ChatGPT's Deep Research can scan dozens of pages about a brand in a single session and still cite almost none of them. David Konitzny at Peec AI logged more than 200 Deep Research runs across ten-plus accounts this year and found the same pattern every time: the agent reads product pages, comparison posts, and reviews at length, then writes its final report from a short list of sources it judges trustworthy enough to name. For a brand tracking AI visibility, that gap between being read and being cited is now a sharper risk than being ignored outright. A page the agent never opens was never going to be cited. A page it opens three or four times and still leaves out of the report says something more specific about what it trusts, and what it doesn't. If you're already tracking how often your brand gets mentioned across ChatGPT, Gemini, and Perplexity, this adds a second number worth watching: how many of those mentions would survive as an actual citation in a longer research session.
What a "deep research" session actually does
Deep Research products inside ChatGPT, Gemini, and Perplexity run a multi-step loop on a single prompt: dozens of search queries, repeated page opens, note-taking, and a final written report with inline citations, instead of the single-pass retrieval behind a normal chat answer.
How often that loop even runs is capped very differently by plan. Perplexity Pro allows up to 20 Deep Research runs a day, near 600 a month at daily use, against five reports a month on the equivalent Google AI Pro plan for Gemini, a 120x difference between subscribers paying close to the same monthly price. ChatGPT's Plus tier sits in between, at ten runs a month, with each one reading what the product describes as “dozens of sources” before writing up a result. None of this measures citation quality; no independent lab has tested the three engines head to head on accuracy. But it is a plain allowance difference, and it decides how often any one engine's version of deep research even gets pointed at a given brand before citation becomes a question at all.
The logs show reading and citing are two separate decisions
In Peec AI's analysis of its own Deep Research session logs, the citation decision happens well after a page has already been read. Reading shapes what the model understands; naming a source is a separate judgment call about whether a specific claim needs backing. Plenty of pages that shaped the final answer never get that second nod.
The numbers show how early the filtering starts. Pages with fewer than 20 links made up about 78% of what Deep Research actually read in full on its first pass; pages with 60 or more links dropped to 33%. The “click” command, the one that would let the agent follow a link off the page it already has open, was used zero times across every logged session; it stayed inside the page, re-reading the same text window without a new keyword about one time in five. When an in-page keyword search failed, the agent tried a different keyword roughly 60% of the time and abandoned the page for a fresh search about a quarter of the time. That pattern echoes something we found looking at the crawl-to-refer ratio across AI bots generally: the ratio of what a bot takes from a site to what it gives back in referral credit is already lopsided before citation accuracy even enters the picture.
A citation gap is the distance between how often an AI engine reads about a brand and how often it names that brand as the source backing a claim. The two numbers rarely move together.
Trust language explains who gets named
When Deep Research justifies a citation in its own reasoning trace, the language clusters around words like "official," "reliable," and "credible." Pages it treats as background rather than evidence get hedged with phrases like "not an official source," and forum or user-generated content almost never earns a citation even when it was read closely.
That pattern tracks closely with Google's E-E-A-T framework for ranking web pages, even though nothing in a Deep Research prompt instructs the model to apply it. The practical read for a brand: being the page a journalist would call the official source on a topic, your own documentation, a page you control and can keep current, matters more to the citation decision than how many times the agent opened your content. We found a related bias looking at how AI engines weigh Wikipedia and Wikidata over a brand's own site: engines default to the entity record they consider independently verified over brand-authored copy, for the same underlying reason. The agent is hunting for confirmation it can cite; raw information alone doesn't satisfy that.
Citations concentrate on a small number of domains, and they're rarely yours
A synthesis of three 2026 datasets covering roughly 600,000 ChatGPT citation events found Wikipedia and Reddit alone account for about a quarter of all U.S. citations, 13% and 12% respectively, while the Wall Street Journal, the New York Times, Bloomberg, and the Financial Times don't crack the top 20 domains at all.
Forbes lands 18th with a 1.38% share, in the same review, published by Josh Bernstein at Something Inc. in August 2026. That concentration means most brands are competing for citation space with reference and discussion sites that have nothing to do with their category, not with each other. A related breakdown we ran on third-party citation sources found a similar pattern specific to B2B brands: most citations referencing a company came from Reddit, LinkedIn, and G2 rather than the company's own domain. The lesson for a brand monitoring its AI Share of Voice is that a strong citation share has to be built partly on pages the brand doesn't own, by making sure it's represented accurately on the sites these engines already trust.
Citation errors compound the gap further
Even when a Deep Research agent does name a source, the citation isn't always accurate. An August 2026 study of three agentic research systems found citation recall, the share of claims correctly backed by a real source, ranged from 58.7% down to just 7.1% depending on the system's architecture.
The paper, from researchers at Bar-Ilan University, UNC Chapel Hill, and UT Austin, traced most errors to the orchestrating agent that drafts the final report rather than the sub-agents doing the actual page reading, with the orchestrator accounting for 84.7% of one system's final-report errors. Two fixes closed part of the gap in testing: giving the orchestrator raw source text instead of a researcher agent's paraphrased notes, and instructing it not to state anything it couldn't cite. Both lifted citation recall by about five points without hurting report quality. For a brand, the implication is blunt. A citation gap in your favor isn't necessarily a sign the agent couldn't find you. It's just as often a sign the sentence connecting your brand to the fact got lost a step or two before the final report was written.
What this means for tracking brand visibility
Treat Deep Research like a research assistant whose first pass decides almost everything. Three concrete levers follow from the logs above:
Keep the pages it's most likely to open under 20 internal links, to land in the 78% that gets read in full.
Put the clearest factual claim about the brand near the top of the page, inside the first few thousand characters the agent actually opens.
Make sure the page reads as a primary or official source rather than marketing copy; the agent hedges hard on anything that sounds like an advertisement for itself.
That last lever is the hardest to pull cosmetically. The agent isn't scoring word count; it's reading tone and structure much the way a skeptical journalist would, deciding whether a page sounds like the record or sounds like a pitch. Earlier VizibleAI research into how often AI engines cite press releases found promotional, announcement-style copy gets cited at a rate close to zero regardless of how factual it is, for the same underlying reason.
Tracking the deep-research gap well also means measuring two different numbers instead of one: how often a brand gets mentioned at all, and how often that mention survives into an actual citation. A 126-million-prompt study covered on this blog found that exact split, a brand's mention rate and its citation rate, moving independently of each other across engines, the same read-versus-cite distinction Deep Research makes visible inside a single session.
Frequently Asked Questions
What is the AI deep research citation gap?
It's the difference between how many pages an AI deep research agent like ChatGPT's Deep Research actually reads during a session and how many of those pages it names as a source in its final report. Peec AI's 2026 log analysis found the agent reads broadly but cites narrowly, often skipping pages that shaped its answer.
Does ChatGPT's Deep Research crawl differently from a normal ChatGPT answer?
Yes. Deep Research runs a multi-step loop, dozens of searches, repeated page reads, and note-taking, before writing a report with inline citations. A standard ChatGPT answer typically pulls from a handful of retrieved snippets in a single pass, without that extended research phase.
Why do AI deep research tools cite Wikipedia and Reddit so often?
A 2026 review of roughly 600,000 ChatGPT citation events found Wikipedia and Reddit together account for about a quarter of U.S. citations, more than any single news publisher. Deep research agents treat recognized, third-party-verified sources as safer to cite than a brand's own marketing copy.
Can a brand improve how often it gets cited in AI deep research reports?
The data points to a few concrete levers: keep key pages short enough to be read in full rather than skimmed, lead with the specific factual claim an agent would need to cite, and make sure the page reads as an authoritative source rather than promotional content, since these agents treat anything that sounds like self-promotion with heavy skepticism.
Are AI deep research citations reliable even when a source is named?
Not entirely. An August 2026 study of three agentic research systems found citation recall, the share of claims correctly backed by a real source, ranged from 58.7% down to 7.1% depending on the system, with most errors traced to the agent drafting the final report rather than the one reading the pages.
What's the difference between a brand's AI mention rate and its citation rate?
Mention rate tracks how often a brand's name appears in an AI engine's answers. Citation rate tracks how often a specific page tied to that brand is named as the source. A large-scale 2026 prompt study found these two numbers move independently, meaning a brand can be mentioned often while still rarely being cited.