How to Run Deep Research With Cited Sources

Quick answer: Ask a research question in plain words, and Searchbase plans the investigation, searches, reads the pages and returns a written report where every claim carries a link to its source. It costs 20 to 40 credits and usually finishes in under 60 seconds. Describe your ideal customer. Searchbase finds them across Europe in one prompt, with the source on every cell.

There is a category of question that a table cannot answer. "Who are the main players in modular kitchen furniture in Italy and how are they positioned" is not a list, it is a picture. Somebody used to spend two days assembling it from search results, competitor sites, trade press and company filings, and the output was a document nobody could check because the sources lived in that person's browser history.

Deep Research is that job, done in a minute, with the sources attached to the sentences.

Step 1. Ask a question, not a keyword

The prompt should read like a brief you would give a junior analyst:

"Research the Italian modular kitchen furniture market. Who are the main manufacturers, how do they position themselves, what price segments do they compete in, which of them are advertising right now, and what has changed in the last year? Cite everything."

Compare that with "Italian kitchen furniture market". The second is a search query and returns a summary of what search engines already surface. The first names the dimensions you care about, and those dimensions become the structure of the report.

Three things make a question good here. Name the geography, because European market research is exactly where generic sources are thinnest. Name the dimensions, so the report has sections you can use rather than a wandering essay. And set the recency, because "what has changed in the last year" forces the agent to look for dated evidence instead of leaning on whatever is best indexed.

Step 2. Read the plan

Before it fetches anything, the agent writes out what it intends to investigate: the sub-questions it will answer, the kinds of sources it will look for, and the order.

Read it. A report is more expensive than a page read and the plan is free, so this is the highest-leverage twenty seconds in the process. If the plan is going to lean entirely on English-language trade press for an Italian market, say "prioritise Italian sources and company sites". If a dimension you care about is missing, add it now.

Step 3. Let it search, read and cross-check

Now it runs, and the sequence is the same one a good analyst uses. It searches, in the local language and in English. It collects candidate sources and opens them: company sites, trade press, public registries, product pages, and where it is relevant, the ad libraries and social profiles. It reads rather than skims headlines. Then it cross-checks, which is the step that separates research from summarising, because a claim that appears on one site and is contradicted on two others should not arrive in your report as a fact.

You can watch this happen. The sources being read appear as they are opened, so by the time the report lands you already have a sense of what it is built on.

Step 4. Read the report with the citations

The output is a written report, sectioned along the dimensions in your question, and every claim carries a link to the page it came from.

Read it by clicking. Pick the three claims you would repeat in a meeting and open their sources. If a claim's source is a competitor's own marketing page, you now know that it is a positioning statement rather than an independent fact, and you can present it that way. That judgement is only available to you because the link is there.

Claims that could not be supported are absent or marked as unverified. The report does not fill a gap with a confident sentence, which is the single most useful difference between this and asking a chatbot the same question.

Step 5. Ask follow-up questions in the same thread

The research is a conversation, not a one-shot document:

"Go deeper on the two brands in the mid price segment. What do their retail networks look like and how are they advertising?"

Follow-ups are usually much cheaper than the first report, because the agent already has the context and is investigating a narrower question.

Step 6. Turn the report into a table

The strongest pattern is the two-step. Research first to understand the shape of the market, then extract:

"Now give me a table of all the manufacturers you found, with company name, website, price segment, region, employee count if public, and whether they are currently advertising on Meta."

The report tells you what to ask for. The table gives you the rows to act on. Doing it in that order produces far better columns than guessing the schema before you understood the market.

What you get back

A written report with sections matching the dimensions you asked about, and a link on every claim.

The same provenance rules apply as everywhere else. Each cited value carries the source page, a freshness badge saying when it was read, and the extraction method. Where the research produces structured data, it arrives as a table with the source on every cell.

Values that cannot be verified stay n/a and claims that cannot be supported do not get made. Private company revenue is the recurring example: for most European private manufacturers, it is not public, so it is not in your report. A report that hands you a revenue figure for a private Italian company either found a filing, in which case there is a link, or invented it.

Export is JSON on every plan, CSV from Starter, the API from Pro and webhooks on Business.

What it costs in credits

A Deep Research report costs 20 to 40 credits. A page read costs about 1 credit and a social profile about 3, which is what the follow-up extraction steps are priced at.

So a research question plus a follow-up plus a 40-row extraction table is roughly 40 for the report, 15 for the follow-up and 40 for the table, call it 95 credits for a piece of work that used to be a two-day assignment.

Free gives 150 credits a month, which is three or four reports, and is a fair test of whether the output is good enough for your market. Starter is €29 a month for 1,000 credits, which is around 25 to 50 reports plus the extraction work around them. Pro is €79 for 4,000 credits with API access. Business is €199 for 12,000 credits with webhooks.

For comparison, one report costs less than a coffee and lands in under a minute. The relevant benchmark is not another software subscription, it is the hourly cost of the person who used to do this.

Limits and honesty

Public sources only. Nothing behind a login, nothing behind a paywall, no private databases, no subscriber-only trade press. If the definitive source for your question is a paid industry report, this will not open it, and it will not paraphrase it from secondhand summaries and present that as the source.

No private financials. Revenue, margins and headcount for private companies are usually not published in most European markets. Those cells stay blank.

It reports what sources say, and marks who said it. A claim from a company's own site is a claim from a company's own site. The report cites it as such rather than laundering marketing copy into fact.

Under 60 seconds is typical, not guaranteed. A broad question across several markets and languages takes longer, because there is more to read.

It is research, not a decision. The report gives you evidence and provenance. Weighing it is your job, and the citations are there precisely so that you can.

Outreach is drafted, never sent. If the research ends in a list of companies to contact, the agent writes the drafts and you review and send them.

Frequently asked questions

How is this different from asking a chatbot the same question? Two things. It goes and reads the live web for this specific question rather than answering from what it already absorbed, and it links every claim to the page behind it. Unsupported claims are left out instead of being written confidently.

How long does a report take? Usually under a minute. Broader questions spanning several markets or languages take longer because there is more to read, and you can watch the sources being opened while it works.

What makes a question too broad? A question with no geography, no dimensions and no time frame. "The furniture market" produces something readable and useless. Name the country, the segments you care about and the period, and the report becomes usable.

Can it research a single company? Yes, and it is a good use. Products, positioning, public hiring, public advertising, public press, all cited. Private financials will still be blank if they are not published.

Does it work in Italian? Yes. The interface is English and Italian, and for a European market the agent searches local-language sources, which is generally where the useful material is.

Can I get the sources as a list? Every claim is linked in place, and you can ask for the source list as its own table with a freshness badge on each entry.

What if two sources disagree? The report says so and cites both. Disagreement between sources is information, and flattening it into one confident sentence would be the wrong answer.

Prices checked on 2026-09-08 on searchbase.org.