How ChatGPT, Perplexity, and Claude Choose Sources in 2026

There is no public ranking algorithm for generative AI engines. There is, however, a growing body of observed behavior — from leaked system prompts, engineer interviews, and structured experiments run across tens of thousands of queries. Here is what we know, as of April 2026, about how the three dominant answer engines choose which sources to cite.

ChatGPT: Bing retrieval + OpenAI re-ranker

When ChatGPT browses, it issues one or more Bing queries, pulls the top 10–20 results, fetches the HTML, and hands passages to an internal re-ranker before the model composes its answer. The re-ranker weighs three things heavily: freshness (dateModified within the last 24 months), passage coherence (paragraphs that read as standalone answers), and schema signals (Article, FAQPage, and HowTo blocks in particular).

Pages that surface in ChatGPT’s citations disproportionately share four traits: a clear H1 that matches the query intent, an early summary paragraph, dateModified schema, and a named author. Pages that never surface almost always miss at least two of these.

Perplexity: hybrid retrieval with a proprietary quality model

Perplexity runs its own web index on top of Brave’s crawler, with a separate academic and news index layered in. What sets it apart is the quality model — a classifier that scores candidate sources for citation-worthiness before the LLM ever sees them.

The quality model is trained heavily on user clicks within the Perplexity UI: users click some cited sources and not others, and the model learns from that signal. In practice, this means two things. First, original reporting and primary data outperform aggregators, even at the same domain authority. Second, sources with visible author names and clear publication dates are clicked more and therefore cited more.

Claude: curated retrieval, authority-weighted

Claude’s retrieval for web citations — whether via Claude’s web search tool or via Claude in Chrome — biases toward a curated trust tier. The engine prefers named-entity sources (Wikipedia, major publishers, academic domains, well-established brand sites) and de-prioritizes content farms.

For a mid-sized WordPress site, the path into Claude’s citation pool is through entity linking: connecting your brand and your author entities to Wikidata, cross-linking from adjacent authoritative domains, and maintaining a consistent Organization schema across the site.

What all three engines reward

Across all three engines, the same five on-page behaviors consistently lift citation rates:

• Passage-level answers. One sentence, then supporting detail. Not a build-up.
• Fresh dateModified. A post from 2022 loses to the same content re-published in 2026.
• Named authors with bios. Anonymous posts are cited a fraction as often.
• Clean schema. Especially Article, FAQPage, and HowTo.
• Fast TTFB. Under 600ms. Slow responses get dropped during retrieval.

These aren’t new — most of them are also classic GEO signals. But their weight is higher in the citation decision than in classic ranking.

What all three engines penalize

• Thin or low-originality content. Posts that paraphrase what’s already in the training data add no new information and are rarely cited.
• Unlabeled AI output. All three vendors have quality classifiers that flag AI-generated patterns. Edit aggressively.
• Paywalls without structured previews. If the crawler can’t read the body, the passage doesn’t qualify.
• Broken or outdated schema. A malformed JSON-LD block can get your page skipped entirely.

The single highest-leverage change

If you do one thing this week, lead every H2 with a one-sentence answer. The rest of the section can unfold however you like. That single habit lifts citations across all three engines because all three extract passages in the same way: first sentence after the heading, then more if needed.

FAQs

Can I game AI citations with prompt injection?

No. All three engines strip instruction-like content from retrieved HTML before composing answers. The defensive filters have been production-grade since late 2024.

Does domain authority still matter?

Yes, indirectly. Higher DA correlates with better inbound citations, cleaner technical baseline, and more editorial signal — all of which AI engines reward. DA itself isn’t a ranking factor, but the things that cause high DA are.

How do I track which engines are citing me?

Manually: run sample queries weekly. Automated: use a citation monitor like GEO Suite’s built-in tracker, or roll your own with the three vendor APIs.