What 657 clicks taught us about earning search traffic in 2026

What 657 clicks taught us about earning search traffic in 2026

In the 92 days to 18 September 2026, Google showed TrustList pages in its search results 652,463 times. People clicked 657 times. That is a click-through rate of 0.10%, or about one click for every thousand times a page of ours appeared. The figures come from our own Google Search Console account, and we have not rounded them in our favour.

TrustList is a directory. It holds about 87,000 published company listings and 12,000 software products, and our sitemap offers Google roughly 120,000 addresses. For a site that size, 657 clicks in a quarter is not a traffic strategy. It is a finding. This article is about what that finding taught us, what we changed, what has not worked yet, and what we would tell anyone else running a large site who is trying to earn visitors from search in 2026.

It is the first of three pieces. The second asks whether SEO is dying or changing, and the third looks at the changing role of search engines as institutions. This one is our own account, and it is the least flattering of the three.

What the numbers actually say

Start with where the 657 clicks went, because it overturned an assumption we had held for a long time.

We had built the site around ranking pages: pages under /best/ that list the firms in a category, a country or a city, such as a page of software development companies in one country. There are 18,333 of them in the sitemap. We assumed they would be the pages people found, because they match the kind of search a buyer types: a service plus a place.

Search Console says otherwise. Over the 92 days:

Page type	Impressions	Clicks	Click-through rate	Typical position
Listing pages (one business or product)	about 33,000	528	1.6%	6 to 9
Ranking pages (/best/…)	about 255,000	67	0.03%	60 to 80
Whole site	652,463	657	0.10%	—

The ranking pages collected most of the impressions and almost none of the clicks. They appear, but at positions 60 to 80, which means page six or later. Nobody reads that far, so the impressions mostly record Google trying a page and the searcher never seeing it.

The listing pages are the opposite. They were shown far less often, but when they appeared they sat on the first page, around positions 6 to 9, and people clicked 1.6% of the time. Four out of five of our clicks, 528 of 657, went to a page about one named business or product.

The reason is not mysterious once you look at the queries. The searches that led to a click were mostly someone looking for a specific company by name, or a specific product. For those searches, a page that is actually about that company is a reasonable result. For a broad search such as a service in a city, we are competing with the firms themselves, with established directories, with review platforms and now with Google's own AI answers. A page that lists fifty firms with a line each is not the best answer on the web for that, and Google's ranking reflects it.

We also looked at geography. Our analytics show that visitors from the United States arrive from Google at a rate of roughly one click a day. The site is UK-based and many listings are UK firms, so that is not surprising, but it corrected another quiet assumption that the size of the catalogue would carry us abroad.

Impressions that are not people

The second lesson was about who is searching.

Search Console has a list of queries where we get impressions but no clicks. We went through 943 of these "gaps" expecting to find content we should write. About 310 of them, roughly a third, were not buyer searches at all. They were quoted searches combining a person's name, a company name and the word "phone". That pattern belongs to contact-scraping tools that query Google to harvest phone numbers and email addresses. No human clicks those results, and there is nothing useful to write for them.

We also found 28 ranking pages that had no listings on them at all, yet between them had collected 15,190 impressions. Google was showing empty pages. That is the kind of thing a site owner does not see unless they go looking, and it is exactly what a quality reviewer notices.

The lesson for anyone reading their own Search Console data: an impression is not a person, and a gap is not an opportunity. Before you plan content from a gap list, remove the queries a scraper would make, and check whether the pages collecting impressions actually have anything on them.

Our logs: more machines than readers

The third lesson came from our server logs, and it changed how we think about who the site is for.

We took one window of our edge proxy logs: 9.3 hours, from 21:19 UTC on 24 September to 06:39 UTC on 25 September 2026. In that time the site served 77,221 HTML page requests (images, scripts and style sheets excluded). This is one overnight window in UK time. It is a sample, not a month, and the mix will differ during the working day. With that caveat, this is what fetched our pages:

Who fetched the page	Page requests	Share
Browser-like user agents	35,401	45.8%
OpenAI GPTBot	15,285	19.8%
OpenAI OAI-SearchBot	5,359	6.9%
Meta crawlers	4,519	5.9%
Moz (SEO tool)	2,324	3.0%
Bingbot	2,180	2.8%
Applebot	1,747	2.3%
Anthropic ClaudeBot	614	0.8%
Googlebot	221	0.3%
Other unnamed bots and scripts	8,793	11.4%

All the AI companies' crawlers and fetchers together made 27,591 requests, 35.7% of the total. OpenAI's GPTBot alone fetched about 69 pages for every page Googlebot fetched. We let these crawlers in deliberately; blocking them is a choice some publishers make and we have decided against it for now.

The "browser-like" row needs a warning of its own. Our analytics show that much of it is an automated browser farm that signs in to the site, so it is not all people. Of those 35,401 browser page views, 76.1% carried no referrer, 23.8% came from our own pages, and 27, which is 0.1%, came from a Google search results page. None came from a social network in that window.

Twenty-seven visits from Google in nine hours, against 15,285 fetches by one AI crawler. Whatever else is true, a site like ours is now read by machines far more than it is visited from search. Those machines have different purposes. OpenAI's own documentation says GPTBot gathers content for training its models, OAI-SearchBot is what surfaces sites in ChatGPT's search answers, and ChatGPT-User makes fetches on behalf of a person asking a question. In our window ChatGPT-User fetched 66 pages. That is small, but each of those was a live question from a person that led to one of our pages.

Thin content and the AdSense refusal

The fourth lesson was the one that hurt, and it came from outside.

On 25 September 2026 Google AdSense declined the site for "low-value content". AdSense is Google's advertising programme for publishers, and approval depends on a review of the site's content. The refusal told us something our own counts had already shown, but we had not acted on it fast enough.

We counted the catalogue the same day. Of the live listings, about 24,350, a quarter of them, carry fewer than 50 words of profile text. 12,178 company listings have no profile text at all. On top of that, 9,620 live companies and 3,416 live products were last updated in December 2020, when the original catalogue was imported. 5,092 software listings show 2020 prices with no date beside them, and when we checked a sample of 30 against the vendors' own pricing pages, only 6 still matched.

Then the websites themselves. On 24 September we drew 1,000 company listings at random and followed each website link. 171 of them no longer led to the business as listed: the domain was parked, for sale, redirected somewhere unrelated, or gone. Scaled up, that suggests a meaningful share of the catalogue describes businesses that have closed or changed.

None of this is unusual for a directory that grew by import. But Google's guidance is plain about what it wants. Its page on helpful content asks whether content provides "original information, reporting, research, or analysis" and whether it offers substantial value compared with other pages in the results. It also lists warning signs, including "using extensive automation to produce content on many topics". Its spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users". We did not build the catalogue to manipulate anything, but 24,000 near-empty pages look the same to a reviewer whatever the intention was. We were, in effect, a thin programmatic site with good intentions.

What we changed

We decided the answer was not more pages. It was fewer, better ones, and being honest about the rest.

Rebuilding thin listings in our own words. For each thin listing we go to the business's own website, read what it says it does, and write a profile in our own words: what the firm offers, who it serves, where it is, and what can be checked. We do not copy the firm's marketing text, and we do not invent anything the site does not support. The first batch was 48 listings rewritten.

Retiring what has closed. Where a business has clearly closed, the listing is retired with a status that says so, and it leaves every ranking. In the first batch, 2 were retired as closed. A page that tells a reader "this firm no longer trades" is more useful than a page pretending it does.

Removing what nothing can be found for. Where we cannot find any trace of a business beyond the listing itself, the listing comes down. That will shrink the catalogue, and we have accepted that.

A curated sitemap. We added a second sitemap, /sitemap-curated.xml, holding 233 URLs: the rebuilt listings and our editorial pages, the pages we stand behind. The main sitemap still lists everything, but the curated one is our statement of what we think deserves a searcher's time.

Ranking pages that carry counted facts. Because ranking pages collected impressions and no clicks, we changed what they say. They now state facts we counted about the firms on them instead of a generic introduction. We also stopped pretending the order comes from reviews: the whole catalogue holds 20 reviews, 14 of them published, against 87,205 company listings, so our rankings are ordered by how completely and verifiably a listing is documented.

A news desk that records things nobody else collects. In September 2026 our news desk logged 28 vendor changes that had both a notice date and an effective date. The median notice period was 31 days. Six gave a week or less, three gave none at all, and eight gave more than 90 days. Four of the 28 switched a feature on by default for existing customers. That is original data. It is small, but it is ours, and it is the kind of thing Google's own guidance says it wants to find.

A launch board with the questions asked. Our launches page recorded 60 launches dated September 2026: 28 AI models and 32 products. Reviewing them, we found many new-product launches were relaunches, or had no named operating company, and several publish no price. We say so on the page rather than listing them as if they were equal.

What has not worked yet

It would be easy to write this as a turnaround story. It is not one yet.

The click numbers above are the current numbers. We do not have a before-and-after, because the rebuild started this month and Google takes weeks to re-crawl and re-assess pages. We will not know for a while whether 48 rewritten listings move anything, and 48 is a small fraction of about 24,000 thin ones.

The ranking pages still rank poorly. Adding counted facts has made them more honest, but a page listing firms in a city is still competing against the firms, against larger directories and against AI answers that summarise all of them. We may find that some of these pages are simply not worth keeping in the index.

AdSense has not re-reviewed the site. We do not know what threshold it will want.

And the AI crawlers that read us so heavily send almost nothing back that we can measure. Cloudflare, which sees a large share of web traffic, published figures in August 2025 comparing how many pages each platform crawls with how many visitors it refers. In July 2025 it counted OpenAI at about 1,091 crawls per referred visitor, while Google's ratio was in single figures. Our own window is consistent with that. We still think being read by these systems is better than being absent from them, but we cannot yet show that it brings buyers to listings.

What we would tell another site owner

If you run a large site, especially one built by import or by template, this is what our quarter taught us.

1. Split your Search Console data by page type before you do anything else. Impressions by type, clicks by type, position by type. We had assumed our ranking pages were our shop window. They were 39% of impressions and 10% of clicks. Your assumption may be just as wrong.
2. Find out which pages earn clicks, and build more like them. For us it was pages about one named business. For you it may be one product, one question, one tool. Put your effort there.
3. Clean the gap list. Remove queries that only a scraper would make: names plus "phone", "email", "contact". Check that pages collecting impressions actually have content.
4. Count your thin pages honestly. Set a threshold, such as 50 words of original text, and count how many pages fall under it. If it is a large share, that is your problem, and more pages will not solve it.
5. Check whether your subjects still exist. A random sample of 1,000 outbound links is an afternoon's work. If a sixth of them are dead, a reviewer will find that too.
6. Rewrite from primary sources, in your own words. Read the business's or product's own site. Write what it does plainly. Leave out what you cannot support.
7. Retire and remove, visibly. A page that says a firm has closed is useful. A page describing a firm that closed four years ago is not.
8. Publish a curated sitemap of what you stand behind. It forces you to decide which pages you would defend.
9. Produce something only you can count. Notice periods, launches that name no company, dead websites in a random sample. Original data is the one thing an AI summary has to cite you for.
10. Read your logs, not just your analytics. Analytics tools are designed to count people and often filter bots. Your logs show who is really reading the site, and in 2026 that is mostly machines.

We will report back when we have a quarter's data from the rebuilt pages. If the numbers do not move, we will say that too. For the wider question of what this means for search as a discipline, read is SEO dying or changing. For how the engines themselves are changing, read the changing role of search engines.

Sources

• TrustList Google Search Console, performance report for the 92 days to 18 September 2026: impressions, clicks, click-through rate, page-type split, query gaps. TrustList internal data, read 25 September 2026.
• TrustList edge proxy logs, 21:19 UTC 24 September to 06:39 UTC 25 September 2026: page requests by user agent and referrer. TrustList internal data, one overnight window, a sample not a month.
• TrustList catalogue counts, 20 to 25 September 2026: thin listings, last-updated dates, undated prices and price sample, dead-website sample of 1,000, reviews, news desk and launch board counts. TrustList internal data.
• Creating helpful, reliable, people-first content: the self-assessment questions on original information and substantial value, and the warning sign about extensive automation. Google Search Central, last updated 10 December 2025.
• Spam policies for Google web search: the definition of scaled content abuse. Google Search Central, last updated 28 August 2026.
• Overview of OpenAI crawlers: the purpose of GPTBot, OAI-SearchBot and ChatGPT-User. OpenAI (read 25 September 2026).
• The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals: crawl-to-referral ratios by platform for January to July 2025. Cloudflare, 29 August 2025.
• Your AdSense account wasn't approved: the programme's wording on sites without enough original, rich content. Google AdSense Help (read 25 September 2026).