The agent that reads thirty sources before it writes a word
For two years, AI search meant a fast answer from a handful of sources. In February 2025 that changed. A new class of tool — deep research agents — will spend minutes browsing dozens of pages and hand back a cited report. A thin page can slip into a quick answer. It does not survive a reader this patient.
the short answer
February 2025 brought deep research agents — OpenAI’s Deep Research, Perplexity’s, Grok’s, building on Gemini’s — that plan, browse dozens or hundreds of sources, and write a long cited report over minutes. Unlike a quick answer, they compare many sources carefully, so thin content that slipped into a short reply gets filtered out. Inclusion now means being a genuinely deep, credible source that holds up against a thorough comparison — earned authority, not surface optimisation.
key takeaways
- In February 2025 a new class of tool arrived: deep research agents (OpenAI Deep Research, Perplexity, Gemini, Grok) that plan, browse dozens or hundreds of sources, and write a long cited report.
- Unlike a quick answer (a few sources, seconds), a research agent reads far more and takes minutes — thin content that slipped into a short reply does not survive the comparison.
- OpenAI Deep Research scored 26.6% on Humanity’s Last Exam, far above prior models; reasoning over many sources took a real step up.
- The earned-authority implication: being included in a report means being a deep, credible source that holds up against a thorough comparison — not just being well formatted.
- Some tools are more susceptible to SEO bias than others, but as agents compare more sources, real depth and credibility win more ground than surface optimisation.
two ways a machine can read the web
In words, so the bars do not carry it alone: the two modes differ in how widely they read before they write. A quick answer takes seconds, consults a few sources, and returns a short synthesis. A deep research agent takes minutes, browses dozens of sources, and returns a long, cited report. The bar lengths stand for how many sources each consults — and the wider that comparison gets, the harder it is for a shallow page to be among the ones chosen.
Why this lands squarely on earned authority
There is a reason this shift favours authority over tactics, and it is worth being precise about. Optimisation works best against a reader in a hurry — an engine that skims, matches, and moves on can be nudged by format, keywords, and structure into citing a page that is not really the best on the topic. A deep research agent takes that hurry away. It reads widely and compares deliberately, which is exactly the condition under which tactics lose their grip and substance starts to dominate. You cannot format your way into being the deepest source in a field of thirty when something is actually reading all thirty. You can only be it. That is the definition of earned authority: a standing that comes from being genuinely better, which is slow to build and cannot be faked at the moment of comparison.
This is why the research agent is, quietly, good news for brands that have done the real work and bad news for those that have not. For years, surface optimisation let thin operators punch above their weight, ranking and sometimes getting cited despite shallow substance. A reader patient enough to compare many sources erodes that advantage. The brand that genuinely owns a question — that has the data, the depth, the demonstrable expertise — was always the one that deserved the citation; the research agent is the first reader systematic enough to reliably give it to them. The work, then, is the work it always should have been: become the source a careful analyst would choose, because increasingly the analyst is a machine that reads everything.
The research agent, in three parts
What a research agent does differently from a chatbot, why thin content does not survive its comparison, and what it rewards instead. Open each layer for the part that changes how you build.
01 What a research agent does differently
A normal AI answer is a sprint: it grabs a few sources and returns a synthesis in seconds. A deep research agent is an expedition. Given a single prompt, it plans an approach — some, like Gemini’s, even show you the plan first — then browses the web on its own, reading dozens or hundreds of pages, following leads, refining its strategy as it learns, and finally writing a long report with clickable citations. It takes minutes, not seconds, because it is doing what a human analyst would do over an afternoon, compressed. The arrival of these tools in February 2025 — OpenAI’s Deep Research, Perplexity’s, Grok’s, building on Google’s from December — was not a faster chatbot. It was a different mode of machine reading: wide, iterative, and patient, where the chatbot is narrow, single-pass, and fast. That difference is the whole story for how your content gets treated.
02 Why thin content does not survive it
A quick answer can be fooled by a thin page, because the engine only had time to glance at a few sources and yours was neatly formatted and easy to lift. A research agent removes that luck. When something reads thirty sources on the same question, the shallow ones are exposed by comparison: the agent has, sitting right beside your page, five others that go deeper, cite more, or simply get it right where you waved your hands. The agent’s whole job is to notice that difference and prefer the better source. So the page that earned a citation in a hurried answer can quietly lose it in a deliberate one — not because it got worse, but because it was always thin and now something is actually reading carefully enough to tell. Depth that did not matter when the reader skimmed becomes decisive when the reader compares.
03 What this rewards
If the agent behaves like a diligent analyst, it rewards what diligent analysts reward, and that is good news for anyone who has been building real authority rather than gaming a surface. It rewards depth: the page that fully answers the narrow question beats the one that broadly gestures at the topic. It rewards original data: a figure only you hold is something the agent cannot get from the other sources, which makes you structurally hard to omit. It rewards clear sourcing: claims an agent can verify are claims it will trust enough to repeat. And it rewards unambiguous credibility — being plainly identifiable as a real authority, with real expertise behind the words. None of this is a new trick. It is the old standard, finally enforced by a reader patient enough to check, which is exactly the kind of reader that rewards earned authority over optimisation.
A wrinkle: not every agent is equally hard to fool
Honesty requires a caveat, because the picture in early 2025 is not uniform. These tools differ in how carefully they read and how resistant they are to manipulation, and at least one of them was noted at launch as more susceptible to SEO-style bias than its peers — more willing to be swayed by content engineered to rank rather than content that is genuinely best. So the clean story of "the patient reader rewards substance" has exceptions: on a given tool, on a given day, an optimised-but-shallow page may still sneak through. The agents are new, they vary, and they will change.
But the exceptions do not reverse the direction, they just slow it. The whole value proposition of a deep research agent is thoroughness, and a tool that can be reliably fooled by surface tricks is a tool that fails at its one job; the competitive pressure is toward reading more carefully, not less. Betting your strategy on the agents staying easy to manipulate is betting against the entire reason they exist. The durable move is to be the source that wins when the reader is careful — and to treat the cases where a shallow page still slips through as a shrinking loophole, not a strategy.
How to be the source an agent includes
If the agent is a diligent analyst, you make yourself includable the way you would earn a place in any serious analyst’s report. Own the narrow question completely: be the page that fully answers the specific thing, not the one that broadly surveys the topic, because in a field of many sources the specific and complete one is what adds value. Bring what the others cannot — original data, a figure or a finding the agent will not get from the twenty-nine pages beside yours, which makes omitting you a real loss to the report. Source your claims so the agent can verify and therefore trust them. And be unmistakably identifiable as a credible authority, so that when the agent weighs whether to lean on you, the answer is easy.
Keep one structural habit through all of it: clarity. An agent that cannot cleanly find and extract your contribution may pass over it even when it is the best, so the depth has to be legible — stated plainly, well organised, easy to lift the relevant piece from. But do not mistake the structure for the substance. The chatbot era let you win on format for a while; the research-agent era hands the advantage back to the brands whose authority is real, because the reader finally has the patience to tell the difference. That is the whole arc of this shift, and it points where the AC Group has pointed for 27 years: be the source worth citing, and the careful readers — human or machine — will find you.
The B2B case: when a buyer commissions a report on your category
For a B2B audience this shift is unusually concrete. A buyer evaluating vendors no longer has to read a dozen comparison pages themselves; they can ask a deep research agent to do it, and get back a cited report on the options in your category — strengths, trade-offs, who fits which situation. That report is built by an agent that read widely and compared deliberately, which means the vendors it names and characterises well are the ones with deep, consistent, credible presence across the sources it consulted, not the ones with the cleverest landing page. Being the brand that such a report describes accurately and favourably is downstream of being genuinely well-documented across your category’s sources — exactly the earned standing that survives a careful reading. The buyers most likely to use these tools are often the most considered ones, deciding on behalf of a team, which makes being the report’s trusted reference worth more than winning a quick click. The win is quieter and more durable: you do not get a visit you can count, you get characterised — accurately or not — inside a document that shapes a shortlist you never see. That is a reason to care while the tools are new and few categories are well covered.
Deep research agents: quick answers
Is this different from AI Overviews or a chatbot answer?
Yes, in degree and in kind, and the difference matters for how you get included. A chatbot answer or an AI Overview is fast and shallow by design: it pulls from a handful of sources and returns a short synthesis in seconds. A deep research agent is slow and broad on purpose: given one prompt it plans an investigation, browses dozens or even hundreds of sources on its own, iterates as it learns, and returns a long, cited report minutes later. The quick answer rewards being easy to lift in a sentence; the research agent rewards being worth including after a thorough comparison against many alternatives. A page that wins the first by being neatly formatted can still be passed over by the second if a deeper, more credible source covers the same ground better. They are different games played on the same content, and the research agent is the harder one to fool.
How do I get cited in a deep research report?
By being the source that holds up when an agent reads you alongside thirty others on the same topic. A research agent is, in effect, running a comparison you do not control: it gathers many sources, weighs them, and includes the ones that add something credible the others lack. So the path to inclusion is to be genuinely better on the specific question — deeper, more specific, better sourced — not merely present. Original data helps enormously here, because a figure only you have is something the agent cannot get from the other twenty-nine sources, which makes you hard to leave out. Clear structure helps the agent find and lift your contribution. But the core is substance: a thin page that would slip into a quick answer gets filtered out when an agent has the time and the sources to notice that someone else said it better. Earn the inclusion by being the best source on the narrow thing, not the most optimised page on the broad one.
Do these agents change my SEO?
They do not replace it, but they do add a harder test on top of it, and they reward different things at the margin. Classic SEO got you found and ranked; that still matters, because an agent often starts from the same web your search rankings live on, and being discoverable is the price of entry. What the agents add is a depth test that ranking alone never imposed: being a top result is not the same as being worth citing in a report, and an agent that reads your page in full against rivals can tell the difference between a page optimised to rank and a page that genuinely answers. Interestingly, some of these tools were noted early on as more susceptible to SEO-style manipulation than others, so the picture is not uniform. But the direction is clear: as agents read more sources more carefully, surface optimisation buys less and substantive authority buys more. Keep the SEO; do not expect it to carry you through a careful reader.
Should I optimise differently for research agents?
Mostly you should optimise more honestly, not differently — the agents reward the things good content was always supposed to have. The temptation with every new surface is to invent a new tactic for it, but a deep research agent is closest to a diligent human analyst, and what persuades a diligent analyst is depth, accuracy, clear sourcing and genuine expertise. So the move is not a separate "research agent optimisation" programme; it is to make your best pages actually authoritative — own a question completely, bring data others lack, cite your own sources, and be unambiguous about who you are and why you are credible. The one structural habit worth keeping is clarity, so the agent can find and extract your contribution cleanly. Beyond that, the agents are quietly enforcing the standard the industry has paid lip service to for years: be the real authority, not the best-optimised imitation of one.
A note on sources and certainty
This is written in February 2025, the month deep research agents arrived in force: OpenAI launched Deep Research on the third, Perplexity followed on the fourteenth, Grok shipped its own that month, and Google’s had been out since December. We have described what they were at launch — agents that plan, browse many sources, and write a cited report, with OpenAI’s scoring 26.6 percent on a hard expert-reasoning benchmark, far above prior models. We have not borrowed the capabilities they gained later: the wider access, the visual browsing, the upgraded backbones all came after this date, and the tools were rough and varied at the start. What is durable, and what this note is really about, is the shift in kind: a reader that compares many sources carefully rewards depth and credibility over surface optimisation. The AC Group has spent 27 years arguing that authority is earned by being verifiably the best source — and the research agent is the first reader systematic enough to keep proving us right.