01How an Answer Picks Its Source
Most questions that a buyer types now get a written reply before they get a list of links, and that reply quotes a page or two. We set out to learn how it picks them.
- Short searches landed on pages that carry their words, since of the 1,571,175 citations in Bing’s list of the searches Copilot ran for the Sustainability Directory from 1 November 2025 to 22 September 2026, 66 % came on searches for which a cited page held every word of the search in its address, and on two-word searches it was 89 %.
- One search went to one page when Copilot ran ethical sourcing practices for 61,289 citations and the overview named for it collected 61,268, and seven more of the larger queries pair off with an entry of their own within 8 %.
- The most quoted pages were named for a search, as 43 of the 49 most quoted in a sample of 98 that we read live on 7 October 2026 carried the words of a listed search in their address, against 16 of the 49 quoted least, and the two groups were built alike.
- A term page drew three times a tag page’s citations, 53.5 on average in Bing’s lists against 17.1, and term pages are the ones that open on their meaning in one line.
- Under 1 % of the pages took a third of the citations, 1,180 of the 138,256 in Bing’s lists, and of the 143 entries with greenwashing in their address, three earned 89 % of all that this group earned.
- Copilot came to the Directory to explain things, on lookups that Bing labelled Informational for 62 % of the references and Research for 18 %, and Commercial for no more than 0.1 %.
- Ranking and quoting overlap in part, for 40 % of the pages Copilot cited never appear in the Directory’s Google data, and on three of the studio’s sites the highest click rate, 2.1 in a hundred, came at places 21 to 30 in Google’s results.
Our Research Question
When an AI answer cites a source, how does it choose the page, and what in a page earns that choice? We put this to the records we hold, and each part below takes up one piece of it.
A firm that pays to be found now has a second list to watch beside its rankings, which is the set of pages an answer quotes, and the usual reports do not show it that list at all. So we read the reports that do, and we read them as they are, with each figure kept to the window it came from. Our largest record is the Sustainability Directory’s, a root site and eleven subject sites, where Bing counted 17,500,282 citations between 1 November 2025 and 22 September 2026.
We matched the searches in Bing’s lists to the pages it cited, and we read 98 of those pages live on 7 October 2026. Beside it there is a firm’s own site, read in September 2026 and left unnamed with its field, and there are three of the studio’s sites, which we read for how often each place in Google’s results is clicked.
Where Google and Microsoft have published how their AI features use a page, we have set what they say beside what our records show. The full account of the Directory’s engines is in the paper Every Engine Keeps Its Own Record.
Search, With the Answer Included
A buyer who reads a written reply may never open a result, because the page it quotes has already done the work. So a firm has a second thing to learn beside its rankings, which is the page a machine quotes when one of its buyers asks, and whether that page is the firm’s own. Google says its AI features rest on the same ranking and quality systems as its search, and that they ask for no special files, markup or tricks; to be shown as a supporting link at all, a page has to be indexed and eligible to appear with a snippet.
On the firm’s site we read in September 2026, 81 % of the citations Copilot gave it came on searches where Google showed the site as well; on the Directory the share was two thirds.
Work done for one of them, then, serves the other.
A firm needs one website, one plan and one budget for both. The pages it already has will carry the answers too, once each of them is the kind of page that an answer can use.
Google’s Index and Bing’s
Each of the AI replies in our records draws on one search engine’s catalogue of the web. Google’s AI Overviews, AI Mode and Gemini read the one that Google keeps; Copilot, the AI summaries in Bing and the partners that Microsoft names in its report read the one that Bing keeps. So an answer engine inherits the map that search has drawn of a firm.
In May 2026 the Directory showed us how close that bond is. From 18 to 28 May, Google’s AI answers drew on its pages between 8,425 and 19,293 times a day, and on 29 May it was 138 times.
When Google removed the Directory from its results, its own AI answers dropped the site the day before the notice came.
Copilot, which reads Bing’s catalogue, went on citing it, 2.03 million times in June 2026 and 2.72 million in the first 22 days of September. Yahoo, DuckDuckGo and Ecosia serve that same catalogue, and together they still sent the Directory 1,047 visits a day from 18 to 23 September.
A page is quoted by the engines that keep it.
02From Question to Short Search
Before it writes a word, an assistant turns what it was asked into a few words of its own. It reads what its index returns and cites its sources, and Bing lists those phrases for every site it quotes.
What a Grounding Query Is
Microsoft’s AI Performance report opened in public preview on 10 February 2026. For each site it lists the phrases that its AI used to fetch the pages it went on to cite, and it calls each phrase a grounding query. It gives them apart from the pages, as a sample, with the number of citations that each one drew, and beside them it counts how many of a site’s pages were cited on an average day.
Google describes the same step for AI Overviews and AI Mode as a query fan-out, in which several related lookups are run across the subtopics of a question and across its data sources. In both cases it is the search that the index answers, and the person’s own words stay with the assistant.
Take one query from the Directory’s record, ifc performance standards, which Bing lists with 14,565 citations. The page whose address carries those words is an explainer on the esg site, and it drew 15,230. Its answer opens on a sentence that an assistant can repeat whole.

Long Questions, Short Searches
On the firm’s site, 85 % of the citations Copilot gave it came on queries of one to three words. An assistant rewrites a person’s question into a few words before it looks, so that the question itself is in no report that a firm can open.
A firm that sizes its AI demand by counting long queries will find almost none, and be wrong.
The Directory’s record says the same at a larger scale, and says it very plainly. Searches of one to three words carried 65 % of the citations in Bing’s list, and searches put as questions carried 12 %. In August 2026 we read the record and saw that the citations had gone to the specific.
They went to narrow terms that Google showed rarely or never, and a third of them came on searches that are missing from the Directory’s Google data altogether.
Bing’s list opens on ethical sourcing practices and greenwashing, at 61,289 and 45,217 references. Below them run thousands of phrases of two to four words, such as anadromous fish migration or water-efficient appliances, and each is a term on which few sites have written a page of its own.

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 |
Named for the Search
We set each of the 15,839 phrases in Bing’s list against the 48,040 cited pages in the same report, and we marked a phrase whenever one of those pages had all of its words in its address. Small words such as what, is and explain were left out of it, and plurals were folded together. Bing never joins a phrase to a page itself.
The match is ours.
Of the 1,571,175 citations in the list, 66 % came on queries that had a page named for them, and it was 79 % on the 200 largest of them. On searches of two words the share was 89 %, on three words it was 86 % and on one word it was 76 %. It fell to 58 % at four words, to 32 % at five and to 14 % at six or more.
That is where a search starts to read like a whole prompt, and no address carries all of its words.
So a short query very often lands on a page that carries its words, and the Directory had written such a page for most of the short queries that Copilot ran.
That is the first half of how a page is chosen, and it is the half that a firm can see in its own report.

The numbers
| Words in the search | Searches | Citations | Share on searches with a page named for them |
|---|---|---|---|
| 1 word | 368 | 172,489 | 75.7% |
| 2 words | 3,826 | 508,733 | 88.9% |
| 3 words | 3,113 | 334,168 | 86.3% |
| 4 words | 2,185 | 139,597 | 57.6% |
| 5 words | 1,913 | 114,850 | 31.9% |
| 6 words or more | 4,434 | 301,338 | 13.9% |
| All | 15,839 | 1,571,175 | 65.6% |
Abbreviations and Misspellings
Some of the largest lookups had words in them that no address held. Copilot ran ikm 5,910 times, and the term page whose title those three letters shorten drew 5,879 citations. It ran bbso sustainability 7,880 times, and the tag page on behaviour-based safety observation earned 3,549 in the root site’s list, a little under half as many.
Both pages sit on the esg site. Each spells out what its search shortens, and for ikm the two counts sit within 1 % of each other, close enough to suggest that the engine made the link, and it made it by meaning, since none of the three letters sits in that address as a word.
Misspelt queries show it from the other side. Copilot ran what si cultural erosion 5,262 times and decomissioning 6,017 times. Both brought the Directory into answers, although neither spelling appears in any of its addresses, and Bing does not say which documents they reached.
For a firm the lesson is narrow and practical, and it costs nothing to apply. A page that writes its terms out in full, under the names its buyers use, is found by the acronyms and by the typing slips as well.
03One Search, One Page
Where Copilot ran a phrase often, a single address of the Directory usually took its citations, and Bing’s two separate tables agree on the counts almost to the unit.
Which Page Wins a Search
Copilot ran ethical sourcing practices 61,289 times, which is more than any other phrase in the Directory’s list, and the overview named for that term drew 61,268 citations. The two numbers come from separate tables in the report. So it is a finding in itself that they sit so close, and the pattern repeats further down.
When Copilot asked for urban wildlife definition, the word definition included, the term page Urban Wildlife answered it, with 6,268 citations on the query and 6,373 on the page. Seven more of the larger queries pair with one page each, and in each pair the two counts are within 8 % of one another. Most of those entries define a single term, at its own address, under the words of the query.
A page that owns one term this cleanly takes nearly every quote that the term brings in. We read it as the plainest evidence in the record of how the choice is made, with one caveat, which is that Bing itself draws no line from a search to a page.
| The search Copilot ran | Its citations | The page named for it | The page’s citations |
|---|---|---|---|
| ethical sourcing practices | 61,289 | Ethical Sourcing Practices, an overview on the root site | 61,268 |
| ifc performance standards | 14,565 | What Are the IFC Performance Standards and Why Are They Important to Financiers, an explainer | 15,230 |
| urban wildlife definition | 6,268 | Urban Wildlife, a term page | 6,373 |
| water-efficient appliances | 6,010 | Water-Efficient Appliances, a term page | 6,292 |
| ikm | 5,910 | the term page that spells the letters out | 5,879 |
| environmental vulnerability | 4,935 | Environmental Vulnerability, a term page | 4,907 |
| upcycling applications | 4,133 | Upcycling Applications, a term page | 4,134 |
| anadromous fish migration | 3,525 | Anadromous Fish Migration, a term page | 3,787 |
Pages That Split a Search
Greenwashing was the second largest phrase, with 45,217 references, and there are 143 cited addresses of the Directory that carry the word. Between them they earned 43,667 citations, and three of them took 89 % of that. The explainer What Is Greenwashing and How Can It Be Identified got 22,459 of them, a page on greenwashing in textiles got 8,779 and a definition of it in ESG got 7,634.
Each of the three answers its own question about the word. The other 140, mostly further explainers and question pages on the same subject, shared the last 11 %.
So the engine chose one page for each angle of the term, and it left the rest with very little. In September 2026 we had read the record the same way, that one plain page for a term is cited where two compete, and the greenwashing pages put numbers on it.

The numbers
| Page | Citations |
|---|---|
| learn.sustainability-directory.com/learn/what-is-greenwashing-and-how-can-it-be-identified | 22,459 |
| lifestyle.sustainability-directory.com/learn/what-is-greenwashing-in-the-context-of-sustainable-textiles-and-how-can-it-be-avoided | 8,779 |
| esg.sustainability-directory.com/learn/what-is-the-definition-of-greenwashing-in-esg | 7,634 |
| esg.sustainability-directory.com/learn/what-is-the-difference-between-greenwashing-and-bluewashing | 1,143 |
| fashion.sustainability-directory.com/learn/how-can-brands-communicate-sustainability-without-greenwashing | 510 |
| esg.sustainability-directory.com/learn/what-is-greenwashing-and-how-do-esg-ratings-attempt-to-mitigate-it | 278 |
| The other 137 pages | 2,864 |
Most Quotes, Few Pages
Across the eleven subject sites, Bing listed 138,256 addresses that Copilot had cited, and there were 5,322,923 citations spread over them. When we ranked them by citations, the first 1,180 carried a third of the total. Most were term pages, 652 of them, and there were 242 question pages and 188 explainers among the rest.
1,180 pages, under 1 % of those Bing listed, carried a third of the Directory’s citations.
Once it has been chosen, a page tends to be chosen again, and more often. On an average day in November 2025 Copilot cited about 7,261 of the Directory’s entries, and in September 2026 it was about 22,861 of them. It quoted each of them about 3.41 times a day at the start and 5.41 times a day at the end.
A page that answers a term well keeps the term, and each new search that turns on it lands there too. That makes the work finite.
A firm needs one plain page for each term and each fact that its buyers’ searches turn on: what it does, for whom, how long it takes, what it costs, where it works, in which languages, and what its methods are called.
04Inside the Quoted Pages
We read 98 of the Directory’s cited addresses as they answered on 7 October 2026, and we set the most cited of each kind beside those that were cited only once or twice.
Four Kinds of Page
The Directory sets its text in a few fixed templates, and those that carry most of Copilot’s citations open in different ways. A term page puts the term in the title and its meaning in one line beneath it, labelled Meaning, and then it offers three depths of reading and a glossary of related terms. Question pages do the same with a question and a one-sentence answer.
An explainer, filed under learn, sets a line of about 18 words under its title and then a short answer of a few sentences in a card. A tag page opens on a paragraph of about 44 words, and then it lists the articles that carry the tag, some 80 to 95 headings of them in our sample. Its markup names it a collection and carries no date and no defined terms, so the one sentence an engine could lift shares the screen with dozens of other titles.
On the root site there is a fifth kind, an overview of one term across several subject sites, and the most cited page in the whole record is one of them. Its first paragraph runs to about 65 words and the page to about 2,700. The most cited of them gathers what six subject sites say of its term, fashion, lifestyle, pollution, product, prism and ESG, under a heading each, and two of the ten we read also mark their questions and answers in the markup.

Meaning in One Line
On the ten most quoted term pages, the line under the title ran from 12 to 20 words, with its label. Question pages ran from 12 to 21 words and explainers from 13 to 21, and on each of them the line says what the thing is before anything else on the page does. On the term page Spatial Distribution the line reads “Meaning → Arrangement of phenomena across space, impacting resource access, equity, and sustainability outcomes”, and on the question page What Is Systemic Inequality it is one sentence of 21 words, which calls it “the embedded, enduring unfairness within society’s institutions, policies, and norms that creates unequal opportunities and outcomes for groups”.
A line like that can be lifted out whole. It needs no paragraph around it to make sense. An assistant can set it inside its own reply, with the source beside it, and the Directory’s term pages are built for exactly that, since they define one thing in one line at their own address; it was those definitions that drew the citations, half of all of them.
There is one more reason that we read into the pattern, and we mark it as our own. An engine pays for every page that it fetches and reads, so a page that answers with certainty in its first line, that sends its text in the HTML and that weighs little is the cheapest sure answer it can find.
The Directory’s quoted pages of those three kinds arrive at 21 to 25 KB of compressed HTML, with every word of the text inside it. There is no script that has to run before a crawler can read them.

Built Alike, Top and Bottom
If the way a page is made decided which page was quoted, then the most and the least quoted pages of each kind would differ, and in our sample they hardly do. The ten most quoted term pages drew a median of 9,821 citations. Nine quoted rarely drew a median of 1.
Yet they opened on lines of 15.5 and 20 words, ran to 2,378 and 2,625 words, and carried the same structured data, an Article and a WebPage with 18 defined terms. It is the same with question pages, with explainers, with tag pages and with overview pages, down to their weight.
What set the two groups apart was the search, and the sample shows it very plainly. Of the 49 most quoted pages in it, 43 carry in their address every word of at least one search in Bing’s list, and of the 49 that were quoted least, 16 do. All ten of the top term pages are named for a listed search, and only one of the nine at the bottom is.
A template that opens on its answer lets a page be used, and how often it is used then follows the demand for its term.
Our own notes of 6 October 2026 held that a certain length, plain tags and a simple build can put a page ahead of fuller articles, and the record fits that in two stages. The one-line meaning gets a page chosen for its term, and the rest of the page, some 2,400 words on a term page, gives an answer more to draw on once it is chosen. Bing’s counts show the first stage and cannot show the second, so we keep the second as our reading.
The sample is small, ten pages a group, and we read it in October, after the window that Bing’s counts cover. Every page in it that carries a date was published between March 2025 and June 2026, so most of them stood as they stand now for the whole of that window, but a firm should hold the likeness as firm and the exact medians loosely.

The numbers
| Kind of page | Group | Pages read | Citations (median) | First paragraph, words (median) | Words on the page (median) | Sent, KB (median) |
|---|---|---|---|---|---|---|
| Term pages | most cited | 10 | 9,821 | 15.5 | 2,378 | 23 |
| Term pages | cited 1 to 5 times | 9 | 1 | 20 | 2,625 | 24 |
| Question pages | most cited | 10 | 4,024 | 18.5 | 2,599 | 25 |
| Question pages | cited 1 to 5 times | 10 | 2 | 15 | 2,500 | 24 |
| Explainers (learn) | most cited | 10 | 6,747 | 17.5 | 774 | 21 |
| Explainers (learn) | cited 1 to 5 times | 10 | 2 | 16.5 | 792 | 21 |
| Tag pages | most cited | 9 | 1,809 | 44 | 3,803 | 36 |
| Tag pages | cited 1 to 5 times | 10 | 2 | 45 | 3,038 | 29 |
| Overview pages | most cited | 10 | 2,888 | 64.5 | 2,742 | 23 |
| Overview pages | cited 1 to 5 times | 10 | 2 | 59 | 2,230 | 21 |
Quotes per Page, by Kind
Across the whole of Bing’s lists the templates part ways, and they part in the order of their first screens. On average a term page drew 53.5 citations and a question page 45.6. An explainer got 33.6, a scenario 23.2 and a news story 19.2, and it was a tag page that got the fewest of all, at 17.1.
Term pages made up 36 % of the pages that were listed, and they drew 50 % of the citations. Tag pages made up 17 % and drew 7 %. In Google’s own clicks it was terms that took the largest share too, with 35 % of them.
We read the gap as what the first screen costs. A tag page states its meaning in a paragraph and then hands the reader a list, so the sentence that an answer wants sits lower and shares the page with dozens of other titles; a term page puts it first, and puts it alone.

The numbers
| Kind of page | Pages | Share of pages | Citations | Share of citations | Citations a page | Median |
|---|---|---|---|---|---|---|
| Term pages | 49,366 | 35.7% | 2,640,456 | 49.6% | 53.5 | 12 |
| Question pages | 23,599 | 17.1% | 1,077,105 | 20.2% | 45.6 | 12 |
| Explainers (learn) | 25,152 | 18.2% | 844,901 | 15.9% | 33.6 | 9 |
| Scenarios | 4,588 | 3.3% | 106,460 | 2.0% | 23.2 | 8 |
| News | 4,275 | 3.1% | 81,969 | 1.5% | 19.2 | 7 |
| Tag pages | 23,093 | 16.7% | 394,960 | 7.4% | 17.1 | 8 |
05Asked to Explain
Bing gives each search it lists an intent and a citation rate, and read together they show what kind of question the Directory was quoted on and how much of each answer it supplied.
Informational, Then Research
Of all the times Copilot cited the Directory, most came on queries that ask what something is. Bing labelled the searches behind 62 % of them Informational, 18 % Research, 5.5 % Learn and Solve and 1.3 % Comparison. Searches for a site by name and searches labelled Commercial each carried 0.1 %, and 1,665 of the 15,839 searches had no label at all.
The Directory sells nothing, so its mix is the mix of a reference; a firm that sells may well see more searches about buying, and the pages quoted on them will be its offer and its proof.
For a firm, the useful reading of this is that an answer quotes a page so that it can explain something. The pages that say what the firm does, how it does it and for whom are the ones within reach of a citation, and they are very often the pages a firm has written last.

The numbers
| Intent, as Bing labels it | Share of citations |
|---|---|
| Informational | 62.36% |
| Research | 17.67% |
| No label given | 6.43% |
| Others | 6.04% |
| Learn and Solve | 5.52% |
| Comparison | 1.25% |
| Creation | 0.35% |
| Navigational | 0.14% |
| Commercial | 0.12% |
| Utility | 0.06% |
| Live Event | 0.02% |
| Planning | 0.01% |
| Conversational | 0.01% |
| Media | 0.01% |
| Local | 0.00% |
A Share of One Answer
For each search, Bing also gives the Directory a citation rate, which its launch post of 10 February 2026 does not define. We read it, as the paper on the Directory’s engines does, as the Directory’s part of the sources that the answers on that search cited. Weighted by citations, the rate was 34 % across all of the searches.
On one-word searches it fell to 21 %. Where a search ran to two words or more, it held near 35 %.
In August 2026 we set down a claim that we could test on this record. Where many sites have written about a term, one site’s part of the answer should shrink, and where it is one of few pages that define a narrow term, its part should grow. The one-word searches, greenwashing among them, are where the most sources compete, and the rate is lowest there.
Of all the figures in this paper it is the one we can least defend, because it rests on a measure that Bing has not explained. If the rate counts something else, such as the share of answers on a search that cited the site at all, then the gap between one word and two would still stand, but the words around it would change, and a firm that reads its own report should hold the rate loosely until Bing defines it.

The numbers
| Words in the search | Searches | Citations | Citation rate, weighted |
|---|---|---|---|
| 1 word | 368 | 172,489 | 20.8% |
| 2 words | 3,826 | 508,733 | 36.3% |
| 3 words | 3,113 | 334,168 | 32.4% |
| 4 words | 2,185 | 139,597 | 36.8% |
| 5 words | 1,913 | 114,850 | 36.7% |
| 6 words or more | 4,434 | 301,338 | 34.5% |
Weekdays at a Desk
Over the whole record, from 1 November 2025 to 22 September 2026, Copilot cited the Directory less at weekends. Its rate there ran at 54 % of a weekday’s. So the person who asked was most likely at work, between two meetings, and that is the reader most firms want to reach.
06Ranked, Quoted, Clicked
Search results, AI answers and the visits that follow each keep their own list of a firm’s pages, and on our records the three lists overlap only in part.
Quoted Without Ranking
Of the 138,256 addresses that Copilot cited on the Directory, 55,347 never appear in the tables of Google’s search data that we archived. That is 40 %. Those tables are capped, so a page that is missing from them may still have been shown, and the true overlap is somewhat larger.
40 % of the pages Copilot cited never appear in the Directory’s Google data.
Half of the citations Copilot gave, 51 %, fell on pages to which Google’s search had sent a click at some point. On the firm’s site, 81 % of Copilot’s citations came on searches where Google showed the site as well, and on the Directory it was two thirds.
So the pages that rank and the pages that get quoted are two lists that share many names, and a report that is built on rankings alone will only ever see one of the two.
Different Pages for Each Engine
Google’s AI features drew mostly on the firm’s glossary, and Copilot drew on its practical guides, which are the ones that show how a thing is done.
One site gave two maps of what it is for.
On the Directory, Google’s AI answers leaned on tag pages, which had 29 % of their appearances, while Copilot’s citations leaned on term pages, which had 50 %.
A reading that looks at one engine describes half of a firm.
An AI-Only List
There are searches that bring Copilot to a firm while Google shows it nowhere. On the firm’s site we gathered those into one AI-only list. It is the clearest map we have of where a firm already holds authority that search has yet to recognise, and the next page is usually best spent there.
The Directory’s own AI-only searches carried a third of its citations, 34 %, on phrases that never appear in its Google query tables. Google’s tables never list them. Its largest show what such a list holds: questions typed whole, abbreviations, narrow terms, strings of keywords and misspellings.
| The search Copilot ran | Its citations | What it is |
|---|---|---|
| explain the concept of genetic conservation and its importance. | 18,853 | a question typed whole |
| bbso sustainability | 7,880 | an abbreviation |
| smart commute | 6,743 | a narrow term |
| ikm | 5,910 | an abbreviation |
| what si cultural erosion | 5,262 | a misspelt question |
| upcycling applications | 4,133 | a narrow term |
| recreational opportunities offered hydroelectric energy advantage | 3,623 | a string of keywords |
| gwms | 2,962 | an abbreviation |
Third Page, Highest Rate
Reports still judge success by position, and most of them quote a curve in which the first result takes about one click in three and the tenth almost none. We measured our own curve on three of the studio’s sites, with every search for the firms’ own names left out and one partner’s too, since people click a name wherever it stands. The window runs through 2025 and 2026.
Across 1,506,928 impressions and 9,746 clicks, a result in the first place was clicked 0.4 times in a hundred, one in the second 1.0 times and one in the third 0.8. At the tenth place the rate was 0.3 and for places 11 to 20 it rose to 0.9, while the highest rate of all, 2.1, came between places 21 and 30; results on the second and third pages drew 30 % of the clicks from 16 % of the impressions. Places 4 to 10 held two thirds of all impressions, 66.5 %, and drew 54.5 % of the clicks, and the first three places drew 11.6 % of the clicks from 9.9 % of the impressions.
Average position is a blunt measure, and a few hundred searches at the lower places can move a rate, so we do not read this curve as a law. Two readings fit it, and both of them point the same way.
A search where a page sits first is often a broad one, with an AI answer or a row of other features above every link, so the first result is seen and passed over. On the third page, a result usually answers a narrow search, typed by someone who wants exactly that page.
So a firm can stop treating the first place as the target. It can aim to be the page that answers, the one a buyer opens because it says just what they asked, whether a list put it first or an answer quoted it.

The numbers
| Average position | Clicks per hundred impressions |
|---|---|
| 1 | 0.4 |
| 2 | 1.0 |
| 3 | 0.8 |
| 4 | 0.6 |
| 5 | 0.7 |
| 6 | 0.7 |
| 7 | 0.7 |
| 8 | 0.5 |
| 9 | 0.4 |
| 10 | 0.3 |
| 11–20 | 0.9 |
| 21–30 | 2.1 |
| 31+ | 0.4 |
07What Makes a Page Quotable
Five traits run through the pages that the record shows being quoted, and what Google and Microsoft have published names most of them, in plain words of their own.
A Fact Stated Once
First, a quoted page opens on one checkable line, the thing it is about put plainly enough to repeat without rewording. For a firm that means what it does, for whom, where and since when, in a line that an answer can lift.
Second, each page owns one question and each question has one page. The Directory’s pairs of search and page show what that is worth. Its 143 greenwashing pages show what crowding costs, since the engine picked three and left all the rest with 11 %.
Our own note of 2 October 2026 put it in a line, that the page which states one checkable fact is the page an answer quotes. The record now carries the numbers behind it.

In the Order of Asking
Third, a quoted page follows the sequence of the buyer’s question, with the term first and then its parts, and each part opens on the words that a reader would type. The Directory’s term pages run that way, from the term to its meaning to three depths and a glossary, and they were the ones quoted most. Microsoft’s guidance for its AI answers asks for clear headings, for tables and for sections of questions with their answers, so that the facts are easy to find and to repeat.
Google asks for the content to sit in text, which is what the Directory’s pages do, with every word of it in the HTML as it is sent.
A page whose answer sits in a picture, a slider or a PDF without text gives an engine nothing at all that it can repeat.
Said Again in Data
Fourth, the page says the same thing in its structured markup, which names who the firm is, what it offers, where it works and in which language; a definition is marked as a defined term and a set of questions as questions.
On the Directory’s term and question pages the markup reaches past the page’s own term. Each related term in the glossary is written into it as a defined term with a one-line description of its own, 18 on a term page and about 25 on a question page at the median, so that one address hands an engine a small dictionary.
Google states that structured markup is not needed for its AI features and that no special schema exists for them, and it asks only that the markup match the visible text. The Directory’s own record agrees with the first half of that, since its most and its least quoted pages carry the same markup. We write it all the same, because it states each fact once, in words that every engine can read without guessing from the layout, and because the one rule that Google sets for it is the rule a quotable page lives by.
The studio’s own term pages carry a defined term whose description is the line on the page, word for word.
Fifth, every page agrees with every other, always. If the services text, the about section and the brochure describe the firm in three ways, an engine has no single fact to quote; a fact becomes quotable once everything the firm publishes says it the same way.

Mistakes in a First Reading
Almost every firm that we look at shows the same few errors, and each of them is cheap to fix once somebody has seen it. Some fill their site with long chat-style text in the hope of matching how people talk to an assistant, but the assistant searches with a few short words, so that text is rarely found for the query it imitates.
Others build pages that are near-identical for each city, for each product variant or for each name of a service. The engine then sees no single owner of the fact, and the answer goes to whichever page states it more plainly, which is often a directory.
Often the year a firm was founded sits in a picture, its services in a slider, its method in a PDF with no text and its prices in a brochure.
What is not in a page’s text and data does not exist for an answer.
Sometimes the site, the profiles and the listings disagree about what the firm does, where it is and since when, and then the answer picks one version or names a rival whose facts agree.
Last comes the habit of measuring with AI-visibility tools, which estimate how often a firm is cited from prompts that they choose and run themselves. A firm can read Copilot’s citations and Google’s AI impressions directly, in its own accounts, and anything else is a model of a model.
08The Question Map
It is a list of what buyers ask, in their own words, with each one put live to the AI engines and with what they answered kept.
Twenty or Thirty Questions
For each item the map records who is named, which page is quoted and which of the firm’s own pages should own the reply. A first map is short, twenty or thirty items, and it is always more useful than its length suggests.
The extract below is a construction, which we made for a firm that sells a technical service across borders.
| The buyer’s question | Asked in | Named in the answer | Quoted page | The firm’s page that should own it |
|---|---|---|---|---|
| who does this, for a firm like mine? | Google AI Mode | two competitors | a directory listing | the service page, its first screen |
| how long does it take? | Copilot | no one | none | a new page that states the timeline plainly |
| what does it cost? | Google AI Mode | one competitor | its price page | the offer page, with the ranges it can state |
| is it done in my language? | Copilot | the firm | the firm’s glossary | the languages page, with each market’s address |
Found on a First Map
There are three things that turn up for most firms at the start. Some questions name nobody, which leaves an open space that a plain page can take.
On others a competitor is quoted from a page that states the fact more plainly than the firm’s own does. And some of them the firm’s pages almost answer already, where moving one line to the first screen is the whole of the fix.
Read Again at Ninety Days
At the end of a quarter we put the same questions to the same engines, in the same browser, and mark each one whose answer moved. A question that did not move tells the firm which page to rewrite next, and the map grows by the questions its buyers add.

Decided in One Room
Nobody writes a quotable page alone. Its one fact has to be the firm’s position, its words the approved words and its data the site’s, and every other page has to agree with it.
That makes it a choice for the whole firm. The question map belongs to the direction a firm sets for itself, which covers the position and its words, a brief for every page with its first screen first and every claim fixed once, and the term map beside the question map. All of it is set in one room with the people who decide, and checked again at ninety days against the same record.
The machine quotes the page that is plainly right. A firm whose pages agree gives it one.
09Method and Sources
Every figure above comes from accounts that a firm already holds or from pages that anyone can open, and each one is set out here with its source and its window so that it can be checked.
Bing’s Lists, Read Twice
Microsoft’s AI Performance report covers Copilot, the AI summaries in its search and some partners, and it gives each site three records. Its daily totals are complete, and for the Directory’s root site they come to 17,500,282 citations in 326 days, from 1 November 2025 to 22 September 2026. Grounding queries and cited pages come as samples.
The root site’s lists hold 15,839 searches with 1,571,175 citations and 48,040 pages with 1,970,256, and the eleven subject sites’ own lists hold 138,256 pages with 5,322,923.
We read each of the lists for a question of its own. Searches, their intents and rates, and their match to pages come from the root site’s lists, which the report gives side by side. Kinds of page and their share of citations come from the subject sites’ lists, which are larger, as in the archive’s own reading of 25 September 2026.
The same page can carry different counts in the two, since each is its own sample. Wherever one page is named we give its count from the root site’s list, as the paper on the Directory’s engines does.
A search counts as named by a page when the page’s address holds every content word of the search, with plurals folded and words such as what, is, explain and definition left out. Bing never joins the two. Every pairing here is ours.
Google’s side comes from the archive’s reading of Search Console, 13 May 2025 to 24 September 2026, whose page and query tables are capped, so a page or a search that is missing from them may still have been shown.

Pages Read Live
On 7 October 2026 we fetched 100 of the Directory’s cited addresses as a browser would and took each one as it arrived. For each of five kinds, term, question, explainer, tag and overview, we took the ten most cited in Bing’s lists and ten drawn at random, with a fixed seed, from those cited one to five times. Two of them, both on the term site, answered 503, as that site has done since its hosting lapsed, and so 98 were read.
From each we took the first paragraph, the words on the page, the structured data, the headings, the weight of the HTML as it was sent and in full, and whether its address carries every word of a search in Bing’s list.
We drew the random ten with Python’s generator set to the seed 3, sent each request with a desktop browser’s user agent and waited 0.6 seconds between requests; the scripts, the 98 copies of the HTML and the table of what each one held sit in the run’s own folder. The six screen images were taken in Chrome at a 1,440 px viewport and cut in height only. We checked each one pixel by pixel for any person, address or account, and we cut away the image band on the overview page, which shows part of a person.
The Firm and the Click Curve
The second site was read in September 2026, in Search Console, in Bing Webmaster Tools with its AI Performance report and in Analytics with its enquiry form’s opens. We never name it or its field, we give its figures as shares, and this paper adds nothing to what it said of that site before.
Click-through by position comes from Search Console on three of the studio’s sites in 2025 and 2026, with searches for the firms’ own names and one partner’s removed, over 1,506,928 impressions and 9,746 clicks grouped by average position. The file that holds it does not name the three sites, and neither does this paper.
Read side by side, these reports answer the question a firm can act on. Setting each page’s AI exposure against its form opens and its searches for the firm’s own name, and comparing the click-through of quoted and unquoted pages at the same place, shows whether a machine’s mention of the firm brings work.
Documents Cited
Google set out its guidance on AI features in two pages, AI features and your website, updated on 10 December 2025, and its guide to generative AI features in Search, updated on 10 July 2026. Microsoft introduced the AI Performance report on the Bing Webmaster Blog on 10 February 2026, with its definitions of citations, cited pages and grounding queries and its advice on content.
Google introduced the Generative AI report in Search Console on Google Search Central. The full account of the Directory’s engines is the paper Every Engine Keeps Its Own Record, and the extract from a question map above is a construction of ours, drawn from the record of no client.
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.

