Something changed in how clients find me this year. More of them arrive saying “I asked ChatGPT who to talk to about this” than “I found you on Google.” If that shift is happening in your market too, the question is no longer just how to rank — it’s how to be the source the AI quotes.
What Actually Changed About Search
For twenty years the deal was straightforward: rank well, get a click. AI search breaks that deal in two ways.
First, a growing share of questions get answered without a click at all. Google’s AI Overviews, ChatGPT’s search, Perplexity — they synthesize an answer at the top and the user’s question is resolved. Your ten blue links never enter the picture.
Second, and this is the part worth optimizing for, those answers cite sources. Being one of the three or four sources a model pulls from is the new front page. The traffic is smaller than classic search traffic, but the visitors arrive pre-qualified — the AI already told them you’re a credible answer to the exact thing they were asking about.
That’s the game now: not ranking first, but being retrievable and quotable.
How AI Search Picks Its Sources
Most AI answer engines run some form of retrieval-augmented generation — the same architecture behind the RAG search engine I built on 1.1 million news articles. Understanding the mechanics tells you exactly what to change.
Roughly, the pipeline is:
- The user’s question gets interpreted, often rewritten into several sharper search queries
- The system retrieves candidate passages from an index — not whole pages, passages
- The best passages get reranked for relevance and reliability
- The model writes an answer from those passages and cites what it used
Two consequences fall straight out of that. The unit of competition is the section, not the page — a single well-written H2 that answers one question cleanly can get cited from a page that ranks eighth. And passages must survive being read alone, because the model sees a chunk of text stripped from your layout, your nav, and your surrounding argument. If a section only makes sense after reading the three above it, it gets dropped.
What to Actually Change on Your Site
Write section headings as the questions people ask. “How much does a custom CRM cost?” beats “Pricing Considerations.” The heading is a strong retrieval signal, and question-shaped headings match question-shaped queries.
Answer in the first two sentences, then elaborate. Lead with the direct answer, then add the nuance, caveats, and examples. Buried answers don’t get retrieved — the retrievable passage is the top of the section.
Make every section self-contained. Restate the subject instead of relying on “it” and “as mentioned above.” Slightly more repetitive to a human reader, dramatically more quotable to a machine.
Use specific, attributable facts. Numbers, ranges, dates, named tools, concrete conditions. “$6,000–$18,000 for a single-workflow build” is citable. “Costs vary depending on requirements” is filler, and models skip filler.
Add a real FAQ section with real answers. Question-and-answer pairs are close to the ideal retrievable format, and FAQ schema removes any ambiguity about which answer belongs to which question.
Keep structured data current. Article, Organization, FAQ, and Product schema; accurate datePublished and dateModified. Freshness is a genuine tiebreaker for anything with a year in the query.
Stay technically crawlable. Server-rendered HTML, fast responses, clean sitemaps. Retrieval crawlers are far less patient with heavy client-side rendering than Googlebot has become — an issue worth checking if your site is a JavaScript-rendered SPA rather than a statically rendered or server-rendered stack.
Building Authority the Way Models Measure It
Retrieval gets you considered. Reranking decides whether you’re cited, and that step leans on signals of reliability.
Consistent entity information matters more than most people expect — the same business name, role description, and specialization stated the same way across your site, your profiles, and anywhere you’re mentioned. Models build an internal picture of who you are from all of it, and contradictions make you a weaker candidate for citation.
Being mentioned in places the model already trusts matters too: industry publications, community answers, comparison roundups, documentation. These have always been good for SEO. In an AI-mediated world they do double duty, because they feed the model’s picture of you even when nobody clicks the link.
Topical depth beats scattered coverage. A cluster of connected articles that thoroughly covers one domain reads as expertise. A single post on twenty unrelated topics doesn’t.
How to Measure It
The honest caveat: attribution here is worse than in classic search, and anyone promising precise AI-search rank tracking is overselling.
What does work:
- Referrer traffic from
chatgpt.com,perplexity.ai, and similar hosts in your analytics — small numbers, unusually good conversion rates - Search Console patterns where impressions climb but clicks don’t, a reasonable proxy for AI Overview inclusion on your terms
- Manual citation checks — take the ten questions your customers actually ask, run them through the major assistants monthly, note who gets cited
- Direct and branded traffic lift, since a meaningful share of AI-influenced visitors arrive by typing your name in later
Judge this channel on lead quality, not volume. The people who arrive after an AI recommended you tend to show up further along than anyone from a cold search.
Should You Block AI Crawlers?
It depends on how your business makes money. If you’re a publisher whose revenue comes from pageviews, uncompensated scraping is a real problem and blocking is a defensible position.
If you’re a service business, a consultancy, or B2B software, blocking the crawlers that produce cited answers removes you from the conversation where your buyers are now making decisions. The sensible position for most businesses: allow retrieval crawlers, keep genuinely proprietary material behind authentication where crawler policy doesn’t apply, and treat citations as the distribution channel they’ve become.
What This Doesn’t Change
The fundamentals didn’t move. Fast, well-structured pages with genuinely useful content still win — AI search just raised the penalty for padding and lowered the payoff for keyword-stuffed pages that never answer the question. If your content was written for people and organized clearly, most of this is editing rather than rebuilding.
The sites that lose ground here are the same ones that were already losing it: slow, vague, and structured for search engines instead of readers. The reasons a website loses customers haven’t changed, they’ve just started applying to machines too.
If you want a straight assessment of how your site looks to AI search — what’s retrievable, what’s invisible, and what’s worth fixing first — let’s talk.