How to Use Claude Research to Build a Supplier Shortlist

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Use Claude Research to Build a Supplier Shortlist.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Use Claude Research to Build a Supplier Shortlist.

To build a cited supplier shortlist, turn on Claude Research with the + button (Pro, Max, Ultra, Team or Enterprise), give it a brief with quantity, specification and budget, and ask for a link and supporting passage for every claim. Open each link yourself, mark claims confirmed, partial or unconfirmed, and request samples only from suppliers that pass.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

A citation is a route to evidence, not proof that the supplier can fulfil your job. The cited page may cover a different product, quantity or service level. Make Claude preserve those differences in the table. An honest “not confirmed” is more useful than a complete-looking shortlist built from incompatible offers.

How do you specify a supplier purchase before using Claude Research?

Specify the purchase by stating the business decision the research should support, then list the material, size, quantity, intended use, delivery requirement and acceptable alternatives in words a supplier can answer. “Find good packaging suppliers” leaves too much open. “Find suppliers able to quote for 600 rigid presentation boxes matching the attached dimensions, with a sample available before approval” defines a job.

Our main illustration is a florist considering presentation boxes for a recurring gift range. The initial order is 600 units, the target all-in budget is $1,200 and the owner wants a sample before committing. These numbers are hypothetical purchasing requirements, not real supplier prices. The exercise ends with a sample shortlist, not an automatic purchase or a claim that any actual supplier meets the brief.

Separate essential conditions from preferences. For this florist, the sample requirement and the internal dimensions are essential. A preferred colour may be negotiable if the box can carry the approved label. Write each essential condition as a yes, no or unknown question. A preference can receive a score later; an unknown essential condition needs confirmation before the candidate is treated as qualified.

Include exclusions that prevent wasted research. For example, exclude offers that only provide a design service when the job needs physical boxes, and exclude packaging intended for a different use. Do not ask Claude to infer suitability from photographs. If the product touches something that needs specialised safety assessment, obtain the appropriate documentation and advice rather than accepting generic marketing wording.

Research suppliers for this purchasing brief.
Business: florist. Item: rigid presentation box.
Illustrative first order: 600. Budget ceiling: $1,200 all-in.
Required dimensions and materials: use the attached approved specification.
Essential: sample before approval; written quantity and delivery terms.
Return 6 candidates, then identify up to 3 worth requesting samples from.
Use primary supplier pages for product and trading claims.
For each claim give URL, page title, supporting passage and caveats.
Write unknown when price, availability or suitability is not confirmed.
Do not contact suppliers or place orders.

The requested six candidates and three samples are working choices for this exercise, not Claude limits. Reduce the candidate count if only a few offers plausibly match. A report that explains why there are only two qualified candidates is more useful than one padded with unsuitable sellers. Keep the full rejected-candidate list so you can see why names disappeared.

How do you turn on Claude Research for a supplier search?

Click + below the chat box and choose Research on a paid Claude plan, and a blue indicator shows it is on. Research is on Claude's paid plans (Pro, Max, Ultra, Team and Enterprise) on the web, Claude Desktop and the mobile apps; Pro costs $20 a month billed monthly or $17 a month on the annual plan. In the older interface web search must also be switched on; Anthropic's web search help says accounts with the new Claude experience have no web search toggle, because Claude searches the web when it helps. Review the current Research instructions if your workspace controls differ.

One Team and Enterprise catch: Research always runs on Anthropic's Opus models, whichever model you picked for the chat. If an administrator has switched those Opus models off for your role, Research will not work, and the fix is a settings change by the owner rather than a better prompt.

Start on the web or desktop when your research needs connected internal sources. Claude's help page says Research started on mobile searches the web without using connected tools. For this public supplier scan, the web alone may be sufficient. If you also need last year's purchase specification, upload its approved version or use a permitted internal source and explicitly identify it.

Review connected access before starting. Claude's help page says Research reaches Gmail, Google Calendar, Google Docs and any other connector switched on for the chat, and uses them without asking for approval each time. A supplier search does not need broad access to unrelated client correspondence, so switch off the connectors you do not need from the + menu before you send the brief. Keep the source scope appropriate to the purchase and avoid entering account credentials or payment information. You are asking for an evidence pack, not delegating authority to buy.

Research draws on the same usage limits as standard Claude conversations and can consume them faster because it retrieves multiple sources. Start with the single product brief rather than combining every purchasing question into one request. Do not assume a fixed per-report charge or a guaranteed number of monthly reports. Check your account's usage before repeating a broad search with only cosmetic changes.

What should a Claude Research supplier table show?

Ask Claude for an evidence table before asking “which is best?”, with each row holding the product reference, quantity basis, advertised price if present, sample arrangement, lead-time wording, terms URL and unresolved requirements. Give each candidate a neutral reference such as Candidate A, alongside its actual supplier identity in your working file. Keep claims in separate columns so one citation cannot appear to support an entire row.

For the illustrative florist, a row might read: “Candidate A; specified box type advertised; sample policy stated; price at 600 not confirmed; production time stated but delivery transit unspecified.” That is a usable research result. “Candidate A; reliable, affordable, fast” is not. The first row tells the buyer exactly what to ask; the second mixes opinions with unsupported conclusions.

A filled-in register for the florist's first pass might look like this. The candidates, wording and statuses are illustrative, but the pattern of partial matches is typical of supplier research.

CandidateClaim in the reportSource typeWhat the cited page actually saysStatus
ABox comes in the required sizeSupplier product pageLists outside dimensions onlyUnconfirmed: ask for internal size
ASamples availableSupplier help page“Samples available for a fee”Confirmed; fee not stated
BMinimum order 100Marketplace listingMinimum of 100 for plain boxesPartial: printed version unknown
CMade from recycled boardSupplier sustainability pageRecycled content in the outer wrapPartial: covers one component

Ask Claude to identify whether a page is the supplier's own product page, its help page, a marketplace listing or an independent review. These sources have different jobs. A supplier page can establish what is advertised, while a review may raise a question to investigate. Neither alone proves future fulfilment. Keep opinions out of the columns intended for contractual facts.

Use checking AI research citations for a wider evidence method. In this purchasing workflow, focus on the conditions that could change your order: specification, quantity, complete cost, sample terms and dates. An unrelated background claim about the supplier's history deserves less attention than an unsupported promise about the product you are about to pay for.

How do you check a Claude Research citation?

Open each cited URL and test whether the page supports the exact claim, because a line that looks complete can rest on the wrong product, quantity or service level. Here is the kind of line a first Research pass can return, shown as an illustration with what a buyer would fix before trusting it.

Candidate B | rigid two-piece box | minimum order 100 [1]
price from $1.40 per unit [2] | samples available [3]
lead time: 5 working days [2]
[1] marketplace listing  [2] supplier product page  [3] supplier FAQ

Three repairs are needed. Source 1 covers the plain version, not the printed box, so the minimum order for 600 printed units is unknown; product variants often matter more than a seller's general claim to offer custom packaging. Source 2 gives $1.40 only at 5,000 units, so the price field at 600 reads “quotation needed”: do not multiply $1.40 by 600 and call it a budget, but keep the tier on file for future volumes. And “5 working days” on that page is production after artwork approval, not delivery, which can decide whether the supplier is viable for a dated event. The line looked complete; after checking, only the sample offer stands, and it still needs a fee and a date.

For every essential requirement, open the cited URL and locate the relevant wording. Read the surrounding qualification, not just the quoted sentence. Record whether the page supports the claim fully, partially or not at all. If the page has changed or cannot be opened, keep the claim unverified until you obtain a usable primary record. Do not treat a search snippet as a current quotation.

  1. Check that the supplier and product variant match the candidate row.
  2. Find the passage supporting the specific requirement.
  3. Read quantity, material, delivery and eligibility qualifications around it.
  4. Record what is missing and who must confirm it.
  5. Keep the source URL and the date you actually checked it.
  6. Recheck terms and pricing when you request the final quotation.

Environmental statements need the same scrutiny. If a supplier says that an option uses recycled material, do not turn it into an independently verified claim about the entire finished product. Ask which component the statement covers and request the supporting documentation relevant to your intended claim. If the owner plans to repeat it in customer copy, have the wording checked before publication.

Give every failed citation a specific repair instruction: “The linked page supports plain boxes only; find a source for the printed option or leave this requirement unknown.” Asking Claude to “double-check everything” often produces another fluent report without resolving the exact gap. Your correction should name the claim, the source mismatch and the acceptable form of evidence.

How do you compare supplier costs fairly?

Compare supplier costs by requesting quotations against the same specification and quantity after the source check, then adding setup and delivery to the unit price. The following calculation is entirely illustrative. Candidate A quotes $1.45 per unit for 600 boxes, plus $90 setup and $120 delivery: $870 + $90 + $120 = $1,080. Candidate B quotes $1.60 per unit and $60 delivery with no setup fee: $960 + $60 = $1,020.

Candidate C quotes $1.20 per unit but requires 1,000 units and charges $100 delivery: $1,200 + $100 = $1,300 cash outlay. Its nominal unit price is lowest, but the first order exceeds the sample $1,200 ceiling. The florist only needs 600 now, so the comparison must show both actual cash outlay and excess stock. Do not disguise the extra 400 units as an immediate saving.

Illustrative offerCash outlayUnits receivedDecision issue
A$1,080600Sample and specification still to pass
B$1,020600Check equivalent material and finish
C$1,3001,000Exceeds first-order budget

Candidate B is $60 below A, but that does not settle the choice. If B's sample fails the specified dimensions, its lower price is irrelevant. Keep mandatory quality checks outside the weighted score. Only score candidates that pass, or explicitly mark a provisional score as conditional. A candidate with an unknown essential condition should never quietly become the winner because its price column is attractive.

For scoring qualified candidates, the owner might propose 40 points for specification fit, 25 for complete cost, 20 for confirmed delivery and 15 for sample and support arrangements. Those are example weights, chosen for this job. Document how each score was earned and run the calculation again if delivery becomes more important. For the later quotation stage, use comparing purchasing quotes line by line.

If you want to build the landed-cost comparison in a spreadsheet, try The Complete Excel Course with AI Integration on Coding Liquids.

How do you request samples and quotes after the research?

Send every candidate the same written request, changing only the product reference, and ask for a written cost breakdown rather than a verbal all-in estimate. Include the approved specification, quantity, artwork requirements, sample purpose and the list of unresolved conditions. The buyer should be able to tell whether the quote includes setup, samples, delivery and any required finishing without translating a different format for each supplier.

Hello [first name],
Please quote against the attached specification for 600 units.
Confirm the product reference, material, internal dimensions and finish.
Show unit cost, setup, sample, delivery and any other charges separately.
Explain when production starts and distinguish production from transit.
Confirm the process if the sample or delivered goods differ from the spec.
Please identify any requirement you cannot meet.
Thanks, [your name]

For a dog groomer researching replacement appointment cards, an illustrative request might specify 2,000 cards, a writable reverse and a sample tested with the pen used at reception. The research could identify print options, but a sample reveals whether the card smudges in the actual workflow. Turn practical requirements into physical checks rather than expecting an online description to settle them.

A driving school sourcing teaching folders has a different test. If instructors carry folders between lessons, the sample check may include page retention and repeated opening, not simply appearance. An illustrative inspection record could say “20 pages loaded; two fell out after repeated handling; reject this binding”. The numbers describe a proposed test, not a claim about an actual product or supplier.

Budget the investigation explicitly. Suppose three samples cost an illustrative $25 each and staff spend 90 minutes inspecting them. The sample cash cost is $75; the staff time is separate. If the final order saving between two viable suppliers is only $60, the samples still may be justified by fit and failure prevention. Do not claim a purchasing saving merely because the cheapest quote looks lower.

How do you save checked Claude Research results into a Project?

Save the checked results by creating a Project for the recurring purchasing category and adding the approved specification, checked claim register, quotation comparison and sample results. Claude documents reusable Project knowledge and instructions; use those to preserve the evidence for later work. To move the research chat itself, click the dropdown arrow next to the chat's name, choose Add to project and pick the purchasing Project in the Move chat window. Check that the resulting workspace contains the records you expect.

Do not promote the whole raw research report to authoritative knowledge without review. Keep a clear distinction between “research leads”, “supplier-confirmed terms” and “owner-approved choice”. A later request to reorder must use the current quotation and approved sample result, not the first web listing. The Project is an organised working reference, while your normal purchasing system remains the record of orders and approvals.

For a photography studio sourcing presentation packaging, a useful saved entry might read “Candidate B sample accepted for the specified print size; price valid only for the quoted quantity; reorder requires reconfirmation.” That is more precise than “preferred supplier”. It helps a colleague understand both the decision and the point at which fresh research is needed. Do not imply that the sample approval covers every product sold by that company.

Assign an owner to recheck facts that can change. Stock availability, quoted price and delivery capacity need checking at purchase time. Your specification may stay stable longer, but an altered product range can still invalidate an earlier fit decision. Record the actual check date for each source rather than stamping the entire report with one reassuring date that hides older evidence.

When should you stop a supplier shortlist?

Stop the shortlist when essential conditions are still unresolved, so count them before sending a purchase request. If the florist still lacks confirmation of the internal dimensions or sample availability, the candidate is not ready for selection. Write “two essential questions unanswered” rather than “high confidence”. A confidence adjective does not tell the owner what remains to be done or who needs to answer it.

Use a final handover with four parts: qualified candidates, rejected candidates and reasons, unresolved questions with owners, and the recommended next action. The next action might be “request two samples”, not “place the order”. Keep the recommendation proportionate to the evidence. When the sample and written terms are satisfactory, follow your ordinary approval process; purchase-order automation addresses that later operational stage.

Set aside a first-run planning allowance of around two hours for briefing, reading citations and preparing requests, then measure the real effort. This is a suggested work budget, not Research's execution time. If checking the report takes longer than researching manually, narrow the brief or improve the table. The output worth keeping is a compact evidence trail that makes the buying decision easier to inspect.

Recognise repeated claims disguised as separate evidence

Several pages can repeat the same unsupported statement. An illustrative supplier profile, directory entry and marketplace description may all contain identical wording about rapid delivery. That is not three independent confirmations of the production slot you need. Trace the statement to its origin and request a written commitment for your specific quantity, specification and approval sequence.

Check identity consistency before exchanging money. Compare the trading identity on the quotation with the identity on the supplier's own site and the account details supplied through your normal verification process. If something differs, pause and contact the supplier through a previously verified channel. Research can assemble the discrepancy; it should not explain it away because the names look similar.

Finally, preserve rejection reasons in plain language. “Rejected because quoted dimensions refer to the outside of the box, not the required inside measurement” prevents the same unsuitable candidate returning next month. “Low score” does not. A good shortlist leaves enough evidence for a second person to reproduce the decision without trusting the wording of the AI report.

If you want help fitting supplier research into how your business actually buys, I offer one-to-one AI consultations for small businesses.

Claude Research for suppliers: questions buyers ask

How do I use Claude Research to build a supplier shortlist?

Turn on Research with the + button on a paid plan, give Claude a purchasing brief with quantity, specification and budget, and ask for a supplier table with a link and supporting passage for every claim. Then open each link yourself, mark claims confirmed, partial or unconfirmed, and request samples only from suppliers that pass.

Can I trust the citations in Claude Research?

Treat a citation as a route to evidence, not proof. The cited page may cover a different product, quantity or service level, so open each link, find the relevant wording and read the surrounding qualification. Record whether the page supports the claim fully, partially or not at all, and keep it unverified if the page cannot be opened.

Is Claude Research available on the free plan?

No. Claude's pricing page lists Research on Pro, Max and Ultra and on Team and Enterprise, but not on Free, although Free does include ordinary web search. Pro costs $20 a month billed monthly or $17 a month on the annual plan, so a one-off supplier scan can also be done with web search and more manual checking.

Can Claude Research read my past supplier emails?

Yes, if Gmail is connected and switched on for that chat, Research can search it alongside the web without asking each time. Start it on the web or in Claude Desktop: Research started on the mobile apps searches only the web, with no warning that connectors were skipped, and it cannot use local MCP servers set up in Claude Desktop.

Does Research use up my Claude limits faster?

It draws on the same usage limits as normal chats but uses them faster, because it runs many searches and reads several sources. Check where you stand in Settings, Usage. On paid plans you can turn on usage credits, billed at standard API rates, if you hit a limit mid-project.

Further reads

Sources: Claude help centre articles Use research on Claude, How can I create and manage projects? and Claude Cowork and chat are one Claude; Claude pricing; Anthropic, Claude takes research to new places.

Want a second pair of eyes on your setup?

Bring the process you're trying to fix. On a 1:1 call we'll work out which AI tools should handle it, which parts stay manual, and how to roll it out without disrupting the team.

Book a 1:1 call with me