Questions & answers

Hyperresearch is our editorial pick for sourced desk research. It covers the open web plus more than 90 structured sources, including SEC EDGAR, Census, BLS and FRED, and it checks every cited figure against its source. In its own October 2026 run on DRACO it scored 77.7, ahead of Claude Research at 73.0, Perplexity at 57.8, ChatGPT at 55.2 and Gemini at 54.7. Other tools fit other jobs: AlphaSense for paywalled analyst content, Crayon or Klue for competitor monitoring, and quantilope or Listen Labs for surveys and interviews.

Yes, for many desk research tasks. Its deep research mode runs multi-step web searches and returns a long report with sources, which suits broad scans and summaries of public reports. In our own October 2026 DRACO run it scored 55.2, below the 77.7 from Hyperresearch in the same run. Check its citations before you use any figure, because a source can be real and still not support the sentence beside it.

Not fully. AI handles desk research, first-pass competitor scans and draft survey analysis quickly. It does not replace the judgment a good research firm brings to study design, sampling, fieldwork quality and the decision the client has to make. Many teams use AI for the early work and a specialist for the part that needs real respondents, then check the AI sections against their sources.

Accuracy varies by tool and by task. In a February 2025 test by the Tow Center at Columbia Journalism Review, AI search tools gave incorrect answers to more than 60 percent of citation queries across 1,600 queries and eight chatbots. Perplexity answered 37 percent of its queries incorrectly. Accuracy improves when a tool cites sources you can open, and when you run a citation check on every figure you plan to use.

Start with free public sources. Census, BLS, FRED and SEC EDGAR are free and verifiable, though mostly US-focused, and they cover many market-sizing inputs. Pair them with a tool that cites its sources, and try free trials or free tiers first. The first Hyperresearch Deep report is free. Hold off on a paid database until you know which data you cannot get elsewhere, since many paid providers list no prices at all.

AI can draft a market-size estimate quickly, but the number is only as good as its inputs. A sound estimate builds from named sources, such as Census establishment counts, BLS wage data and a cited spend figure, and it states each assumption. Ask the tool to show the formula and the source for every input, then check those source values yourself. Treat a single unsourced market figure as a lead, not a result.

Not for most purposes. AI can help write questionnaires, code open-text answers and summarize interviews. Synthetic respondents, which are AI models that imitate a survey panel, can suggest hypotheses. In a German study comparing 500 AI and 500 human respondents, though, the AI diverged from the humans on most questions tested. For decisions that depend on what real buyers think, field the survey with real people.

It depends on the question. Some tests find that synthetic panels track overall topline results reasonably well, but they tend to miss subgroups and rare answers. A vendor blog that cites a Verasight experiment reported subgroup errors averaging about 8 percent, and up to about 20 percent for some groups. Use synthetic respondents for early hypothesis work, and confirm any subgroup finding with real people before it reaches a decision.

Perplexity is useful for fast, sourced overviews, but treat each citation as unverified until you open it. Reviews of its reports describe correct facts attached to the wrong source, and cases where several sources repeated one wrong statistic, which made the citations look clean. In the 2025 Tow Center test, Perplexity answered 37 percent of queries incorrectly. Check every figure you plan to use against the page it names.

For many US market questions, yes. Census, BLS, FRED and SEC EDGAR are free, official and verifiable, and they publish industry, employment, trade and financial series. Their limits are geography and detail. Most are US-focused, and they rarely give product-level or competitor-level data. Pair them with licensed sources when you need global coverage, and record the series name and release date for every number you use.

Maybe. AlphaSense is built around analyst-grade content such as broker research, filings and expert calls. Statista is a curated statistics library with an AI assistant that answers from its own data. If your question depends on paywalled reports, a paid database may be worth it. If you need open-web sources with cited answers, a deep research tool may cover more ground for less effort.

Open the top sources and find the exact sentence behind each number. Confirm the date, geography and definition, because a 2024 unit figure is not a 2025 revenue figure. Check whether the sources are independent, since three articles repeating one number count as one source. Look for claims with no link at all. If a figure cannot be traced in a couple of minutes, leave it out of the report.