One Prompt, One Site Audit: Deep Research Done Right

A deep research site audit prompt is a different artifact from a chatbot question. Every major vendor now ships a deep-research mode — OpenAI, Anthropic, Google, Moonshot — and each one runs a multi-step agentic loop that plans sub-questions, issues dozens of searches, reads sources, and synthesizes a structured, cited report over minutes rather than seconds. Point that loop at a website instead of a research question and you get something close to a first-pass consultant's audit — if the prompt is scoped correctly.

That conditional is the whole story. In our own audit work we have found that an unscoped ask — "audit this website" — tends to collapse into a narrow technical-SEO checklist, because that is the shape of query these tools most resemble from their default examples. The fix is not a better model. It is a better prompt: one that explicitly widens scope past SEO, names four distinct ask-categories, demands URL-level evidence, and forces the agent to admit what it cannot measure.

This guide covers what deep-research modes actually are (with each vendor's documented operating envelope), the exact prompt structure we use, why explicit scoping should change the output shape, the blind spots no crawl can overcome, and how to read the resulting report critically. Every worked example uses an invented composite services website — no real audited company appears anywhere in this post.

A standard chat completion answers directly from one forward pass plus whatever a single search round returns. A deep-research run is a compound process. OpenAI's Deep Research API documentation describes a model that autonomously plans sub-questions, uses tools like web search and code execution, and produces a final structured response — with tasks that can take several minutes and a recommended background mode so clients don't time out. A 2025 academic survey of the field (arXiv: 2506.18096) defines these agents by dynamic reasoning, adaptive long-horizon planning, multi-hop retrieval, iterative tool use, and structured analytical reports — explicitly contrasting that compound process with static single-pass generation.

The performance gap is measurable, with caveats. OpenAI's own launch announcement reported 26.6% accuracy on Humanity's Last Exam — a hard, multi-hop research benchmark — versus roughly 10% for the best single-pass frontier models at the time (a vendor-run benchmark from early 2025, so treat the exact figures as OpenAI-reported). Anthropic's engineering write-up on its multi-agent research system reports that an orchestrated lead-plus-subagents setup outperformed a single-agent baseline by 90.2% on Anthropic's internal research eval, and traced 80% of the performance variance in the BrowseComp eval to token and effort spent. Those are internal, vendor-reported evals on general research tasks — not audit-specific tests — but the mechanism they quantify (dynamic follow-up, cross-source verification, adaptive planning) is exactly the mechanism a whole-site audit needs.

That sentence, from Anthropic's Research feature announcement, is the cleanest vendor statement of why this matters for audits. A site audit is inherently multi-hop: you cannot judge a services page without also reading its siblings, the navigation that leads to it, and the competitors ranking beside it. A single chat turn structurally cannot do that follow-the-thread work. A deep-research loop is built for it.

The four major deep-research implementations differ in architecture, documented scale, and — critically for audit work — whether a human checkpoint sits between the prompt and the crawl. The table below compiles what each vendor's own documentation states, at the time of writing. It is a spec snapshot, not a ranking: the leading public benchmark we reviewed scored these tools on 2025-era model snapshots that have all since been superseded, and the models behind them change frequently.

Two cells deserve emphasis. First, Google's numbers are the most concrete operating envelope any vendor publishes: per the Gemini API documentation, a standard run issues an estimated ~80 search queries and a Max run ~160 — which tells you roughly how many pages of a site a single audit run can genuinely visit. Second, Kimi-Researcher represents a genuinely different design philosophy — a single RL-trained trajectory rather than an orchestrated workflow — a useful reminder that "deep research" is a product category, not one architecture.

Cost tracks the loop, not the answer. These modes are metered on tokens, and a run that reads dozens of sources consumes multiples of a chat turn — Anthropic reports roughly 4× for single-agent research and ~15× for multi-agent. Independent testing in mid-2025 found real per-call costs for OpenAI's heavier deep-research model running into the tens of dollars on some queries — a dated, third-party figure we cite as historical context only; check the live pricing pages before budgeting. For a fuller product-level walkthrough of Google's flow, see our playbook for Deep Research Max mode — that post covers the product workflow; this one covers the prompt-and-deliverable method.

Here is the method as we practice it. The audit prompt has two load-bearing parts: a scope-widening opener that licenses the agent to look beyond SEO, and four explicit ask-categories that structure what comes back. Everything else — evidence requirements, ranking, the closing top-ten list — hardens the deliverable. The same principle drives our argument that the brief is the product: the specification you write is the single biggest lever on what an agent produces.

Three details matter more than they look. The opener — "act as a senior website consultant, not an SEO tool" — reframes the agent's role before any category is named. The evidence clause ("name the exact URL(s), quote the evidence you saw") converts vague observations into checkable claims, which is what makes the report verifiable at all. And the forced admission — "state clearly that impact is your estimate, because you cannot see analytics or revenue data" — bakes the tool's biggest limitation into its own deliverable, so nobody downstream mistakes a reasoned guess for a measured priority. If the audit surfaces structural findings — a slate of missing pages, weak topic grouping — pair the report with topic-cluster architecture when planning the rebuild.

Be clear about the epistemic status of this claim: it describes our practice, and it cannot be independently benchmarked, because it comes from private audit work. But the reasoning is straightforward, and two public facts support it.

First, the reasoning. None of the four vendors' documentation reviewed for this piece offers audit-specific prompting guidance — every documented example is a research or fact-finding task. A tool whose defaults are shaped by research-style queries, pointed at a website, will gravitate toward the website-shaped query it knows best: a technical-SEO inspection. That is not a capability failure; the same agent reasons perfectly well about form friction or confusing navigation when asked. It is a scoping failure. The instruction "you may and should also cover UX, intuitiveness, and information flow" exists to overwrite the default frame before it sets.

Second, the supporting evidence. Google's app-level Deep Research flow makes planning a mandatory, user-visible step: the agent writes out its research plan and waits for you to confirm or edit it before any crawl runs. That design choice is a vendor telling you, in product form, that what the agent decides to investigate is determined up front and is worth a human checkpoint — in other words, that scope set at prompt time shapes everything downstream. Anthropic's finding that token and effort spent explains 80% of performance variance in its BrowseComp eval points the same direction: where the agent spends its effort budget is the outcome — and the prompt is what allocates it.

This is the section to hand to anyone who wants to treat an AI-generated audit as final. The limits below are not model weaknesses that the next release fixes — they are consequences of how crawling works.

Start with the crawl itself. Reporting from SALT.agency documents that AI crawlers including GPTBot, Claude-Web, and PerplexityBot do not execute JavaScript — they read the initial server-delivered HTML only. On a client-rendered site, content injected in the browser is invisible to most AI agents outright. Sitemaps have their own mechanics: per Search Engine Land's sitemap guide, the protocol caps a single file at 50,000 URLs / 50MB; standard sitemap best-practice guidance separately excludes noindex pages, error URLs, and orphan (unlinked) pages. A sitemap entry is a claim by the site owner about what exists — not a validated snapshot of what the page contains.

Then the part that follows by construction, which we present as reasoned analysis rather than a sourced quote: your analytics, CRM and deal data, revenue per page, internal search logs, and support-ticket volume all live in systems a public crawl has no access route into. They are login-gated, unlinked from any public URL, or not web-exposed at all. Professional information architecture practice also relies on live-participant methods — card sorting and tree testing — that reveal users' actual mental models; a crawl of page content cannot substitute for either. The table below makes the boundary concrete.

One practical consequence: when the agent flags a category hub as weak, it is judging the page's content and links — it cannot know whether that hub converts. That diagnosis still tends to be directionally useful; we go deeper on the pattern in our diagnostic on thin category-hub pages. And for the mechanical layer the crawl does handle well, run a full technical SEO audit checklist alongside the judgment-heavy deliverable this method produces.

How do you tell a genuine audit from a plausible-sounding one? The research community's own numbers are the best starting point. DeepResearch Bench — a 2025 preprint benchmark of 100 PhD-level research tasks across 22 fields, written by domain experts — scores deep-research agents on two axes: report quality (comprehensiveness, depth, instruction-following, readability) and citation trustworthiness. Its published results showed a meaningful spread across vendors at the time — evidence that deep-research modes are not interchangeable — though every model in that snapshot has since been superseded, so the specific standings are history, not guidance.

Two of its findings transfer directly to audit work. The benchmark scores citation volume and citation accuracy as separate axes — an agent can produce many citations or correct ones, and the two do not automatically travel together. And its automated scoring framework reached 72.56% overall alignment with expert human judgment across the benchmark's tasks — a direct admission that even the best current automated evaluation of research quality only roughly agrees with a human expert. If automated judges cannot fully grade these reports, neither can you skip the human review of one.

That line, from Infomineo's analysis of AI errors in consulting deliverables, describes the exact failure mode an unverified AI audit invites — the output reads confident whether or not it is right. The same piece adds a subtler warning: "As consultants invest time and effort in prompting, refining, and shaping AI-generated output, the result increasingly feels validated through iteration." It isn't. Polishing a report is not the same as checking it. Frontier models still hallucinate at rates that are nowhere near zero, concentrated in exactly the long-horizon agent tasks an audit run is — our benchmarking of documented 2026 hallucination rates covers the numbers. Practically: spot-check a sample of the report's URL-level claims against the live site, verify quoted evidence is real text, and treat every unverifiable claim as unwritten.

One more lens: the audit's UX commentary should map onto recognized practice, not invented criteria. Nielsen Norman Group's ten usability heuristics — visibility of system status, error prevention, recognition over recall, and the rest — remain the standard expert-review checklist, and NN/g is explicit that they are broad rules of thumb rather than mechanical pass/fail tests. If the agent's UX findings rhyme with those heuristics, it is reasoning inside established practice; if they don't, ask why.

The deliverable this method produces is a first-pass consultant's report: cross-category findings with URL-level evidence and an explicitly estimated ranking. Getting from there to a roadmap is a data-join exercise. Take the agent's top-ten list, attach your analytics (does the flagged page actually get traffic?), your CRM (do the leads it should produce actually convert?), and your internal search and support logs (do users actually report this friction?) — then re-rank. In our experience the agent's list and the data-informed list overlap substantially but never perfectly, and the differences are where the real prioritization judgment lives.

The one-prompt audit is also one instrument among several — the skill is routing each job to the right one. This is the workflow we productize in our agentic SEO service; if you want the build-it-as-a-service angle instead, see packaging autonomous audits as a service.