How AI Search Decides Who to Cite — and Who Disappears

How AI Search Decides Who to Cite — and Who Disappears

The Numbers Behind the Shift

At Google I/O on May 19, 2026, Google announced that AI Mode had passed one billion monthly users just one year after launch, with queries more than doubling every quarter. It also made Gemini 3.5 Flash the default model powering AI Mode globally, and shipped what it called the biggest upgrade to the search box in 25 years.

The traffic side of that story was already measured before the milestone. Pew Research Center tracked the real browsing behavior of 900 U.S. adults and found that when a Google search produced an AI summary, people clicked a traditional result link in only 8% of visits — versus 15% when no summary appeared. Links inside the AI summary were clicked in just 1% of visits. And 26% of sessions ended entirely on the AI-summary page, compared with 16% on a traditional results page. Google has disputed the study's methodology, but no one — including Google — is arguing that clicks are going up.

So the honest framing for a small business is this: fewer people will click through to your website from an informational search than at any point in the web's history. The question that matters now is not "how do I rank #1" but "how do I become one of the sources the AI answer is built from — and how do I win the smaller number of clicks that remain."

How an AI Answer Actually Gets Built

This is the part most coverage skips, and it is the part worth understanding, because the AI is not writing answers from memory — it is running searches and reading pages, live.

When you ask AI Mode a question, it breaks your question into several smaller sub-queries and runs them in parallel — Google calls this query fan-out. Each sub-query pulls candidate pages from the same index normal search uses. The model then reads passages from those candidates, composes an answer, and attaches citations to the passages it leaned on.

Three things follow from that mechanism, and each one is actionable:

First, if Google can't crawl and render your page, you don't exist in this system. Not penalized — absent. The retrieval step can only choose from what's in the index. Slow pages, broken rendering, and accidental crawler blocks now cost you twice: once in traditional rankings, and again in every AI answer you could have been cited in.

Second, the model cites passages, not domains. It is not selecting "a trustworthy brand"; it is selecting a specific chunk of text that directly answers a specific sub-question. A 400-word section with a clear heading that states its answer in the first two sentences is more citable than a 3,000-word page where the answer is smeared across twelve paragraphs.

Third, the model prefers sources that say something checkable. When candidate passages disagree, generic content loses to content with specifics — numbers, dates, names, first-hand observations — because specifics can be corroborated against other retrieved sources.

The Part the Panic Posts Are Missing

Citation selection is the first ranking system in twenty years that doesn't structurally favor the biggest site. Traditional rankings compound authority: more links beget more rankings beget more links. Passage-level citation resets some of that. A five-page local contractor site with a genuinely specific page — real prices, real timelines, real photos of real jobs — can be the cited source for a question that a national aggregator answers with boilerplate.

We wrote about this dynamic when AI Overviews first started reshaping small business traffic, and the pattern has only sharpened since: the sites losing hardest are the ones whose content could have been written by anyone. Aggregators, thin listicles, "ultimate guides" assembled from other guides. The AI answer is that content now, done faster.

The other overlooked fact: the clicks that survive are worth more. Pew's data says fewer people click through — but the ones who do have already read a summary and still wanted the source. They arrive closer to a decision. A site built to convert that visitor — clear offer, fast pages, obvious next step — loses less revenue than raw traffic numbers suggest.

What Loses Ground in an AI-First Results Page

The losers share a profile:

• Content written for a keyword instead of a question. Pages that circle a topic without ever committing to an answer give the model nothing to cite.
• Answers buried below the fold. Long intros, recycled context, "in this article we will cover" — the model reads for the passage that answers, and these pages make it dig.
• No original information. If every fact on your page exists on fifty other pages, the citation goes to whichever source is cleanest — and it usually isn't you.
• Technical debt. Pages that render poorly without JavaScript, load slowly, or block crawlers in ways the owner never intended. We've seen sites unknowingly blocking the exact user agents that feed AI answers.

What a Citable Site Looks Like

Practical moves, in priority order:

1. Write answer-shaped sections. One question per heading, the direct answer in the first sentence or two, supporting detail after. This is good for human skimmers and it is exactly what passage retrieval selects for.
2. Publish something only you can publish. Your prices. Your process. What a job actually costs in your county. Results you measured on your own work. First-party information is the one input the AI cannot synthesize from elsewhere — it can only cite it. It's the same principle behind building digital authority: credibility comes from verifiable specifics, not volume.
3. Keep the technical layer clean. Server-rendered content, working structured data, fast pages — the kind of unglamorous upkeep ongoing site maintenance exists to cover. Speed already shapes how visitors judge you; now it also gates whether the retrieval step sees your page at all.
4. Name who is behind the content. Retrieval corroborates sources against each other, and pages with identifiable authors, a real business address, and consistent details across the web corroborate cleanly.
5. Build for the post-summary visitor. Assume everyone arriving has already read an AI answer. Your page's job is to be the depth the summary couldn't provide, and to make the next step obvious.

Why This Is Structural, Not a Trend

Google is not running an experiment it might roll back. One billion monthly users, a new default model, and the search box rebuilt around AI input — that is infrastructure investment, not a feature test. The answer layer is now permanently between you and most of your potential visitors.

That changes what a small business website is for. Its job used to be capturing attention at the top of the funnel. Its job now is being the credible, specific, technically clean source that the answer layer draws from — and converting the smaller, warmer stream of visitors who click through anyway. Sites built as brochures do neither. Sites built as sources do both.

The Sitora Take

Most of the advice circulating right now tells small businesses to produce more content. We think that's exactly backwards. The AI answer layer punishes volume and rewards verifiable specificity — and specificity is the one asset a small local business has more of than any content farm: real work, real numbers, real ground truth about your trade in your town.

The technical half matters just as much and gets ignored because it's invisible. A site that renders cleanly, loads fast, and structures its content for retrieval is now competing in every AI answer in its niche. A site that doesn't is invisible in a way no amount of writing fixes.

👉 Get a website built to be cited, not just seen or book a strategy call with Sitora.