TL;DR: More than two thirds of searches never reach a website, and the share is still climbing — the answer is now the click. I measured my own project, bewertiq.org, against the framework: the page renders one h1, zero h2, zero tables, zero lists and 397 characters of visible text, with Organization plus WebSite schema and a SearchAction that is declared but does not work. The correction nobody applied yet: Google's FAQ rich result died in May 2026, so the FAQPage markup that every checklist still recommends no longer produces a SERP feature. What still works: answer-first paragraphs, semantic density, entity linking and an llms.txt that the AI crawlers actually read.
The scenario
bewertiq.org is a static single-page application that turns 140,067 business profiles and 28.4 million reviews into a transparency score. It ships as one JavaScript bundle — no backend, no server-side rendering. That architecture is fast and cheap, but it has a property the marketing never sees: when an AI answer engine or a crawler reads the page, it reads the shell, not the data. The data only exists after the browser executes the bundle.
So I ran the same checks on my own project that I run on client sites, to see how visible the site actually is to the systems that now decide whether a human ever visits it. The results were worse than I expected, and they are the perfect worked example for what zero-click search does to a page that was not built for it.
The pattern
Two thirds of queries end without a click
SparkToro and Similarweb measured the US clickstream for January to April 2026 and published it in June: 68 percent of Google searches end without any click to an external site. That is up from 60 percent in 2024 and roughly 45 to 50 percent in 2016. Germany and France sit at 62 to 65 percent — the same trend, slightly delayed. When a query shows an AI Overview, the zero-click share jumps to 83 percent, against a 60 percent baseline for queries without one. Comscore counted AI Overviews on 39.4 percent of US desktop searches in June 2026, up from 25.8 percent earlier in the year. Pew measured the click-through directly: 8 percent of people click a result when an AI Overview is present, versus 15 percent when it is not.
The asymmetry that matters for a business: being cited inside an AI Overview raises your organic clicks by 35 percent and your paid clicks by 91 percent. The battle moved from ranking to being the source the answer quotes.
The SATISFICER and the MAXIMIZER
Every query is answered twice: once for the SATISFICER, who wants the answer now and stops at the overview, and once for the MAXIMIZER, who opens the page to compare, verify and dig deeper. Zero-click SEO optimises for both. The SATISFICER decides from the answer box — your 40 to 60 word summary is your pitch. The MAXIMIZER arrives at your page and needs structure immediately: headings, tables, lists. A page that serves only one of them serves half the traffic.
The answer-first framework
Write 40 to 60 words of direct answer directly under every h2 and h3, before any setup or storytelling. Answer engines lift these blocks verbatim. The block has to be extractable: one concrete claim, numbers where possible, no pronoun references like "this tool" that lose meaning when quoted out of context. Every claim in the block needs its evidence right below it — the table or the list — because the crawler now has something to cite twice.
Semantic density: tables and lists are citations
A table is the most extractable structure a search engine and an AI summariser can read: one row per fact, one column per dimension. Ordered and unordered lists do the same for sequences and options. A page with zero of these structures has nothing to cite. My measurement of bewertiq.org came back at zero h2, zero h3, zero tables, zero lists — 397 characters of visible text in total, all of it the hero. The site is a data product that ships no data in its HTML.
Entity signals: schema and the author
Structured data is how you tell the machine who you are. Organization schema with knowsAbout and areaServed connects the brand to the topic and the geography. Article and BlogPosting schema connects content to an author — and an author with a real profile URL is the E-E-A-T signal answer engines weigh when they decide whether to quote you. bewertiq.org ships Organization and WebSite schema only. No Article, no BlogPosting, no Person, no author link.
A correction: the FAQ rich result is gone
Every checklist from 2024 still recommends FAQPage schema for the FAQ rich result. Google removed that SERP feature on May 7, 2026. Search Console dropped the FAQ filter and report in June 2026 and the API data in August 2026. FAQPage remains a valid Schema.org type and harmless to publish, but it no longer produces a rich result — the remaining consumers are retrieval crawlers that read the questions as content. Do not delete it; just stop expecting a SERP feature from it.
llms.txt: the file the answer engines read
llms.txt is the convention published on llmstxt.org: a Markdown file at the root, or any subpath, that summarises what a site is about and lists its key pages. The only required element is an h1 with the site name. A blockquote summary follows, then file lists under h2 headings, one link per line. Google does not consume it — but the AI crawlers and retrieval agents do, which is exactly the audience the zero-click shift created.
Click-SEO versus Zero-Click-SEO
| Dimension | Click-SEO | Zero-Click-SEO |
|---|---|---|
| Goal | Rank and get the click | Be the answer, then get the click |
| Content type | Long-form articles | Answer-first blocks plus depth |
| Keyword focus | High-volume head terms | Question phrases and entity terms |
| Technical focus | Speed, indexation, links | Structured data, semantic structure, llms.txt |
| Success metric | Clicks and sessions | Impressions, citations, branded search |
Measuring it
Clicks are now a lagging indicator. The leading ones: Search Console impressions versus clicks on branded queries, the share of impressions that carry an AI Overview, branded search volume over time — people who saw the answer and looked for you afterwards — and assisted conversions, where the brand touch happened inside an answer rather than on the page.
Findings
The example site, bewertiq.org, measured on 2026-11-07:
| Severity | Finding | Evidence |
|---|---|---|
| HOCH | Zero semantic structure: one h1, no h2, no h3, no table, no list, 397 characters of visible text | DOM measurement of the served page |
| HOCH | Declared SearchAction does not work: every query returns the same 20,541-byte shell | /?q=test returns the identical homepage |
| MITTEL | No Article, BlogPosting or Person schema, no author link — entity signals end at Organization and WebSite | The two JSON-LD blocks on the page |
| MITTEL | No llms.txt, no Markdown twins for the AI crawlers | GET /llms.txt returns the SPA shell |
| NIEDRIG | FAQPage schema absent — acceptable now that the rich result is gone, but the questions still lack any markup | No FAQPage type in the JSON-LD |
| GOOD | Title, description and Open Graph tags complete; Organization schema carries knowsAbout and areaServed | Meta tags and JSON-LD block zero |
The answer path
Reproducing it in the lab
Every probe below is a read-only measurement of a public page — no payloads, no writes.
# 1. what the crawler sees: the shell
curl -s https://www.bewertiq.org/ | wc -c
# <- 20541 bytes, one build stamp, no data
# 2. semantic density
curl -s https://www.bewertiq.org/ | grep -c '<h2'
# <- 0
curl -s https://www.bewertiq.org/ | grep -cE '<table|<ul|<ol'
# <- 0
# 3. visible text
curl -s https://www.bewertiq.org/ | sed -e 's/<[^>]*>/ /g' | tr -s ' ' | wc -c
# <- 397 characters of visible prose
# 4. structured data
curl -s https://www.bewertiq.org/ | grep -c 'application/ld+json'
# <- 2 blocks: Organization and WebSite, nothing else
# 5. the dead SearchAction
curl -s 'https://www.bewertiq.org/?q=test' | shasum -a 256
# <- identical hash to the homepage: the declared search is not honoured
# 6. the answer-engine file
curl -s -o /dev/null -w '%{http_code}' https://www.bewertiq.org/llms.txt
# <- 200, but it serves the SPA shell, not a Markdown summary
# assertions the mirror checks, before vs. after:
# h2 count 0 -> 6+
# tables + lists 0 -> 8+
# visible prose 397 chars -> several thousand
# JSON-LD blocks 2 -> 5 (Organization, WebSite, Article, Person)
# llms.txt shell -> Markdown with h1 + file lists
# SearchAction dead -> /?q= returns real search results
Detecting it in your own estate
- Count your own semantic density first: h2, h3, table, ul, ol per page. If the crawler sees none, the answer engine quotes someone else.
- Check what your JSON-LD actually contains — Organization, WebSite, Article, BlogPosting, Person with a profile URL. Two blocks is a start, not a strategy.
- Test your SearchAction for real: submit the query the schema advertises and compare the bytes. A declared action that returns the homepage is worse than no action.
- Open Search Console and split impressions with and without AI Overviews; track branded search volume monthly. That is the zero-click report card.
- Publish llms.txt and Markdown twins for your key pages — one link per line under h2 file lists, an h1 with the site name, a blockquote summary. Validate with the Lighthouse audit.
- Keep FAQPage schema if you have real questions, but stop counting it as a rich-result strategy — the feature is gone since May 2026.
Remediation
# immediately
1. add a 40-60 word answer block under every h2 and h3, claim first, evidence second
2. convert the core facts into real tables and lists - one row per fact
3. ship Article and Person schema with an author profile URL on content pages
# within one week
4. publish /llms.txt with h1, blockquote summary and h2 file lists
5. add Markdown twins for the top landing pages
6. fix or remove the SearchAction - the declared query must return results
# within one month
7. track impressions with and without AI Overviews, branded volume, assisted conversions
8. re-measure semantic density after every content change
What you lose if you skip this
- Two thirds of your queries already end without a click. On queries with an AI Overview it is 83 percent — if the answer does not cite you, those users never see your name.
- The cited source earns the click instead: organic clicks rise 35 percent for brands the overview quotes. Exclusion from the answer is a compounding traffic tax.
- The SATISFICER is lost on the SERP, the MAXIMIZER bounces on your page — a structureless page converts neither.
- Branded search volume stays flat while the competitor's grows, because "saw it in the answer" is how people now discover a brand first.
- The data you already own never becomes an answer, because it only exists in a JavaScript bundle no answer engine executes.
Takeaways
- The click is the reward for being the answer, not the starting point. Zero-click share is 68 percent and rising, 83 percent on AI-Overview queries.
- Measure your own page before you optimise anyone else's: one h1, zero h2, zero tables and 397 characters was the state of my own product.
- FAQPage rich results died in May 2026 — schema stays valid, the SERP feature does not. Update the checklists.
- llms.txt is not for Google. It is for the retrieval agents that replaced the ten blue links.
- Semantic structure is content: tables and lists are the citations an answer engine lifts.
- Re-run the measurement monthly. The framework is cheap to apply and the only cost of skipping it is invisibility.