Deep research on every account in your market.
S1 agents are SalesOne’s deep research agents. Not a database, not a quick AI lookup: they search widely, read the sources, check each fact at its source with its date, judge fit with a reason for every reject, find who owns the problem and write a point of view. A second pass reviews the work.In our tests, database-sourced lists had about 1 in 4 clear misfits; S1 about 1 in 20.
Updated October 7, 2026. SalesOne research team.
What does a deep research agent do on one account?
Seven steps, in the order the work happens, with what each one measured across our production runs. S1 agents do the homework. Your team builds the relationship.
01
Search widely
Press wires, filings, trade press, sponsor news and the sources your segment names, newest events first.
31
searches, on average, for every account delivered
02
Read the sources
The pages that matter are opened and read, through one shared reader that paces each site and keeps to public pages.
7
pages read, on average, for every account delivered
03
Check each fact at its source, with its date
A buying signal counts only when it has a source link, is about this company, still holds, and is dated inside its window.
375 of 375
buying signals link their source; 98% are dated to the day, and the newest on a typical account was 59 days old at delivery
04
Judge fit, with a reason for every reject
Every rule in your profile is checked. A company turned down keeps the rule it failed and the evidence.
4
companies judged, on average, for each one kept; all 802 turned down keep their reason, 88% citing dated evidence
05
Find who owns the problem
Current people only, from leadership pages, filings and press releases. Anyone who has left is caught and held back.
4.2
people per account; 95% with a dated source, 85% from sources beyond professional-network profiles
06
Write a point of view
What changed, the problem it creates, who owns it and why now, in two sentences tied to the dated facts.
147 of 147
reviewed accounts carry a written reason to reach out
07
A second pass reviews the work
Before delivery, a review pass re-checks every buying signal against its source and looks for fresher news.
40
buying signals corrected, 58 fresher ones added and 3 accounts pulled, across 147 accounts
Then
The next run starts from this one
Companies already covered, every reject and its reason, and your ratings go into the next research brief.
From four production research runs, October 2026: 249 accounts delivered, 802 turned down and 1,057 people found. The point-of-view and review-pass figures come from the three runs that had a review pass (147 accounts). These describe what the research recorded; they were not blind-checked. How we measure S1 research.
One company, researched three ways
The same sample company through a contact database, a quick AI lookup and S1 deep research. The first two are faster and cheaper. Only one tells you why now, who owns it and where each fact came from.
A sample company, Halvorsen Castings, researched three ways. A contact database returns a record at once: industry, size, revenue, location, a contact who is no longer the COO and an intent topic, with no source or date for any field. A quick AI lookup reads three pages and writes a short general summary with one cited link and no dates. S1 deep research runs about 31 searches and reads 7 pages, records each fact with its source and date, turns down a neighboring company with the rule it failed, finds four people who own the problem, catches that the former COO has left, writes a point of view, and a second pass corrects one buying signal's date.
How do databases, AI lookups and deep research compare?
Twelve angles, by category rather than by product, between what each is good at and what it is not built for. Where an alternative does better, the table says so: databases on speed and cost per contact, general deep research on one hard question, persistent account agents on watching accounts over time.
| Angle | Contact database | Quick AI lookup | General deep research | Persistent account agents | S1 deep research agents |
|---|---|---|---|---|---|
| Good at | Contact details at scale, fast and cheap. | One well-defined field across thousands of rows, quickly. | One long, cited report on one hard question. | Watching the accounts you give it and noting what changes. | The step before: which companies in a market belong on the list, each with a dated reason, and a verdict for every one turned down. |
| Where candidate companies come from | Better here: Fastest to a long listFilters over records it already holds: industry, size, location, technology. | The list you upload or pull from a database. It researches the rows it is given. | Whatever the one question names. Not built to list every company in a market. | The accounts you give it, usually a named list or your CRM. | Starts from dated events in filings, press wires and trade press, and builds the list against your profile. |
| Depth on each account | A fixed record: company facts, technology and contacts. | One question per column. More depth means more columns, each billed. | Better here: Deepest on a single questionMany pages read for one long report on one question. | Research that builds up on each tracked account over time. | About 31 searches and 7 pages read for every account delivered, all on the same fields. |
| Checking each fact | Record-level checks, such as email validation. No source shown for each fact. | Varies. The best tools check each criterion with a cited link; most return free text. | Cited prose. Independent audits find that a share of citations do not support the claim. | Varies by maker. Some keep a reasoning trace for each decision. | Each buying signal keeps its source link, and a second pass re-checks every one before delivery. |
| Freshness and dates | Refreshed on the provider’s cycle. No date on each fact. | As fresh as the page read that day. The source date is rarely shown. | Current when it runs. Nothing is re-checked unless you ask again. | Kept current on the accounts it tracks. Dates on each fact vary by maker. | Every buying signal carries the date it happened; 98% are dated to the day. |
| Sensing what changed | Company-level topic intent and news alerts. Broad and low cost. | Only when you run the column again. Some newer agents track changes per account. | None. Each report is one run. | Better here: Best at watching accounts over timeWatches the accounts it is given between your requests. | Every account begins with a dated event inside its time window, and accounts whose window is open are flagged. Research runs when your team asks. |
| Qualification, with reasons | Filters and scores, without a written reason. | Possible if you write the rubric yourself. | An opinion in prose, not a verdict against your profile. | Works accounts already chosen, so turning companies down is rarely its job. | Judged against every rule in your profile. Every reject keeps the rule it failed and the evidence. |
| Buying committee | Better here: Widest raw contact coverageContacts by title and department across very large record counts. | Usually one person per prompt, such as the head of operations. | At least one maker limits compiling details about individual people. | Some identify stakeholders on each tracked account. | About 4 current people per account who own the problem; 95% carry a dated public source. |
| Point of view | None. Data only. | A personalized line per row. | A long report, not a reason to reach out. | Account summaries and suggested next steps. | A short point of view per account, tied to its dated facts. |
| Human approval | Not applicable: it supplies data. | Depends on the tool. Some send messages on their own. | Not applicable: it writes a report. | Depends on the tool. Some act on deals on their own. | A person approves every message before it goes. |
| Feedback into the next run | None for your account. | You edit the prompt. | None across runs. | Builds on what it learned about each account. | Your fit ratings, and the reasons behind them, go into the next research brief. |
| Memory of what was covered | The same records for every customer. | Rows in a table. Most runs start fresh. | None. Each report stands alone. | Memory for each account it tracks. | Remembers every company checked for you, matched by website and name, and never researches one twice. |
| Cost model and output | Better here: Lowest cost per contactSeats and credits per contact. Output: rows. | Better here: Low cost per rowCredits per row and per column. Output: cells. | Per report, or a plan with a monthly quota. Output: prose. | Platform plans; cost per account is rarely published. Output: updated account records. | Priced on qualified accounts; companies turned down are explained and not charged. Output: researched accounts. |
| Not built for | Saying which accounts fit and why now, with a source and date for each fact. | Qualifying against your profile, or keeping a reason for each company left out, unless someone builds it. | Every company in a market, in the same checked fields. | Deciding which companies in a market belong on the list. | Watching your accounts between runs, or a phone number in seconds. |
Categories, not named products, from the makers’ own documentation, October 2026. The citation audits are DeepResearch Bench (2025), DeepTRACE (2025) and the Columbia Journalism Review’s Tow Center study of AI search (2025). For named tools, see the best AI tools to research accounts.
How well does S1 research hold up?
Each figure comes with its method. Where a test was small or the other side won, it says so. The figures each production run records are in the steps above.
A blind test against a leading AI research API
Same client profile on both sides, the same known companies excluded. The records were mixed and anonymized, then checked against their sources by reviewers who did not know which side produced each one.
58 of 61
accounts from S1 agents were usable, against 27 of 39 from the API.
Usable accounts
S1 agents95%AI research API69%People per company, on a scale of 0 to 5
S1 agents3.9AI research API1.6People backed by a source beyond professional-network profiles
S1 agents83%AI research API30%Where the API won
Buying signals that held up when checked
S1 agents75%AI research API96%
The API was also faster, 5 to 11 minutes a run, and cheaper. A smaller test on our own profile, 10 accounts a side, tied on usable accounts.
Method: one B2B consulting client’s profile (US mid-market and enterprise manufacturers, financial services and insurance), October 3, 2026. 61 accounts from S1 agents against 39 from the API. The reviewers were AI checker agents working from the cited sources, not people, and they came from the same model family as the S1 agents. S1 had about an hour and 15 agents; the API ran at its default effort, about $1 a run. One profile, so read it as a direction, not a guarantee. Full method.
Against database-sourced lists
Databases are the fastest way to a long list. What they cannot do is say which companies fit and why now.
1 in 4 vs 1 in 20
is roughly how many records were clear misfits: database-sourced lists against S1 research.
20 months
since the median company record in one company-data API was last refreshed, across 499 records.
Internal tests, Sept 28 to Oct 3, 2026: 569 records from a contact database (70) and a company-data API (499), sorted on their returned data, against 71 S1-researched accounts checked blind against sources in two tests. Different samples and methods; the S1 sample is small. Method.
Why general deep research is not built for this
General deep research tools are excellent at one long report on one question. Researching every company in a market into the same checked fields is a different job, and an independent benchmark measures exactly that.
5.1%
of tasks fully completed by the best general agent on WideSearch, a benchmark of collecting every entity in a set into a table (the best agent tested; dedicated deep research products were left out because they returned reports instead of tables).
2.4%
average success for the best agent on DeepWideSearch, when finding each entry also took multi-step reasoning.
Sources: WideSearch, arXiv 2508.07999, August 2025 (opens in a new tab); DeepWideSearch, arXiv 2510.20168, October 2025 (opens in a new tab). Both measure general AI agents; neither tested S1.
SalesOne production research runs and internal tests, September and October 2026. Client name withheld. Every checked figure comes from AI checker agents working from cited sources; none is a human audit. How we measure S1 research.
Which AI models do S1 agents use?
S1 agents run on frontier AI models, chosen by blind testing, with SalesOne’s own orchestration and rules the agents cannot skip. In the blind test, a small model invented people and companies, so it is never used for research.
SalesOne orchestration
- A lead agent splits each request into assignments by segment and region.
- Up to 10 research agents work in parallel, each on its own accounts.
- A review pass re-checks every buying signal; a committee pass fills missing seats.
- Every agent works from fixed SalesOne research briefs plus your profile, playbooks and fit ratings.
Rules no agent can skip
- Before research starts, every company is checked against what was already researched for you and your exclusion lists; existing SalesOne clients are blocked again at publish.
- No account accepted or turned down without a written reason.
- A person counts as ready only with a dated source from the last 12 months.
- Public pages only: no logins, no getting past CAPTCHAs, no social or professional-network pages.
How does S1 know what changed?
S1 starts from what changed: every account begins with a dated event inside its time window, and accounts whose window is open are flagged.
Each buying signal has a response window, for example 90 days after a new executive starts. SalesOne puts accounts with an open window first and shows the date to reach out by.
Research runs when your team asks for accounts, and before outreach a fresh check looks for dated facts from the last 12 months. S1 does not watch your accounts between runs.
What does each run build on?
The companies already covered, every reject and its reason, and your team’s fit ratings, which go into the next research brief.
About 7,000
Companies remembered so far across SalesOne
Known companies, exclusions and existing clients are skipped, matched by website and by name, so nothing is researched twice for you, even under a different website.
20+ checked
Reject reasons that suggest a change
Once a target group has 20 or more checked companies, SalesOne counts the reasons they were turned down and suggests a change to your profile. A person confirms it before the next run uses it.
Under 2%
Searches that have stopped finding companies
When the last 20 searches in a segment and region find under 2% new companies, that search is retired and sampled again a month later.
Your team’s fit ratings, with their reasons, also go into the next research brief. Replies and call outcomes are captured in SalesOne, on each account’s history; they do not change the research.
Questions buyers ask
What are S1 agents?
S1 agents are SalesOne’s deep research agents. They work every account in your market the way a senior researcher would: search widely, read the sources, check each fact at its source with its date, judge fit against your profile with a reason for every reject, find the people who own the problem and write a point of view. A second pass reviews the work. In production runs they averaged about 31 searches and 7 pages read for each account kept.
Is SalesOne a contact database?
No. SalesOne is not a contact database: it does not sell records to search or export. S1 agents research and qualify each account from dated, sourced public evidence, and contact details are looked up only for the people your team picks.
How is this different from an AI research column in a spreadsheet tool?
An AI column asks one question per row, which is quick and cheap. S1 agents work the whole account: about 31 searches and 7 pages read for every account delivered, every rule in your profile checked, the buying committee found and a point of view written, then a second pass re-checks every buying signal. The output is a qualified account, not a cell.
How do S1 agents check facts?
A buying signal counts only when it has a source link, is dated to the day (to the month when the source gives no day), is about this company and not a parent or a customer, still holds, and falls inside its window. Before delivery, a second pass re-checks every buying signal against its source. In three runs it corrected 40 signals and added 58 fresher ones across 147 accounts. The checks are done by agents; no human audit has been run yet.
How fresh is the research?
Every buying signal carries the date it happened. In our runs, the newest signal on a typical account was about two months old at delivery. Research runs when your team asks for accounts, and before outreach a fresh check looks for dated facts from the last 12 months. S1 does not watch your accounts between runs.
What does deep research cost?
SalesOne is priced on qualified accounts, not database rows or credits. Each account is delivered with its deep research. Companies S1 agents turn down are researched too, explained and never charged. Plans are quoted on a call, based on your market and the number of accounts you need.
Related reading
- Account research
Product
- Buying committee mapping
Product
- What each run builds on
Product
- How pricing works
Pricing
- Best AI tools to research accounts before outreach
Comparison
- ZoomInfo alternatives: cheaper data or deeper research
Comparison
- How we measure S1 research
Method
See your own market, researched.
Bring your offer and one customer you would like more of. We will research a sample of your market and show you the accounts S1 agents accept, the ones they turn down, and why.