AI SEO: Citations, Sources & Optimization

What changed in 2024-2026

Three things shifted at the same time. Google rolled out AI Overviews and the broader Search Generative Experience, moving a real share of high-intent queries off the ten blue links and into a generated answer panel. ChatGPT shipped a built-in search mode that browses the web and cites sources alongside its answers. Microsoft folded Bing Chat into Copilot and brought the same surface to Windows, Edge, and Microsoft 365. The net effect is that an SEO programme aimed only at “rank for the keyword, earn the click” now leaves money on the table. The new surface is be the cited source: even when no click follows.

The technical implications are concrete. Crawlers we did not have to think about three years ago: GPTBot, OAI-SearchBot, Google-Extended, ClaudeBot, PerplexityBot: now decide whether and how a domain is represented inside an answer engine. Robots-policy decisions made for one of them implicitly affect the others. llms.txt has emerged as a plain-English manifest some engines respect and others do not. And structured data: Article, FAQPage, Organization, Product, BreadcrumbList: now feeds two consumers instead of one: classical SERP features, and the answer-engine grounding pipelines that prefer machine-readable claims over inferred ones.

How AI engines pick which sources to cite

Each engine ranks sources differently, but the signals overlap enough to author against. In our work across the five engines we track, the durable signals are:

• Crawlability for the engine's own user-agent. If GPTBot is blocked in robots.txt, a page does not enter ChatGPT's grounding index, full stop. The same applies per engine. Auditing this is the single highest-ROI AI-SEO task and most sites still get it wrong.
• Topical authority on a single, named entity. Engines prefer pages that are clearly “about” one thing. A glossary entry titled exactly “Performance Max” outperforms a 5,000-word omnibus that mentions Performance Max in one paragraph.
• Direct, declarative answer in the first 200 words. Quick-answer blocks like the one at the top of this page are cited disproportionately because the engine can lift them verbatim with confidence.
• FAQ schema with question text that matches real PAA queries. The match does not have to be exact; engines normalise. But the question-shape matters more than the keyword density.
• External corroboration. An entity that is described consistently across our domain, Wikipedia, and the primary platform's own docs is more “known” to the engine than one that lives only on our site.
• Freshness on volatile topics. Engines deprioritise pages with stale dateModified on subjects that move fast: ad-platform policy changes, AI features, pricing.

The Unled AI SEO playbook

1. Map the entity. Pick one named entity per page: a platform, a tactic, a vertical: and write to it. Our M5 intent matrix and the M7 keyword matrix linked below are how we do this internally.
2. Open with a quick answer. 80-160 words. Plain language. No marketing. Engines lift this; humans skim it.
3. Layer the body in three passes. First pass: definition + scope. Second pass: how it works (named mechanisms, not metaphors). Third pass: the operational view: what we do, how we measure, what fails.
4. Mark up with schema. Article + FAQPage on long-form, Product on account / programme pages, Organization sitewide. Use real, verifiable values; AI engines downrank schema that contradicts visible content.
5. Audit the crawler matrix. Maintain an allow-list per engine in robots.txt and llms.txt. Decide brand-by-brand; the choice is rarely uniform.
6. Cite primary sources externally. Linking to the platform's own documentation increases the engine's confidence that we know what we are talking about. Linking to competitor blogs decreases it.
7. Re-baseline monthly. Engines change their citation behaviour on a 30-90-day cycle. A page that was cited in March may be ignored in June with no edits. The fix is measurement (M14), not panic.

Per-engine guides

Each major answer engine gets a dedicated page covering its crawler, its ranking quirks, and the optimisation moves we use:

• ChatGPT SEO: OpenAI's search mode, OAI-SearchBot, and how to be cited in chat answers.
• Gemini SEO: Google AI Overviews, Google-Extended, and the overlap with classical Google SEO.
• Copilot SEO: Microsoft Copilot, the Bing index, and the Microsoft 365 answer surface.
• Claude SEO: Anthropic's ClaudeBot, web-aware Claude, and citation patterns.
• Perplexity SEO: PerplexityBot, Perplexity Pages, and the citations-first answer model.

What this is not

AI SEO is not a guarantee of citation. There is no “submit URL” flow that puts a page into ChatGPT's grounding set. There is no method that bypasses an engine's safety filters, and we do not sell one. There are no “hidden prompts” or invisible-text tricks worth recommending: every engine we track has either trained against them, ignores them at inference time, or both. The work that compounds is the boring work: clean structure, honest content, the right schema, the right crawler permissions, and the patience to measure on a monthly cycle.