Picture one analyst asking an AI agent for a market briefing before a meeting. While they make a coffee, the agent drafts a research plan, runs a long series of searches, opens page after page across news sites, trade titles and reference sources, and returns a structured report with a source list at the end. From the publisher's side, that session looks like a burst of ordinary page requests. From the analyst's side, nobody needed to click anything. That is AI deep research: an autonomous, multi-step browsing and synthesis process that turns many publisher pages into a single answer, and that sits almost entirely outside the click-based model publishers have monetised for twenty years.
What is AI deep research?
Deep research is a mode in AI assistants where the model, given an open-ended question, plans and runs its own research process. It searches, reads pages, spots gaps, searches again, and then writes a long-form report with citations. The user sets the question and, in some products, reviews the plan. The agent chooses the sources.
The main products are now well established. OpenAI's Deep Research launched in February 2025 and was upgraded on 10 February 2026 to run on GPT-5.2, with a full-screen report viewer and a control that lets users choose which websites a report draws on (Technobezz, 11 February 2026). Google's consumer Gemini Deep Research launched in December 2024; its developer agent followed in December 2025, and on 21 April 2026 Google launched two evolved versions built on Gemini 3.1 Pro: Deep Research, tuned for speed, and Deep Research Max, which uses extended compute for slower background jobs. Both are available in public preview through the Gemini API (Google blog, 21 April 2026). Perplexity and Anthropic offer their own research modes, and Claude's research capability has been available since 2025.
How is deep research different from a normal AI answer?
A standard AI answer consults a small number of sources and surfaces a few inline links. Deep research is different in three ways that matter to publishers.
First, depth. A session runs for minutes rather than seconds and can consult many more pages than a single chatbot reply. None of the vendors publishes an official per-session page count, and third-party estimates vary widely, so treat any single figure with caution. What is not in doubt is the design intent: breadth of coverage is the product.
Second, format. The output is a report, often long, with sources listed in a reference section or side panel. The user's goal is to avoid reading the sources themselves.
Third, control. Users can now steer the agent toward particular sites. OpenAI's February 2026 update lets users add or prioritise specific websites. Google's April 2026 release lets users switch web access off entirely and have the agent work only from custom data supplied through remote MCP servers or uploads, and names FactSet, S&P Global and PitchBook as partners on those data connections. Source selection is becoming a product feature, which means being on a user's preferred list matters as much as ranking.
What does deep research mean for publisher traffic and revenue?
It deepens a decline already under way. Chartbeat data reported by Press Gazette shows Google organic search referrals fell by about a third globally, and 38 per cent in the US, in the year to November 2025. The Reuters Institute's Journalism, Media, and Technology Trends and Predictions 2026 report, based on 280 media leaders in 51 countries, found respondents expect search referrals to fall around 43 per cent over the next three years.
Deep research adds a specific mechanism on top. Content is read in volume, yet the visit that would have carried an ad impression, a newsletter sign-up or a subscription prompt never happens, because the report is the destination. The cited URLs become an appendix rather than a journey. There is also a measurement problem: the page fetches arrive as server requests, often without a referrer, so they do not appear in browser-based analytics. Our explainer on why AI bots do not show up in Google Analytics covers why.
Can publishers see or block deep research agents?
Partly, and the picture differs by vendor.
Google's agent identifies itself. The Gemini-Deep-Research user agent is operated by Google and described by the Known Agents directory as the agent that collects resources for Gemini's Deep Research feature. It is classed as user-triggered rather than a training crawler. The directory notes it is expected to follow robots.txt but has too little observed activity to report measured compliance.
OpenAI's position is less clear for this use case. Its bots documentation lists ChatGPT-User for certain user actions and states that, because those actions are user-initiated, robots.txt rules may not apply. The page does not name a separate agent for Deep Research, so publishers cannot assume a robots.txt rule will govern it.
Anthropic says its bots honour robots.txt, and that Claude-User may access websites when a person asks Claude a question. Blocking it stops that retrieval, and Anthropic notes this may reduce visibility in user-directed search.
The practical consequence is that robots.txt is a request, not an enforcement layer, and user-initiated agents are the category where it is weakest. For more on that limit, see our piece on whether robots.txt can stop AI crawlers. Verified-identity approaches such as Web Bot Auth are designed to close that gap, but adoption is partial.
Do deep research tools pay publishers?
Not by default. A citation in a report carries no payment. A few routes exist, each with limits.
Licensing deals pay some large publishers, but they are bilateral and out of reach for most. Perplexity runs a revenue-share programme through its Comet Plus subscription: it pays 80 per cent of the revenue to participating publishers, with payouts based on human visits, search citations and agent actions (Press Gazette, October 2025). It is one of the few schemes that counts citations rather than only clicks, but participation is by agreement and the total pool has not been published. Pay-per-crawl approaches such as Cloudflare's let publishers price access to content, but they depend on the agent being identified and willing to pay.
What can publishers do about deep research?
There are three realistic postures, and many publishers will combine them.
The first is to restrict: block specific agents in robots.txt, at the CDN or through verified-bot controls, accepting that this forfeits citation and any referral from those tools. The second is to license or join a programme where the economics are known. The third is to monetise the read itself. Because a deep research fetch is served at the CDN layer, that is where it can be identified, and where a revenue event can be attached to it without waiting for a human visit that is no longer coming. This is the approach blankspace takes: detecting agent retrievals at the edge and enabling contextual brand information to accompany the content in the AI response. Whether it suits a given publisher depends on their traffic mix and brand-safety requirements.
Whichever route is chosen, the first step is visibility. Publishers who cannot see which agents are reading which pages cannot price, restrict or license anything. Server logs are the starting point; our guide to finding AI crawler traffic in server logs walks through it.
Frequently asked questions
When did AI deep research tools launch?
Google's Gemini Deep Research launched in December 2024, and OpenAI's and Perplexity's versions followed in February 2025. Google released its Deep Research agent to developers in December 2025 and launched upgraded Deep Research and Deep Research Max agents on 21 April 2026. OpenAI moved its version to GPT-5.2 on 10 February 2026.
How many pages does a deep research query read?
It depends on the question and the product. No vendor publishes an official figure, and third-party estimates range from tens of pages to a couple of hundred for complex queries. The reliable point is that a session reads far more pages than a standard chatbot answer, and does so without sending a matching volume of visits.
Does deep research send referral traffic to publishers?
Rarely. Reports list sources, but the user commissioned the report to avoid reading them. Click-through is therefore low, and publishers should not expect citations to replace lost search referrals. Measure it in your own analytics rather than assuming.
Can publishers block deep research agents?
Sometimes. Google's Gemini-Deep-Research agent is expected to follow robots.txt, and Anthropic says its bots honour it. OpenAI says robots.txt rules may not apply to user-initiated ChatGPT-User actions. Blocking also removes the publisher from that tool's reports, so it is a trade-off rather than a free fix.
Is being cited in a deep research report worth anything?
Citation builds visibility and may influence a user's later choices, but it carries no payment on its own. Its value is strongest where the publisher also has a way to capture revenue from the read, such as a licensing arrangement, a programme like Perplexity's Comet Plus, or monetisation at the CDN layer.