Rep AI Giving Wrong Answers? How to Fix Incorrect Responses

Share
	Illustrated ecommerce support AI correction workflow showing a warning-marked answer, a source review with magnifying glass, and a corrected response confirmed by green checks.

uick answer: If Rep AI gives a shopper an incorrect answer, open AI Training > Test & Train, reproduce the question on the appropriate channel, and inspect Sources under the response. Correct the inaccurate Shopify data, store page, FAQ, document, URL source, or promotion that contributed to the error. For a specific question and answer, use Correct the AI to create a Custom FAQ. If the problem concerns behavior rather than a fact, adjust the relevant AI Instructions and save them. Then repeat the original question, try several alternative phrasings, and check the channels your store actually uses.

A wrong answer is not necessarily evidence that the underlying model has invented a fact. Your store might be publishing two conflicting return policies, a product metafield might not be enabled, or an email-specific instruction might differ from your chat instructions. Each cause requires a different fix. The most reliable process is reproduce → inspect → correct the source or rule → retest.

This guide focuses on troubleshooting incorrect or misleading replies in Rep AI. It explains how to investigate them, what each correction tool changes, and what to verify before considering the issue resolved. The examples are hypothetical merchant scenarios designed to illustrate the documented controls; they are not tests performed in a live Rep AI account.

First, identify what kind of answer went wrong

A customer may report that “the AI was wrong,” but that description can cover several independent problems. Classifying the symptom early prevents you from applying a generic instruction to an issue caused by the product catalog.

Observed problemLikely area to investigateAppropriate starting point
The AI quotes an outdated shipping or returns policyWebsite page, URL source, Custom FAQ, or fileInspect Sources and fix the outdated material
The AI says it does not know a routine answerMissing information in its knowledge baseAdd a precise FAQ or appropriate knowledge source
The AI gives an incorrect size, material, or stock statusShopify product data, inventory, or metafieldsValidate the Shopify record and enabled product fields
The answer is factually correct but promises too muchGlobal or channel-specific instructionsReview the relevant AI Instructions
Chat replies correctly, but email or Instagram does notDifferent channel configuration or contextTest the affected channel separately
A response misrepresents an attached fileAttachment-reading result in email simulationCheck whether the attachment was actually read
The assistant suggests a product the merchant wanted hiddenCatalog eligibility, recommendations, store settingsCheck the product's recommendation availability
A repeated error remains unexplainedReal conversation transcript and source detailsReview Conversations and report a problem

Do not treat the table as an automatic diagnosis. It is a map for investigation. An outdated statement may come from more than one source, and a recommendation can be affected by relevance, inventory, market eligibility, or the customer's existing cart.

Wrong information, missing information, and unwanted behavior are different

A wrong answer contains an incorrect claim: for example, saying that returns are permitted for 60 days when the merchant's current policy says 30. A missing answer occurs when the assistant cannot find enough information to respond confidently. Unwanted behavior might be excessively confident wording, an inappropriate tone, or an unsupported promise of delivery by a particular date.

The distinction matters because a Custom FAQ is useful for a concrete, repeatable question, while a broader instruction governs how the AI should respond to a class of situations. Neither tool can compensate for unreliable product data without investigating the underlying Shopify record.

Step 1: Reproduce the incorrect answer in Test & Train

Start with the original customer wording rather than rewriting the question into something easier for the AI. Small differences in phrasing, product names, or context may explain why the problem appeared in one conversation but not another.

  1. Sign in to the Rep AI console and open AI Training > Test & Train.
  2. Choose the simulation corresponding to the reported problem: Chat, Email, or an available connected Meta channel.
  3. Enter the customer's original question. If you are simulating an email, include its subject where required.
  4. Record the answer that appears, including any qualifiers, missing details, or unsupported promises.
  5. Open Sources below the reply and examine what information was used.
  6. Note the incorrect claim and the source you believe should establish the correct answer.

Rep AI documents Test & Train as a private testing interface: preview questions are not sent to real customers. The simulation can also be available for a connected channel even where live AI Answering is not enabled. This makes it useful for checking corrections without deliberately exposing shoppers to a suspected error.

There is an important access distinction. Email simulation depends on a supported connected helpdesk; Meta simulations require connected platforms. Do not assume every account has every tab or that all supported conversations use identical instructions.

How to Exclude Products from Rep AI (Shopify Guide)
Learn how to exclude products from Rep AI individually or in bulk, and understand why Available, Add-to-cart, Browsable and Shopify Markets settings have different effects.

Reproduce more than a single isolated sentence

Some mistakes appear only after an earlier message changes the context. For example, the initial question “Do you ship internationally?” might receive a correct general answer, while a follow-up “Then will it arrive in Toronto on Friday?” could elicit a promise the store cannot verify.

Reconstruct the shortest meaningful conversation that reproduces the problem. Capture the question, the preceding context, and the observed reply. Later, use this same sequence as a control test so you can determine whether your change actually helped.

Step 2: Use Sources and Explanation to locate the problem

A correction should begin with evidence. In Test & Train, inspect Sources beneath the message. Depending on the answer, the cited material may include a product record, Custom FAQ, URL source, file, or other store knowledge. Check whether the material itself is correct, current, complete, and applicable to the shopper's situation.

Real customer messages can also be examined in Conversations, where Rep AI documents transcript review and, where available, Sources and Explanation. These views help you understand which materials contributed to the response and how the answer was presented. Interface wording and the location of controls may differ between documentation versions or individual screens.

Do not interpret the presence of a source as proof that the answer is true. A source might be outdated, another source might contradict it, or the question might concern a country or product variant to which the information does not apply.

Check for contradictions before adding another FAQ

Imagine a store that updated its returns policy last week. The current website says 30 days, but an older uploaded PDF still states 60 days. Adding a third FAQ saying “30 days” may improve one test, yet it leaves conflicting information in the knowledge set.

A cleaner fix is to identify the authoritative current policy, update the relevant store page, and correct or remove obsolete sources. Where a precise Custom FAQ is needed, ensure it matches that same policy. Do not deliberately maintain several incompatible versions and assume one will always win.

Historical conversations need special care. Rep AI describes cases where an explanation for an older conversation may be reconstructed or unavailable. A reconstructed explanation can assist an investigation, but it should not be described as a complete, original technical trace of everything the model did at the time.

Step 3: Correct the underlying source of truth

The best correction is often outside the chat box. If a shopper asks about a return window, the store's policy page must be correct. If the AI gives the wrong fabric composition, start with the product data. If a promotion has expired, check the promotion and any supporting content that still advertises it.

Work through the content source relevant to the error:

  • Storefront website: update an inaccurate policy, shipping page, product description, or public information page.
  • Shopify catalog: correct the product, variant, inventory, or other source fields where the claim originated.
  • Custom FAQ: edit or replace a question-and-answer pair that contains an old or ambiguous answer.
  • File Source: replace, revise, or remove obsolete uploaded information.
  • URL Source: correct the destination content and verify the knowledge source is updated or rescanned as appropriate.
  • Promotions: check campaign terms, dates, restrictions, and stale references.

After making a change, return to Test & Train and check whether the assistant uses the corrected information. A change in Shopify or on the website is not a validated fix until you observe the behavior you wanted in a relevant test.

Choose one authoritative answer for time-sensitive policies

Shipping times, refunds, discounts, and availability change. A merchant should decide which internally approved source governs each claim, and keep other documents aligned with it. For a complex policy, specify meaningful conditions rather than publishing a short answer that becomes inaccurate outside its most common case.

For example, “Orders can be returned within 30 days” may omit whether the clock starts at purchase or delivery, whether customized products are excluded, and whether the item must be unused. The right correction preserves such qualifications if they are part of the actual store policy. Never invent them for the sake of a polished response.

Step 4: Correct a specific answer with Correct the AI

Rep AI's Correct the AI workflow is designed for a question whose answer should be changed. It provides a direct route from an incorrect response toward a reusable, explicit answer.

In Test & Train, reproduce the issue and select Correct the AI. Review the displayed question and the wrong answer, enter the answer that the merchant has verified, and save the correction using the available interface control. Rep AI documents the resulting correction as a Custom FAQ. It can subsequently be used as knowledge across the supported response contexts.

This distinction is essential: using a button labeled “Correct the AI” is not evidence of fine-tuning the underlying model weights. The documented operation creates a question-and-answer knowledge entry. Its usefulness depends on the quality of the content and on whether the system retrieves it in the relevant context.

A similar correction workflow is documented on ordinary messages in Conversations. When a question is reformatted or normalized during correction, read the edited wording before saving. A revised question that accidentally removes a product, country, or order-status condition can turn a narrow answer into an overbroad one.

Example: correcting a shipping claim

Suppose an assistant says, “Every international order arrives in two business days.” The merchant's approved policy does not support such a promise. A useful correction could answer the actual customer question by explaining that delivery estimates vary by destination and shipping method, and by directing the shopper to the merchant's verified delivery information.

Before adding the correction, verify what the published policy really states. If the public page itself contains the two-day claim, fix that page first. Then test a general question, a specific-country follow-up, and an urgent-delivery request. Passing only the exact original question is not enough to demonstrate a robust correction.

Step 5: Create or edit Custom FAQs directly

You do not have to wait for a wrong answer to create a precise knowledge entry. Rep AI documents the path AI Training > Knowledge Base > Custom FAQs for adding and managing question-and-answer pairs.

Write the question in language a customer would naturally use. Put the approved answer in a form that is direct but not broader than the source permits. For product-specific information, check the available product-page targeting options in the current console; Rep AI's historical guides differ in how many products can be selected, so avoid relying on an assumed universal limit.

The save control may appear with wording such as Update or Publish depending on the workflow and documentation version. Follow the button displayed in your account rather than treating a label from an older screenshot as permanent.

Why an accurate FAQ may still fail to answer a particular question

A FAQ can exist without being selected for every variation of a question. Its wording may be too narrow, it may conflict with other knowledge, or its product context may not match the current shopper's request. Rep AI's training documentation recommends retesting the original question and nearby phrasings, and allowing brief processing time when a change has just been saved.

If the correction is not applied, inspect Sources again, look for conflicts, and improve the FAQ question or answer. Avoid stacking several almost-identical contradictory FAQs to force a result. Record what actually changed and what happened in the test.

Rep AI Products Not Syncing With Shopify? Fixes
Shopify products missing from Rep AI? Learn to separate catalog sync issues from recommendation settings, inventory, market exclusions and missing product data.

Step 6: Use AI Instructions for behavior, not as a substitute for correct facts

A factual FAQ answers a specific question. An AI Instruction tells the assistant how to behave across applicable conversations. The latter is better suited to situations such as avoiding unsupported delivery guarantees, asking for clarification when a customer omits a critical detail, or keeping customer-service language appropriately measured.

Rep AI documents Global AI Instructions as well as channel-specific instructions in its testing environment. Depending on the connected channels, these may include Email, Instagram, Facebook, and WhatsApp. Channel instruction lists are managed separately, and edits must be explicitly saved; do not assume that typing in the editor commits the change.

Suppose the AI accurately reports a policy but says “I guarantee this will arrive tomorrow” when the information does not justify certainty. Review whether a relevant instruction encourages overconfident answers. A carefully scoped instruction might require acknowledging that delivery estimates depend on the checkout option and destination. The wording should reflect approved business policy, not a made-up universal rule.

Do not assume editing a global instruction automatically resolves every conflict in a channel-specific rule. Test the channel where the issue appeared. A chat response can look correct while an email simulation behaves differently because it has different instructions or context.

Step 7: Investigate unanswered questions through Missing Information

A missing answer and a false answer need different follow-up procedures. Rep AI documents Insights > Missing Information as a place to examine questions that its AI could not answer due to insufficient knowledge. Entries can include frequency, the originating page, categories, and statuses such as Open and Resolved.

For an unanswered recurring question, review its context, determine the merchant-approved response, and use the documented Answer workflow to add usable information. This may create or feed a FAQ. Prioritizing questions that recur across real customer conversations is a sensible editorial practice because the same missing knowledge can affect several shoppers.

However, Missing Information is not a complete database of every incorrect statement. A confident but wrong answer might never enter this list. If you are investigating a factual mistake, always examine Test & Train and the actual conversation rather than concluding that the absence of a missing-information entry means everything is accurate.

When to add an additional source instead of a short FAQ

Some questions require a detailed reference: size charts, complex shipping zones, product-care instructions, or multi-condition policies. A concise FAQ can point to the main facts, but the underlying material must also be accurate, structured, and accessible. Consider the knowledge source type that best represents the content and how often it changes.

A static product-care document might be appropriate as a maintained file, whereas a frequently changing policy may be easier to manage on a public page that is kept current. These are workflow recommendations, not a claim that one source type is universally preferred by the Rep AI engine.

Step 8: Check Shopify products, inventory, and metafields

Not all wrong answers originate in the knowledge-base editor. Rep AI uses connected Shopify product information. Before teaching it a contradictory answer, verify what Shopify itself says about the item.

For a wrong stock claim, examine the product's available quantity, whether inventory tracking is enabled, and the Shopify setting for continuing to sell when out of stock. Rep AI's troubleshooting guidance treats these as relevant checks when an available item is described as unavailable. The underlying Shopify screens and field names may evolve, so verify the current merchant admin when following an older documented procedure.

For a wrong fabric, dimensions, or other custom product property, check whether the data is stored in a Shopify metafield that Rep AI is configured to read. The documented interface is AI Sales > Catalog > Product Metafields. Rep AI describes an activation step rather than automatic use of every possible field, with up to 10 product metafields and 5 variant metafields enabled under the documented configuration.

Some variant-metafield types, including certain rich text, money, JSON, file, and reference types, are not available as usable values in the documented integration. Activating a field does not create a missing value in Shopify. It also does not transform that field into a universal product-search filter: Rep AI describes metafield knowledge as contributing to responses, not as a mechanism for filtering every catalog search.

Recommendations are not identical to factual claims

If the AI fails to recommend a particular item, check the product's eligibility and the recommendation context before diagnosing “wrong facts.” Relevant conditions may include whether the product is published, in stock, restricted for the shopper's market, already in the cart, or specifically excluded from recommendations.

The task of fixing products missing from the synced catalog belongs to product-sync troubleshooting. Deliberately excluding certain items belongs to catalog availability controls. Keep those tasks separate from correcting the wording or factual content of an AI answer.

Step 9: Refresh outdated URL sources, documents, and promotions

A knowledge source can be technically present but factually stale. When Sources identifies an old policy page or document, decide whether to update the original content, replace the source, or remove a duplicated obsolete record.

Rep AI documents adding website pages, Google Docs, and Google Sheets through URL Sources. For Google-hosted documents, the documented access setup requires Anyone with the link viewing access and permission for the relevant download, copy, or print operations. That makes these links unsuitable for confidential customer lists or private store records. Do not expose sensitive information merely to make an AI source readable.

For Google Sheets, the documented URL-source scan reads visible tabs together and ignores hidden tabs. It has a 35,000-character total limit for that Sheet, beyond which the scan can fail. This limit belongs to that particular URL-source workflow; it is not a general Rep AI product-catalog limit.

After editing a supported Google document or sheet, the published guidance describes resubmitting the same link with Add URL to refresh the source. Check the scan outcome rather than assuming the content was ingested correctly. If the assistant continues using old promotional terms, inspect the Promotions knowledge source and any other pages repeating the offer.

Why a clean correction sometimes needs content cleanup

A typical store has information spread across product pages, policy pages, files, promotional landing pages, and FAQs. If the same fact appears five times and only one copy gets updated, inconsistent answers may persist.

Maintain a small record of the authoritative source, its owner, and its last update for high-impact subjects. Remove or revise redundant obsolete statements when the business changes a rule. The goal is not to maximize the number of knowledge entries; it is to make the information the AI may use coherent.

Sync Shopify Products with Rep AI: Catalog & Updates
Rep AI syncs Shopify products automatically after your store is connected. Learn how to check the catalog, verify changes, and distinguish sync from recommendation and cart settings.

Step 10: Troubleshoot incorrect email attachment answers

If the error concerns an attached invoice, photo, form, or PDF in a simulated email, check what the simulator actually read. Rep AI documents email-attachment states including Attachment read, Cut short, Attachment not read, and No reading returned.

Its Test & Train email simulator documentation specifies up to five attachments per message, 25 MB each, with supported image and document formats. It also describes a 10,000-character email-body limit for the simulation and requires a subject for an initial email. These are simulation-specific limits. Do not claim they define every production message limit or every file-ingestion pathway in Rep AI.

If an attachment is cut short, the answer might omit a condition near the end of a document. If it was not read, the AI should not be expected to quote its contents accurately. Try a smaller or clearer test document where appropriate, inspect the reading status, and retest. If the actual workflow still fails, preserve the evidence for support instead of asserting a cause you cannot confirm.

A screenshot of a damaged product can also be ambiguous. A useful QA outcome may be that the assistant appropriately asks for clarification or defers to a human rather than inventing a detailed conclusion about the image.

Step 11: Test each connected channel on its own

A response that works in website chat does not necessarily establish the same behavior in email, Instagram, Facebook, or WhatsApp. Rep AI documents separate tests and instructions for supported connected channels, and some channel types are not interchangeable.

Use the following review order:

  1. Reproduce the original question on the affected channel.
  2. Review the knowledge source used there, where exposed.
  3. Check that channel's specific instructions and save any changes.
  4. Retest the factual claim and the tone of the answer.
  5. Repeat on other channels that carry the same customer question, if they are connected.
  6. Record the result for each channel separately, including not tested where access is unavailable.

The latest documented test surfaces distinguish connected Meta direct-message flows from public comments. Rep AI describes an Instagram public-comment test in a separate area; it does not document equivalent Facebook public-comment simulation in the same way at the time covered by the dossier. Do not treat an Instagram DM test as proof that every public-comment workflow behaves identically.

Check consistency without erasing useful channel differences

It can be reasonable for a website chat answer to be brief and for an email answer to include additional context. The factual claim should still align with the store's approved policy. Consistency means agreement on essential facts, not forcing every channel to use identical copy or creating unsupported promises in the name of personalization.

Step 12: Review real conversations and report unresolved issues

Conversations is the appropriate place to review actual customer exchanges rather than relying only on simulations. Find the message, read the transcript around it, and inspect any available Sources or Explanation data. Rep AI also documents corrective actions and Report a problem for issues that require vendor review.

The position of Report a problem can vary; newer documentation describes it within or near the Explanation view when available, while some interfaces expose it elsewhere in the message actions. Avoid promising that every account or historical transcript will show the same controls.

When reporting a repeatable problem, prepare a concise record: the date and channel, the exact relevant question, the AI's reply, the expected answer with its approved source, the sources shown by Rep AI, and the tests already attempted. Include screenshots if useful, but remove unnecessary personal information. This is a recommended troubleshooting record, not a mandatory vendor-defined support form.

A conversation marked unhelpful is worth reviewing, but that label does not establish that the AI made a factual error. A customer may dislike a correct answer, or the response may be confusing rather than false. Classify it before changing any knowledge or instructions.

A practical fix matrix for common Rep AI mistakes

SituationWhat to check firstBetter corrective actionHow to validate
Incorrect delivery estimateCurrent store policy and SourcesUpdate policy content and any conflicting FAQGeneral and country-specific shipping prompts
False return eligibilityReturn policy, exclusions, dated filesCorrect the underlying policy and targeted FAQStandard and exception-based return scenarios
Wrong product materialShopify product data or selected metafieldsFix the product source or expose the approved fieldExact product and near-identical variant prompts
Unavailable item described as in stockShopify inventory and trackingCorrect inventory settings/data as appropriateTest an in-stock and an out-of-stock item
Old discount still quotedPromotions and outdated URL/file sourceUpdate or remove obsolete campaign contentAsk for current discounts and expired offer terms
AI invents a delivery guaranteeRelevant channel instructions and policyCorrect source and behavioral rule as neededTest urgent-delivery questions on that channel
One channel differs from anotherChannel-specific instructionsEdit and save the affected channel instructionsRepeat the same facts on both channels
Added FAQ does not appearFAQ wording, scope, conflicts, processingImprove FAQ and clean up contradictory sourcesOriginal question plus paraphrases
Email response ignores attached PDFAttachment-read stateRepair the test input or escalate reading issueRepeat with a supported, readable document
Recurring unanswered questionInsights > Missing InformationAdd a verified answer or sourceTest original and similar questions

The final step is always a test of the actual response, not merely confirmation that the settings screen accepted the edit.

For high-impact customer claims, use a lightweight written record. This is an editorial testing method, not a proprietary Rep AI requirement.

Before the change:

  • Copy the customer's relevant question and the answer that was wrong.
  • Record the affected channel and enough surrounding context to reproduce it.
  • Write the correct merchant-approved answer and identify its authoritative source.
  • Inspect Sources and mark the root cause confirmed, suspected, or unknown.

During the change:

  • Correct the underlying source first where it is wrong.
  • Record whether you edited a Shopify field, website page, FAQ, file, URL, promotion, or instruction.
  • Check other sources for contradictions and save changes in the relevant interface.
  • Avoid making several unrelated edits simultaneously; otherwise you may not know which change affected the result.

After the change:

  • Rerun the original question exactly.
  • Try two or three plausible customer paraphrases and at least one relevant follow-up.
  • Test the affected channel and any other important connected channels separately.
  • Check Sources again where available; record the observed output rather than assuming success.
  • Mark the outcome pass, partial, fail, or needs support.
  • Keep an eye on future real conversations for recurrence of the same claim.

A worked example: two contradictory refund answers

A hypothetical store updates its refund eligibility from one set of conditions to another. The current website has the new text, but an old FAQ still contains the previous exception. A shopper asks, “Can I return an opened item?” Rep AI replies with the old eligibility wording.

The wrong approach is to add a broad instruction saying “always tell customers returns are available.” That would create a new risk and does not remove the outdated condition.

The better approach is to inspect Sources, confirm the FAQ or policy fragment contributing to the answer, update the obsolete source, and create a precise approved FAQ only if needed. The operator then tests an unopened item, an opened item, a personalized product, and an ambiguous case, using the real store policy for each expected answer. Only observed correct results count toward resolution.

A worked example: inaccurate product dimensions

A shopper asks whether a particular backpack fits a laptop of a certain size. The assistant replies using dimensions that do not match the merchant's current catalog. Rather than supplying made-up dimensions in a FAQ, the operator first checks Shopify's product record and any relevant metafield. If that metafield is not enabled for Rep AI, the proper next step may involve its catalog configuration, provided the field type is supported.

The operator then retests the specific product, a size-related paraphrase, and a similar-looking different product. The test should not be considered successful merely because the assistant sounds confident: its answer has to match the verified product information.

Common mistakes to avoid when fixing Rep AI answers

Adding a FAQ before checking Sources. A FAQ may improve one test while the wrong website page remains live. Fix the incorrect source when possible.

Treating instructions as factual databases. An instruction such as “be helpful” cannot establish a reliable, current refund policy or product size.

Assuming the AI retrains itself after every conversation. Rep AI's documentation describes explicit knowledge, instruction, and correction workflows, not automatic retraining from every customer message.

Confusing a missing answer with a false answer. The Missing Information area can reveal knowledge gaps, but it is not a complete list of confident mistakes.

Generalizing a simulation limit to all channels. The attachment limits described for test email and the character limit for a Google Sheet URL source belong to different product workflows.

Assuming every source has been fully read. Verify URL scans, attachment statuses, and source references instead of inferring that a document was ingested perfectly.

Relying on a single successful test. Customer questions are variable, and channel configurations may differ. Use a repeatable, small set of tests with written expected answers.

Inventing a guaranteed accuracy rate. A correction is a measurable configuration change, not proof of a universal percentage improvement or permanent elimination of mistakes.

Final takeaway

When Rep AI gives a wrong answer, trace the information before changing the behavior. Start in Test & Train, inspect Sources, repair incorrect store knowledge or Shopify data, use Custom FAQs for precise corrections, and use AI Instructions where the issue is a general response rule. Then test the original question, alternate phrasings, and relevant channels.

The goal is not simply to make one answer look better in a preview. It is to establish a repeatable correction process based on current merchant-approved facts, consistent knowledge sources, and observed responses—without claiming that any configuration change makes mistakes impossible.

Frequently asked questions

How do I correct a wrong answer in Rep AI?

Go to AI Training > Test & Train, reproduce the customer's question, and inspect Sources. Fix an inaccurate source of knowledge. If a specific answer needs correcting, use Correct the AI, which Rep AI documents as creating a Custom FAQ. Test the question again before marking the issue resolved.

Does Correct the AI retrain the underlying language model?

The documented correction mechanism creates a Custom FAQ. That is a knowledge-base correction, not evidence that the underlying model weights were fine-tuned. It may improve responses that use the new answer, but it does not establish a guarantee for every future question.

Why is Rep AI still giving the wrong answer after I added a FAQ?

Check whether the FAQ's question matches the shopper's intent, whether it is scoped to the right context, whether another knowledge source contradicts it, and whether the change has had time to be processed. Reproduce the question, inspect Sources again, and try paraphrases rather than repeatedly adding duplicate FAQs.

Can I train Rep AI to answer a question it currently cannot answer?

Yes, the documented knowledge workflows include Custom FAQs, URL Sources, File Sources, and instructions for appropriate scenarios. For recurring unanswered questions, Insights > Missing Information can help find gaps. Use verified merchant information, and remember that a newly added source still needs testing.

Does Rep AI learn automatically from every customer conversation?

Its documentation describes explicit correction and knowledge-management workflows and does not support claiming that every conversation automatically retrains the AI. Review actual conversations, make deliberate changes where justified, and validate them.

Can Rep AI answer differently on chat, email, and WhatsApp?

The system has distinct connected-channel contexts and documented channel-specific instructions. Test the same factual issue separately where those channels are enabled. Email and Meta simulations also have connection prerequisites; do not assume every account exposes every test option.

Where can I see which source Rep AI used?

Open Sources under a response in Test & Train. In actual conversations, the Conversations view may also expose Sources and Explanation. Availability and placement can vary, and some older explanations may be reconstructed or unavailable.

Will a Custom FAQ override every conflicting document?

Do not rely on a universal override guarantee. Rep AI's guidance emphasizes keeping the knowledge base coherent. Correct or remove obsolete material as well as adding a precise answer, then test the actual result across different phrasings.

What should I do if Rep AI gives the wrong inventory information?

Begin with the Shopify product and inventory settings, including available quantity and relevant tracking or backorder behavior. For custom product details, check whether the correct metafield is enabled and supported in Rep AI. A wrong Shopify value should be fixed at its source rather than disguised with an unrelated FAQ.

Why did Rep AI ignore a Google Sheet or a PDF?

Check the specific ingestion or reading workflow. A Google Sheet used as a URL source has documented access requirements and a 35,000-character total limit; hidden tabs are not read in that workflow. In simulated email, inspect attachment statuses such as Cut short or Attachment not read. These are not interchangeable limits.

Can I test Rep AI answers without sending a message to a customer?

Rep AI documents Test & Train as a private simulation. Connected Email and Meta options depend on the relevant integration, and a channel can be testable even when AI Answering is disabled for live traffic. Always distinguish a test result from a live production observation.

What if I cannot identify why the AI answered incorrectly?

Preserve the original question and relevant conversation, examine Sources and any available Explanation, verify the authoritative merchant information, and collect the outcomes of your tests. Then use the documented Report a problem route or support channel with a concise reproduction. Do not claim the root cause is known if the evidence remains inconclusive.