When a user runs a deep research query, an AI agent begins an autonomous browsing session that lasts anywhere from five to thirty minutes. It issues dozens of search queries, visits tens or hundreds of web pages, reads and reasons across the content it finds, and then distils everything into a structured report - with citations buried at the bottom of a document the user rarely scrolls to the end of. Each of those page visits lands on a publisher's server as a legitimate read: content is consumed, processed in the model's context window, and woven into a response the user receives without ever seeing the original source in a browser.
That is the deep research model, and it represents the sharpest version of the trend that has been reshaping publisher economics since 2024. Where AI Overviews and standard chatbot answers typically consume one or two sources per query, deep research tools are designed to consume many, and to synthesise rather than cite. For publishers, it is content consumption at scale with traffic and revenue returns close to zero.
What is AI deep research?
Deep research refers to a class of AI feature in which a model is given an open-ended question, then autonomously plans and executes a multi-step research process - searching the web, visiting pages, reading content, identifying gaps, and repeating - before writing a long-form report in response. The process is agentic: the model decides which sources to visit and in what order, without the user directing each step.
Google Gemini Deep Research launched in December 2024, available initially to Gemini Advanced subscribers. OpenAI and Perplexity both launched their own deep research products in February 2025, with Perplexity's accessible on a freemium basis. By March 2026, Perplexity had updated its product to route tasks across more than twenty frontier AI models simultaneously. All three products are now widely available and actively used.
A single deep research query can involve the agent consulting twenty or more sources in a single session, with complex queries routinely drawing on hundreds of pages. One documented session lasted twenty-eight minutes and referenced twenty-one distinct sources. The Wikipedia article on ChatGPT Deep Research notes that deep research is explicitly designed for queries that require coverage across hundreds of web pages, distinguishing it from standard agentic web search which typically uses a few dozen.
How is deep research different from a standard AI answer?
A standard AI chatbot answer - whether from ChatGPT, Gemini, or Perplexity in its default mode - typically consults a handful of sources, pulls a short summary, and surfaces one or two inline citations the user might click. Click-through rates from AI citations are low: SparkToro data from January 2026 found that roughly twelve to eighteen per cent of Perplexity citations produce a click.
Deep research produces a fundamentally different output. The report can run to twenty-five or fifty pages. Citations are listed in a reference section rather than embedded as live links the user encounters mid-read. The user's intent is to receive the synthesised report, not to browse its sources. The practical result is that click-through from deep research citations is lower than from standard AI search citations - and standard AI citations already represent a substantial drop from traditional search referrals.
The scale of content consumption per query is also categorically different. Where a standard AI chatbot answer might visit two or three pages, a deep research query visits tens or hundreds. Research on AI agent web traffic published in 2026 found that LLM-based agents generate ten to sixty times more web traffic than human users for identical queries. The traffic lands, but it does not convert into an ad impression, a session, or a revenue event.
What does it mean for publisher traffic?
Publisher referral traffic from search was already declining sharply before deep research became mainstream. Chartbeat data shows global organic Google search traffic fell thirty-three per cent between November 2024 and November 2025, with the US decline reaching thirty-eight per cent over the same period. The Reuters Institute's Journalism, Media, and Technology Trends and Predictions 2026 report - based on a survey of 280 media leaders from 51 countries - found that publishers expect search engine referrals to fall a further forty-three per cent within three years, with a fifth of respondents expecting losses above seventy-five per cent.
Deep research accelerates this trajectory in two ways. First, it increases the volume of content consumed per query without increasing referral traffic, widening the gap between what publishers produce and what they earn. Second, it normalises a research workflow in which the user never visits a publisher site at all - the report is the destination, and the source URLs are an appendix.
The traffic that deep research does send is almost impossible to detect with standard analytics. Like other AI crawler activity, deep research agent visits typically arrive without a referrer header and may use rotating user agents. A publisher may see a spike in what looks like direct or unknown traffic without knowing it originated from dozens of deep research queries that consumed their content extensively.
Do deep research tools credit publishers in their reports?
Yes, but in a way that rarely translates into traffic. Deep research reports typically include a numbered or bulleted reference list at the end, and the text may contain inline citations linking to source URLs. However, users who commission a deep research report are doing so precisely because they want to avoid the work of reading sources themselves. The report is the product. The citations are a signal of thoroughness, not an invitation to browse.
OpenAI's Deep Research, which now runs on GPT-5.2, lists sources alongside the final report. Google Gemini Deep Research presents sources in a sidebar. Perplexity's version, which updated in March 2026 to route subtasks across more than twenty models, displays citations inline. In all cases, the format is designed to give the user confidence in the report rather than to drive traffic back to the original publisher.
There is no compensation mechanism attached to these citations. A publisher whose article is consulted twenty times in a single Gemini Deep Research session receives the same payment as a publisher whose article was not consulted at all: nothing.
What can publishers do about it?
The core problem is structural. Deep research agents read publisher content at the CDN layer - the moment the HTML is served - and that read is where the value transfer happens. No JavaScript runs, no ad auction fires, no session begins. The read is invisible to the publisher's analytics and generates no revenue under any current advertising or licensing model.
Three responses are available to publishers at the moment.
The first is access control: using robots.txt, Cloudflare's Content Signals Policy, or Web Bot Auth-verified agent tokens to restrict or block deep research agents. This prevents consumption but also forfeits any citation or referral traffic the tool might otherwise send.
The second is licensing: seeking a direct deal with the AI provider to permit deep research use in exchange for a fee. In practice, as the Reuters Institute data and existing reporting on AI licensing deals makes clear, only a small number of very large publishers have access to these deals, and the economics are often unfavourable even for those who do.
The third is edge monetisation: detecting agent reads at the CDN layer and generating a revenue event from the read itself, without depending on the human visit that no longer comes. This approach treats the read - rather than the click - as the monetisable unit. It is the direction blankspace has built toward: intercepting Live Search Agent traffic at the edge, identifying the agentic context, and enabling a brand fact to accompany the content into the model's response.
Frequently asked questions
When did deep research AI tools launch?
Google Gemini Deep Research launched in December 2024. OpenAI Deep Research and Perplexity Deep Research both launched in February 2025. By early 2026, all three were widely available, and Perplexity had released an updated version routing tasks across more than twenty AI models simultaneously.
How many pages does an AI deep research query visit?
It varies by query complexity, but a single deep research session typically visits tens of source pages, with complex queries drawing on hundreds. One documented ChatGPT Deep Research session referenced twenty-one sources across twenty-eight minutes of processing. Research published in 2026 found LLM agents generate ten to sixty times more web traffic than human users for equivalent queries.
Does AI deep research send referral traffic to publishers?
Rarely and in small volumes. Citations are listed in reference sections of long-form reports, and users who commission deep research reports do so to receive the synthesis rather than to browse sources. Click-through from deep research citations is lower than from standard AI search citations, which are themselves much lower than traditional search click-throughs.
How does deep research differ from a standard AI Overviews or chatbot answer?
AI Overviews typically synthesise one or two sources into a short summary on a results page. Standard chatbot answers visit a handful of sources and surface inline citations the user encounters mid-response. Deep research is an agentic, multi-step process that autonomously browses dozens to hundreds of sources over five to thirty minutes and produces a long-form structured report - often twenty-five pages or more. The scale of content consumption per query is an order of magnitude higher, but the referral traffic generated is not.
Can publishers opt out of being used in deep research?
Partially. Robots.txt directives, Cloudflare's Content Signals Policy, and emerging standards like Web Bot Auth can be used to block or restrict specific AI agents. However, many deep research tools use the same user agents as general web crawlers, making targeted exclusion difficult in practice. Full opt-out also means forfeiting the citations and the small volume of referral traffic these tools do occasionally generate.