Why Direct Contact Got Valuable: Zero-Click, AI Crawlers, and Content Tiers

Running a newsletter was, for a long time, not something I had to evaluate stage by stage. In earlier years, working inside an agency or a startup, the list, the sending plan, the content development, and the campaign structures were handled by a team; what fell to me was measurement and decisions. Now I handle this and similar processes myself.

The reason I put the newsletter back on my agenda was not only that the work changed hands. Two other developments point in the same direction.

First, alongside the everyday use of AI tools, direct contact gained value: an address that can be reached without a platform’s ranking, a model’s summary, or an algorithm’s decision sitting in between is one of the few things left. Second, my observation is that social media fatigue has opened ground again for targeted email that people actually enjoy following. Instead of a post lost in a feed, a text waiting in an inbox that has to earn being opened. So for my own projects and for my clients I needed a verifiable basis for these decisions. I started by reading the current discussions, the research, and the cases.

Direct contact does not mean email alone, either. Community spaces (Telegram and Discord groups, forums), publishing and comment spaces (online publications and social media posts), and events (online sessions, in-person meetups) fill a similar gap. What distinguishes a newsletter comes down to three things: it can be run by one person; it does not require everyone to be present at the same time, meaning the reader opens it on their own schedule; and the list sits on your own side rather than inside a platform. This post covers only one of these, the newsletter; the others are out of scope, at least for now. And the newsletter I mean is not a classic marketing campaign but curated email built on the reader’s interests and behaviour, written to invite a reply.

References From Two Places, the Setup on My Side

Two separate layers need separating here, because the comparison gets muddled easily. On tone and approach my reference is Substack: curated, personal, written to invite a reply. Structurally the reference is automation tools like ActiveCampaign, Klaviyo, and Kit; the work there is building flows that advance on metadata. In Klaviyo those same flows are mostly wired to store events, so the equivalent is on the ecommerce side, not on this site. And the paid subscription and access restriction that are Substack’s own structural business already exist here, belonging to the site’s membership layer rather than to the newsletter; a topic for another post.

Instead of the hosted email automation tools named above, I built the setup myself: the list and metadata sit in Listmonk on my own server, Inngest runs the flow, Resend handles delivery. I have written up the installation and the service selection separately. The reason was not only cost: when the list sits inside an email service, all you hold is the panel that service shows you, and there is no way to verify who joined when or whether departures were recorded at all. For anyone sending commercial email from Türkiye there is a second constraint: the İleti Yönetim Sistemi (İYS), the national message management system. Consent records have to be filed with İYS, and that usually runs through a solution provider. I could not find a ready-made equivalent for that flow on email platforms based abroad; to build it I would have had to sync both sides over an API myself, carrying sign-up, confirmation, and unsubscribe events to the platform and to İYS alike. On top of that sit the KVKK obligations: where the data is held, how long it is kept, how a deletion request is executed. All of this is technically solvable, but hard to solve in a setup where the list does not sit on your own side. In your own setup, binding the consent flow to the sending flow is possible; I have described how in a separate post on email automation for ecommerce. It is also the reason I can see some of the measurement problems below at all.

Where Is the Traffic Going, and Who Is Arriving?

Why direct contact gained value needs unpacking a little. The answer to whatever someone wants to know is increasingly delivered inside the search results page or the chat window. Google shows the answer on its own page while citing the source; on the chat side the topic stays in that window, the model fetches the page when it needs to, but the result comes back in the same window. The name for this is zero-click: a question is asked, an answer arrives, and even when sources are listed, in most cases no click happens. It is not an entirely new behaviour; the box that answers directly above the search result has been called position zero or the answer box for years, and AI summaries are that space expanded. On scale, SparkToro’s study based on clickstream data puts 68 percent of US searches in the first four months of 2026 as ending without a click. The term zero-click web, which treats this as a property of the web in general rather than of one search engine, also comes from SparkToro; Amanda Natividad coined it, building on the same team’s 2019 measurement. In that measurement the figure was 49 percent, in 2024 it was 60.45 percent, today it is 68 percent. So this is not a single snapshot but a trend spread across years.¹

And the reason is not only that the answer is delivered right there. Moving between sites has acquired its own cost: pages generated automatically and without much thought purely to capture traffic, pages whose body does not deliver what the headline promised, popups and notification permission prompts on arrival, ads breaking up the text, interfaces that load late. Working out whether the information you want is on a given page usually means scanning the whole thing, and that scan does not pay off on every visit. Staying in the chat window skips that cost entirely. The first item on that list has a formal name too: pages produced at scale for ranking are defined in Google’s spam policies as scaled content abuse, and third-party content leaning on another site’s established ranking strength as site reputation abuse.² So behind the avoidance of site visits sits the state of the sites themselves as much as the ease of getting an answer.

Visits do not stop entirely, of course. A link inside the summary gets clicked, or a visitor scrolls past the top result and comes through from further down. But clicks fall overall, and that fall is not caused by the site appearing less often in search results: impressions hold while clicks drop. The name for this is the great decoupling, impressions and clicks coming apart. The need for information is not shrinking; what is shrinking is that need being met by visiting a site. And a click happening does not mean the open web was reached either: in the same series’ 2024 measurement, roughly a third of US clicks go to Google’s own properties, and out of every thousand searches only 374 reach a site outside Google.³

There is a measured version of this on the publisher side. Across 19 Digital Content Next members over an eight-week window, the median year-over-year change in Google referral traffic was a 10 percent drop overall, 7 percent for news brands, and 14 percent for non-news brands. In evidence submitted to the CMA in the United Kingdom, one publisher’s click-through rate on a popular query fell from 5.1 percent to 0.6 percent while ranking and impressions held; another saw traffic to first-ranking articles fall 25 percent while search visibility rose 7 percent.⁴ So the decoupling is not an oddity of individual sites but a pattern measured the same way in different places.

There is no sign the market is shrinking, either. In March 2025 Google stated that more than five trillion searches are made annually; the last figure it had given before that was two trillion, in 2016.⁵ What chat interfaces add to this picture has not been collected reliably, the two sides’ numbers are produced by different methods, and not all prompts are search in nature. The reason is not substitution, as is often assumed. In Semrush’s study on 260 billion rows of clickstream data, people who start using ChatGPT for the first time do not reduce their Google search sessions, and in fact slightly increase them; the same result holds across 500 days of tracking.⁶ So people are not abandoning search for chat interfaces; search continues, but the click does not happen. What is certain is this: what changed is not the volume but how much of that volume reaches a site.

In 68,879 queries Pew Research tracked across a 900-person panel in March 2025, the link click-through rate was 8 percent on results pages with an AI summary and 15 percent on those without; clicks on the sources shown inside the summary ran at 1 percent.⁷ It should be added that these two groups are not the same queries. In Ahrefs’ measurement across 146 million SERPs, an AI summary appears on 20.5 percent of all keywords, but on 57.9 percent of question-form queries and 46.4 percent of queries of seven words or more; for single-word queries the figure is 9.5 percent.⁸ So the queries that trigger a summary are a systematically different set, and that set already produces different click behaviour. The direction is clear, but how much of the gap comes from the summary and how much from query composition is not.

Google reads the picture the other way round: in a statement published in August 2025 it says total organic click volume to sites has stayed roughly flat year over year and that the quality of the clicks arriving has risen.⁹ The data behind that claim was not published, so it is not verifiable, but it is worth noting that publisher-side measurements contradict the company’s own reading. The product direction supports it as well: in May 2026 Google announced that AI Mode had passed one billion monthly users and that queries had more than doubled every quarter since launch; the same announcement introduced generative UI, which builds an interface per query, and search agents running in the background.¹⁰ So answers being produced without a page visit is not a side effect but a product decision. The company’s internal assessment does not match its public messaging either: in an October 2024 Google document submitted as evidence in the DOJ case, executives treat the erosion of search traffic as inevitable and write that they would rather it shift to Gemini than to ChatGPT.¹¹ The verifiable part is on the revenue side: Alphabet’s quarterly results published in April 2026 report Search revenue up 19 percent and query volume at an all-time high.¹² The two pictures do not cancel each other out; queries and revenue can grow while the share of clicks left for sites falls.

How far that share has fallen is measured too. Ahrefs measured the click-through rate of the first position across 300,000 keywords: on results with an AI summary the rate is 58 percent lower. In the same team’s measurement a year earlier the gap was 34.5 percent, so it is widening.¹³ The same asymmetry exists on the chat side, but the size of the share there depends on what is being counted. Counting search engines alone, ChatGPT’s search-like query volume is about 12 percent of Google’s.¹⁴ Adding searches on ecommerce and social platforms, a measurement across 41 domains puts all AI tools at 3.2 percent, with Google alone at 73.7 percent.¹⁵ On visits actually sent to sites the difference is far sharper: in Ahrefs’ panel tracking 107,300 sites, 27 percent of incoming traffic in July 2026 came from Google and 0.35 percent from all AI assistants combined; ChatGPT alone accounts for 0.25 percent.¹⁶

Every measurement in this section is US- and English-weighted and drawn from large samples. It would be wrong to assume the same ratios hold exactly on a niche site publishing in Turkish; their function here is to show direction, not magnitude.

What This Site’s Data Says, and What It Does Not

Let me look at where my own Search Console data fits this picture, because the series I have is split in two. I moved this site to a new version in January 2026, and the transition was not only technical. On the content side I ran a thorough cull against the KPIs I had set. My criterion was never to bring in traffic; the aim was to offer reliable, applicable information directly related to the work I do, and to put my projects, products, and services in front of the right audience. Daily impressions ran around ten thousand until mid-December 2025, fell to a few hundred during the transition window, and the new structure finally began to settle in the second week of January.

Alongside that process there are clear changes in how arriving visitors behave. A visit is not made to browse the site but to reach the section holding a specific piece of information. That detail matters later, because the answers to where the form should go and why it goes unnoticed come from here.

Part of the incoming traffic is no longer human, either. Cloudflare’s classification splits these visits by behaviour into three: search, which collects content in order to answer with it later; training, which crawls to train a model; and agent, which acts in real time on behalf of a person.¹⁷ The third is the most visible on this site: when a topic comes up, the agent behind the chat interface fetches the page, the content is read, and sometimes summarised back to the user. OpenAI draws the same distinction across its own bots, with user-triggered fetching as a separate category.¹⁸ The common thread: none of these visits sees the form, touches the form, or signs up. When a real person does arrive at the site from an assistant, that has its own name, AI referral traffic, and that is the only side that sees the form. The industry term for the third category is agentic browsing; a chat interface standing in for a search engine is an answer engine. The AEO and GEO acronyms derived from these should be used carefully, being vendor-side labels more than practices with defined measures.

This imbalance has been put into numbers as well. What Cloudflare calls the crawl-to-refer ratio divides the number of pages a platform fetches by the visits it sends back. The values Cloudflare published for January and July 2025 show the gap.¹⁹

Platform	January 2025	July 2025
DuckDuckGo	0.1:1	0.3:1
Google	3.8:1	5.4:1
Yandex	15.5:1	21.4:1
Microsoft	38.5:1	40.7:1
Perplexity	54:1	194:1
OpenAI	1,217:1	1,091:1
Anthropic	286,930:1	38,065:1

The difference between the top and bottom halves of the table is a difference of magnitude: on search engines the ratio stays in single or double digits, while on model crawlers it climbs into the thousands and, at the extreme, hundreds of thousands. That the numbers swing this much month to month makes using any single value as fixed data meaningless; what is fixed is the difference in magnitude itself.

So the site has two audiences and only one of them can sign up. The pool facing the newsletter is both shrinking and shifting in composition. Giving up was an option too; I chose instead to extract every usable signal from the traffic that remains.

What follows from this picture also depends on the site’s revenue model, and that distinction is necessary for reading the rest of this post. On publications built on impressions, traffic means revenue directly; there is no substitute for what is lost there, and that is not this post’s subject. Models resting on ad clicks are seeing that ground narrow too, and here is how that looks in numbers: in Seer Interactive’s analysis of 3,119 search terms across 42 organisations, the paid click-through rate on queries with an AI summary fell from 19.7 percent to 6.34 percent, while over the same period average cost per click across industries rose 12 percent year over year.²⁰ On a niche site the goal was never impressions: the target is a small number of right visitors who turn into a product or service sale. That is the criterion on this site, so traffic is not a success metric in itself; the question to ask is whether the traffic belongs to the right audience and whether the connection with that audience holds. There are examples of the two diverging: HubSpot’s blog traffic shed millions of monthly visits while revenue kept growing, because the channels that influence a purchase decision no longer send visits.²¹

That distinction has been tested against data as well. When Cyrus Shepard manually classified more than 400 winning and losing sites and compared them against twelve months of traffic change, the strongest differentiator was offering your own product or service: 70.2 percent of winners do, against 34.6 percent of losers. The losers are mostly news, information, and affiliate sites. A narrow topical focus and owning your own assets separate the two groups in the same direction.²² It is a correlational study and establishes no causation, but the split by revenue model shows up here too.

There is one more consequence on the publisher side. Content that took time, detailed cases, and original data become candidates for a training set the moment they are left open to crawlers. Publishers who do not want that pull those pieces off the open web; into email, into closed groups, or behind an access wall. 404 Media did this in early 2024 and stated the reasoning plainly: their articles were being scraped and republished by bots, they were watching search results fill with content produced at scale, and they wanted to reach readers through a channel they owned. What they wanted for that was an email address.²³ There are concrete examples on the loss side too: All About Berlin, a Berlin guide run for eight years, wrote that it lost 75 percent of its traffic after AI summaries, and in the same piece described how some publishers in similar positions moved content behind a paywall while others stopped updating altogether.²⁴ The decision is as editorial as it is technical: which knowledge stays open to everyone, and which goes only to a group that has shown interest. This site has three tiers as well: open posts, posts requiring membership or a paid subscription, and text sent directly by email. So restricting access is not only a defensive reflex; it is part of the interaction layer itself.

There is a content quality side to the discussion too. The share of AI-generated text on the open web has begun to be measured: a study on a Common Crawl sample puts it at 51.7 percent of English-language articles as of May 2025, and another finds some amount of AI text on three quarters of newly published pages.²⁵ This does not mean model training data has become synthetic at the same rate, because training sets are filtered; what is being measured is a change in the composition of the pool. The risk arising from it has a name: model collapse, where models trained recursively on generated data lose the tails of the distribution and outputs collapse toward the mean.²⁶ The practical consequence is that faulty or superficial content does not only mislead a reader, it becomes the next model’s input. This is why a human stays in the loop for what gets published here: claims are tied to their sources, and model output is not published without verification. It is also one of the reasons part of the content stays closed, so that a piece that took verification effort does not become indistinguishable inside a pile of unverified material.

The Technical Side of Closing Access

The access decision has an implementation side too, and the options widened over the past two years. The oldest tool is robots.txt, but on its own it can only say crawl or do not crawl; it cannot separate what the crawled content will be used for. Content Signals came out to fill that gap: search, model training, and being used as model input can each be stated separately in the same file. This site’s preference is written there, open to search, closed to training.²⁷

The second option is on the infrastructure side. Cloudflare has for a while been separating model crawlers by behaviour and offering bulk blocking, and has added a layer on top called pay per crawl: you can choose between giving a crawler free access, blocking it, or charging per request. When a charge is set, the crawler receives HTTP 402 and the price is announced in a header. As of August 2026 the feature is still in closed beta, so not a widespread revenue line; its value here is turning access from a binary switch into something that can be priced.²⁸

None of this replaces the connection built with a reader. Managing how content is crawled is the defensive side; on the other side sits an address a reader left of their own accord, and these are different questions. However well the defensive side is built, what brings someone back to the site is that address.

How the Layer Works

And what gets used is not the sign-up count but the interaction itself. Publishing a post works one way: content goes out, and whether anything comes back stays unclear. As a general observation, most posts on content sites go without comments, and a significant share of the comments that do arrive come from spam sources rather than real visitors.²⁹ It is why I removed the comment section on this site as well. Meaningful interaction comes through social media, but there it stays scattered who shared what and what it was a response to. Comment volume on social media is not stable either, and its direction varies by platform: average comments per post went from 66 to 50 on TikTok and from 24 to 20 on Instagram between 2024 and 2025, while rising from 17 to 22 on Facebook.³⁰ A number that cannot be controlled and moves in opposite directions from platform to platform offers no ground to build a process on. Email does not replace these channels; it is a supporting layer that gathers up the scatter.

What carries the layer is metadata. For each person who signs up, which post they came from, which cluster they fall into, which language they chose, and the source information at the moment of sign-up are stored; what gets sent is determined by that. On the Listmonk side a separate flow advances for each person, in steps starting from the point where they signed up. I do not send standard announcements or campaigns; what goes out is not a single send landing on the whole list at once but that person’s next step.

The real return is not measurement either, it is the replies. I end the emails with a call to action, and the replies that come back show two things at once: that the emails are genuinely being followed, and that the suggestions are being applied. There is a closeness here that publishing does not allow (I do not run a comment section), the tone can be a notch warmer, questions arrive directly, and requests and offers can be tailored. The subject of new posts often comes out of this as well. For a solo founder this process is genuinely valuable.

The numbers on the email side support this too, but they have to be read together with their denominator. Over the last 15 days, 15 emails in total went out, welcome emails included, to 7 people who signed up this month (August 2026). Of those 15 emails, 13 were opened and links were clicked in 5. Writing it as a rate would be misleading: at this denominator the confidence interval for opens spans 62 to 96 percent and for clicks 15 to 58 percent, and a single open moves the picture by several points.

One more thing worth noting for reading the open metric correctly: Apple Mail Privacy Protection pre-fetches images on the user’s behalf for people receiving email through Apple Mail, so those people count as having opened even when they have not.

So the job of these numbers is not to make a performance claim but to show that the flow works and the list is not dead. The feedback on the newsletter page shows the same.

What Is Next

Every decision up to here sits behind the layer: which content stays open, which channel gets run, where the data lives. What the reader sees is none of that, only the newsletter form inside the post. For someone who does not sign up, the flow, the metadata, and the reply loop behind it have no equivalent at all.

Where the form goes, what it looks like, and what it says is a separate post; I have collected the decisions and the tests in a piece on the newsletter form as an interaction layer. How the flow advances, the sending strategy, and metadata-driven segmentation are left to later posts.

Footnotes

1. SparkToro, “In 2026, Less than One Third of Google Searches Still Send a Click”. January-April 2026, US, Similarweb’s desktop and mobile clickstream panel. Not Google’s own data but a panel-based estimate. The year-over-year comparison is also not a single continuous series: the 2016 and 2019 figures come from the now-closed Jumpshot panel, 2024 from Datos, and 2026 from Similarweb, and Fishkin explicitly calls this apples to oranges. The data does not cover Google’s mobile app, only in-browser searches; since zero-click is more common in the app, the real figure is probably higher. ↩
2. Google Search Central, “Spam policies for Google web search”. The policy definitions: scaled content abuse, “generating many pages primarily to manipulate rankings rather than to help users”; site reputation abuse, “third-party content published on a site primarily because of the host site’s established ranking signals”. The term enshittification, describing the gradual degradation of platforms at users’ expense, is used to describe the same picture; it belongs to Cory Doctorow and, being polemical, is only named here. ↩
3. Rand Fishkin, “2024 Zero-Click Search Study”, SparkToro, 2 July 2024. Datos panel, September 2022 to May 2024. In the EU the figure is 360 clicks to the open web per thousand searches; Google sends more traffic to its own properties there than in the US. ↩
4. Jessica Davies, “Google AI Overviews linked to 25% drop in publisher referral traffic”, Digiday, 15 August 2025. The DCN measurement is limited to 19 publishers and the May-June 2025 window; the PPA’s evidence to the CMA rests on member publishers’ own reporting and has not been independently audited. ↩
5. Google, “AI, personalization and the future of shopping”, 3 March 2025: “We already see more than 5 trillion searches on Google annually.” The company’s own internal data, not an independent measurement. The previous statement was two trillion in 2016, with no official figure given in between. ↩
6. Luke Harsel, Anna Yudina and Aleksandr Drozdov, “ChatGPT Is Not Replacing Google, It’s Expanding Search”, Semrush, 11 August 2025. US desktop devices only; Semrush is both a tool vendor and the panel owner. A second measurement in the same direction appears in SparkToro’s August 2025 report: more than 95 percent of US devices continue to use traditional search engines, and the growth rate of AI tool adoption has been slowing since late 2024. ↩
7. Athena Chapekis and Samuel Bestvater, “Google users are less likely to click on links when an AI summary appears in the results”, Pew Research Center, 22 July 2025. Browser data from 900 participants was tracked in March 2025; an AI summary appeared on 12,593 of 68,879 queries. A panel-based measurement, so specific to US users and to that month. The same study also reports where summaries appear: 8 percent of one- or two-word searches trigger one, rising to 53 percent for searches of 10 words or more and 60 percent for queries beginning with a question word. The two compared groups are therefore not the same query set, and Pew claims no causation. ↩
8. Ryan Law and Xibeijia Guan, “What Triggers AI Overviews? 86 Factors and 146 Million SERPs Analyzed”, Ahrefs, 10 November 2025. 99.9 percent of keywords triggering an AI summary are informational; for navigational queries the figure is 0.9 percent and for local searches 7.9 percent. ↩
9. Elizabeth Reid, “AI in Search is driving more queries and higher quality clicks”, Google, 6 August 2025. The statement says total organic click volume has stayed roughly flat year over year and that click quality has risen. The underlying data set was not shared and the definition of a “quality click” is the company’s own, so the claim is not independently testable. ↩
10. Elizabeth Reid, “A new era for AI Search”, Google, 19 May 2026. The company’s own announcement, not independent verification. ↩
11. Danny Goodwin, “Google Search traffic decline is inevitable, execs say”, Search Engine Land, 12 May 2025. The document was published in the DOJ filings; it is a summary of an internal October 2024 meeting, not an official company statement. ↩
12. Alphabet Inc., “Alphabet Announces First Quarter 2026 Results”, SEC 8-K Exhibit 99.1, 29 April 2026. Google Search & other revenue rose from 50.7 billion dollars to 60.4 billion. Sundar Pichai’s statement: “Search had a strong quarter with AI experiences driving usage, queries at an all time high, and 19% revenue growth.” ↩
13. Ahrefs, “AI Overviews Reduce Clicks by 34.5%”, April 2025, and for the updated measurement “AI Overviews Reduce Clicks (Update)”, December 2025 data. The current study compares 150,000 keywords with an AI summary against 150,000 without and finds a 58 percent lower click-through rate in first position. Worth noting these are studies run by a tool vendor; the post also states that independent studies put the gap anywhere between 47.5 and 90 percent. ↩
14. Ahrefs, “ChatGPT Has 12% of Google’s Search Volume but Google Sends 190x More Traffic to Websites”. The ratio is calculated after isolating search-like prompts; not all prompts count as searches. In OpenAI and Harvard’s joint review of 1.5 million conversations, a quarter of conversations are purely search in nature. ↩
15. Rand Fishkin, “Search Happens Everywhere”, SparkToro, 3 March 2026. Datos desktop panel, 41 domains, 2025. The same study carries one more caveat: almost all Google visitors run a search, while only about half of ChatGPT visitors enter a prompt, so treating visit counts as a usage indicator is misleading. ↩
16. Ahrefs, “AI vs Search Traffic”, July 2026 snapshot. The panel consists of 107,300 sites using Ahrefs Web Analytics, so it rests on voluntary participation and represents sites doing active marketing. In the same snapshot Bing is at 1.23 percent and DuckDuckGo at 0.26 percent. ↩
17. For category names and definitions, Cloudflare, “Bots: AI bot categories”. Cloudflare classifies bots by behaviour rather than under a single “AI bot” label. For the two-audiences framing, also Matt Butcher, “Your Website Now Has Two Audiences: Humans and AI”, Akamai, 17 August 2026. The latter is a post containing product promotion and its market projections are not verified against a primary source; it is cited here only for the framing. ↩
18. OpenAI, “Bots”. The four documented bots and their categories: GPTBot for training crawls, OAI-SearchBot for search crawls, ChatGPT-User for user-triggered fetching, OAI-AdsBot for ad verification. ↩
19. Cloudflare, “The crawl before the fall… of referrals: understanding AI’s impact on content providers”, 1 July 2025, and for current values Cloudflare Radar AI Insights. The ratio is calculated by dividing HTML requests from a platform’s crawlers by page referrals from the same platform. Because it varies across a very wide range depending on platform and period, Cloudflare’s published January and July 2025 values are given rather than a single current figure; the source is “The crawl-to-click gap”, 29 August 2025. That extreme values like Anthropic’s 286,930:1 in January drop by an order of magnitude within the same year shows how quickly the ratio changes with platform policy. ↩
20. For the paid click-through rate, the Seer Interactive analysis is relayed in Bluepear’s Search Engine Land piece “Why CPC keeps rising”, 26 March 2026; the piece itself is sponsored and sells a product at the end, so the relayed figure should be treated as second-hand. For the cost-per-click increase, an industry compilation: Digital Applied, “Google Ads Benchmarks 2026”, 5 April 2026; the raw data was not shared, so it should be taken as a directional indicator. The same compilation notes that conversion rates rose as well: the clicks that remain are more expensive but carry more intent. ↩
21. Rand Fishkin, “Traffic Is Down; Revenue Is… Up?”, SparkToro, 28 January 2025. A reading built on one company’s chart, not a controlled comparison. The same author’s “Inimitable Product is the New Make Great Content” reaches the same conclusion and counts email among “the last owned channels big tech hasn’t yet extracted the value from”. ↩
22. Cyrus Shepard, “5 Data-Backed Features Of Websites Winning Google in 2026”, 9 April 2026. The sample is more than 400 sites, traffic change taken from the Ahrefs API over a twelve-month window, features classified manually, Spearman correlation used. The author separately states that the findings are correlational and establish no causation. ↩
23. 404 Media, “Why 404 Media Needs Your Email Address”, 26 January 2024. The reasoning in the piece falls under three headings: content being scraped and republished, search results filling with content produced at scale, and being able to reach readers directly. In their own words, “we need to be able to reach our readers directly on a platform that we own and control.” One publisher’s decision, not a measurement of the wider industry. ↩
24. Nicolas Bouliane, “AI is killing All About Berlin”, 26 May 2026. One publisher’s own account, not a measurement; cited here to show the concrete shape of the loss. ↩
25. Graphite, “More Articles Are Now Created by AI Than Humans”: a sample of 65,000 English-language URLs from Common Crawl, with articles counted as AI-generated when more than half the text is flagged as AI; the figure was 2.2 percent in January 2020 and 51.7 percent in May 2025, and has been flat since late 2024. For the new-pages side, Ahrefs, “What Percentage of New Content Is AI-Generated”: in April 2025, 74.2 percent of 900,000 new pages contained some AI text, but only 2.5 percent were entirely AI and 25.8 percent entirely human. Both measurements rest on AI detectors, whose error rates are high; this is why Axios could read the same data as “AI has not overtaken human content”. ↩
26. Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson and Yarin Gal, “AI models collapse when trained on recursively generated data”, Nature, 631, 755-759, 2024. The scope of the finding is contested: critics argue the same collapse is not observed in scenarios where data is accumulated rather than replaced (arXiv:2410.12954). An author correction was also published for the paper. So the use here is not a claim of inevitable collapse but that the accumulation of unverified content carries a cost. ↩
27. The preference file is public at ceaksan.com/robots.txt; for the syntax used, contentsignals.org. Content Signals allows the search, ai-train, and ai-input preferences to be stated separately inside robots.txt. Worth noting that it is a statement of preference, not a technical block. ↩
28. Cloudflare, “Introducing pay per crawl”, 1 July 2025, and the AI Crawl Control documentation. Requests where payment is required receive HTTP 402, the price is announced in the crawler-price header, and the crawler side can set a ceiling with crawler-max-price. In closed beta as of August 2026; this may change. ↩
29. This is my own observation and does not rest on published longitudinal data. In 2006 Akismet opened a public statistics page showing daily spam and legitimate comment counts (“Spam Stats”, 22 May 2006); the last working archive snapshot of that page dates to October 2010, and akismet.com/stats/ now redirects to a product page. Disqus does not publish a comparable series either. If verifiable data on this turns up, it will be added here. ↩
30. Socialinsider, “Social Media Benchmarks For 2026”, across 70 million posts. The figures belong to brand and business accounts, not personal ones. The percentages in the report’s summary do not match its own table, so raw averages were used rather than percentages. ↩

Key Takeaways

• 01 Zero-click, the decoupling of impressions from clicks, and the share taken by AI crawlers together shrink the pool of visitors reaching a site and change its composition. A site now has two audiences, and only one of them produces interaction that carries value.
• 02 Cloudflare's crawl-to-refer ratio shows that the visits returned per page fetched vary by orders of magnitude between platforms. The number swings month to month, so it should not be treated as a fixed value.
• 03 Avoiding a site visit is not only about the answer being delivered in a chat window. Content produced at scale, popups and notification prompts, ads, and slow interfaces all raise the cost of scanning a page.
• 04 The accumulation of unverified content carries a cost, and on the publisher's side that cost turns into an access decision: part of the work that took effort moves off the open web into email, closed groups, or behind a membership.
• 05 The reason for a self-hosted setup is not cost but verifiability. For anyone sending commercial email from Türkiye, the second reason is İYS and KVKK: hosted platforms based abroad have no ready-made way to bind the consent flow to the sending flow, so both sides have to be synced over an API.

Frequently Asked Questions (FAQ)

+ What do zero-click, the great decoupling, and crawl-to-refer ratio describe?

All three measure different parts of the same picture. Zero-click is when the answer is delivered on the search results page or in a chat window and no click ever happens. The great decoupling is impressions holding steady in search results while clicks fall, the two coming apart. Crawl-to-refer ratio is Cloudflare's measurement: the number of pages a platform fetches divided by the visits it sends back. The first two describe the human visitor side, the third the model crawler side.

+ If traffic is moving to chat interfaces, why build a newsletter?

It depends, and an answer given without that distinction is misleading. For a publication that monetises through impressions and produces many pages across a wide range of topics, what is lost is revenue directly; a newsletter is not the fix there, it is a different problem. Models that depend on ad clicks are seeing that ground narrow too. On a niche site the goal was never impressions in the first place: it is the right visitor, one who turns into a service or product sale. That is the criterion on this site. Traffic is falling, direct contact is what is being gained; a real visitor who makes it to the site becomes more valuable, and the inbox is the one place where no algorithm or closed window stands in between.

+ Why keep part of the content closed?

Two reasons. First, content and original data that took time to produce become candidates for a training set the moment they are left open; a piece that took verification effort should not end up indistinguishable inside a pile of unverified material. Second, deciding which knowledge stays public is an editorial call. This site has three tiers: open posts, posts that require membership, and text that goes out by email only.

+ Why your own setup instead of a hosted service?

There are two axes and they should not be conflated. On tone and approach my reference is Substack; structurally it is what automation tools like ActiveCampaign, Klaviyo, or Kit do, which is flows driven by metadata. Rather than buying the second, I built it: the list and metadata live in Listmonk on my own server, the flow runs on Inngest, delivery goes through Resend. That keeps it verifiable who joined and left and when, and it makes the İYS requirement on the Türkiye side solvable.