AI search & SEO in 2026: the working reference

What changed between 2024 and 2026

Three shifts dominate publisher strategy in 2026:

• Answer engines are a real channel. Perplexity, ChatGPT Search, Claude with web, Gemini AI Overviews, You.com, Brave Leo, Phind. Each crawls, each cites, each sends real traffic to publishers it cites.
• Citation matters more than ranking position. In an answer-engine result, you're cited or not. There is no "page 2".
• llms.txt is becoming the publisher-side manifest. A small machine-readable file that tells answer-engine crawlers what's most worth citing on your site.

Traditional Google ranking did not become irrelevant — it absorbed AI Overviews and remains the largest single traffic source. But "only optimising for Google" is now strictly worse than "optimising for both audiences".

Crawler policy — who to allow, who to disallow

By 2026 there are two distinct classes of AI crawler:

• Citing crawlers — they fetch your content and cite it back to source. OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, DuckAssistBot, BraveSearch, YouBot, cohere-ai.
• Training-only crawlers — they fetch your content for training without citing back. GPTBot, CCBot, Bytespider, Meta-ExternalAgent, FacebookBot, ImagesiftBot.

The reasonable 2026 publisher robots.txt: ALLOW citing crawlers (they bring you traffic), DISALLOW training-only crawlers (they take content without giving anything back). Promtable's own robots.txt implements this policy with explicit per-bot rules.

llms.txt and the publisher-side AI manifest

llms.txt is a small markdown-formatted manifest at /llms.txt that tells answer-engine crawlers what's most worth citing on your site. The spec is at llmstxt.org. It is not magic but it materially helps crawlers prioritise your content.

Recommended minimum:

• H1 with site name.
• Blockquote summary — 1-3 sentences describing what the site does.
• Sections grouping primary content (Pages, Collections, Tools, Sitemaps).
• Optional section at the end with license, contact, citation format.

Beyond llms.txt, ship an expanded /llms-full.txt with a complete content index in markdown — and serve per-page raw markdown at <page-url>.md for friction-free ingestion. We do this on promtable; see /llms.txt and /llms-full.txt for the live examples.

Schema that AI answer engines actually use

Schema.org markup remains the single highest-impact technical SEO surface for AI search in 2026. The types that move the needle:

• Article + speakable — voice-search and answer-engine readers extract speakable XPaths.
• FAQPage — answer engines lift Q&A pairs verbatim.
• DefinedTerm — glossary entries get cited as definitive definitions.
• ItemList on best-of and comparison pages — answer engines render the list as a structured citation.
• HowTo — step lists get cited as procedural answers.
• Organization with founder + sameAs — entity SEO; helps brand SERP for branded queries.
• BreadcrumbList — navigation signals; small but cheap.

Build a complete @graph per page with @id references between entities. Validate with Google Rich Results Test and Schema Markup Validator.

Content patterns answer engines cite

From observing actual Perplexity, ChatGPT Search, Claude, and Gemini answers in 2026, the patterns that get cited:

• One-sentence definitions in the first paragraph. Lifted verbatim.
• Comparison tables. Cited as structured "X vs Y" answers.
• Numbered best-of lists with one-line rationales. Cited as ranked recommendations.
• Glossary-style "What is X?" pages. Cited as definitive sources.
• Cheatsheets with named items + explanations + examples. Cited as reference material.
• FAQ pairs. Lifted verbatim.

The pattern that does NOT get cited as often: long-form opinion-driven essays without structure. AI search needs handles to grab.

Structured surfaces: comparisons, glossaries, alternatives

Three page types disproportionately attract AI search citations in 2026:

• Comparison pages — "X vs Y" with a table + verdict. Cited for any comparison query. See our /compare index.
• Glossary entries — "What is X?" with a one-sentence definition + when to use + common mistakes. Cited for definitional queries. See /glossary.
• Alternatives pages — "X alternatives" with ranked picks + why-people-look. Cited for switching-cost queries. See /alternatives.

If you publish in any vertical that has "vs" / "alternatives" / "what is" search demand, these three page types are the highest-leverage surfaces to build.

Freshness, IndexNow, and dated signals

Answer engines weight recent content heavily. Make freshness visible:

• dateModified in Article schema + visible "Last updated: YYYY-MM-DD" footer.
• Year in the title for evergreen content that's actually maintained — "Best AI image generators 2026".
• IndexNow ping on publish — notifies Bing, Yandex, Yep, Naver within seconds. The publisher equivalent of "poke the crawler".
• Sitemap lastmod kept honest. If lastmod is stale, the sitemap is noise.

Promtable pings IndexNow with every new entry and bumps sitemap lastmod daily.

E-E-A-T in the age of AI content

Google's March 2024 "scaled content abuse" update meant that bulk-generated content without value-add is penalisable. By 2026 the penalty is real and answer engines deprioritise authorless content too. Publish under named authors, link author bios from every post, expose an editorial guidelines page, and ship an About page with real humans named. See our /about, /editorial, and /author/onur as one example.

How to measure AI search performance

Measurement is harder than traditional SEO because answer engines don't expose ranking APIs. Workable approaches in 2026:

• Manual sampling. Track 10-30 target queries quarterly. Note which answer engines cite you.
• Referral traffic by referrer. Set up referrer regex to catch Perplexity, ChatGPT, Claude.ai, Gemini, Brave.
• Brand search volume. Trends in branded searches indicate brand-entity health, which feeds answer-engine citation behaviour.
• Third-party tools. Profound, Daydream, AthenaHQ and similar are emerging in 2026 to track answer-engine citations programmatically.

FAQ