In the first four months of 2026, 68% of Google searches in the United States ended without a single click. In 2024 that number was 60.45%.

For publishers, referral traffic from Google fell 38% year over year. And when an AI Overview sits at the top of the page, 8% of users click an organic result, against 15% when no summary appears.

Yet total US organic traffic fell only 2.5% in the same window. There was no collapse. There was redistribution, and somebody is getting those clicks.

What changed

The contest used to be about position. You published, you ranked, and the rank turned into clicks.

Now there is an extra step between your text and the reader, and it is not yours: a system reads the page, summarizes it, and decides who gets cited. Ranking first no longer guarantees being the source the summary picks, and a citation from position five can be worth more than the top spot.

Hence the 43% of surveyed companies already running some form of GEO/AEO strategy, optimizing for generative and answer engines, against close to zero a year earlier. It is still SEO. What changed is the reader: alongside people and crawlers, there is now a third audience that consumes the whole page and gives back three sentences.

The damage was not evenly spread. Health, finance, and education lost between 11 and 23 percentage points of organic click share, and those are exactly the niches where the question has a short, factual answer. If your content exists to answer "what is X", the summary does that better than you do.

What an AI system needs in order to cite you

Nothing exotic, and the list converges fairly well among people who write with this in mind.

The main claim has to show up early. A system that summarizes needs to find the answer, not infer it from twelve paragraphs. Burying the conclusion at the bottom was a habit built for holding readers on the page, which is the behavior that stopped working.

The structure has to be machine-readable: headings that say what the section holds, figures with units and dates, and markup declaring what the page is. BlogPosting JSON-LD with author, publish date, and modified date is what lets a system decide whether your text still holds in August 2026.

The dates have to be true. Content with no date, or with a date that does not move when the text does, gets treated as uncertain.

Every language needs a declared canonical URL. If the same article exists in three languages and the versions do not point at each other as alternates, they compete instead of compounding. Correct hreflang is what stops the Spanish version from cannibalizing the English one.

And agents need an explicit path. robots.txt still matters, and llms.txt has become the convention for telling a model what lives on the site.

Where this touches CastorStack

I wrote this partly because it demonstrates what it describes. It is published in three languages, and each version declares the other two as alternates. You can check that in the hreflang of this page.

That comes out of the data model, not a plugin. In CastorStack every post is one row per language, and translations of the same article share a translationKey. That key generates the alternates: you write the Spanish version and the hreflang across all three adjusts itself, with no hand-maintained mapping table, which tends to be where this kind of configuration rots.

The rest of the plumbing comes along. sitemap.xml and robots.ts are generated from the real content, there is an llms.txt, every post carries BlogPosting and breadcrumb JSON-LD, meta title and meta description are editable per article and per language, there are Open Graph and Twitter cards, the canonical URL is per language, and incremental revalidation runs every 60 seconds, so publishing does not require a rebuild.

The articles live in the API and are edited from the admin panel. That weighs more than it sounds: when you find out the opening sentence of your piece is not being extracted cleanly, fixing it has to cost two minutes, not a pull request and a deploy.

What this does not solve

None of it buys a citation. There is no markup that forces a model to pick you as a source, and I would distrust anyone selling that.

What correct plumbing does is smaller and more reliable. When a system does decide to read your content, it will not find three versions of the same article fighting each other, a date that contradicts the text, or a page with no signal about what it is. It removes the technical reasons to discard you. Having something worth citing stays your problem, and that is the part you cannot install.

The uncomfortable thing about 2026 is that both of those got more expensive right as building software got cheaper. Writing the CRUD is now the cheapest item on the list, and being found is one of the most expensive.

Sources

  • The Digital Bloom, Organic Traffic Crisis Report, 2026 Update
  • GoodFirms, AI SEO Statistics 2026: SERP Visibility & Zero-Click Trends
  • Position Digital, 100+ AI SEO Statistics and Insights for 2026