When AI Agents Do the Research, Not Just the Answer

The One Thing

Getting quoted and getting used are not the same bar.

Perplexity just merged Deep Research into an agent that writes its own searches, reads dozens of pages across the open web, and hands back a finished report — not an answer with a citation. Almost nothing published today is written for that second bar.

Perplexity just merged two of its tools into one.

Ask it a question now, and it doesn't just search and answer. It now:

1. Plans a research project

2. Runs the equivalent of thousands of searches against it

3. Hands you back a finished report, deck, or dashboard

No second step. No copying an answer into another tool to make it useful.

This is the kind of behavior that's changing how “getting found” works.

For two years, the goal has been getting cited” — earning the sentence an AI assistant quotes when it answers a question.

That's still real, and it still matters.

But watch what Perplexity just built: an agent that doesn't stop at answering. It goes and builds the thing the person actually wanted — the comparison, the report, the shortlist — by reading dozens of pages across the open web and deciding, one by one, what's usable and what gets left out.

Getting quoted in an answer and getting pulled into someone else's finished work product are not the same bar. Almost nothing published today is written for the second one.

Key Takeaways

• Perplexity's Computer now runs Deep Research end-to-end — one query handles research, analysis, and deliverable generation, up from a two-tool process (Perplexity, 2026).
• The mechanism is “Search as Code.” The model writes and runs its own retrieval program — the equivalent of thousands of parallel searches — across the open web and a user's own files.
• 78% of B2B software buyers already use Deep-Research-style tools somewhere in vendor research; 41% do so regularly (G2, 2026).
• Single-vendor comparison tasks succeed 87% of the time in one panel study — but 54% of users still trust manual search more than agentic results.
• Getting cited and getting used inside a synthesized report are different bars, and almost no content is written for the second one yet.

What actually changed with Perplexity Computer?

Here's the news, straight from Perplexity's own account of it:

Deep Research — the tool that runs multi-step research and writes a cited summary — is now built directly into Computer, the system that turns research into a finished deliverable.

Before this, you ran Deep Research in one place, then moved your findings into Computer to build something from them.

Now one query does both - and the big thing is that most don’t realize it’s happening under the hood every time someone puts a question into Perplexity.

The mechanism behind it is called Search as Code.

Instead of running one search and reading the results, the model writes a program that designs and runs its own retrieval — the rough equivalent of thousands of targeted searches, executed in parallel, adjusted on the fly when something comes up short.

It pulls from the open web and a person's own connected files at the same time, filters and removes duplicates before anything reaches the model doing the writing, and comes back with a PDF, a deck, a dashboard, or a working document.

How Search as Code works

1

Writes its own search code, instead of running one query

2

Runs the open web and your own files together

3

Filters and dedupes before anything reaches analysis

4

Produces a finished deliverable, not an answer

Perplexity says accuracy, depth, and citation quality all went up once Deep Research moved into Computer, measured against three benchmarks built for exactly this kind of task.

One outlet covering the release reported the specific jump on one of those benchmarks — a measure of agentic browsing skill — from 40.7% to 83.8%. That's a different tier of capability running against your content today.

Is this different from what I already told you about OpenAI Presence?

Yes, as you can imagine in AI, things change constantly. One company releases a tool/feature that does one thing, and the other outdoes it.

OpenAI Presence, which I wrote about a few weeks back, builds agents that speak for a company using only that company's own private material — its docs, its policies, its internal systems.

It was never going to read your blog, and I said so at the time: that agent doesn't reach outside its own walls.

Perplexity's Computer does the opposite. It goes out into the open web on someone else's behalf — a stranger, a prospective buyer, someone with no relationship to you — and decides what's worth pulling into their project.

One agent never leaves the building. The other one is walking through every open door on the internet, including yours, deciding what to bring back.

If Presence was about your content feeding your own machine, this is about your content surviving someone else's.

B2B buyers are already the audience for this

G2 surveyed 1,076 B2B software buyers in March of this year and found that seventy-eight percent already use a Deep-Research-style tool somewhere in how they evaluate vendors.

Forty-one percent do it regularly, not as a novelty.

78%

of B2B software buyers already use a Deep-Research-style tool somewhere in vendor research

41% do it regularly, not as a novelty (G2, 2026 — 1,076 buyers surveyed). This isn't a someday audience.

Picture what that means for you.

Someone evaluating options in your category doesn't open ten tabs and read your homepage anymore.

They ask an agent to build them a comparison, and the agent goes and reads your homepage, your pricing page, your case studies, and your competitors' equivalents, then writes its own version of the comparison you never got to make yourself.

So you're not losing a click. You're losing the pen.

What makes content survive being used, not just cited?

Getting quoted requires one good, clean sentence.

Getting used inside someone else's report requires something sturdier: your content has to hold up when it's pulled out, compared against three competitors, and represented in someone else's table.

Cited vs. used

Same page. Two different readers, two different bars to clear.

Written to get cited

Needs

One strong, quotable sentence

Reads well to

A human skimming an answer

Positioning

Stated once, in prose

Survives a quote, not a comparison

Written to survive a task

Needs

Claims traceable to a named source

Reads well to

An agent building a comparison table

Positioning

Consistent everywhere it appears

Survives being pulled apart and reused

A study tracking a rolling panel of over eight thousand users through last year found that agents succeed at single-vendor comparison tasks eighty-seven percent of the time.

Now, it’s worth flagging that this is one vendor-commissioned panel, not a peer-reviewed result, so treat the number as directional, not gospel.

Within that caveat, the pattern it found lines up with common sense: agents do best with informational tasks, worst with anything requiring creative synthesis, and depend heavily on whether a claim traces back to something verifiable.

Content with clear, traceable, comparison-ready structure gets used correctly.

Content that only reads well to a human — vague claims, no sourcing, positioning buried in a paragraph — gets skipped, flattened, or misrepresented.

The honest caveat with all of this

Here's the thing to work on before you get too comfortable with any of this.

The trust gap

The same panel study found that 54% of users still trusted a manual search more than the agent’s answer — even when the agent finished the task. High completion doesn’t equal high trust, and it shouldn’t.

That same study found that 54% of people using these agentic tools still trusted a manual search more than the agent's answer, even when the agent finished the task.

High completion doesn't equal high trust, and it shouldn't.

An agent doing a good job of assembling a report is not the same as the report being right, and until someone checks it, it's still just a fast guess with good formatting.

That's the argument for making sure a human is the one who reviews what an agent builds before it becomes someone's decision — the same case I've been making about content generally, aimed one level higher up the chain now.

What to do about it this quarter

Four moves, none of them exotic.

1. Build real, authentic comparison content — a real table with real criteria, not three paragraphs describing why you're different. An agent building a comparison for someone else favors pages already shaped like the thing it's trying to produce.

2. Make every claim traceable. A number with a named source survives being checked. An adjective doesn't.

3. Keep one version of your pricing, your positioning, and your process, everywhere they appear. An agent pulling from three pages that disagree will pick one, and you won't get a vote in which.

4. Run the test yourself. Ask a research agent to compare your category the way a buyer would, and read what it built. That's the fastest way to see whether you're the source it used or the one it worked around.

You’re not losing a click.
You’re losing the pen.

— Brad Bartlett

None of this replaces the discipline behind getting cited in the first place — it sits on top of it.

The content that gets quoted correctly and the content that gets used correctly inside someone else's report turn out to be the same content, built the same way, just judged by a stricter reader.

Citation was the first bar your content had to clear.

This is the next one, and it rewards the same instinct — say something specific, prove it, keep your story straight — just tested by a machine that's no longer satisfied with quoting you. It wants to use you.

Work With Me

Would your content survive being pulled into someone else's comparison?

My GEO Visibility Audit scores your content for exactly that — across ChatGPT, Perplexity, and Google’s AI Overviews — and hands you a prioritized fix list, specific enough to give a writer.

Ten business days. Fixed price. No sales call to get started. The full fee credits toward any project after.

Book a GEO Visibility Audit →

Fiverr Pro vetted  ·  4.9 stars  ·  1,600+ client reviews

Sound human. Get cited.  ·  Made with 💙 in kcmo

---

Frequently Asked Questions

What is Perplexity Computer?

Perplexity Computer is a system that coordinates multiple AI models to research a question, analyze findings, and produce a finished deliverable — a report, deck, dashboard, or document — from a single query, using both open-web search and a user's connected files.

Is this different from Perplexity Deep Research?

Deep Research is now built into Computer rather than being a separate step. Previously a user ran Deep Research in one place, then moved to Computer to build a deliverable from the findings; now one query handles both.

Does this affect ChatGPT and Google too, or just Perplexity?

Perplexity is the most publicly documented example right now, but OpenAI, Google, and Anthropic are all building toward agents that act on the open web on a user's behalf, not just answer a single query. The underlying shift — content needs to survive being used inside a task, not just quoted — applies regardless of which agent runs it.

How is writing for an agent's task different from writing to get cited?

Getting cited means an engine judged one claim quotable. Getting used inside a task means an agent pulled your page into a multi-step process — a comparison, a report, a shortlist — and represented you correctly alongside other sources. That requires traceable, comparison-friendly structure, not just a good quotable sentence.

Does a high task-completion rate mean the result is trustworthy?

Not automatically. One 2026 panel study found agents completed single-vendor comparison tasks 87% of the time, but 54% of users still trusted manual search results more than the agent's. Completion and trust are different measurements — a good reason to keep a human checking the output, not just the process.