1. LandOffer
  2. Blog
  3. All articles
  4. Build a Gmail-to-Sheets Application Receipt Inbox with a Human Review Step

Job search

Build a Gmail-to-Sheets Application Receipt Inbox with a Human Review Step

Original synthetic import and review distinguish individual messages, application evidence, duplicate retrieval and marketing false positives.

9 min readLandOffer team

Build a receipt inbox that stages evidence for review, rather than a script that changes your application count whenever an email contains “application.” Keep each received message identifiable, put extracted company and role fields in proposal columns, and require a person to confirm what the message establishes. A receipt can support an application record; marketing, recruiter correspondence and repeated messages should not create new applications automatically.

LandOffer publishes this guide and offers job-search tools. We checked Google’s public Apps Script documentation on October 21, 2026 and tested original synthetic records locally. We did not read a real mailbox, authorize a script, deploy a trigger or write a live spreadsheet. Use this design while finding recent technical jobs on LandOffer.ai, with account permissions and live deployment treated as separate implementation decisions.

A candidate moves incoming evidence through a review tray before adding a verified record to a job-search ledger.

Give the receipt inbox a narrower job than your application tracker

Use two logical tables. Receipt Inbox records individual messages and their review decisions. Applications records the candidate’s own application identities, such as an internal application key paired with an employer and requisition. Several messages can support one application. One company can have several roles. A thread is useful context, but it is neither a single email nor a reliable application identifier.

For Receipt Inbox, start with mailbox key, Gmail message ID, thread reference, message date, sender, subject and a short evidence excerpt. Then add proposed employer, proposed requisition, proposed application key, extraction version and review state. Keep reviewed employer, reviewed requisition, reviewer note and reviewed time in separate columns. An empty proposal is an acceptable result; an invented match is not.

The first state is Pending review for every staged message. The importer must not fill Reviewed application status or increment an application total. A person can later approve a receipt link, classify correspondence as action-only, reject a false positive or leave the message unresolved. Preserve that decision when the next import runs, so a new batch cannot erase the distinction you already checked.

Do not store a complete resume, every attachment or an entire recruiting conversation just to prove a receipt arrived. Retain the smallest useful excerpt and a private source reference that lets the authorized reviewer return to the message. Sharing the spreadsheet also shares whatever evidence you copied into it; an email’s original audience does not automatically control a separate spreadsheet.

Collect messages individually even when search returns threads

Google documents GmailApp.search(query, start, max) as returning Gmail threads. Inspect messages within each returned thread rather than assigning the thread’s first subject to every message. A matched conversation can contain your own reply, a draft or unrelated follow-up. A search selects candidate conversations; each message still needs its own direction check and review.

GmailMessage exposes individual IDs, sender and recipient fields, date, subject and plain-text body. Use the individual message ID in your stored evidence key. Google’s thread reference explicitly warns that a thread ID can vary with its contents. Keep it for context, while deduplicating imports by mailbox key plus message ID. Do not confuse this Apps Script message ID with an employer’s requisition or an email header’s Message-ID.

Check direction at the individual-message level. Exclude drafts and messages sent by the candidate’s known addresses and aliases from this received-message workflow. Review actual recipient and delivery context when direction is ambiguous; forwarded messages, mailing lists and delegated mailboxes need explicit handling. A sender-address comparison alone is not a verified direction classifier for those cases. An archived received message is still received, so an inbox-only condition would omit evidence unnecessarily. Apply your intended message-date bounds to individual messages too: expanding a returned thread can expose older correspondence outside the collection period. Preserve message dates separately from import time. A run timestamp describes when the collector looked, not when you applied or when the employer acknowledged receipt.

Start with a small collection containing both known receipts and messages that merely look relevant. A subject containing “application received” can occur inside quoted text or a promotional invitation to apply. Sender display names and apparent domains are not, by themselves, proof of employer identity. Preserve provenance and leave business interpretation to review.

Worked import: eight observations become five pending records

Everything in this example is fictional. Tessa’s local fixture uses mailbox key candidate-demo and invented message IDs. She has two possible applications at Lantern Ridge: backend role LR-731 and developer-tools role LR-744. The extractor proposes requisitions only from explicit LR- identifiers in synthetic text. It does not claim to understand arbitrary employers or email layouts.

Observation Synthetic evidence Completed staging disposition
m101, thread tA Receipt says application received for LR-731 One Pending review record
m101 again Same message returned in another batch Skip as the same evidence key
m102, thread tA Recruiter asks for availability; quotes an earlier receipt Separate Pending review record, not another application
m103, thread tB Receipt explicitly identifies LR-744 Separate Pending review record
m104, thread tC “Application tips” promotion mentioning LR-731 Pending review; no status inferred
m105, thread tD Update mentions both LR-731 and LR-744 Pending review with ambiguous role proposal
m106 Candidate’s draft response Exclude from received-message staging
m107 Candidate’s sent response Exclude from received-message staging

The completed result is five pending records, one duplicate skipped and two direction exclusions. There are still zero automatically approved applications. Tessa’s extractor retains both proposed identifiers for m105 instead of choosing the first one. The two messages in tA remain separate evidence rows, preserving the recruiter request beside the receipt.

We executed this original fixture locally, including a repeat import. Replaying the same observations adds zero rows and leaves existing review fields unchanged. We also checked that reordered observations produce the same evidence-key set. This validates the stated offline rules, not Gmail retrieval, Apps Script authorization, Sheets writes or real-message extraction accuracy.

The distinction matters when you resume after a failure. If a batch was written but the run’s progress marker was not saved, a repeat scan should find the stored message keys and avoid appending them again. Do not use “unread” as the processing marker: reading a message is unrelated to whether its evidence was safely stored and reviewed.

A reviewer compares two distinct role folders against a mixed group of receipt and promotional messages.

Worked review: five messages support two applications

Tessa now reviews the original synthetic evidence, not just the proposals. She checks her own stipulated application record and the explicit receipt wording. Her decisions below are completed fictional adjudications, not outputs from a live mailbox or an automatic classifier.

For m101, she confirms that LR-731 is the backend role she applied to and links the receipt to internal application APP-L31. Her note records that the receipt acknowledges receipt of that application; it does not establish recruiter review, interview selection or every later task’s completion. For m103, she makes a separate link to APP-L44, the developer-tools application.

She classifies m102 as Action only. Its new content asks about availability. Quoted receipt wording does not make this a new submission, and the thread’s shared identity does not erase the action request. She retains its association with APP-L31 without adding a third application or sending any reply.

She rejects m104 as Not a receipt: the promotional message offers advice and suggests applying. An identifier in that message is discovery context, not proof that she submitted anything. She leaves m105 On hold because two role identifiers appear without enough evidence to assign this update to one application. The automation’s ambiguity remains visible rather than becoming a confident but unsupported match.

Her final review totals are two Receipt approved, one Action only, one Not a receipt and one On hold. The approved evidence supports two distinct internal application keys. These five message decisions are not five applications. If another approved receipt later points to APP-L31, it can add corroborating evidence without creating another application record.

This design also handles incomplete history honestly. If a genuine acknowledgment arrives for a role absent from the tracker, hold the application link until the candidate reconciles the missing record. Do not manufacture a submission date from the email’s date: acknowledgment timing can differ from when the candidate submitted. Store the message date and reviewed submission date as separate facts.

Implement the boundary before adding scheduled runs

An Apps Script implementation would page through selected threads, inspect each message, check stored evidence keys and build proposed rows. It would then write only new pending records to the receipt table. Keep the application-update operation separate from this collector. No mail-sending, reply, archive, delete or mark-read action is needed for this dataflow.

Google’s authorization guide explains that service access requires authorization and that scope detection examines code. The documented GmailApp search and message-reading methods list the broad https://mail.google.com/ scope. A collector’s intended read-only behavior therefore does not mean this particular service grants only narrow read access. Review the actual requested permissions before any future deployment; the documentation review here granted none.

For Sheets, Range.setValues expects a two-dimensional array matching the target range. Its documentation also says a value beginning with = is interpreted as a formula. Treat imported subjects and excerpts as untrusted text and define a literal-text encoding policy before a live write. Our offline CSV stores synthetic text only; creating a CSV is not proof that opening it in Sheets preserves every value safely. Test the selected write path separately with harmless formula-like strings before importing real mail.

Overlapping executions introduce another problem: two runs can both decide a message is new. Google’s LockService supports guarded sections, but obtaining a lock object is not acquiring it; the documented lock acquisition methods must succeed. Within a future guarded write section, reread stored keys before appending. A script-level lock does not coordinate unrelated script projects or manual spreadsheet edits, and it does not create a database transaction.

Review failures without overwriting decisions

Record extraction version and run outcome so a partial import remains diagnosable. On failure, stop and retain the last confirmed evidence state. A retry may rediscover old messages, which is why stored message keys matter more than a claim that each batch runs once. Do not clear reviewer notes merely to rerun extraction after improving a pattern.

If the same evidence key arrives with contradictory fixture content, our local check rejects that conflict instead of silently overwriting the first record. Distinct IDs carrying similar receipt text are a different issue: retain the evidence separately, then let review decide whether both concern one application. Content similarity alone cannot safely merge two real submissions to different roles.

Keep the review queue small enough to inspect. For a reviewed row, retain the original proposal alongside the corrected application key and a brief reason. If a later extraction version suggests a different requisition, stage that discrepancy for another decision rather than replacing the approved link. A new parser version is a proposal change, not fresh employer evidence. A user who cannot explain why a proposed role matched should leave it unresolved. Missing receipts do not prove an application was never submitted, and a completed import does not prove the mailbox contains every acknowledgment. Retrieval coverage, extraction accuracy and business interpretation need separate checks.

Start with the pending inbox and its evidence keys, then review the resulting links against your own application history. The useful result is a traceable message-to-application relationship with uncertainty preserved; a tidier inbox does not establish a stronger hiring outcome. Continue your search through LandOffer.ai while keeping receipts, required tasks and recruiting outcomes distinct.

Sources and Further Reading