# Research ~500 target accounts a month with an AI assistant and Google Sheets > Replace the account-research columns you'd build in Clay with batched, cited web research in Claude or ChatGPT, a strict JSON output schema, a scoring rubric and a 10% hallucination spot-check. It does not replace verified contact data — you still need Apollo, ZoomInfo or similar for emails and phones. - Outcome: Each month, turn a list of ~500 target companies into a qualified, cited, de-duplicated account list with a fit score and a personalisation angle for the best ones. - Replaces: Clay (partial) - Uses: Claude (optional) — Web search / Research mode with citations to research each batch and return JSON.; ChatGPT (optional) — Alternative: web search or deep research with citations to research each batch.; Google Workspace — Google Sheets as the master account list, de-dup formulas and QA log.; Apollo.io (optional) — Verified business emails/phones for Tier A accounts only. Any licensed contact-data provider works (ZoomInfo, etc.). - Extra monthly cost: $0 — $0 extra if you already pay for Claude or ChatGPT with web search and already have a contact-data source. Usage limits on research features vary by plan, so heavy months may hit caps. Contact data is NOT included: if Clay was also your email-finding tool, budget for a contact-data provider separately. - Difficulty: moderate; skill: basic; setup: 2–3 hours (rubric, prompt tuning on 2 test batches, sheet setup) - Evidence: editorial; illustrative example: true; last reviewed 2026-10-13 - Canonical: https://onelesstool.com/workflows/prospect-research-with-ai-assistant ## Good fit for - Founders, SDRs and small sales teams researching tens to a few hundred accounts per week - Account-based outreach where a cited 'why now' matters more than list size - Teams that used Clay mainly for AI research and scoring columns, not for waterfall contact enrichment - Anyone who can spend a few hours a week on batched research and spot-checks ## Not a fit for - Teams that need verified emails, mobile numbers or firmographics at scale — that's a contact-data product, not an AI assistant - Thousands of accounts per month or fully automated enrichment on a trigger — manual batching doesn't scale; consider an API pipeline or keep Clay - Workflows that depend on scraping LinkedIn or other sites whose terms forbid automated collection - Anyone not willing to spot-check sources — unchecked AI research will contain fabricated or misattributed facts ## Before A Clay workbook imports a company list, runs enrichment and AI research columns per row (company summary, recent news, hiring signals, a scoring formula, an AI-written first line) and pushes results to a CRM or sequencer. It's fast and automated but billed in credits/actions, and the bill grows with every column and re-run. ## After A Google Sheet holds the master list. Each working session you paste 10–25 new, de-duplicated companies into a saved research prompt in Claude or ChatGPT with web search on. The assistant returns strict JSON (one object per company, every claim with a source URL), which you paste into the sheet. Formulas flag duplicates and score totals; you open the sources for 10% of rows (and every Tier A) before trusting the batch. Only Tier A accounts go to Apollo/ZoomInfo for verified contacts. ## Steps ### 1. Write your rubric before you research anything Copy the **Qualification rubric** artifact and replace the bracketed descriptions with your real ICP: core industries, employee sweet spot, the signals that actually predicted past wins. Test it on **10 past customers and 10 lost deals**. If the rubric doesn't rank most past customers above the losses, fix the rubric now — no prompt will rescue a vague ICP. Keep `reachability` partly human: warm paths come from *your* CRM and network, not from the AI. ### 2. Set up the master sheet Create a Google Sheet with tabs **Input**, **Accounts**, **Manual**, **QA**. - **Input:** `Company name | Domain | Normalised domain | Already researched?` with the formulas from the *De-duplication and QA* artifact. - **Accounts:** one column per field in the JSON schema, plus `Batch ID`, `Researched on`, `QA result`, `Owner`, `Status`. - Use **Data → Data validation** dropdowns for `tier` and `Status` (`New / Checked / Contacted / Disqualified`). To paste JSON into columns, ask the assistant to also output the batch as **TSV** in a second code block (the batch prompt below does this), then paste into the first empty row of Accounts. ### 3. Run research in batches of 10–25 In Claude (turn on **web search**, or use **Research** on Pro/Max/Team/Enterprise) or ChatGPT (web search or **deep research**), paste the **Batch account research** prompt with 10–25 companies. Why small batches: - Quality drops and the model starts skipping or merging companies in long lists. - Research modes consume plan limits faster; Anthropic notes Research uses the same limits as normal chats but can use them up faster. - A bad batch of 20 is cheap to re-run; a bad batch of 200 isn't. **Throughput reality:** 500 accounts ≈ 25 batches of 20. At 15–30 minutes per batch including QA, budget **8–20 hours a month**. Save the prompt in a Claude Project or ChatGPT Project/custom GPT so every batch uses identical instructions. ### 4. Validate structure, de-duplicate, then spot-check sources Follow the *De-duplication and QA* artifact for every batch: 1. **Structure:** row count in = out; `fit_total` = sum of scores; every row has ≥1 source; no emails/phones. 2. **Identity:** if `resolved_domain` ≠ input domain, the model may have researched a namesake — move to **Manual**. 3. **Spot-check 10%** (minimum 3 rows, plus *every* Tier A): open each source URL and confirm the claim is on the page and the date matches. Log PASS/FAIL in **QA**. 4. If more than 1 in 10 fails, re-run the batch through the **Verify claims** prompt or redo it — don't fix rows by hand and trust the rest. The rule *“no URL, no claim”* is what makes this usable: an uncited signal is capped at 1 point by the rubric. ### 5. Get contact data the compliant way — only for Tier A An AI assistant **does not provide verified contact data**. Emails it 'finds' are often guessed patterns. For Tier A accounts only: - Look up the buyer in **Apollo, ZoomInfo or another licensed provider**, or use contact details the company publishes for business enquiries. - **Do not scrape LinkedIn.** LinkedIn's User Agreement (section 8.2) prohibits scraping or copying profiles and using bots or unauthorised automated methods. Read profiles manually if you must; don't automate it. - **GDPR (EU/UK contacts):** document your lawful basis (usually legitimate interests) and give the Article 14 notice — who you are, why you have their data, where it came from, their rights — **at first contact** at the latest. - **CAN-SPAM (US):** applies to B2B email too. Accurate headers and subject, a postal address, a clear opt-out honoured within 10 business days. Other countries (e.g. Canada's CASL) can be stricter — check where your recipients are. ### 6. Write openers from cited signals and close the loop Use the **Personalisation from cited signals** prompt on Tier A rows only. It must quote the signal and URL it used, so you can check it in seconds. After a month, compare reply and meeting rates by tier. If Tier B converts like Tier A, your rubric is too strict; if Tier A doesn't convert, re-weight `need_signal` and `timing`. Then decide what to cancel: if you only used Clay's research columns, downgrade it; if you also relied on its contact waterfall, keep or replace that part with a contact-data subscription. ## Prompts ### Batch account research (strict JSON) Research 10–25 companies with citations and score them against your rubric. ```text You are a B2B research analyst. Use web search. Research each company below for a seller of {{our_offer}} whose ideal customer is: {{icp_description}}. Companies (name, domain): {{company_list}} Scoring rubric (score each 0–3): {{rubric}} Rules: 1. Every factual claim must come from a page you actually opened. Put its URL in source_url / sources. If you cannot find a source, write "unknown" and add the gap to `unknowns`. Never guess employee counts, funding or dates. 2. Prefer the company's own site, press releases, job boards and reputable news from the last 12 months. Today is {{today}}; ignore signals older than 18 months. 3. If a name is ambiguous, research the company at the given domain; set match_confidence accordingly. If you can't confirm it's the same company, use "low" or "not_found" and don't score it. 4. Do NOT output personal email addresses, phone numbers or home locations. Do not use LinkedIn content obtained by scraping. 5. reachability: score only 2 (buyer role publicly identifiable) or 1 (unclear); I'll add warm paths myself. 6. fit_total must equal the sum of fit_scores. Tier: A 11–15 with need_signal ≥2; B 7–10; C 3–6; otherwise disqualified with a reason. Output exactly two code blocks and nothing else: (a) a JSON array that validates against this schema: {{json_schema}} (b) the same rows as TSV with a header row, columns in schema order, signals flattened as "type: claim (date) URL" joined by " | ". Return one object per input company, in input order, even if not found. ``` ### Verify claims (second pass) Have the assistant re-check a batch whose spot-check failed, claim by claim. ```text Act as a fact-checker. For each row in the JSON below, open every source_url and check whether the page explicitly supports the claim, for the right company and date. {{batch_json}} Return a table with columns: input_name | claim | source_url | verdict (SUPPORTED / NOT FOUND ON PAGE / WRONG COMPANY / WRONG DATE / PAGE UNAVAILABLE) | corrected claim with new URL, or "remove". Then return the corrected JSON array with unsupported claims removed, fit_scores lowered where a removed claim was the basis (cap uncited criteria at 1), fit_total and tier recomputed. Do not add new claims without a URL. ``` ### Personalisation from cited signals Draft a first line for Tier A accounts that a human can verify in seconds. ```text Write a first line for a cold email from {{sender_name}} at {{our_company}} to the {{buyer_role}} at {{company_name}}. Use ONLY these cited signals: {{signals_with_urls}} Our relevant proof point: {{proof_point}} Rules: max 30 words; reference exactly one signal factually (no flattery, no 'I noticed', no invented details); no claims about the person. Give 3 options, and under each write the signal and URL it relies on. ``` ## Templates and scripts ### JSON output schema for a research batch (schema) ```json { "$schema": "https://json-schema.org/draft/2020-12/schema", "title": "AccountResearchBatch", "type": "array", "items": { "type": "object", "additionalProperties": false, "required": ["input_name", "input_domain", "resolved_domain", "match_confidence", "what_they_do", "industry", "hq_country", "employee_range", "signals", "fit_scores", "fit_total", "tier", "personalization_hook", "unknowns", "sources"], "properties": { "input_name": { "type": "string", "description": "Exactly as given in the input list" }, "input_domain": { "type": "string" }, "resolved_domain": { "type": "string", "description": "Domain of the company actually researched. Empty if not found." }, "match_confidence": { "enum": ["high", "medium", "low", "not_found"] }, "what_they_do": { "type": "string", "maxLength": 300, "description": "One or two plain sentences" }, "industry": { "type": "string" }, "hq_country": { "type": "string", "description": "ISO country name or 'unknown'" }, "employee_range": { "enum": ["1-10", "11-50", "51-200", "201-500", "501-1000", "1001-5000", "5000+", "unknown"] }, "signals": { "type": "array", "maxItems": 5, "items": { "type": "object", "additionalProperties": false, "required": ["type", "claim", "date", "source_url"], "properties": { "type": { "enum": ["hiring", "funding", "leadership_change", "product_launch", "expansion", "tech_stack", "partnership", "other"] }, "claim": { "type": "string", "maxLength": 200 }, "date": { "type": "string", "description": "YYYY-MM or YYYY-MM-DD; 'unknown' if not stated" }, "source_url": { "type": "string", "format": "uri" } } } }, "fit_scores": { "type": "object", "additionalProperties": false, "required": ["icp_industry", "size", "need_signal", "timing", "reachability"], "properties": { "icp_industry": { "type": "integer", "minimum": 0, "maximum": 3 }, "size": { "type": "integer", "minimum": 0, "maximum": 3 }, "need_signal": { "type": "integer", "minimum": 0, "maximum": 3 }, "timing": { "type": "integer", "minimum": 0, "maximum": 3 }, "reachability": { "type": "integer", "minimum": 0, "maximum": 3 } } }, "fit_total": { "type": "integer", "minimum": 0, "maximum": 15 }, "tier": { "enum": ["A", "B", "C", "disqualified"] }, "disqualify_reason": { "type": "string" }, "personalization_hook": { "type": "string", "maxLength": 240, "description": "Must reference one of the cited signals" }, "unknowns": { "type": "array", "items": { "type": "string" } }, "sources": { "type": "array", "minItems": 1, "items": { "type": "string", "format": "uri" } } } } } NOTE: this schema intentionally has NO fields for personal email addresses or phone numbers. Verified contact data comes from a contact-data provider, not from an AI assistant. ``` ### Qualification rubric (0–15) (template) ``` QUALIFICATION RUBRIC — edit the descriptions to match YOUR ideal customer profile (ICP). Score each criterion 0–3. Total 0–15. The AI proposes scores; a human confirms Tier A. 1. icp_industry 3 = core target industry ([your core industries]) 2 = adjacent industry we have won before 1 = plausible but no references 0 = outside our market → disqualified 2. size (employees, from the company's own site or a cited source) 3 = inside our sweet spot (e.g. 51–500) 2 = just outside (e.g. 20–50 or 501–1,000) 1 = far outside but has a relevant division 0 = unknown after research (do NOT guess) 3. need_signal (evidence they have the problem we solve) 3 = explicit, dated, cited signal in the last 6 months (job post for the role we replace, public complaint, initiative announced) 2 = indirect signal (growth, new market, tech stack that implies the need) 1 = generic fit only 0 = signal suggests they've already solved it (e.g. announced a competitor) 4. timing 3 = trigger event in the last 90 days (funding, new leader in the buyer role, expansion) 2 = trigger event 3–12 months ago 1 = none found 0 = known freeze (layoffs, acquisition pending) — note it 5. reachability 3 = we have a warm path (customer, investor, partner in common) — fill from YOUR records, not the AI 2 = clear buyer role exists and is publicly identifiable 1 = buyer role unclear 0 = no way in found TIERS A = 11–15 and need_signal ≥ 2 → human review, then contact-data lookup, then personal outreach B = 7–10 → nurture list / lighter-touch sequence C = 3–6 → park; re-check in 6 months disqualified = icp_industry 0, or a hard exclusion (existing customer, competitor, sanctioned country, under legal minimum size) RULE: any score based on a claim without a source URL is capped at 1. ``` ### De-duplication and QA checklist with Sheets formulas (checklist) ``` DE-DUPLICATION AND QA — run on every batch before it enters the master sheet A. De-duplicate (Google Sheets, master tab "Accounts") 1. Normalise the domain in a helper column (assume raw domain/URL in column B): =LOWER(REGEXREPLACE(REGEXREPLACE(TRIM(B2),"^(https?://)?(www\.)?",""),"[/?#].*$","")) 2. Flag duplicates against everything already researched: =IF(COUNTIF($C$2:$C, C2)>1, "DUPLICATE", "") 3. Before each run, remove input rows whose normalised domain already exists in Accounts (=ISNUMBER(MATCH(C2, Accounts!C:C, 0)) on the input tab). This avoids paying for the same research twice. 4. Also check resolved_domain vs input_domain: if they differ, the AI may have researched a different company with a similar name → set status "CHECK MATCH". B. Structural checks (instant, every row) [ ] JSON parsed without edits (if not, re-run the batch; do not hand-fix silently) [ ] Row count out = row count in [ ] fit_total equals the sum of fit_scores [ ] Every signal has a source_url; every row has ≥1 source [ ] No email addresses or phone numbers anywhere in the output (search for "@" and digit runs) — delete if present [ ] match_confidence "low"/"not_found" rows are moved to a Manual tab, not scored C. Hallucination spot-check (10% of rows, minimum 3 per batch, ALWAYS include every Tier A) For each sampled row, open every source URL and confirm: [ ] The page loads and is about this company (not a namesake) [ ] The claim is actually stated on the page (not inferred) [ ] The date matches Record result in a "QA" column: PASS / FAIL + note. If >1 of 10 sampled rows FAIL → re-run the whole batch with the verification prompt; don't trust the rest. D. Compliance gate before any outreach [ ] Contact details came from a licensed provider or the person's own published business contact page — not scraped from LinkedIn [ ] EU/UK contacts: lawful basis (usually legitimate interests) documented; GDPR Art. 14 notice given at first contact [ ] US commercial email: CAN-SPAM basics in the template (accurate From/subject, postal address, working opt-out honoured within 10 business days) [ ] Suppression list (unsubscribes, existing customers, competitors) applied ``` ## Limitations - No verified contact data: AI assistants don't maintain a B2B contact database. Emails they produce are often guessed patterns; use a licensed contact-data provider. - Manual batching: 10–25 companies per run. 500 accounts a month is 25–50 runs and roughly 8–20 hours of human time, versus near-zero in an automated Clay table. - Research quality depends on what's public. Private companies, small firms and non-English markets return many 'unknown' fields. - AI research makes mistakes: namesake companies, outdated news presented as current and claims not actually on the cited page. The 10% spot-check reduces but doesn't eliminate this. - Research features have plan-based usage limits that can change; a heavy month can hit them. - No native CRM/sequencer push — results move by copy/paste or CSV import. ## Risks - **reliability** (high): Hallucinated or misattributed facts can end up in outreach and damage credibility. Mitigation: no URL, no claim; spot-check 10% plus every Tier A; re-run failing batches. - **compliance** (high): Outreach to people still triggers GDPR (EU/UK: lawful basis, Article 14 notice by first contact) and CAN-SPAM (US: applies to B2B; FTC lists penalties of up to $53,088 per violating email). Using the AI doesn't change who is the data controller — you are. - **compliance** (medium): Scraping LinkedIn (directly or by asking an agent to) breaches LinkedIn's User Agreement section 8.2 and can get accounts restricted. Keep research to public web pages and licensed data. - **sensitive_data** (low): Company lists and ICP notes are commercially sensitive. Use a business/team plan or check your consumer plan's training settings before pasting customer or pipeline data. - **functionality_loss** (medium): You lose Clay's automation: scheduled refreshes, enrichment waterfalls across providers and pushes to CRM/sequencers. ## Verification checklist (run before cancelling) - [ ] The rubric ranks at least 7 of 10 known past customers above at least 7 of 10 known lost/unfit deals in a back-test. - [ ] Two full test batches (≥40 companies) came back as valid JSON with one row per input, in order, with no emails or phone numbers. - [ ] Spot-check of those batches: at least 90% of sampled claims are SUPPORTED when you open the source URL; every Tier A row passed a full source check. - [ ] Side-by-side on 25 accounts: the AI's tiering agrees with your existing Clay scoring (or a human's judgement) on at least 20. - [ ] The de-dup formula catches a deliberately re-pasted batch (all rows flagged DUPLICATE) and normalises 'https://www.example.com/about' to 'example.com'. - [ ] Contact data for Tier A accounts is coming from a licensed provider or published business contacts, and your outreach template includes the CAN-SPAM elements and (for EU/UK) an Article 14 notice. - [ ] You timed a real month's workload (or a week × 4) and the hours are acceptable versus what you pay for Clay. ## Savings methodology Savings = Clay subscription you can drop or downgrade − any extra spend (normally $0 if you already pay for Claude/ChatGPT with web search). Contact-data costs are excluded on both sides: if Clay was also your email finder, you need a contact-data subscription either way, so subtract that from the savings. Labour is the hidden cost: at 15–30 minutes per 10–25 companies, 500 accounts is roughly 8–20 hours a month — price your time before deciding. Example (illustrative): A two-person outbound team that used a Clay Launch plan mainly for research and scoring columns on ~500 accounts a month. $167/mo → $0/mo. Assumptions: ILLUSTRATIVE: Clay's pricing page on 2026-10-13 showed $167/month as the Launch headline price but also lower 'starts at' prices and different figures in its FAQ, so we treat $167 as an example, not a verified price. Use your own invoice. They already pay for Claude or ChatGPT with web search; no new subscription. Contact data comes from an Apollo/ZoomInfo seat they already have; that cost is unchanged and excluded. 8–20 hours a month of research and QA time is not converted to dollars. ## Sources - Clay pricing: https://www.clay.com/pricing (accessed 2026-10-13) - Using Research on Claude: https://support.claude.com/en/articles/11088861-using-research-on-claude (accessed 2026-10-13) - LinkedIn User Agreement: https://www.linkedin.com/legal/user-agreement (accessed 2026-10-13) - FTC CAN-SPAM Act compliance guide: https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business (accessed 2026-10-13) - GDPR Article 14: https://gdpr-info.eu/art-14-gdpr/ (accessed 2026-10-13) - ChatGPT deep research FAQ: https://help.openai.com/en/articles/10500283-deep-research-faq (accessed 2026-10-13)