01What the Fee Bought, in Numbers
For sixteen months a firm paid an agency each month to look after its search, and when its own accounts were read beside the studio’s, they gave seven findings.
- Google showed the firm’s pages far more often than anyone chose them, and from 10 April 2025 to 9 August 2026 it showed them 1,967,075 times for the 4,429 visits that it sent, and the number of visits it sent in a month fell by 44 % in ten months.
- Most of what it sent was demand the firm already had, since 1,203 of the 1,624 visits whose query was reported carried the firm’s name, and in the last 28 days 43 of 50 did.
- Places rose while visits fell, as the average position of the queries without the name went from 43 to 31 over the three months to 12 August, and 163 of them were in Google’s top ten, against 97 over the sixteen months.
- The AI answers drew on the site about 23 times for each visit that Google sent, as there were 13,871 uses of it in the 84 days to 9 August 2026 and only 597 visits.
- All 74 leads carried one and the same source tag, so the record that the firm itself kept from 28 February to 6 July 2026 could not say where a single one of its buyers came from.
- Machines filled the denominators, so that when the tags on the Sustainability Directory began to wait for a real scroll, tap or key press on 17 September 2026, 97 % of the sessions it had recorded fell away.
- A dated record prints its own misses, and by 7 October 2026 three of the ten forecasts that Allomedo’s tracker had filed were already wrong, with the drop of 3 October in it hour by hour.
Our Research Question
This paper asks what a firm buys with the fee it pays each month for its search. It reads the firm’s own accounts with the searches for its name set apart, and it asks what record would let the owner see the answer every month. Most of the evidence comes from one firm, which stays unnamed here along with its industry and its markets, and the rest comes from sites the studio built and runs, the Sustainability Directory, Sentiention, Allomedo and the three sites behind our click curve, so their records are ours to print in full.
Every figure in it carries its source and its window. The last part, Method and Sources, says how each record was pulled, read and checked, so that another firm can repeat the work on its own accounts.
Twelve Months on a Retainer
Its site was well built, and the owner kept the agency while asking us for a reading of the company’s own record beside it, to learn what the monthly fee was buying. A retainer of this kind usually brings a summary of rankings and traffic each month, a list of recommendations and a tally of whatever new pages or links were made. What it seldom measures is which pages bring enquiries, or from whom, and it almost never says, with a date and a number, whether last quarter’s change moved anything at all.
We wrote our answer down on 22 August 2026. The retainer had kept the demand that the business already had in good order, and nobody had ever pointed the work at buyers who had not yet heard of it.
Totals, With Their Windows
Google keeps 16 months of a site’s search history (Google Search Central), and the reading took all of it that there was. The firm’s export of 12 August 2026 runs from 10 April 2025 to 9 August 2026, which is 487 days; for the AI engines it was what their own reports held on that same day, which came to 84 days for Google and 180 for Copilot.
| What was counted | Figure | Window |
|---|---|---|
| Visits from Google | 4,429 | 10 April 2025 to 9 August 2026 |
| Times Google showed the site | 1,967,075 | 10 April 2025 to 9 August 2026 |
| Visits per showing | 0.23 %, one in 444 | 10 April 2025 to 9 August 2026 |
| Visits whose query Google reported | 1,624, or 37 % | 10 April 2025 to 9 August 2026 |
| Visits known to have come by the firm’s name | 1,203, or 27 % | 10 April 2025 to 9 August 2026 |
| Visits from Google in a month | 369 at the peak, 205 at the end, down 44 % | September 2025 to July 2026 |
| Uses of the site in Google’s AI answers | 8,391 | 18 May to 9 August 2026 |
| Citations of the site by Copilot | 12,971 | 11 February to 9 August 2026 |
Only 37 % of those visits came with a query that Google reported. It kept the rest back for privacy, so most of what people typed to reach the site never showed up in anyone’s report.
Of the ones it did report, 74 % carried the name on the door. Across all of them, reported or not, 27 % are known to have come that way, and the true share is very likely higher.
Over 16 months of its own accounts, the firm’s pages were shown 444 times for each visit that Google sent them.
02What the Year Bought
Most of the visits came from people who already knew the firm, and the queries that would bring it a new buyer were barely reached.
The Name in the Search
Of the 1,624 visits whose query Google reported, 1,203 carried the brand of the company and 421 did not. A query with the brand in it comes from demand that the business already has, which was built by its sales team, its trade fairs, its reputation and its customers’ memory; the rest come from people who want what it offers and have not yet heard of it. So a retainer that raises the first kind and leaves the second where it was is keeping in order the demand that was there already.
That leaves 421 in sixteen months that the record can tie to someone who did not type the name. They thinned out, too. There were 45 of them in the three months to 12 August, about 15 a month, and 7 in the final 28 days.
Of all the figures in this paper the 421 is the one most likely to move, because Google holds back rare queries. The 2,805 visits it withheld here may well hide more searches of that kind than the reported ones show, and nothing in the record says how many.

The numbers
| Query behind the visit | Visits | Share of all visits |
|---|---|---|
| A query with the firm’s name | 1,203 | 27.2 % |
| A query without the name | 421 | 9.5 % |
| A query Google withheld | 2,805 | 63.3 % |
| All visits from Google | 4,429 | 100 % |
Month by Month
When the export’s daily rows are added up, the visits that Google sent the firm stayed between 218 and 308 in each of May, June, July and August 2025, reached 369 in September and held above 300 until December. From January 2026 on they eased, to 294, 265 and 247 in the three months up to March; there was one rise, to 308 in April, and then they were 228, 249 and 205 by July.
The number of times that Google showed the site had its own peak earlier, at 250,956 in July 2025. It was down to 136,662 by September and to 51,632 by July 2026, which is 79 % below that best month. The reading first called the drop in visits a collapse, and its own checker changed the word, since no single date carries it.
It is a slide of 44 % across ten months. Each step is small, from a few visits to a few dozen, and the size of it only ever shows when the year is set against its best month.

The numbers
| Month | Visits from Google | Times shown | Visits per thousand showings |
|---|---|---|---|
| May 2025 | 248 | 196,150 | 1.3 |
| Jun 2025 | 218 | 210,263 | 1.0 |
| Jul 2025 | 308 | 250,956 | 1.2 |
| Aug 2025 | 229 | 234,880 | 1.0 |
| Sep 2025 | 369 | 136,662 | 2.7 |
| Oct 2025 | 345 | 93,504 | 3.7 |
| Nov 2025 | 343 | 103,132 | 3.3 |
| Dec 2025 | 305 | 117,507 | 2.6 |
| Jan 2026 | 294 | 76,731 | 3.8 |
| Feb 2026 | 265 | 74,511 | 3.6 |
| Mar 2026 | 247 | 58,861 | 4.2 |
| Apr 2026 | 308 | 63,114 | 4.9 |
| May 2026 | 228 | 79,759 | 2.9 |
| Jun 2026 | 249 | 58,674 | 4.2 |
| Jul 2026 | 205 | 51,632 | 4.0 |
Rankings Up, Visits Down
Among queries without the brand, the average position of the firm’s pages improved from 43 over the whole sixteen months to 31 over the three months to 12 August 2026, and then it slipped back to 35 in the final 28 days. In those three months there were 163 such terms in Google’s top ten and 207 more at places 11 to 20, against 97 and 129 over the whole of the period.
And of the arrivals that Google did report, the share that came without the brand shrank with each newer window, from 26 % of them across the sixteen months to 14 % in the final 28 days. In those 28 days there were 50 reported visits to the site. Only 7 of them came without its name.
Our reading listed in full the 463 terms that already stood at places 4 to 20, which it called the shortest path to more visits. The agency had done honest work by the measure its reports used, and the places its reports tracked plainly went up. But read without the split, a year like this one looks like progress, and the owner goes on paying for a climb that brings very few new buyers.

The numbers
| Window to 12 August 2026 | Average position, queries without the name | Such terms in the top ten | At places 11 to 20 | Reported visits | Without the name | Share without the name |
|---|---|---|---|---|---|---|
| Sixteen months | 43 | 97 | 129 | 1,624 | 421 | 26 % |
| Twelve months | 36 | 109 | 152 | 1,264 | 273 | 22 % |
| Three months | 31 | 163 | 207 | 229 | 45 | 20 % |
| 28 days | 35 | 87 | 148 | 50 | 7 | 14 % |
The Answers It Fed
Copilot cited the pages of the firm 12,971 times in the 180 days from 11 February to 9 August 2026. Over the 84 days from 18 May, Google’s AI used them 8,391 times, and that was all that each of the two reports held when the accounts were read. In those same 84 days there were 5,480 citations from Copilot and only 597 visits from Google, which makes it 23 uses for each visit.
The count for Copilot sits in Bing Webmaster Tools and the one for Google in Search Console, where any company can read both of them for itself. A summary of rankings and visits has never had a line for either.

The numbers
| 18 May to 9 August 2026 | Count |
|---|---|
| Uses in Google’s AI answers | 8,391 |
| Citations by Copilot | 5,480 |
| Visits from Google | 597 |
Asked live, the AI engines answered the firm’s buyers from the pages of its competitors, and the firm fed answers that never named it.
03How a Monthly Report Is Made
Two of Google’s tools and the agency’s own log are what such a summary is built from, and every rate in it rests on the way those tools count.
Four Lines and Their Sources
Rankings come first in most summaries, and most of the time they are Search Console’s average position. In Google’s own definition that is the highest place a link to the site held, averaged over each query it appeared for, and the click-through rate is clicks divided by impressions (Search Console Help). Traffic is the console’s clicks or the sessions Analytics recorded, and the recommendations and the tally of new pages and links come from the agency’s own log of its work.
Each of those counts carries a property the summary seldom states. An average position moves with the mix of queries behind it, and it can improve when Google stops counting the showings far down its results, as Zeo found across most sites in September 2025. It can climb while visits fall, too, as this firm’s did, from 43 to 31.
The console reports the query behind only 37 % of the clicks that it counts here. And Analytics counts every session its tag records, people and machines alike.
None of the four lines ever splits a visit by the name in its query, ties it to the enquiry it became, counts the AI answers that used the page, or states what share of the sessions were machines. Those missing counts, with a dated forecast beside them, are the five lines of the sheet in One Page a Month.

Machines in the Visits
From 17 September 2026 the tags on the Sustainability Directory waited for a real scroll, tap or key press before they would load. The sessions that Analytics recorded fell from about 692,700 a day in the first half of September to about 20,200 a day from 18 to 23 September. So 97 % of what it had recorded as visits had never touched the page.
Direct, which is the channel Analytics gives to the visitors that it cannot place, fell from 680,399 a day to 14,584. Search barely moved. Its arrivals went from 5,744 a day to 5,161, and those from AI assistants held as well, at 86 and then 111, which means that they were people.
Any rate that a report divides by all sessions, engagement or the share who wrote in, was divided mostly by agents before that day. We separate agents from people before any engagement figure is read, and a report that skips that step always measures the agents along with the people.

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 |
04How Search Changed Under the Report (2025 to 2026)
The way that Google counts lies beneath every figure in a monthly summary, and over the months of this firm’s retainer the records we keep show that it shifted more than once.
From Ten Links to an Answer
Backlinko’s curve, the one most reports quote, has the first result clicked 27.6 % of the time on average, and it gives the top three 54.4 % of all the clicks (Backlinko). It rests on Search Console data for 1,312,881 pages and 12,166,560 queries that Semrush provided, and it was last updated on 16 April 2025. In it the first result is ten times as likely to be clicked as the tenth.
We measured our own curve on three of the studio’s sites, in their own accounts, across 2025 and 2026. Each search for the firms’ own names and one partner’s was taken out, because people click a name wherever it stands. Across the 1,506,928 impressions and 9,746 clicks, the first place drew 0.40 clicks in a hundred, the second drew 0.99, the third 0.85, and the tenth no more than 0.33.
Places 11 to 20 drew 0.87, which is more than the first did. And places 21 to 30 drew 2.08, which was the highest rate on the whole curve.

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 |
Where the Clicks Land
Set in five groups by place, the studio’s curve spreads what it was shown and what it drew unevenly. Places 4 to 10 held two thirds of the impressions, 66.5 %, and drew 54.5 % of the clicks. And those at 11 to 30 held 15.8 % of the one and drew 29.5 % of the other, nearly twice their share of what was shown.
Taken together, places one to ten drew 0.56 clicks in a hundred and places 11 to 30 drew 1.21. The first place held 3.1 % of the impressions and 1.9 % of the clicks; places 21 to 30 held 4.4 % of the one and 14.1 % of the other.
A result in first place is often there for a broad query, where Google sets its own features and an AI answer above all the links, so it is seen and passed over. One on the third page of results is often there for a narrow query, typed by someone who wants exactly that page, though we have tested neither reading.
Sentiention shows the same gap between being seen and being chosen. In the first three weeks that its domain property holds in Search Console, from 13 September to 4 October 2026, it was shown 113,630 times and clicked 536 times, or 0.47 in a hundred, as we read it on 4 October; Google fills in the latest two or three days of such a window afterwards, so that rate can still move. We hold this curve loosely, and a buyer should too.
An average position is a blunt measure, and a few hundred queries at the lower places can move a rate. The safest use of our curve is as a check, and any business can run it by pulling the same table from its own console with its brand taken out.

The numbers
| Average position | Impressions | Share of impressions | Clicks | Share of clicks | Clicks per hundred |
|---|---|---|---|---|---|
| Places 1 to 3 | 149,926 | 9.9 % | 1,132 | 11.6 % | 0.76 |
| 4 to 10 | 1,001,484 | 66.5 % | 5,312 | 54.5 % | 0.53 |
| 11 to 20 | 171,598 | 11.4 % | 1,499 | 15.4 % | 0.87 |
| 21 to 30 | 66,219 | 4.4 % | 1,378 | 14.1 % | 2.08 |
| 31 and beyond | 117,701 | 7.8 % | 425 | 4.4 % | 0.36 |
September 2025, the Hundred Results
On 9 October 2025 Zeo published its analysis of a change that Google had made about a month before, when it disabled the setting that returned 100 listings on one page (Zeo). In the data that Zeo had, 87.7 % of the sites lost impressions and 77.6 % of them lost unique keywords, their average positions looked better, and their clicks did not change at all. Zeo gives no sample size.
We compare a site with itself, one year against the one before, in clicks only. Clicks were the one count that held still.
The firm in this paper had its best month for visits, 369, in September 2025, which was the month of the change. Its showings fell 62 % from that month to July 2026, and 79 % from their own peak in July 2025. Over the same span its average position for queries without its name went from 43 over the sixteen months to 31 over the latest three.
Part of both movements may be Google’s change, which in Zeo’s data lowered impressions and flattered positions across most sites. The company’s own record has no way to separate the two.
Divided by showings, its visits went from 1.2 in a thousand in July 2025 to 2.7 in September and 4.0 in July 2026. A summary that printed that rate would have shown it more than tripling, in a year when the visits themselves fell by 44 %.
But visits from Google are clicks, and those were what Zeo found unchanged, so the fall from 369 to 205 stands fully outside it. If a firm’s report shows better places and fewer showings from that month on, it is the clicks that are always the line to read.
The Answers Grew
Bing opened its first report on where Copilot cites a publisher’s pages on 10 February 2026, and the history in it reached back to November 2025. On the Sustainability Directory network, Copilot cited its pages 743,409 times in November 2025, the first month that was on record, and it cited them 2,719,538 times in the first 22 days of September 2026. In between, the count dipped in the spring and then set new highs in June, August and September.
Google’s AI answers come into the record in May 2026, when they drew on the Directory 172,938 times, at between 8,425 and 19,293 a day from 18 to 28 May. On 30 May 2026 Google issued a manual action against the whole domain, and its AI answers had dropped the Directory the day before the notice. June held 340 such uses, and after June there were almost none.
Over those months the clicks that Google sent to the network peaked at 370,755 in March. The action stopped them. Bing kept on sending between 31,000 and 39,000 clicks a month from February, and Every Engine Keeps Its Own Record tells that part of it whole.
The engines also chose their own pages. Those that define one term drew half of the citations from Copilot and a third of the clicks from Google. The top 1 % of the pages drew 35 % of all the citations, and on weekends the count ran at about half the weekday level, which is how people asking at work behave.
Copilot mostly searches in terms, too. Of the searches Bing listed as the ones it ran to ground its answers on the Directory, only 12 % were phrased as questions, so we size the demand that reaches the AI engines the way we size search demand, from the terms people type. What a citation is worth to a firm that sells is still an open question, and none of the records in this paper ties one to an enquiry.

The numbers
| Month | Copilot citations | Google’s AI answers |
|---|---|---|
| Nov 2025 | 743,409 | |
| Dec 2025 | 787,568 | |
| Jan 2026 | 890,102 | |
| Feb 2026 | 1,299,961 | |
| Mar 2026 | 1,852,508 | |
| Apr 2026 | 1,655,612 | |
| May 2026 | 1,406,185 | 172,938 |
| Jun 2026 | 2,031,858 | 340 |
| Jul 2026 | 1,926,793 | 5 |
| Aug 2026 | 2,186,748 | 0 |
| Sep 2026, 1 to 22 | 2,719,538 | 7 |
Of the pages Copilot cited on the Directory, 40 % had never appeared in Google’s data at all.
05Where Buyers Actually Write
An enquiry begins on the firm’s own form, in the words of the person sending it, and that is where a reading ends its count.
Step Zero
Before anyone says what a company should do, a reading measures what its presence has already done. It begins by joining four records, page by page. Search Console shows what Google shows, for which queries and on which pages, and Analytics shows what visitors do and which pages turn into enquiries.
Bing is the second engine and the index Copilot answers from. The AI answers are the buyers’ own questions asked live, each one kept with whether it names a firm and which pages it quotes.
Before anything is read, the history is archived, because the console drops its oldest day every day. A business that never saved it has no way to set this year against the last one it had.
For the firm in this paper the joins come down to a handful of numbers. Of its 4,429 visits, Google withheld the query behind 2,805 and reported it for 1,624. And only 421 of those came without the firm’s name, so a split by name could tie fewer than one visit in ten to a buyer who did not already know the firm.

Enquiries by Page
Three things are counted for each address on the site: how many visitors opened the enquiry form, how many sent it, and how they said they found the firm. Such a form, one that asks that last question and keeps the address the visitor arrived on, is cheap to add, and with it a quarter’s work can be judged by what it brought. Before any of those counts becomes a rate, we split the visitors into people and agents, since on the Directory the agents were most of what Analytics recorded.
The firm’s own lead record shows what is lost without such a form, since every one of the 74 leads that reached it from 28 February to 6 July 2026 carried the same source tag, the one for search. So the field held a filing habit. No lead in it can be traced to a search, a page or a referral.
We used the tag as the team’s own record that enquiries came in through the site, and built no claim on it. A form that kept the page and asked how the buyer found the firm would have given each of those 74 leads a source of its own.
The reading printed all of the 74 leads that the firm had shared and the 72 first messages its buyers wrote, word for word.
First Messages, Word for Word
Enquiries need to be read as well as counted. Each one is marked once as a fit, a near fit or one the firm should not serve, in the terms its sales team would use to qualify a buyer.
Over a year, those marks show which pages bring the buyers it wants, which bring the wrong ones, and which questions on the site draw people it should not serve at all. An owner recognises that measure at once, because it is the one a sales team already uses.
Of the 74 leads, 26 were quoted. The largest group, 30 of the leads, ended because the buyer sent no files or specifications for the firm to price from. Beside them, 19 of the leads carry no reason, 18 were won or are still open, 6 found that the price did not fit, 5 were empty and two asked for work that the firm does not do; a lead can carry two of those tags.
Where a buyer named the exact service, the quote nearly always went out, for 8 of the 9 leads in one service and for 8 of 8 in two others taken together. The 35 leads that named no service fared worst. None of them was quoted, and three buyers wrote that the form itself had failed on them.
Their first messages carry the buyers’ own words for the work and for the files they hold, and those were words that the site did not lead with. Three of the 72 raised confidentiality before anyone had answered them.

The numbers
| What the team recorded | Leads |
|---|---|
| No files or specifications to price from | 30 |
| No reason recorded | 19 |
| Won, or still open | 18 |
| The price did not fit | 6 |
| An empty enquiry | 5 |
| Outside what the firm does | 2 |
| All leads (a lead can carry two tags) | 74 |
When the Brand Comes Back
More queries for the firm’s own name usually mean that something outside Google has worked, a trade fair, a campaign or a recommendation. If they grow after a change to the site, a likely reason is that the change made the company easier to recall for people who first met it through a query without its name. That is the effect a retainer should be able to show, and only a count split by name and by page will find it.
The same split by page tells a business whether the pages the AI engines quote are the pages its buyers write from. Where they turn out to be different pages, a firm paying for visibility in the answers is paying for pages that bring no enquiry, and it is the count by page that tells it so.
06What a Reading Leaves Behind
What we delivered stayed with the firm, written to travel inside it and to its agency, with no studio branding on any page.
Documents of Record
First came what is true and how we know it. It arrived as seven files, each on one subject.
- The map of the work, six pages on how it was checked and what it decided.
- What the engines show, from Google, Bing and the AI answers, and what the firm’s visitors do there.
- Buyers’ own words, all the leads that the firm had shared and the first messages its buyers wrote, word for word.
- Who else answers, the 160 competitors that we had verified, each of them read in a structured way, 195 reads in all.
- The site, all 288 of its pages, each one scanned.
- One more market, the one the firm wanted next, read the same way.
- Ten opportunities, each with its own test.
All seven share one cut-off, 12 August 2026. Six decisions came out of the work, each standing on at least two of the files, each traced to a table and each table to a file.
Fifteen Findings Attacked
Every number in the files was recomputed from its raw export before anything was written on them, 22 base figures in all. All 22 matched. Then each of the 15 headline claims in the search report went to an independent checker whose one instruction was to disprove it from the source files.
Of the 15, 7 held as written. The other 8 held with a qualification. Each was corrected in the pages where it stood, with the correction printed beside it, so that a reader sees what the claim was and what the check made of it.
Two of the corrections changed how the firm reads its year, and the first was the drop in visits, which had been written as a collapse until the checker found no date on which it broke; it became a softening of 44 % over ten months. On Google’s AI answers, the first pass had read each page before the answer block loaded, so that side was measured again with a pass that waited for a real interaction, and both passes were kept.
Smaller ones show the grain of the checking. One claim had all 45 visits from AI assistants in the year falling in its last three months, and the checker could confirm 28 of them in the last 28 days and not the rest, so the three-month version was dropped. Another count, 8,390 uses in Google’s AI answers, came back as 8,391 and was put right.
Three questions stayed open, stated as open, each with its own number. One is that the console’s page export and its chart export disagree by a few percent, so the reading never adds page-level showings up to a total for the site.

Direction, as Delivered
Then came the work that the firm could act on, which was sent in August 2026. It runs to a deck of 73 pages and an annex of seven documents, 129 pages in all. The brand takes 14 of them, the voice 20, the briefs for the site’s pages 37, the search work 19 and the visual system 13, with a working annex and the page copy beside them.
With it came 1,316 frames, each of them named, described and tagged, which had been cut from 375 source frames in 14 themes, where about 500 had been promised. Ten experiments are named in the deck, each with its measure, and the 90-day clock for them runs from the day that the first pages go live. A working session with the firm was written up as The Ten Moves, a pack of 36 pages.
The record holds the firm’s state as it stood in August 2026, so that each change can be checked against it over the next three to six months. Every figure in it can be checked by the owner without us.
Two Weeks Beside the Agency
Our standard reading runs on read-only access, delivers the record by day 7 and the full reading by day 14, and changes nothing on the site in that time. It ends in two reports. In the search report a firm finds where its demand comes from, which pages turn it into enquiries and what changed and when, with every number recomputed from its source.
The AI visibility report lists which buyer questions get an answer, which firms are named, which pages are quoted and which of the firm’s own pages should own each answer. With them come the question list, which matches each of the buyers’ questions to the page that should answer it, in the order worth doing, and the record itself, archived before the engines drop its oldest months. A reading sits beside whoever runs a firm’s search and reads what their work buys, and where the agency is already doing the right work, it gives the agency the figures to show that.

07One Page a Month
Recurring search work holds up when each piece of it is named, dated and checked against the first record, so that the owner can see what it bought.
A Sheet the Owner Reads
Everything the usual summary lacks fits in five lines. Each of them is dated, and each can be checked against the first record.
- Demand, split in two, the clicks and the enquiries from queries with the firm’s name and from queries without it, month on month.
- Enquiries by page, the page a visitor arrived on before writing, and what the form says about how they found the firm.
- The answers, the buyers’ own questions asked live each month in Google’s AI Mode and in Copilot, with who was named and which page was quoted.
- A stated share of visits filtered out as automated, so that every rate is measured against people.
- Forecasts, dated, what last month’s change was expected to move, by how much and by when, and whether it did.
A monthly fee is fine on those terms. The rule we hold for recurring work is that none of it goes unmeasured, and a fee that meets it is one an owner can defend.
We hold our own sites to that sheet. Every morning a tracker reads the Search Console, Bing and Analytics of Allomedo, keeps the report of each run, logs every change that reaches the site with the date it was made, and files each forecast with the day on which it is to be checked. It has run since 26 September 2026, at 04:30 Poland time, and it changes nothing on the site that it reads.

Allomedo’s First Fourteen Days
Allomedo was built on 21 September 2026, and its sitemap, which listed 347,825 addresses with no errors and no warnings, went to Google on 22 September. Its clicks from Google rose to 232 on its third day, fell to 54 and 83 over the first weekend and came back to 243 on Monday 28 September. The times it was shown passed 10,000 a day on three days of that fortnight.
When the first weekend fell, the tracker’s readout of 28 September wrote down two readings and one test. Either a business site was having its first weekend, or Google had cut the new site’s exposure; Monday 28 September would decide which, and our own deployments, the site’s health and the Analytics tag were ruled out first. Monday brought 243 clicks.

The numbers
| Day | Clicks | Impressions |
|---|---|---|
| 22 Sep 2026 | 14 | 304 |
| 23 Sep 2026 | 145 | 5,307 |
| 24 Sep 2026 | 232 | 10,421 |
| 25 Sep 2026 | 193 | 9,411 |
| 26 Sep 2026 | 54 | 4,820 |
| 27 Sep 2026 | 83 | 5,702 |
| 28 Sep 2026 | 243 | 10,793 |
| 29 Sep 2026 | 219 | 11,239 |
| 30 Sep 2026 | 222 | 9,656 |
| 1 Oct 2026 | 184 | 7,825 |
| 2 Oct 2026 | 103 | 6,524 |
| 3 Oct 2026 | 4 | 344 |
| 4 Oct 2026 | 5 | 369 |
| 5 Oct 2026 | 9 | 236 |

Forecasts Kept as Filed
Each prediction carries the day it was made and the day it is checked, and no cause is claimed before its test. Ten were filed, on 26 and 28 September. Every morning’s run prints each one beside what the data says, and the run of 6 October measured them over the settled days from 27 September to 3 October, the run of 7 October over 28 September to 4 October.
| Made | Checked | Prediction | Result |
|---|---|---|---|
| 26 Sep | 3 Oct | Clicks settle at 150 to 350 a day | 151 a day to 3 Oct, held at the floor; 140 a day to 4 Oct |
| 26 Sep | 3 Oct | Impressions settle at 7,000 to 15,000 a day | 7,440 a day to 3 Oct, held; 6,679 a day to 4 Oct |
| 26 Sep | 3 Oct | Googlebot is back, with at least 8 of the 33 panel pages crawled after 23 Sep | 5 pages, missed |
| 28 Sep | 6 Oct | Weekdays 28 Sep to 2 Oct each reach 120 clicks or more, and the weekend of 3 and 4 Oct falls to 30 to 80 | 243, 219, 222, 184 and 103, then 4 and 5, missed on Friday and on the weekend |
| 28 Sep | 28 Oct | No visitor from an AI assistant through October | 3 sessions already, wrong |
The weekday forecast missed twice. On Friday the site had 103 clicks, and on the weekend it had 4 and 5. It carried its own alternative, that if every weekday stayed under 80 clicks Google had cut the site’s exposure on 25 September, and none of them fell that low.
The visitors from AI assistants came early, two of them from Gemini and one from ChatGPT. Five more forecasts were still open on 7 October.
| Made | Checked | Prediction | On 6 Oct | On 7 Oct |
|---|---|---|---|---|
| 26 Sep | 10 Oct | Pages with impressions in a settled week reach 20,000 to 40,000 | 19,406 | 17,583 |
| 26 Sep | 10 Oct | Product-snippet errors on the panel fall from 4 to 0 | 4 | 3 |
| 26 Sep | 24 Oct | Device records keep at least 85 % of clicks | 93 % | 93 % |
| 26 Sep | 24 Oct | Sourcing searches stay under 10 % of visible query impressions | 4.6 % | 4.4 % |
| 26 Sep | 24 Oct | Bing stays near zero until the https sitemap is added | 21 impressions in 30 days | 42 in 30 days |
On 7 October the page count had slipped from 19,406 to 17,583, further away from the range it was meant to reach, and the snippet errors were down from 4 to 3. Each forecast stays printed as filed until its check date, and one rewritten after the fact can never again show what a month bought.

The numbers
| Day (Pacific) | Clicks | The forecast of 28 September |
|---|---|---|
| 22 Sep 2026 | 14 | |
| 23 Sep 2026 | 145 | |
| 24 Sep 2026 | 232 | |
| 25 Sep 2026 | 193 | |
| 26 Sep 2026 | 54 | |
| 27 Sep 2026 | 83 | |
| 28 Sep 2026 | 243 | 120 or more |
| 29 Sep 2026 | 219 | 120 or more |
| 30 Sep 2026 | 222 | 120 or more |
| 1 Oct 2026 | 184 | 120 or more |
| 2 Oct 2026 | 103 | 120 or more |
| 3 Oct 2026 | 4 | 30 to 80 |
| 4 Oct 2026 | 5 | 30 to 80 |
| 5 Oct 2026 | 33 | |
| 6 Oct 2026 | 14 |
By 7 October 3 of the 10 forecasts filed on 26 and 28 September had missed, and each is still printed beside its data.
The Fall of 3 October
At 01:00 Pacific time on the first Saturday of the month, 10:00 in Poland, the showings that Google gave Allomedo dropped to about 13 an hour. Over the weekdays before, from 28 September to 2 October, it had shown the site 381 times an hour, and over the first weekend 219. The last hour of 2 October held 164 showings, the next hour 88 and the one after it 11.
Clicks went with them. There were 4 on 3 October and 5 on 4 October, at average positions of 20.8 and 23.1 against 9.6 to 11.4 in the week before. Google had shown between 4,484 and 5,000 of the site’s pages on each weekday of that week, and on the Saturday it showed no more than 306 of them.
Neither day was a draft. Both were read on three mornings, and the second read and the third agreed, at 4 clicks and 5. Monday 5 October stood at 33 clicks on its second read, with about 32 showings an hour, which was a twelfth of what a weekday hour had held the week before.
The tracker’s panel of 33 fixed addresses moved in the same week as well. Google reported 2 of them as unknown to it on 1 October, 13 of them on 3 October and 16 by 7 October, and not one of the panel’s pages shows a crawl after 2 October.
The change log holds nothing that reached the site after 27 September, none of the ten forecasts named such a drop, and no readout of it had been written by the run of 7 October. One reading that fits it was already set down in the readout of 28 September, that a new site often gets a few days of broad exposure from Google and then less. By the tracker’s own rule that reading waits for a dated test before anyone calls it the cause.
So the record holds the drop with its hour, its depth and the pages it reached. It names no cause yet. A summary written at the end of October would show a poor month, and it would say nothing of the hour on which the month turned.

The numbers
| Day (Pacific) | Read the next morning | Read two mornings later | Read on 7 October 2026 |
|---|---|---|---|
| 30 Sep | 0 | 215 | 222 |
| 1 Oct | 98 | 175 | 184 |
| 2 Oct | 19 | 103 | 103 |
| 3 Oct | 1 | 4 | 4 |
| 4 Oct | 0 | 5 | 5 |
| 5 Oct | 9 | 33 | 33 |
08Questions for the People You Pay
Any owner can start without a new supplier, by putting these to whoever runs the firm’s search, and the replies, or their absence, are a first reading.
Bought on Trust
Economists have a name for a service whose quality the buyer cannot judge, even after paying for it. Darby and Karni called such qualities credence qualities in 1973 (Journal of Law and Economics), and Dulleck and Kerschbamer reviewed the markets for them in 2006 (Journal of Economic Literature). Their review opens on the owner’s worry, that doubt about an expert’s work outlasts the repair and the bill, because the buyer still cannot tell whether the service was given or the treatment was needed.
A search retainer is such a service. The firm cannot see the searches, and the engines change under everyone’s feet; the summary comes from the very party that is being judged. So the reading keeps the diagnosis apart from the treatment and writes it to be checked, against files the firm itself holds, with a date on every forecast.
Nothing in a reading depends on the studio carrying it out. Whoever does the work, the firm’s agency, its own team or the studio, is measured against the same dated record, and that record turns a matter of trust into one the owner can check.
For the Next Meeting
- Which pages brought enquiries last quarter, and how many? The count wanted is of enquiries, each with the page it arrived on, and visits are a different number.
- What share of our clicks from Google come from queries that contain our name? If it is most of them, the work is keeping demand the firm already had.
- Which queries without our name do we appear for, and which do our buyers type that we miss? Ask for the list itself, every term with its position and its clicks.
- When our buyers’ questions go to Google’s AI answers and to Copilot, who is named, and which pages are quoted? Ask for it as a table that can be run again next month.
- What did last quarter’s change move, with a date and a number? A change that nobody measured always sits on the books as a cost.
- How much of our traffic is automated, and is it filtered out of the summary? If nobody knows, every rate in it is measured against machines.
- Can we have the raw data? The exports from Search Console, Bing and Analytics belong entirely to the firm, and a summary that cannot be checked against them is an opinion.
09Method and Sources
Tools count in their own ways, and a figure read without knowing the way can be wrong by a wide margin.
The Firm’s Own Accounts
All the figures about the first firm come from exports that were taken on 12 August 2026. Search Console gave its web search by day, query, page, country and device over four windows, which were the sixteen months, the last twelve, the last three and the last 28 days, and its Generative AI report gave the 84 days that it held. From Bing Webmaster Tools there were six months and 28 days of search, and there were 180 days of Copilot’s citations, with the queries it ran and the pages it cited.
Analytics gave twelve months of the sessions it had recorded, by channel, landing page and source, with what its link to Search Console held. Its lead record gave 74 rows, each of them with the buyer’s first message, and all 288 of the site’s pages were crawled and read. The AI answers were asked live, with each buyer question put to the engines and each answer kept with the firms it named and the pages it quoted.
For this paper we went back to the daily rows and summed them once more. Each of the totals came back from them exactly as it stands here, the 4,429 visits, the 1,967,075 showings, the 8,391 uses in Google’s AI answers and the 12,971 citations by Copilot. When those same rows are added up by calendar month, they give the monthly figures in the charts.
Withheld Queries and the Brand Filter
To protect privacy, Search Console leaves out searches made a very small number of times and those that contain personal or sensitive information (Search Console Help). So a share of every site’s clicks arrives with no query attached; for the firm in this paper it was 63 % of the visits. A name also has variants, the short form, the old name, the misspellings, the product names only the firm uses and the address typed as a search, and a split that catches only the exact name always counts part of the brand’s demand as new.
Since 20 November 2025 Google’s console can make the first cut on its own, with what it calls the branded queries filter (Google Search Central). Google counts a query as branded when it holds the brand name, a variation or a misspelling of it, or one of the brand’s own products or services; an internal system that is assisted by AI does the sorting, and Google says that some queries may occasionally be misidentified.
The filter is in the Performance report for top-level properties that have enough queries and impressions, and it has been open to every eligible site since 11 March 2026. What it misses has to be built once by hand, from the firm’s own words: every name, product and person that a buyer could only know from the firm.
Allomedo’s Daily Tracker
One job reads Search Console, Bing Webmaster Tools and Analytics at 04:30 Poland time every morning, and it does so through their APIs, read-only. Every pull is kept as it came, under its date, each series goes into its own table, and each day’s report is kept beside the others, so that any figure can be read as it stood on any morning.
In Google’s words, the newest figures in the Performance report are still being collected and might change in the next few hours (Search Console Help). Our tracker sees the last two or three days keep rising for about three days after they end. Each morning’s run falls at about 19:30 of the day before in Pacific time, while that day is still running.
Allomedo’s clicks for 22 September to 2 October 2026 came to 1,690 when the tracker read them on 4 October and to 1,692 on 6 October. On 30 September the first read printed 0 clicks, two mornings later the day stood at 215 and it settled at 222, so a figure read the morning after is always a draft.
The tracker’s working note adds that Search Console counts days and hours in Pacific time and Analytics counts days in the property’s own time zone, so a day in one report is seldom quite the same day in the other. In the tracker’s own files the hours the console returns carry a −07:00 offset, which agrees with the note, though we found no Google page that states it.
Allomedo’s tags load only after a visitor’s own first scroll, tap or key press, and they run without cookies until consent. So Google reports far more clicks than Analytics shows sessions, and in the week from 26 September to 2 October 2026 the sessions from Organic Search came to about a quarter of the clicks that Google counted. A rate that divides a number from one tool by a number from the other is always wrong unless it says which day, which clock and which share of visitors each number counts.
Direct is the channel Analytics gives to every visit it cannot place: people who typed the address, people who followed a link in an email or a chat, people whose browser held back where they came from, and a great deal of traffic that is automated. On the Directory nearly all of it was automated, as the interaction filter showed.

The numbers
| Day | Google clicks | Analytics sessions from Organic Search |
|---|---|---|
| 26 Sep 2026 | 54 | 16 |
| 27 Sep 2026 | 83 | 14 |
| 28 Sep 2026 | 243 | 54 |
| 29 Sep 2026 | 219 | 56 |
| 30 Sep 2026 | 222 | 59 |
| 1 Oct 2026 | 184 | 52 |
| 2 Oct 2026 | 103 | 32 |
| The seven days | 1,108 | 283 |
Records and Their Windows
Nothing in this paper names the first firm, or its industry, its markets or its languages. All the other figures come from the studio’s own properties or from the public pages that are linked here, and the reading itself is Search.
| Record | Source | Window |
|---|---|---|
| The firm in this paper | Its Search Console, Bing Webmaster Tools and Analytics exports, its lead record and the AI answers asked live, as the reading’s documents of August 2026 read and checked them; shares by name are of the visits whose query Google reported, 37 % of all | Search Console 10 April 2025 to 9 August 2026, data cut off 12 August 2026; Generative AI 18 May to 9 August 2026; Copilot 11 February to 9 August 2026; leads 28 February to 6 July 2026 |
| Clicks by position | Three of the studio’s sites in their own Search Console, with searches for the firms’ own names and one partner’s taken out; the shares computed from the same record | 2025 and 2026 |
| Sustainability Directory | Analytics for sessions, Bing Webmaster Tools for Copilot citations and the searches behind them, Search Console for Google’s clicks and AI answers; Every Engine Keeps Its Own Record holds all of it | 1 to 15 against 18 to 23 September 2026; 1 November 2025 to 22 September 2026 |
| Sentiention | Search Console, read on 4 October 2026 | 13 September to 4 October 2026 |
| Allomedo | The studio’s daily tracker on its Search Console, Bing and Analytics: the daily and hourly series, the panel of 33 addresses, the forecasts as filed on 26 and 28 September, the change log, the readout of 28 September and each morning’s report from 1 to 7 October | 22 September to 6 October 2026, as the run of 7 October at 04:34 held them |
| Google’s change of September 2025 | Zeo’s analysis, published on 9 October 2025 | around 10 to 12 September 2025 |
| Definitions and the curve | Google’s Search Console Help on impressions, position and clicks, the Performance report and Search Console data; Google Search Central on the 16 months (8 January 2018) and the branded queries filter (20 November 2025, updated 11 March 2026); Backlinko, updated on 16 April 2025 | as published |
| Credence goods | Darby, Michael R., and Edi Karni, “Free Competition and the Optimal Amount of Fraud”, Journal of Law and Economics 16 (1), 1973, 67 to 88; Dulleck, Uwe, and Rudolf Kerschbamer, “On Doctors, Mechanics, and Computer Specialists: The Economics of Credence Goods”, Journal of Economic Literature 44 (1), 2006, 5 to 42 | 1973 and 2006 |
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.

