The old SEO question was: can I rank for this query? The new one, increasingly, is: will I be cited when an AI answers this query? It's a different question with different answers, and most of the SEO literature is still catching up.
Generative search engines — Google's AI Overview, Perplexity, ChatGPT search, Bing Copilot — read pages, synthesise an answer, and link to source URLs. The citation slot is the new ranking slot, and the URLs that earn it are not always the URLs at position #1.
Here's what we keep seeing in pages that get cited, and what that means for how you should write.
What we know about how AI search picks sources
There's no published algorithm. There are observable patterns from a year of citation analysis, leaked papers, and direct testing. The honest summary:
- Classic SEO matters first. Pages need to be indexable, fast, and ranking somewhere on page 1-3 of organic. AI engines almost always cite from this candidate pool.
- Within that pool, AI engines re-rank by extractability. A page at position 8 with a clearer answer often beats a page at position 2 with a fuzzier one.
- Authoritative signals still help. Domain age, citation count, and known-publisher status are all heavy inputs. New domains don't get cited unless their content is unusually specific or unusually first.
What this means in practice: don't abandon SEO basics for AI optimization. Do classic SEO and make your content extractable.
The patterns in cited content
Across maybe a few hundred AI Overview / Perplexity citations I've examined, the same shapes recur:
1. Answer-first paragraphs
The cited paragraph almost always answers the question in its first 1-2 sentences, then provides nuance. Like this:
The cutoff for a meta description in Google is approximately 920 pixels of width on desktop, which works out to 150-160 characters of average text. Mobile is wider, around 1200 pixels.
Compare to a non-answer-first version:
Meta descriptions are an important part of on-page SEO. Many factors affect how Google displays them, including pixel width, query intent, and the presence of rich snippet markup. The general consensus among SEO professionals is...
Same eventual information. But the AI extractor needs the answer in sentence one. Burying it loses the citation.
2. Concrete numbers and specifics
Pages with concrete data (specific numbers, named studies, dated benchmarks, comparison tables) get cited at much higher rates than pages with vague claims. AI engines are tuned to prefer falsifiable specificity over hedged generality, partly because it's easier to verify and partly because it's more useful in an answer.
Phrases that suppress citation: "many," "often," "typically," "industry experts agree." Phrases that earn citation: "27%," "as of Q1 2026," "in a 2024 Ahrefs study of 4M pages."
3. Original framing and contrarian takes
When a query has 50 pages saying the same thing, AI engines pick a few — usually the most authoritative ones. When a query has one page saying something different and credible, that page gets disproportionate citation share because it adds information the model can't get elsewhere.
This is good news for smaller sites with strong opinions. A 1500-word post arguing against a common SEO trope can outperform a 5000-word post restating the trope, if the contrarian post is well-argued and well-evidenced.
4. Q&A and definition-first structure
Pages structured as questions (H2s) followed by direct answers (first paragraph) cite higher than flowing essay-style content. This is partly schema/extractability — the structure makes it obvious which paragraph answers which question — and partly because users phrase queries as questions, so questions in your content match user queries lexically.
If your content is essay-style, you can usually retrofit by:
- Converting H2s into questions
- Putting the answer in the first paragraph after each H2
- Keeping subsequent paragraphs as supporting evidence
5. Tables, lists, and structured comparisons
A comparison table of three options is more extractable than three paragraphs comparing three options. AI engines preferentially pull tables and short lists because they survive extraction without losing structure.
If you're writing a "best of" or "vs." piece, a table is the citation-friendly format. Prose explanations can come after.
What classic SEO still does for you
Some quick reminders that get lost in the AI-optimization conversation:
- Without indexing, no citation. Crawlable, fast, no JS-only content, proper canonical and robots.
- Without ranking, low citation likelihood. AI engines generally cite from the top organic results.
- Without structured data, less context. FAQPage and Article schema both help models parse your page (see the FAQ schema post for the nuances).
- Without internal links, weak topical authority. Models reward sites that look like topical experts.
Doing AI-optimization while neglecting these is like polishing a car with no engine.
A pre-publish checklist for AI-friendly content
For each new piece you publish:
- Open with the answer. First 1-2 sentences after the H1 should answer the page's central question.
- One number per claim. Where you'd write "many users," write "62% of users (Stripe study, 2024)."
- Question-shaped headings. Where it makes sense, frame H2s as the questions readers actually search.
- Direct answer paragraphs. Each H2 has its answer in paragraph 1, evidence in paragraph 2+.
- At least one table or list. Tables especially survive extraction better than paragraphs.
- Original synthesis. Don't restate consensus. Add your data, your contrarian take, or your specific case study.
- Schema markup. FAQPage and Article where applicable.
- Sources and dates. Cite where your claims come from. Models trust sourced claims more.
This isn't different from "good writing." Good writing has always answered questions clearly. AI search just penalises bad writing more visibly.
What we don't know
Honesty section: nobody outside Google, OpenAI, Anthropic and Perplexity knows exactly how citation selection works. The patterns above are observed correlations, not confirmed causation. Algorithms shift. The advice that's useful in 2026 may be obsolete by 2027.
Two safe bets for the longer term:
- Genuinely useful, specific, original content keeps performing across algorithm changes. It's the only durable strategy.
- Optimization tactics that are essentially "write more clearly" — answer-first paragraphs, structured Q&A, sourced numbers — are unlikely to backfire.
Skip the tactics that feel like keyword-stuffing for AI ("As an AI language model would say..." prompts in your content, etc.). Those will age the way doorway pages did.
Tools
- AI Snippet Writer — generates titles and meta descriptions tuned for both classic SERP and AI extraction. Concrete, direct, and intent-matched.
- Schema Generator — JSON-LD for the schema types AI engines actually parse: FAQPage, Article, Breadcrumb, Product.
- Snippet Optimizer — your snippet is the model's first impression of your page. Tight titles and clear descriptions correlate with higher citation rates.
The new search isn't about ranking on a page. It's about being quotable in an answer. Same principles, different surface, slightly different writing discipline.