Search finds answers. Intelligence builds understanding.

What is the difference between search and intelligence?

Search answers a question at one moment; intelligence builds an understanding of how a subject moves over time. A search tool, even a sophisticated AI one, takes a question in and returns an answer that was true when it ran. Understanding comes from tracking a topic continuously: knowing what it looked like before, what is new, what is merely repeated, and what the change means for you.

The two are often treated as better and worse versions of the same thing. They are different modes, with different strengths, and the difference is about time.

What does a search answer actually capture?

A search answer captures a snapshot: what the retrieved sources said, read through one model, at the moment you asked. Nothing about that is a flaw. It is the contract of the format, and for many questions it is exactly what you need.

The limitation appears when the subject is moving. Consider a hypothetical policy analyst following a proposed rule at a federal regulator. She asks an AI search tool for the status of the rule on a Tuesday and gets a clear, cited summary: comment period open, two industry groups opposed, a final decision expected later in the year.

Three weeks later she asks again and gets another clear, cited summary. It is well written. It reads as current. What it does not tell her is which parts are new since her last question, because the tool has no idea she asked before or what she was told.

A snapshot also carries a timestamp you cannot see. The answer blends a page updated yesterday with an article from months ago, and the prose flattens both into the present tense.

Why can't two snapshots tell you what changed?

Two snapshots cannot reliably tell you what changed because the camera moves between them. When you compare two search answers about the same topic, the differences come from several places at once:

• The world changed. This is the only difference you care about.
• Retrieval changed. A different set of pages ranked for the query, so different facts were available to the model.
• Phrasing changed. You worded the question slightly differently, and the answer followed.
• The model changed. The tool was updated, or simply generated a different summary of the same material.

From the outside, these four are indistinguishable. A fact that disappears from the second answer may have stopped being true, or it may have dropped below the retrieval cutoff. A claim that appears for the first time may be news, or it may be an old claim that a newly indexed article repeated.

That last case matters most. Repetition masquerades as novelty in search, because a restatement published this week looks fresh to a system with no record of having seen the original. Telling "new" from "said again" requires a memory that a search query does not have.

What does building understanding over time require?

Building understanding over time requires holding the lens still while the world moves. In practice that means four things a one-shot answer never has to provide:

1. A baseline. A fixed statement of what was known at the start: the rule as proposed, the positions of the main parties, the expected timeline.
2. Deltas. Each new item is compared with the baseline and reported only as the difference: "the regulator extended the comment period by thirty days," not a fresh summary of everything.
3. Memory. The baseline updates as deltas accumulate, so the system knows that the extension was already reported and does not announce it again when a trade publication covers it a week later.
4. A record of no change. Days and weeks when nothing moved are kept as observations, not discarded as empty.

The fourth point is the least intuitive and often the most useful. In a trajectory, "nothing changed" is a measurement. If the regulator said it would decide by the end of the quarter and the quarter passes with no decision, the absence is the news. If a thesis depends on a competitor not entering a market, every month without an entry is evidence for it. A snapshot cannot report a non-event, because it has no expectation to compare against.

Researchers who evaluate monitoring systems have formalized this idea. The TREC Real-Time Summarization track, run by NIST, scored systems that pushed social media updates on topics users cared about. Its track overview describes "silent days", days with no relevant posts, and under some of its metrics a system earned a perfect score for those days only if it pushed nothing. Recognizing that nothing happened was treated as a skill, not a gap.

How is an AI research agent different from AI search?

An AI research agent differs from AI search in depth, and a continuous research agent differs from both in duration. The distinction is easiest to see across three levels.

• AI search interprets a question and writes a synthesized answer from retrieved pages. Google describes its AI Mode as breaking a question into subtopics and issuing many searches at once, a technique it calls query fan-out.
• Deep research agents run longer, multi-step investigations. OpenAI describes deep research in ChatGPT as an agentic mode that plans its research, searches, evaluates sources and returns a documented report with citations, usually after several minutes. Perplexity describes its Deep Research as performing dozens of searches and reading hundreds of sources before writing a report. Google's Deep Search in AI Mode follows the same pattern.
• Continuous research agents keep a question open after the first answer. They read a defined set of sources on a schedule and report changes against what they already know.

The first two levels are a real advance in depth. A deep research report can replace hours of manual reading. But depth and duration are separate axes. A report built from hundreds of sources is still a photograph, taken once. Running it again next month produces a second photograph, with all the comparison problems of any two snapshots.

Why is continuous research harder than one-shot AI search?

Continuous research is harder than one-shot AI search because each run depends on every run before it. A search can fail and be retried with no lasting harm. A continuous system carries its mistakes forward. The difficulties are specific:

• Choosing the baseline. Someone has to decide what counts as known at the start, and a vague baseline produces vague deltas.
• Judging novelty against history. Every item must be checked against what was already reported, including the same event described in different words by a different source.
• Tracking entities that drift. Products get renamed, companies merge, a proposal gets a new docket number. A system matching on names loses the thread.
• Absorbing source churn. Sites change layout, feeds break, accounts go quiet. A silent source and a quiet topic look the same unless the system checks.
• Containing compounding errors. One misread item that enters the baseline distorts every comparison after it.
• Paying for quiet days. The sources have to be read on every cycle, including the many cycles that produce nothing to report.

Information security learned this lesson long before AI research tools existed. NIST's guidance on information security continuous monitoring notes that an initial authorization to operate rests on evidence available at one point in time, while systems and their environments keep changing. Its answer is ongoing assessment, with the frequency of each check set deliberately. The same reasoning applies to any subject that moves after the first assessment.

When is search the right tool, and when is continuous research?

Search is the right tool when the answer is stable or needed once; continuous research is right when the subject moves and you will act on the movement. A useful test has two questions: will the answer be different next month, and would that difference change what you do?

Search fits when either answer is no:

• Learning the shape of an unfamiliar field before deciding whether to follow it.
• Checking a definition, a specification or a historical fact.
• Comparing options for a single decision, such as a one-time purchase.
• Retrieving a document you know exists.

Continuous research fits when both answers are yes:

• A pending decision by a regulator, a court or a standards body.
• A competitor whose pricing, hiring or product changes affect your plans.
• A research area where preprints arrive faster than you can read them.
• A company you hold, or one whose results move your own.

The two modes also work in sequence. A deep research report is a good way to write the initial baseline. Continuous research is what keeps that baseline true afterward. Teams comparing tools that monitor your market can use the same test to decide which of their questions belong to which mode.

What changes when research becomes continuous?

When research becomes continuous, the unit of output changes from the answer to the change. You stop asking "what is the status?" and start receiving "here is what moved since you last knew." The question you used to retype every few weeks becomes a standing one, and your attention goes to the deltas instead of to rereading the background.

It also changes what you can trust. A continuous record lets you see how a view was formed: which development came first, which claim was repeated, which expected event never arrived. Kindal is built on this model: you choose the sources, and it reads them and writes a brief only when something in them changes, with each line linked to where it came from.

Search will keep getting better at answering questions, and it should. Understanding asks for something search was never designed to hold: the memory of what was true before, and the patience to notice when it stops being true.