Audit Literature Reviews with ChatGPT

Build a reusable ChatGPT prompt to create and audit cited literature reviews.

Introduction

30 Second Summary

A confident answer can sound trustworthy even when its sources do not support its claims. Polished writing makes that gap easy to miss.

In this project, you will build a reusable research prompt in ChatGPT that produces a cited literature review about retrieval practice in higher education. You will test three claims against their sources before turning the corrected prompt into a template for future research.

What You'll Build

You’ll run your finished prompt with Search and see a structured literature review that you can audit claim by claim.

By the end of this project, you'll have:

A baseline scorecard: See which research controls a vague prompt leaves to chance.
An auditable review: Open the sources behind a structured literature review and trace its claims back to evidence.
A repaired research template: Correct unsupported claims and reuse the evidence rules for future academic questions.
Secret Mission: Adapt the template to research spaced practice while keeping it distinct from retrieval practice.

Are there any prerequisites?

You need a ChatGPT Free account with internet access. You also need Microsoft Edge on a macOS or Windows device.

Step 1 confirms that Search works before you begin the review.

Before We Start

Before We Start...

A clear commitment gives you one research outcome to test throughout the project. This is where you define what you are building and why evidence checks matter.

Confirm Your Research Workspace

Your commitment gives every prompt change a fixed research target. The next requirement is a workspace that can connect a claim to inspectable evidence.

ChatGPT Free can use Search on the web. You will prove that a cited result opens in Microsoft Edge.

In this step, get ready to:

Sign in to ChatGPT Free through Microsoft Edge.
Enable Search in a new conversation.
Verify a cited test result against the title of its linked source page.

Sign in to ChatGPT Free

Signing in connects Search to your account. The conversation workspace gives you a clean starting point for the test.

Press Cmd+Space (macOS) or the Windows key (Windows) to open your search bar.
Type Microsoft Edge into the search bar.
Press Enter to open Microsoft Edge.
Enter https://chatgpt.com in the address bar.
Press Enter to load ChatGPT.
Complete the sign-in flow using your ChatGPT Free account.

You should see the ChatGPT conversation workspace with its message composer. Your account is ready for a clean Search test.

Having trouble signing in?

Reload ChatGPT in Microsoft Edge.
Confirm that Microsoft Edge can load another website.
Restart the sign-in flow from the ChatGPT page.

Ask for help with the step that is blocking you: Help me troubleshoot signing in to ChatGPT Free through Microsoft Edge.

Enable Search in a new conversation

Search adds current web sources to a ChatGPT response. A new conversation keeps this access check separate from the research work you create later.

Why Search for this test?

Search provides a fast web result with citations you can inspect. It fits this short workspace test.

Deep research supports multi-step synthesis. Access to it is limited on ChatGPT Free.

Select the new conversation control at the top of ChatGPT's sidebar.
Select View all tools in the message composer.
Select Search.
Confirm that Search remains selected in the message composer.

Search is now enabled for your new conversation. The workspace is ready to retrieve a cited academic result.

Test a citation in Microsoft Edge

A useful workspace test starts with a cited ChatGPT answer. Its citation should open a readable source page.

Copy the test question below into the ChatGPT message composer:

Find one recent academic review about retrieval practice in higher education. Give me one sentence about its scope and include a citation.

What does this test prove?

The question asks for a recent academic review. It requires a citation you can inspect.

Submit the question from the message composer.

You should see a sentence about the review's scope. The answer should include a citation.

No citation in the answer?

Confirm that Search is selected in the message composer.
Start another new conversation from the ChatGPT sidebar.
Submit the exact test question again.

Get focused help if the response still has no source link: Help me find why ChatGPT Search returned no citation for my academic review test question.

Before you open the citation, do you think its page title will match the review described in the answer?

Select the citation attached to the sentence about the review's scope.
Switch to the source page that opens in Microsoft Edge.
Read the source page title.
Compare the title with the claim in the cited sentence.

You should see a page title related to retrieval practice or the review's stated scope. That confirms your workspace can move from AI-generated prose to a source you can inspect.

Does the source page fail to open?

Return to the ChatGPT conversation in Microsoft Edge.
Select the citation again.
Reload the source page if Microsoft Edge opens a blank page.

Ask for help if the linked page still does not load: Help me troubleshoot a ChatGPT Search citation that does not open in Microsoft Edge.

Your Search workspace can now retrieve a cited result and open its evidence. Next, you will test a deliberately vague research prompt to reveal what fluent prose can leave uncontrolled.

Test a Vague Prompt

Your ChatGPT Search workspace is ready in Microsoft Edge. You can now test what a vague prompt produces before adding stronger research controls.

A fluent review can still be unauditable when its prompt leaves source selection undefined. This baseline exposes that gap before you engineer a stronger research process.

In this step, get ready to:

Generate a baseline literature review from a vague prompt.
Score the review against five research controls.
Record the score plus the most important missing control.

Generate the baseline review

A useful baseline keeps the starting request deliberately short. You can then judge which qualities came from the prompt itself.

Return to ChatGPT in Microsoft Edge from earlier.
Use the control at the top of ChatGPT's left sidebar to start a new conversation.
Select View all tools below the message box.
Select Search.
Copy the baseline prompt below:

Write a literature review about retrieval practice in higher education.

What does this prompt ask for?

The prompt names a topic plus an educational setting. Its brevity gives you a consistent starting point for the audit.

Paste the baseline prompt into the ChatGPT message box.

Before you submit it, do you think this single sentence can guarantee an auditable review?

Submit the baseline prompt.

You should receive a literature review that may look polished at first glance. Your deliberately weak baseline is now ready to test.

Didn't get a Search response?

Confirm that Search is active in this new conversation.
Select View all tools.
Select Search.

Help me troubleshoot why my Search-enabled ChatGPT conversation did not generate the baseline review.

Score the baseline review

A scorecard turns a first impression into a repeatable audit. Each point represents a control that the original prompt required plus a feature the response delivered.

What counts as a point?

Score each feature against the original prompt as well as the response. A feature earns one point only when the prompt required it plus the response delivered it.

If ChatGPT supplied a useful feature on its own, record zero for that item. Voluntary additions cannot make the prompt reproducible.

Before you score it, how many of these controls do you think the one-sentence prompt can guarantee?

Record yes or no for a stated higher-education population plus a stated date scope.
Record yes or no for identifiable academic source types.
Record yes or no for citations attached directly to substantive research claims.
Record yes or no for disagreements or limitations separated from the conclusions.
Record yes or no for missing evidence reported explicitly.
Give each yes one point.
Add the points to calculate your score out of five.

Your score should expose the central weakness: the prompt defines no reproducible source-inclusion process. Features that ChatGPT volunteered remain impossible to guarantee from the prompt alone.

Record your score in the checkpoint below.
Name the most important missing prompt control in the same response.

That diagnosis is complete. Next, you will turn the missing controls into a citation-audited research prompt.

Build an Auditable Research Prompt

Your baseline review may sound convincing. However, its vague prompt never defined how ChatGPT should choose evidence or connect sources to claims.

An auditable prompt contract makes that evidence-selection process visible. OpenAI's framework provides Goal, Context, Output, and Boundaries. You will add explicit evidence rules before Search gathers sources.

In this step, get ready to:

Define the research scope in a reusable prompt contract.
Generate a structured review with explicit evidence rules.
Check the response for all six required output sections.

Build the prompt contract

Each part of the contract closes a gap exposed by your baseline scorecard. The complete prompt defines which research belongs in the review before ChatGPT starts writing.

Start a new conversation in ChatGPT.
Select View all tools below the message box.
Select Search.
Prepare your reusable prompt contract by copying the text below:

# Goal
Answer this research question: What does academic research published from 2015 to the present say about the effect of retrieval practice on learning outcomes in higher education?

# Context and search scope
Use web search. Focus on university, college, tertiary, or health-professions education. Prioritize peer-reviewed systematic reviews, meta-analyses, review articles, and applied higher-education studies. Use 4 to 6 academic sources. Exclude marketing pages, unsourced summaries, and studies that only address school-age learners unless they provide necessary context and are clearly labeled.

# Evidence rules
Do not invent citations. Attach a citation to every substantive research claim. For each source, report its title, year, study type, learner population, relevant finding, important limitation, and link. Separate what a source directly reports from your interpretation. If the source type, population, or support for a claim cannot be confirmed, label it “needs verification” rather than guessing.

# Output
1. Research question and scope
2. Evidence table
3. A 350 to 500 word synthesis organized by themes
4. Disagreements and limitations
5. What's missing or still uncertain
6. Source-quality check

# Boundaries
Distinguish retrieval practice from spacing or distributed practice. Do not treat an abstract, search snippet, or citation as proof until the linked page supports the claim. Keep conclusions proportional to the evidence.

How does this prompt contract work?

The Goal locks the project to one research question.
The Context and search scope defines the eligible learner populations.
The search scope also defines the preferred academic source types.
The Evidence rules connect every substantive claim to a citation.
The evidence rules label unconfirmed details as needs verification.
The Output requires a fixed six-part response.
The Boundaries keeps retrieval practice separate from spacing or distributed practice.

Before you submit the contract, which missing baseline control do you expect to become most visible in the new response?

Paste the complete prompt into the ChatGPT message box.
Send the prompt.

ChatGPT shows its Search activity while it gathers sources. Keep the conversation open until the response finishes.

You will see a literature review organized around the prompt contract. Its citations and evidence table give you specific claims to inspect.

Response missing citations?

Confirm that Search remains selected in the conversation.
Check that your pasted prompt includes the final Boundaries section.

Help me troubleshoot a Search response that ignored my prompt's citation rules.

Check every required output section

The six output sections act as acceptance criteria for the review. The response passes this step when every section is present with the requested level of detail.

Before you inspect the response, which required section do you think Search was most likely to miss?

Confirm the response includes Research question and scope.
Confirm the response includes an Evidence table with 4 to 6 linked academic sources.
Confirm the response includes a 350 to 500 word synthesis organized by themes.
Confirm the response includes Disagreements and limitations.
Confirm the response includes What's missing or still uncertain.
Confirm the response includes a Source-quality check.

✔️ All six sections are present

That is a major upgrade from the baseline. Your review now exposes its evidence structure for checking.

ⓧ A required section is missing

A missing section means the response has not fully followed the contract. A focused follow-up repairs the specific gap without restarting the review.

Choose one missing section from the checklist.
Ask ChatGPT to add only that section.
Check the revised response for the requested section.
Repeat the focused follow-up for each additional missing section.

Still missing a section?

Help me write a focused follow-up for the missing section.

Compare your prompt with the complete contract below:

✔️ My prompt matches

Your prompt contract is complete. The structured review is ready for its citation audit.

ⓧ Review the full prompt

# Goal
Answer this research question: What does academic research published from 2015 to the present say about the effect of retrieval practice on learning outcomes in higher education?

# Context and search scope
Use web search. Focus on university, college, tertiary, or health-professions education. Prioritize peer-reviewed systematic reviews, meta-analyses, review articles, and applied higher-education studies. Use 4 to 6 academic sources. Exclude marketing pages, unsourced summaries, and studies that only address school-age learners unless they provide necessary context and are clearly labeled.

# Evidence rules
Do not invent citations. Attach a citation to every substantive research claim. For each source, report its title, year, study type, learner population, relevant finding, important limitation, and link. Separate what a source directly reports from your interpretation. If the source type, population, or support for a claim cannot be confirmed, label it “needs verification” rather than guessing.

# Output
1. Research question and scope
2. Evidence table
3. A 350 to 500 word synthesis organized by themes
4. Disagreements and limitations
5. What's missing or still uncertain
6. Source-quality check

# Boundaries
Distinguish retrieval practice from spacing or distributed practice. Do not treat an abstract, search snippet, or citation as proof until the linked page supports the claim. Keep conclusions proportional to the evidence.

You have turned a vague request into an inspectable research workflow. The next audit can test the review against rules that were set before Search began.

Your evidence structure is locked in. Next, you will test three cited claims against their linked sources.

Audit and Repair the Review

Your structured retrieval-practice review now has all six required sections. It looks complete enough to trust at first glance.

However, citations from ChatGPT Search can be incomplete, outdated, or incorrect. A fluent review becomes auditable only after you compare its claims with the linked evidence.

This step turns three source checks into a repaired review. The finished response also gives you a corrected prompt template for future academic questions.

In this step, get ready to:

Audit three substantive claims against their cited sources.
Classify the support for each claim.
Repair the review using your audit findings.

Audit three cited claims

A citation audit tests whether a source supports the research claim attached to it. Your audit uses four checks for each claim.

Switch back to the structured retrieval-practice review from the previous step.
Select one substantive claim from the evidence table or thematic synthesis.
Open the citation attached to that claim.
Compare the source page title plus year with the details in the review.
Confirm that the source page represents an academic publication.
Identify whether the studied population fits higher education.
Compare the source text with the strength of the claim.
Repeat the four checks for two more substantive claims.

You should now have three claim-source comparisons based on pages you personally opened.

What Does Claim Strength Mean?

Claim strength describes how broad, certain, or causal a statement sounds. A narrow result from one population cannot support a universal conclusion.

An abstract supports only the information visible in that abstract. Recording limited access protects the review from claiming that you inspected evidence you could not reach.

Record the audit findings

Each comparison now needs a clear judgment. The smallest correction keeps as much supported information as possible without stretching the source.

What Goes in Each Audit Entry?

The claim field captures the substantive statement exactly as the review presents it.
The source field records the matching title plus year.
The publication check records whether the page represents an academic publication.
The population check records whether the learners fit the higher-education scope.
The support check records whether the accessible source text justifies the claim's strength.
The status field uses supported, partly supported, or not supported.
The correction field states the smallest change needed to align the claim with the source.
The access field records any abstract-only access limitation.

A supported claim can record that no correction is needed. A partly supported claim usually needs narrower wording about its population, effect, or certainty.

Return to the review conversation from earlier.
Place your cursor in the message composer.
Create three matching audit entries by copying this structure into the message composer:

Claim:
Source title and year:
Academic publication:
Higher-education population:
Support status:
Smallest correction:
Access limitation:

Complete one copy of the structure for each audited claim.
Compare all three entries with the field list above.

You should see three audit entries in the composer. Each entry should connect one claim to an opened source plus a support decision.

Repair the review from your audit

The audit gives ChatGPT specific evidence for revising the review. Your follow-up converts each support decision into a visible correction.

Add the revision instruction beneath your third audit entry by pasting this text:

Revise the review using this citation audit. Remove or weaken unsupported claims, preserve genuine disagreements, label anything unresolved as needs verification, and finish with the corrected reusable prompt template.

What Does This Follow-Up Enforce?

Unsupported claims must disappear or use weaker wording.
Genuine disagreements must remain visible in the synthesis.
Unresolved evidence must carry the needs verification label.
The final template must preserve the review's evidence controls.

Before you submit the message, which audited claims do you expect the repaired review to preserve, weaken, or remove?

Submit the combined audit findings plus revision instruction.

You should receive a repaired review followed by a corrected reusable prompt template.

Compare each repaired claim with its corresponding audit entry.
Confirm that unsupported claims were removed or weakened.
Confirm that genuine disagreements remain visible.
Check every unresolved item for the needs verification label.
Inspect the corrected template for the original scope rules.
Inspect the corrected template for the evidence rules.
Inspect the corrected template for the six output sections.
Inspect the corrected template for source-quality plus proportional-conclusion rules.

The final review should now contain only supported claims at their justified strength. The reusable template should preserve every control that made those claims auditable.

Does the Revision Conflict With Your Audit?

Send a focused follow-up that names the surviving claim plus its audit status.
Ask ChatGPT to restore any missing prompt control without changing the verified findings.

For help finding the mismatch, ask: Help me compare my revised review with my citation audit.

You have completed the hard part: every repaired claim now traces back to evidence you personally checked. Your reusable template carries the same audit rules into future research questions.

Secret mission

Transfer the Prompt to a New Question

Adapt your corrected citation-audited prompt to research spaced practice among university students. You will test whether its evidence controls survive the topic change by scoring the output and auditing one cited claim.

Clean Up Your Resources

This project runs in ChatGPT Free through Microsoft Edge, so it adds no ongoing cost. Decide whether to keep, archive, or delete the project conversations stored in your ChatGPT account.

Resources you used:

ChatGPT conversations containing your baseline review, structured review, citation audit, and optional transfer test.

Keep everything running

No action is needed. Choose this if you want your reviews, citation audits, and reusable prompt available for future research.

Leave the project conversations in your ChatGPT history.
Reopen the corrected retrieval-practice review when you want to reuse its prompt template.
Keep the spaced-practice transfer test for comparison if you completed the Secret Mission.

Pause - I'll come back to this later

Archiving removes a conversation from the regular sidebar while keeping it in your account. Choose this if you want a tidier history without losing your work.

Open the ChatGPT chat history sidebar.
Locate one of your project conversations.
Select the more options menu beside the conversation title.
Select Archive.
Repeat these steps for each remaining project conversation.

Delete - I don't want to use this again

Deleting these chats is permanent, so confirm you no longer need the project history. OpenAI schedules each deleted chat for permanent removal within 30 days after it disappears from your account view.

Open the ChatGPT chat history sidebar.
Locate one of your project conversations.
Select the more options menu beside the conversation title.
Select Delete.
Confirm the deletion when prompted.
Repeat these steps for each remaining project conversation.

Nice Work!

Excellent work! You built a reusable academic research prompt in ChatGPT Free that turns a focused question into a structured literature review with citations you can inspect.

What you learned:

Used a five-point baseline scorecard to expose the research controls missing from a vague prompt.
Built a citation-audited prompt contract with defined scope, evidence rules, uncertainty labels, and six required output sections.
Performed a three-claim citation audit to compare each claim with its source, population, and justified strength.
Completed the optional Secret Mission by transferring the template to spaced practice and auditing one claim from the new review.

Ready to quiz yourself?