AI RESEARCH AGENT
Ask a Research Question. See the Evidence Behind the Answer.
Enter the question you want to investigate. nqzai prepares it as an unsent research draft for you to review after signup.
Illustrative result
What an evidence-led research draft can cover
- QuestionThe decision or comparison to investigate
- EvidenceSources and gaps to examine before answering
- ReviewA prepared draft you control before it runs
This is an example of the workflow shape, not a result from your input.
What an AI research agent actually does
Ask a normal chatbot why one page outranks another and it answers from memory. Ask a research agent and it goes and looks. The loop has four moves, and it repeats them until the question is answered rather than until a counter runs out:
- Plan. Break the question into what would actually settle it — what do those pages say, who is ranking now, what does our own history show.
- Gather, in parallel. Search the web and fetch pages at the same time rather than one after another.
- Read and re-plan. New evidence changes the next question. A page that turns out to be a pricing page, not a product page, sends the next search somewhere else.
- Answer — and declare the gaps. State what the evidence supports, and list separately what could not be checked.
Every page nqzai reads comes back in the same shape: title, meta description, heading tree, FAQ presence, internal and external link counts, word count, main text. That consistency is the point — two pages in an identical format can be diffed, whereas two free-text summaries can only be compared impressionistically.
Where it matches a frontier chat model — and where it does not
Being honest about this is more useful than claiming parity. On a pure open-web question, a frontier assistant is excellent and nqzai is comparable: same loop, same parallel tool calls, same iterative refinement. Three real differences:
- They see more of the web; nqzai sees more of you. A general assistant has no access to your Search Console, your audit history, or which of your contacts bounced. Those are the numbers that decide most marketing questions.
- nqzai declares its gaps structurally. Frontier answers tend to hedge in prose — “likely”, “hard to quantify without tools like Ahrefs” — while still sounding uniformly confident. nqzai renders an explicit list of what it could not verify, generated from the tools that actually failed.
- nqzai can act. The same agent that finds the problem can draft the outreach, fix the page brief, or run the audit. A chat assistant hands you a conclusion and stops.
The part a chat model cannot copy: your own measured history
This is the whole advantage, and it is not a model capability — it is a data position. nqzai has been measuring your domain, so when a question touches it the answer comes from records rather than inference:
- Search Console performance per page and per query — real impressions, clicks and positions, not a guess about what “probably” ranks.
- On-page and indexation audit snapshots with dates, so “this was fine last week” is checkable.
- AI-visibility history across answer engines, with the prompt set that produced it.
- Contact and campaign outcomes — who was verified, what bounced, what was replied to.
Because that evidence is stronger than one live fetch, nqzai answers questions about your pages from those measurements instead of re-reading the page. Competitor pages, which nobody has measured for you, get fetched live.
“What I could not check” — on purpose, every time
An AI answer built on four sources where one failed looks exactly like an answer built on four that worked. So nqzai renders the difference:
• example.com/pricing — this domain does not allow automated reading, so it could not be researched
• “competitor pricing 2026” — no results came back
Anything above that depends on these is unverified.
That block is generated from the tools that actually ran, not written by the model when it feels like mentioning it. A search that ran and returned nothing counts too — the question went unanswered even though nothing errored, and that is the commoner and quieter failure.
What it cannot do
Named plainly, because a capability page that lists only strengths is the thing this product exists to argue against.
- Sites that refuse automated readers stay unread. A 403 or a bot challenge is a hard boundary; nqzai names the domain rather than working around it.
- It does not see pages visually. No screenshots, no layout or design judgement — it reads structure and text.
- It will not fetch your own domain. Deliberate: the stored measurements are better evidence, and it says so instead of pretending it looked.
- Live rank positions for arbitrary domains are not tracked beyond what a search returns at that moment.
Worth knowing: JS-heavy sites are largely a non-issue. Of 13 marketing pages we tested across React and Next.js stacks, 12 returned full server-rendered content — unsurprising, since a page that cannot be read by a crawler cannot rank in the first place.
How it compares
| Tool | Reads live pages | Searches the web | Knows your measured history | Declares what it missed | Can act on it |
|---|---|---|---|---|---|
| nqzai | Yes | Yes | Yes — Search Console, audits, contacts, campaigns | Yes — rendered, not optional | Yes — drafts, audits, outreach |
| Frontier chat assistants | Yes | Yes | No | In prose, when it chooses to | No |
| AI-visibility monitoring tools | No | Scheduled prompts only | Their own metric only | n/a | Mostly no |
| Classic SEO suites | Crawls your site | Index data, not live | Your site’s crawl history | n/a | Recommendations only |
Comparison reflects nqzai’s own measured behaviour as of August 2026; verify competitor capabilities against their current documentation.
Ask it something you actually want to know
The agent decides for itself whether a question needs the web, your stored data, or both. Nothing to configure.
This page covers how answers get their evidence. The other half of the same layer — the eight tools that question a request before they spend your tokens — is where most of the money is saved.
New here? How nqzai works covers the agent, the token model, and what runs where.
Frequently asked questions
What is an AI research agent?
An AI research agent is an assistant that goes and looks things up mid-answer instead of replying from memory. It plans what it needs, calls tools — a web search, a page fetch, a database read — often several at once, reads what comes back, decides whether that settled the question, and looks again if it did not. That loop is the whole difference: the answer is assembled from evidence gathered during the conversation, not recalled from training data.
How is this different from asking ChatGPT, Claude, or Grok the same question?
For a question about the open web, not very — the loop is the same shape and those models run it well. The difference is what each side can see. A general assistant knows nothing about your Search Console history, which of your pages were audited last week, which contacts bounced, or how your last campaign performed. nqzai holds those measurements because it took them. So "why is my page losing to theirs" gets your real numbers on one side and a live read of their page on the other. A general assistant has to guess at your half, and it will guess fluently.
Does it actually read pages, or does it guess?
It fetches. A page read returns that page's real title, meta description, heading tree, FAQ presence, internal and external link counts, word count and main text — parsed from the response, never summarised by a second model that could invent a detail. Searches return live positions, titles and URLs. Every page comes back in the same shape, which is what makes two of them genuinely comparable.
What happens when a website blocks it?
It says so, in those words. Some sites refuse automated readers with a 403 or a bot challenge, and that is a permanent boundary on the analysis rather than a glitch worth retrying. The answer names the domain that refused and lists it under "What I could not check", so any conclusion leaning on that page is visibly unsupported instead of quietly weaker than it looks.
Why does it not read my own site the way it reads a competitor page?
Because it already has better evidence for your property. Your Search Console performance, on-page audit snapshots and AI-visibility history are on file — measured over time, not inferred from one HTML response. So questions about your own pages are answered from those measurements, and live fetching is reserved for pages nobody has measured for you.
What does a research question cost?
Measured on live runs: roughly 87,000 to 163,000 tokens for a full two-page competitive comparison, including the searches, the page reads and the reasoning. One web search is about 5,000 tokens. Signup includes 1 million tokens, so the free allowance covers several complete investigations before any top-up.
Can it be wrong?
Yes, and the design assumes it. Every answer that hit a wall ends with a rendered list of what could not be checked, written from what actually executed rather than from the model's own account of its work — it cannot leave that list out. Read any answer this way: the parts resting on the listed gaps are unverified, and everything else traces to something fetched.
Does it work for questions that are not about SEO?
The loop is general — search, read, compare, conclude — so it handles competitor teardowns, pricing checks, "what does this company actually do", and prospect research before outreach. It is strongest where the question touches data nqzai already holds for you: your outreach results, contact verification, rankings, and AI-answer visibility.