have never seen from an audience of people.
Before the filter, Direct sessions grew from 4.5 million in January 2026 to 27.8 million in July, and almost all of them were automated.
And the search counts swelled with them, as Analytics filed 507,000 sessions under Organic Search in June and 482,000 in July, which were months in which Google sent the Directory one click and then none at all.
On a chart that rise would have looked like success. A report built on those numbers would have been confidently wrong twice, on how many came and on how engaged they were.
The few true strangers in the record were readers. Of the 84 hosts that were not the Directory’s own, most were Google Translate’s, which means people reading its pages in another language. They had 32,749 sessions from January to September 2026, against 144,473,448 on the Directory’s own 23 hosts.
Analytics Against the Engines
Automated traffic came in with search referrers, and so it was filed as if search had sent it. In June 2026 Analytics gave Bing the credit for 268,989 sessions, while Bing’s own count for that month was 32,440 clicks.
And it gave Google the credit for 21,560 sessions in a month in which Google sent the Directory one click.
In January the two had agreed, 24,049 sessions against 24,278 clicks. The gap came with the bots.
Each engine’s own report is the only reliable count of what that engine sent.

The numbers
| Month | Sessions credited to Bing (Analytics) | Clicks Bing counted (Bing Webmaster Tools) |
|---|---|---|
| Jan 2026 | 24,049 | 24,278 |
| Feb 2026 | 42,487 | 32,243 |
| Mar 2026 | 57,098 | 36,499 |
| Apr 2026 | 54,401 | 31,489 |
| May 2026 | 94,878 | 32,735 |
| Jun 2026 | 268,989 | 32,440 |
| Jul 2026 | 251,220 | 31,579 |
| Aug 2026 | 114,687 | 38,790 |
| Sep 2026 | 104,923 | 35,180 |
Who Still Sends People
Once the filter was on, 5,164 visits a day came to the Directory from search, four months after Google’s action. Of those, 97 % came through Bing’s index, from Bing itself and from the engines that serve it. Only 135 came from Google.
And AI assistants added 111 visits a day. That is the Directory as it stands, read by people.

The numbers
| Search engine | Visits a day, 18 to 23 September 2026 | Index it serves |
|---|---|---|
| Bing | 3,976 | Bing’s |
| Yahoo | 847 | Bing’s |
| 135 | Google’s | |
| DuckDuckGo | 109 | Bing’s |
| Ecosia | 91 | Bing’s |
| Other engines | 6 |
08Taught by the Record
The Directory was written with AI before AI answered questions, and the answer engines still cited it after Google had stopped showing it.
Quoted by the Answers
These two years left a short list, which we now apply to every firm we read. A page that says plainly what a thing is gets quoted, so the definition is what we write first: the term understood before it is used.
Each page also says what it is in structured data, so no engine has to guess whether it is reading a term, a question or a collection. And a firm belongs in both indexes, because when Google’s closed on the Directory, Bing went on sending it 31,000 to 39,000 clicks a month. There is a cost on the other side.
What the Directory says now travels inside other people’s answers, under the answer’s name and without its own. So a definition that is loose or wrong goes out a million times as loosely.
Penalised at Scale
With AI, pages cost little to make, and the bill comes later, when a reviewer reads a site’s size as a fact about how it was made, and for its part the Directory took up each new Gemini model within weeks of its release.
The rules for publishing with AI at scale came after the tools, and they applied to everything already published.
But moving the pages does not reset the verdict. Google’s spam policies treat new subdomains or sites made to carry on a violation as circumvention. A firm that answers a manual action with a new name only adds a line to the same file.
Even a humanizing pass changes only how each sentence reads, and it leaves the count of pages where it was. The one control that a firm keeps is the pace at which it publishes, and the reason that each of its pages exists.
Five Checks for a Firm
Any business can run the same reading on its own accounts in an afternoon, and it needs nothing like the Directory’s size to turn something up.
- Open Bing Webmaster Tools, verify the site if it is not verified yet, and read its AI Performance report for how often Copilot cites the firm’s pages, on which searches and which pages. A firm may learn here, for the first time, that an assistant already quotes it.
- Read Google’s two reports side by side, since the Performance report in Search Console shows the clicks and its Generative AI report shows where Google’s AI drew on the firm’s pages. A page can matter a great deal in one of them and count for very little in the other.
- Ask the buyers’ questions live in Google’s AI Mode and in Copilot, the five that a buyer asks before choosing, and write down who is named, who is quoted and whether it is the firm at all.
- Count the people in Analytics by the share of sessions that never scroll, tap or press a key. If that share is large, then every rate in the monthly report is a rate that was measured on machines.
- Export the record from Search Console, its sixteen months today and then again every month after that, because the history that a firm will want in a year is the one that is disappearing now.
Standing Rules
It was this case that gave us our method, and every reading we make now holds to four things it taught us.
- We read every engine on its own, Google, Bing, Google’s AI and Copilot alike, because what they disagree on is itself information; on this network a dashboard of Google alone would have shown a dead site, while Bing’s showed 17.5 million citations.
- The record is archived first, with every source saved before anyone interprets it, because Search Console keeps only sixteen months and the history that is most useful to a firm is the first to disappear.
- We take the bots out before counting, so visits are read only after automated traffic is filtered out, and on this network it had been almost all of the traffic.
- Pages stay few and first-hand. We never make them in bulk to rank, which is the one thing that every engine in the end punishes, and they grow in number only from data that the firm holds and that no one else has.
The same lesson changed what we build. A satellite is now small and made for one market or for one question, it runs on the firm’s own engine, and it sends its readers home.
Terms are defined once, in one sentence first, because the definition is what assistants quote, and every finding in a reading carries the date on which it will be measured again, against the same record.
For most firms Google is still the largest window. A firm is found through several at once, each with its own record and its own judgement. The record belongs to the firm only if the firm keeps it.
09Method and Sources
Every figure here is what one engine reported about its own behaviour, read from the Directory’s accounts and archived in September 2026, before the engines let them go.
Saved Before It Expired
Our first rule for any reading is that the record is archived before anyone interprets it, and this case is why. Search Console keeps sixteen months of history and drops a day every day, and its page and crawl reports keep only three months; the record that Bing keeps of where its AI cites a site only begins in November 2025. Analytics properties that are consolidated or deleted take their history with them.
The Directory’s twelve earlier properties went to the bin on 12 September 2026, to be deleted for good on 17 October, and a deleted property goes after 35 days.
A firm that wants to know what happened has a window, and the window closes. We saved the Directory’s in four pulls on 24 and 25 September 2026, each written so that a second run only fills the gaps.
- Search Console’s API, for the domain property and for each of its twelve site properties, with six search types for each of them. It gave us the whole history by day, by country and by device, then the pages, queries and appearances month by month, and the pages week by week for each subject. In all that came to 6,831 tables and 8.8 GB, from 13 May 2025 to 24 September 2026.
- Google’s own screens, for the reports that Search Console shows only in a browser, which are its manual actions, page indexing, crawl statistics, links and AI answers. We copied each card with Google’s own stylesheets and fonts, and printed it as a vector PDF and as an image 3,200 pixels wide.
- Bing Webmaster Tools’ API, for all twelve of the sites, with every report it gives on a site over the 487 days from May 2025. With them came its AI Performance report from 1 November 2025, which has the citations by day, every grounding query and every page that was cited.
- Analytics’ Data API, for the one property that holds all the sites from January 2026, read by host, which came to 5,194 tables and 3.0 GB. The twelve earlier properties go back to April 2024, and we read them from the bin before the date set for their deletion, another 4.8 GB.
In all it came to about 17 GB. Each script asks for one table at a time, writes it to its own file and skips any file it already has. So a pull that stops on a quota starts again where it stopped.
Bing refused three of our requests for coming too often, at one a second, and each time the job waited five minutes and then carried on.
Three readings came out of the archive. First came a table with a row for each day and a column for every source, from which the monthly counts, the weekday rates and the daily averages in this paper are taken. A second reading sorts every page by the first part of its address, such as /term/ or /question/, with a numbered archive page known by its /resource/ ending, and adds up each engine’s clicks, appearances and citations by kind.
The third ranks each engine’s pages by their count and measures how much of the total the top 1 % and the top 10 % hold. It also sets Copilot’s grounding queries against every query Search Console lists for the same sites, and a search counts as a question when it opens with a word such as what, why, how or explain. No figure here is an estimate by an outside tool.

Records and Their Windows
Search Console gives the numbers for Google, for the domain and for each subject site; those for Bing and for Copilot come from Bing’s own tools, and the visits from April 2024 to September 2026 come from Google Analytics.
As it keeps only sixteen months, Google’s own figures begin in May 2025, and before that the visits from search are the ones that Analytics files as Organic Search. What the engine was in 2024 and 2025 we read from the Directory’s own archive of publishing scripts, each dated by its file.
| Source | What it holds | Window |
|---|---|---|
| Search Console, domain and 12 site properties | Clicks, impressions and position by day, page, query, country, device and appearance | 13 May 2025 to 24 Sep 2026 |
| Search Console’s screens | Manual actions, page indexing, crawl statistics, links | As they stood on 24 Sep 2026 |
| Search Console, Generative AI | Appearances by day, page, country and device | 18 May to 24 Sep 2026 |
| Bing Webmaster Tools | Clicks and impressions for 12 sites | 487 days from May 2025 |
| Bing, AI Performance | Copilot’s citations by day, its grounding queries and cited pages | 1 Nov 2025 to 22 Sep 2026 |
| Analytics, the consolidated property | Sessions by host, channel, source and day | Jan to Sep 2026 |
| Analytics, 12 earlier properties | Sessions by site, channel and source | Apr 2024 to Dec 2025 |
| The Directory’s publishing scripts | The engine’s prompts and models, each file dated | 2024 and 2025 |
Checking Each Figure
We never read a number from one source alone where a second could test it. Analytics and Search Console agree on Google in March 2026, with 424,954 sessions against 370,755 clicks, and Analytics and Bing agree in January 2026, with 24,049 against 24,278; from May on they part, because of the bots.
There is a cap on Search Console’s tables. One request gives back about 63,000 rows, so the page table for March 2026 held 53 % of that month’s clicks and its query table only 15 % of them. When this paper says a page was never in Google, it means that it was never in the Google data we have.
Bing lists the grounding queries and the cited pages as a sample, and its daily totals are complete, so every count of citations here comes from the totals and every share of the queries or the pages comes from the sample. Google’s AI answers by property add up to 176,923 against 173,290 for the domain, two counts Google keeps apart, about 2 % from each other. We corrected two of our own earlier figures for this paper.
A note of ours gave Google’s AI answers as 15,000 to 19,000 a day from 18 to 28 May; the daily table shows 8,425 to 19,293. Another said each subject reached a thousand visits a day within six to eleven months, and five of the eleven did.
A visit from an AI assistant is one that Analytics recorded with that assistant as the referrer. Until the filter went on, bots carried these referrers as well, so the charts of the assistants are good for the trend and less so for the exact counts.
Launch Dates and Charts
We take each outside event from the company that announced it. Google put its AI Overviews out in the United States on 14 May 2024 and in more than 100 countries on 28 October 2024, ChatGPT search launched on 31 October 2024, and Google opened AI Mode to everyone in the United States on 20 May 2025. Bing put its AI Performance report out in public preview on 10 February 2026.
The charts are drawn like Google’s own reports, in our colours, and the numbers in them are printed under each one of them. Our drawings come from the same files, and the screens are Google’s and Bing’s own, cut only where a caption says so and checked so that no address, account or person shows on them.
Nooné, Noo–Studio’s AI writer, wrote this paper from the studio’s own records, and every figure was checked against its source before it was published. Noo–Studio publishes it and answers for it. How the studio uses AI.

