State of AI citations 2026What changed across the four major engines — and what is still honestly unknown.
ChatGPT, Claude, Perplexity, and Gemini all produce citations in 2026, but on very different terms — different surfaces, different consistency, different relationships to robots.txt and user-triggered fetches. This report describes the documented vendor surface, the observable product behaviour, and where the two diverge. Each finding is labelled documented or observed.
By Martin Yarnold · UpdatedThe four major engines, by citation surface
Each engine surfaces citations on a different UI contract. Understanding the surface is a precondition for measuring the citation rate accurately. The columns describe the documented surface, observed consistency, and how measurable the engine is for an outside observer.
| Engine | Citation surface | Consistency | Measurability |
|---|---|---|---|
| ChatGPT | Source cards on Search-mode answers; absent on pure model-knowledge answers. | Inconsistent — depends on whether retrieval triggered. | Medium — citations visible when retrieval ran; unclear which path produced any given answer. |
| Claude | Inline citations in retrieval mode (web access or document context); absent outside retrieval. | Consistent in retrieval mode; absent otherwise. | Medium — clear when invoked with explicit retrieval context, fuzzy in casual chat. |
| Perplexity | Numbered source list on every answer (product design contract). | High — citations always visible. | High — easiest engine to verify against, citations are a UI contract. |
| Gemini | Sources on AI Overviews and grounded answers; UI-dependent disclosure. | Variable — depends on AI Overviews surfacing for the query. | Medium — citations visible when AI Overviews trigger; less visible in chat surfaces. |
Five signal shifts in 2026
The signals operators care about each have a documented vendor contract and an observable product behaviour. The two are not always aligned, and 2026 was the year several of those gaps became operationally important.
Autonomous crawler robots.txt compliance
User-triggered fetcher behaviour
Google-Extended scope
llms.txt adoption
Provenance hashing
What is honestly unknown
None of the four major engines publishes its citation ranking algorithm. Anyone claiming to know the exact weights — "Perplexity scores backlinks at X%", "ChatGPT prefers schema type Y" — is inferring from observable output, not citing a vendor contract. The honest read for 2026 remains: the structural signals operators control are reachability, entity clarity, content depth, author signals, and freshness. Beyond that, the engines differ in ways no outside observer can definitively quantify.
The other persistent unknown is how the user-triggered fetcher stance will shift if regulators or major publishers push back harder than they have to date. The vendor docs are clear about current behaviour; whether that current behaviour holds in 2027 is an open question.
Frequently asked questions
What counts as an AI citation in this report?
A citation is an inline source attribution in a generated answer that an end user can click or read to identify the source. That includes Perplexity's numbered source list, ChatGPT Search's source cards, Claude's in-text references when invoked with retrieval, and Gemini's "sources" disclosure on AI Overviews. It excludes pure paraphrasing without attribution and excludes invisible retrieval steps inside the model's pipeline that the user never sees. The unit being measured is the end-user-visible citation, because that is what drives traffic and trust.
Which engines cite most consistently in 2026?
Perplexity is the most consistent — every answer carries a numbered source list as a product design contract. ChatGPT in Search mode produces source cards on retrieval-augmented answers but is inconsistent on pure model-knowledge answers. Claude in retrieval mode (Claude.ai web access, or with explicit document context) attributes cleanly; outside retrieval it summarises without inline citations. Gemini surfaces sources on AI Overviews and on grounded answers but the surfacing is heavily UI-dependent. Treat Perplexity as the most measurable citation surface; the others all have observable citation behaviour, but with different consistency.
Is being cited the same as ranking high?
No. Search-engine ranking and AI-engine citation are different products with different selection criteria. A page can rank well in Google Search yet not be cited by ChatGPT for the same query, and vice versa. The signals overlap (reachability, content quality, schema clarity) but the systems weight them differently and the fetch + selection paths are not the same. Sentinel measures citation rate per engine separately from search rank because the two diverge often enough to make a single composite metric misleading.
How do user-triggered fetchers (ChatGPT-User, Perplexity-User) complicate measurement?
Each major engine operates two distinct kinds of bot. Autonomous crawlers (GPTBot, OAI-SearchBot, PerplexityBot) honour robots.txt — measurable, controllable. User-triggered fetchers (ChatGPT-User, Perplexity-User) are documented by their vendors as user-initiated requests; OpenAI's bot docs note that robots.txt rules may not apply to user-initiated requests in the same way as to autonomous crawling, and Perplexity has stated Perplexity-User generally does not treat robots.txt as binding. This means measuring "did engine X cite our site?" requires distinguishing the two paths. A site blocking PerplexityBot can still be cited via Perplexity-User; a site present in ChatGPT Search citations may or may not be in OpenAI's training corpus depending on which bot fetched it.
How did Google's AI Overviews and Google-Extended behave in 2026?
Google's AI Overviews became the default presentation for many query types in 2026, with sources surfaced via the "sources" disclosure pattern. Google-Extended remained a separate, AI-specific robots.txt usage-control token: it controls whether Google may use crawled content for Gemini Apps and Vertex AI Gemini training and grounding. Disallowing Google-Extended does not remove pages from Google Search and does not by itself remove pages from AI Overviews / AI Mode in Search — those are Search features controlled via Googlebot plus preview controls (nosnippet, data-nosnippet, max-snippet, noindex). The most common 2026 misconception was treating Google-Extended as the AI Overviews kill switch; it is not.
Did llms.txt adoption matter in 2026?
llms.txt is a machine-readable markdown summary file at /llms.txt that AI engines can fetch to understand a site's authoritative summary, key pages, and citation guidance. Adoption by major brands rose visibly in 2026 — most large publishers, several enterprise SaaS sites, and a growing number of UK e-commerce brands now serve a /llms.txt. The vendor side is quiet about whether they use it as a ranking input, so treat it as documentation discipline rather than a documented citation lever. Sites that ship a clean /llms.txt also tend to ship clean schema, fast TTFB, and complete metadata — the underlying hygiene matters more than the file alone.
How does this report measure all of this?
Sentinel runs a fixed query set across the four engines weekly, records per-engine which sources are cited, and aggregates the citation rate per query as a time series. The numbers in this report describe what Sentinel observed during 2026 against that query set, with each finding labelled as either vendor-documented (sourced from each vendor's public crawler / model docs) or observational (Sentinel's own measurement). No claim is made about internal retrieval algorithms or ranking weights — those are not publicly published. The report is intended as a defensible reference for operators, not a ranking guide.